Ir_cover.dvi MANUAL

User Manual: MANUAL manual pdf - FTP File Search (14/20)

Open the PDF directly: View PDF .
Page Count: 210

Download
Open PDF In Browser	View PDF

NISTIR 6269

Finite Element and Finite Di erence Programs for Computing the Linear Electric and Elastic Properties of Digital Images of Random Materials

Edward J. Garboczi

Building and Fire Research Laboratory
Gaithersburg, Maryland 20899

United States Department of Commerce
Technology Administration

National Institute of Standards and Technology

NISTIR 6269

Finite Element and Finite Di erence Programs for Computing the Linear Electric and Elastic Properties of Digital Images of Random Materials

Edward J. Garboczi

N T OF C O M

M

IT

E

D

ER

UN

ICA

E

D EP

E
TM

C
ER

AR

December 1998
Building and Fire Research Laboratory
National Institute of Standards and Technology
Gaithersburg, Maryland 20899

ST

ATES OF

AM

U.S. Department of Commerce
William M. Daley, Secretary

Technology Administration

Gary R. Bachula, Under Secretary for Technology
National Institute of Standards and Technology
Raymond Kammer, Director

ABSTRACT
This manual has been prepared to show some of the theory behind, and the practical
details for using, various nite element and nite di erence computer programs that have
been developed for computing the e ective linear properties of random materials whose
microstructure has been stored in a 2-D or 3-D digital image. Thirteen di erent computer
programs are described, including nite element conductivity and elastic programs, nite
di erence programs for D.C. and A.C. conductivity, nite element elastic programs that
include thermal strains (eigenstrains), and three auxiliary programs for Gaussian quadrature
and phase percolation. All the programs are written in FORTRAN 77, and operate on an
arbitrary digital image that is read from a le. Arbitrary symmetric conductivity tensors
and arbitrary elastic moduli tensors can be used in the nite element programs. In the
nite di erence programs, the conductivity tensors must be diagonal. Only linear elastic
and linear electrical conductivity problems are considered. The programs can of course be
extended to other problems that have a similar mathematical basis.
Keywords: Building technology, computer modelling, concrete, electrical conductivity,
eigenstrain, elastic moduli, nite di erence, nite element, linear properties, microstructure, random materials, thermal elasticity, transport properties.
Cover picture: Showing, for a horizontal applied eld, the horizontal currents around a
circular inclusion of conductivity 10, embedded in a matrix of conductivity one, computed
using a nite di erence program. The magnitude of the currents go from red = high to black
= low, according to the associated color bar.

iii

Contents

Abstract
List of Figures
List of Tables
1 Introduction
2 Finite element theory
2.1
2.2
2.3
2.4

General aspects and node labelling scheme
Electrical conductivity . . . . . . . . . . .
Elastic moduli . . . . . . . . . . . . . . . .
Thermal strains (eigenstrains) . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

iii
vii
ix
1
4

. 4
. 6
. 12
. 14

3 Finite di erence theory

19

3.1 General description, comparison to nite element method . . . . . . . . . . . 19
3.2 Derivation of nite di erence form of Laplace's equation . . . . . . . . . . . 20
3.3 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4 Explanation of key subroutines and program usage
4.1
4.2
4.3
4.4
4.5

Subroutine ENERGY . . . . . . . . . . . . . . . . .
Subroutine DEMBX . . . . . . . . . . . . . . . . .
Subroutine CURRENT/STRESS . . . . . . . . . .
3-D : 2-D relations for programs . . . . . . . . . . .
Actual program operation . . . . . . . . . . . . . .
4.5.1 Electric and elastic nite element programs .
4.5.2 Eigenstrain programs . . . . . . . . . . . . .
4.5.3 Finite di erence programs . . . . . . . . . .

5 Exact solutions for testing programs
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9
5.10
5.11

De nition of e ective properties . . . . . . . . . . .
Series and parallel . . . . . . . . . . . . . . . . . . .
Small contrast of properties . . . . . . . . . . . . .
Keller and Mendelsen 2-D result for conductivity .
Field uctuation result . . . . . . . . . . . . . . . .
Equal shear modulus . . . . . . . . . . . . . . . . .
Intrinsic properties for spheres, circles, and cubes .
Vegard's law and Goodier result for thermal strains
Hashin and Rosen thermal strain result . . . . . . .
Mackenzie result for pressurized pore space . . . . .
CLM Theorem . . . . . . . . . . . . . . . . . . . .

v

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

25

25
25
25
26
27
27
29
29

31

31
32
33
35
35
36
38
43
46
46
47

6 Other possible uses of programs
6.1
6.2
6.3
6.4

Fixed voltages and displacements . . . .
Removing periodic boundary conditions .
Fixed currents and forces . . . . . . . . .
Surface energies . . . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

7 Making and analyzing images and histograms
7.1
7.2
7.3
7.4
7.5
7.6
7.7

General features . . . . . . . . . . .
Finite element electrical problems .
Finite di erence electrical problems
Finite element elastic problems . .
General features of histograms . . .
Examples of histograms . . . . . .
Phase percolation in images . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

8 References
9 Listing of programs and computer requirements
9.1 Computer details, access information
9.2 Memory requirements . . . . . . . . .
9.3 Listing of key programs . . . . . . . .
9.3.1 ELECFEM3D.F . . . . . . . .
9.3.2 ELAS3D.F . . . . . . . . . . .
9.3.3 THERMAL3D.F . . . . . . .
9.3.4 DC3D.F . . . . . . . . . . . .
9.3.5 AC3D.F . . . . . . . . . . . .
9.3.6 GAUSS.F . . . . . . . . . . .
9.3.7 BURN3D.F . . . . . . . . . .

vi

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

48
48
49
51
52

55

55
55
56
57
57
58
61

62
65

65
65
66
67
88
116
169
181
193
196

List of Figures
1
2
3
4
5
6
7
8
9
10
11
12
13

14

Graphic view of cubic pixel = tri-linear nite element, showing the 1-8 labels
of the vertices. The i,j ,k axes coincide with the x,y,z axes. . . . . . . . . . .
Schematic view of the structure of the Hessian matrix when thermal strains
are used (u stands for elastic displacements,  represents the macrostrains). .
Illustration of nodes near a boundary in the nite di erence method. . . . .
Illustration of how periodic boundary conditions are implemented in the nite
di erence programs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Showing  vs. 2 when 2 di ers only slightly from 1 . The points are nite
element data, and the straight line is Brown's exact expansion to second order
in the contrast (2 , 1 ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Showing the checkerboard microstructure, with dark gray being phase 2 and
light gray phase 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Showing the e ective conductivity and average of the electric eld magnitude
squared in phase 2 for the checkerboard, as a function of 2. The points are
numerical nite element results, and the lines are the exact results discussed
in the text. The system size was 128  128. . . . . . . . . . . . . . . . . . .
Intrinsic conductivity for a 15 pixel diameter sphere embedded in a 403 unit
cell, as a function of the ratio of the sphere conducticity to the matrix conductivity. Finite element and nite di erence data and the exact result are
compared. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Intrinsic elastic moduli for a 15 pixel diameter sphere embedded in a 403 unit
cell, as a function of the ratio of the sphere Young's modulus to the matrix
Young's modulus. The three sets of data show [K], [G], and the exact result,
which is the same for both intrinsic moduli. . . . . . . . . . . . . . . . . . .
Intrinsic conductivity for a 103 cube embedded in a 403 unit cell, as a function
of the ratio of the cube conductivity to the matrix conductivity. The three
sets of data (circle, square, line) compare the nite element method, the nite
di erence method, and Eyges' [25] data, respectively. . . . . . . . . . . . . .
Thermal stress for inclusions with the same elastic moduli as the matrix, but
di erent thermal eigenstrain. Stress is high (top) to low (bottom) in color
bar). Left: ,xx, right: negative of the trace of the stress tensor. Images
from top: Ellipse, 5-pointed star, 6-pointed star. . . . . . . . . . . . . . . . .
Showing the resistance between two nodes, separated by 20 lattice spacings,
of a simple cubic network, normalized by the resistance of one bond, as a
function of the thickness of the network. . . . . . . . . . . . . . . . . . . . .
Image of the horizontal current magnitudes, with the applied electric eld in
the x-direction in all images. The inclusion is phase 2. Left: 2 = 10, right:
2 = 0:1, and both images had 1 = 1:0. Top: Finite element solution. Middle: nite di erence solution. Bottom: exact solution, no periodic boundary
conditions. Color bar shows high (red) to low (black) current scale. . . . . .
Current distribution for the same circular inclusion problem as in Fig. 13,
calculated by all three methods ( nite element, nite di erence, and exact
calculation, 1 = 1:0, 2 = 10. . . . . . . . . . . . . . . . . . . . . . . . . . .
vii

5
18
22
24
34
36
37
40
41
43
45
50

56
59

15 Current distribution for the same circular inclusion problem as in Fig. 13,
calculated by all three methods ( nite element, nite di erence, and exact
calculation, 1 = 1:0, 2 = 0:1 . . . . . . . . . . . . . . . . . . . . . . . . . . 60

viii

List of Tables
1
2
3
4
5
6
7
8
9
10
11
12
13

Finite element (fem) labels for within a single pixel labelled (i,j ,k). The 
i,
etc. values are with respect to the node labelled (i,j ,k). . . . . . . . . . . . .
Formulas for components of [@Nr =@xi ]  nir matrix . . . . . . . . . . . . . .
Neighbor labelling in 2-D and 3-D . . . . . . . . . . . . . . . . . . . . . . . .
Relation between nite element labelling and neighbor labeling . . . . . . . .
Values of r variable for system . . . . . . . . . . . . . . . . . . . . . . . . .
Components of Lpq . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
rp vectors for the 8 pixel corners . . . . . . . . . . . . . . . . . . . . . . . .
Size dependence (L  L system) of e ective conductivity and eld average
for checkerboard, as determined by nite element method. Each individual
"check" of the checkerboard is L=2  L=2. The conductivity ratio 2 =1 = 10.
E ective bulk modulus for checkerboard{equal shear moduli case. . . . . . .
Intrinsic conductivities for sphere, comparing the nite element and nite
di erence techniques. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Intrinsic elastic moduli for sphere, d = 15 in 403 system . . . . . . . . . . . .
Intrinsic elastic moduli for circle{e ect of digital resolution, using nite element method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Memory requirements in terms of bytes per pixel for programs discussed in
this manual. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

6
7
9
10
12
14
15
37
38
40
42
42
66

1 Introduction
The e ective properties of composites have been studied analytically for a long time, back to
the days of Maxwell, who rst solved for the e ect that a single spherical inclusion, with a different conductivity from the matrix, has on the overall or e ective conductivity [1]. Random
materials like concrete, powder compacts, etc. are also composites, random agglomerations
of di erent property material at various length scales. Because of their randomness, their
properties cannot be computed analytically, but instead require numerical computation. To
compute the e ective properties of such materials requires knowledge of the microstructure.
The only practical way such information is obtained is either through microscopy, x-ray microtomography [2, 3], or models [4]. Whatever the method, the microstructural information
is almost always in 2-D or 3-D digital image form, collections of discrete square or cubic
pixels in which each pixel can, in principle, be a di erent phase of the material. Hence the
need to have computer programs that are specialized to work on digital images.
This manual describes the theory behind, and practical instructions for using, various
nite element and nite di erence programs specialized to operate on two and three dimensional digital images of materials. The digital images can be completely general, representing
any material with any number of distinct phases. The nite element programs are set up
assuming that each phase is characterized by an arbitrary symmetry conductivity tensor,
elastic moduli tensor, and eigenstrain or thermal strain tensor. The nite di erence electrical programs are set up for a conductivity tensor that is diagonal, but which can have
di erent terms along the diagonal for the x, y, and z directions. The digitalimages do not
have to have the same dimensions in each direction.
The three basic problems considered are the following: 1) the e ective conductivity
of a material composed of di erent conducting and non-conducting phases, either at zero
frequency (D.C.) using either a nite element or a nite di erence program, or at nite
frequency (A.C.), using a nite di erence program, 2) the e ective elastic moduli of a material composed of phases with di erent elastic moduli tensors, and 3) the e ective thermal
strain/eigenstrain of a multi-phase material with di erent elastic moduli and eigenstrain in
each phase. The material digital image is assumed to have periodic boundary conditions
(see Section 6 for how to remove these conditions).
The digital images of the microstructures analyzed can be generated within the programs
themselves, or can be obtained elsewhere and simply read into the property programs. In
the programs described in this manual, the digital images are always read into the program.
In the discussion in Section 2, about nite element theory, the language is for 3-D. There is
a separate section later in this manual (Sec. 4.4), which explains the quite simple relations
between 2-D and 3-D. The 2- D programs in this package were created directly from the 3-D
programs, and so are nearly identical in structure to the 3-D versions.
There are 10 main programs described in this manual, and available for the user. They include D.C. nite element electrical programs in 2-D and 3-D (ELECFEM2D.F, ELECFEM3D.F
), nite element elastic programs in 2-D and 3-D (ELAS2D.F, ELAS3D.F), nite element
elastic programs that incorporate thermal strains (eigenstrains) in 2-D and 3-D (THERMAL2D.F, THERMAL3D.F), and nite di erence electrical programs for a.c. or d.c. problems in 2-D or 3-D (AC2D.F, AC3D.F, DC2D.F, DC3D.F). In what follows, any mathematical variable will be in italics. All program and subroutine names will be in capitals. Also,
1

a variable with two subscripts, say wmn, may be written as w(m; n), which is the way it
appears in the programs. As much as possible, variable and subroutine names are the same
from program to program. The notation and structure of the conjugate gradient routine
(in subroutine DEMBX) has been made to follow that used in Numerical Recipes, chapter
10 [5]. Also, in the following, the term "vector" is used interchangeably for 3-D and 2-D
vectors, for arrays that have as many entries as there are variables in the problem, and for
quantities that have both magnitude and direction. The meaning will be clear from the
context. Enough details are given in this manual so that, hopefully, any user can rewrite
and change the programs to adapt them to speci c problems, and add new capabilities.
All programs should be run in double precision (8 bytes or 64 bits per real variable).
Running the nite di erence routines in single precision may give reasonable answers, but
double precision is more trustworthy. Running the nite element routines in single precision
will not work, except perhaps for very small systems, on the order of 1000 nodes. All the
routines are written in simple FORTRAN 77, making use of only a small subset of the
language. These routines should therefore be portable across di erent platforms. Those
users who wish to proceed directly to how to run the programs, can skip to Sec. 4.5,
where details of how to use the programs are discussed. Another similar package, which
is for 2-D only but has a graphical interface, is the OOF system, which can be found at
http://www.ctcms.nist.gov/ (click on "Software").
In addition to these 10 main programs, there are also three auxiliary programs, one for
computing Gaussian quadrature weights and points, and two for computing phase percolation
in an arbitrary 3-D or 2-D image.
The layout of the remainder of this manual is the following. Section 2 brie y derives
and de nes the theory behind the nite element programs, electric, elastic, and elastic with
thermal or eigenstrains. Section 3 discusses the theory behind the nite di erence programs,
both A.C. and D.C. The early parts of Section 4 explains the workings of key subroutines of
both the nite element and nite di erence programs, followed by a discussion of how to use
the programs and which variables need to be set by the user. Section 5 then presents the
theoretical results of many exact solutions in electrical conductivity and elasticity, accompanied by tests or examples showing how well these programs can reproduce these results.
These exact results are helpful in testing progams such as the ones presented in this manual,
since they are for non-trivial microstructures and/or choices of individual phase paramters.
Section 6 then presents the theory showing how these programs may be modi ed to solve
many other problems of interest, along with some useful examples of program use. Section
7 explains how to make images and histograms from the local eld results using any of the
programs, and gives some examples, and Section 8 is the reference section. Section 9, the
last section, gives computational details, access information for the programs, and includes
listings of ve of the 10 main programs, and two of the auxiliary programs.
I would like to thank Prof. A.R. Day, of Marquette University, for setting up the rst version of the conjugate gradient routine for the nite element problems, Prof. P.M. Duxbury,
of Michigan State University, for giving me a version of what eventually was developed into
the basic 2-D nite di erence electric program, Prof. M.F. Thorpe, for getting me started in
algorithms applied to digital images, J.F. Douglas, for suggesting some of the example problems for exact composite results, D.P. Bentz, of the NIST Building Materials Division, who
has been a collaborator for almost a decade on digital-image-based microstructure problems,
2

and S.A. Langer, for critically reading this manual and suggesting many changes.

3

2 Finite element theory

2.1 General aspects and node labelling scheme

The theory on which the nite element programs described in this manual are based is very
simple. The essential idea is that a variational principle exists for the linear elastic and
linear electrical conductivity problems. For a given microstructure, subject to applied elds
or other boundary conditions, the nal voltage or elastic displacement distribution is such
that the total energy stored in the elastic case, or the total energy dissipated, in the electrical
conductivity case, is extremized, such that the gradient of the energy with respect to the
variables of the problem (voltage or elastic displacement) is zero. The variable En is used
for both of these cases, and is called an energy in all the following text, even though, in the
case of electrical conductivity, it is really an energy dissipation per unit time or power. To
minimize En, a function of many variables um , the various partial derivatives must equal
zero,
@En = 0
(1)
@u
m

for all values of m. In all the programs, the sum of the squares of all elements of the gradient
vector, whose m'th element is just the partial derivative in eq. (1), is determined during the
solution or relaxation process. The solution of the problem is considered to be reached when
this sum is less than a given small value, so that the condition in eq. (1) is approximately
satis ed for all m. This value, denoted gtest in all the programs, should be chosen small
enough so that the answers obtained, the currents or stresses in the pixels, are no longer
changing signi cantly with further relaxation.
A labelling scheme must be de ned for a single element or pixel in order to derive the
nite element equations for that pixel. "Pixel" and "element" will be used interchangeably
throughout this manual. In the nite element method, each node attached to a corner of a
pixel (8 in 3-D, 4 in 2-D) has a separate label within that pixel. Since in the nite element
method, the energy is de ned within each pixel using only the nodes attached to that pixel,
it is important to be able to refer easily to these attached nodes. The (i; j; k) label for a
pixel, which gives its position in a three-dimensional lattice, is the same as the (i; j; k) label
for the node numbered 1 in the pixel. Table 1 gives this single-pixel labelling scheme, in
terms of the 
i, 
j , and 
k positions of the nodes with respect to (i,j ,k). Figure 1 shows
this labelling scheme graphically, and also de nes the coordinate system used. The (i,j ,k)
axes coincide with the (x,y,z) axes, respectively.
The basic derivations for the electric case and the elastic case will now be reviewed, with
the pixel length taken to be unity. Most books on nite elements will have much of this
derivation. It is given here in order to clarify the structure of the programs described in this
manual. In all the discussion below, r and s run from 1 to 8, and indicate the node of the
nite element being considered, while p and q run from 1 to 3, and indicate the Cartesian
coordinate (1 = x, 2 = y, 3 = z) of vector quantities. The convention is used throughout
the rest of the manual that if a subscript is repeated, it is assumed to be summed over.

4

Z,k
8

Y, j

7

5

6

4

3

1

2

X, i

Figure 1: Graphic view of cubic pixel = tri-linear nite element, showing the 1-8 labels of
the vertices. The i,j ,k axes coincide with the x,y,z axes.

5


i 
j 
k fem label (3-D) fem label (2-D)
0
1
1
0
0
1
1
0

0
0
1
1
0
0
1
1

0
0
0
0
1
1
1
1

1
2
3
4
5
6
7
8

1
2
3
4

Table 1: Finite element (fem) labels for within a single pixel labelled (i,j ,k). The 
i, etc.
values are with respect to the node labelled (i,j ,k).

2.2 Electrical conductivity

Variables in programs

ur = voltage at the r'th node in a pixel
pq = conductivity tensor (di ers in general from element to element)
E~ = (Ex; Ey ; Ez ) = external electric eld applied to image
e = (ex; ey ; ez ) = local eld at a point (x; y; z) inside a pixel
Drs = sti ness matrix in a pixel
Nr (x; y; z) = shape array for cubic pixel
The purpose of this kind of nite element method is to express the total energy of the
system in terms of the voltages only at the nodes, thus e ectively discretizing the continuum.
There are many ways in which to do this. We choose here to use a linear interpolation
method, in what is often called a tri-linear scheme. Essentially we write out the energy of
one pixel, which involves an integral over that pixel, and then combine the pixels of the
whole system into a global energy functional. The elds inside the pixel are expressed as
functions of (x; y; z) via a linear interpolation scheme in terms of the nodal voltages. The
integral is carried out, and we end up with an expression for the energy of the pixel that
is quadratic in the nodal voltages. The global energy is then also a quadratic functional of
the nodal voltages. This expression is minimized with respect to the nodal voltages, using a
conjugate gradient scheme. Using the set of nodal voltages found, the interpolation scheme
can then be re-used to nd the average current, total energy, etc. that can be used to de ne
the e ective conductivity and other quantities of interest (see Section 5.1).
Focus on a single pixel for now. De ne V (x; y; z) to be the voltage within the given pixel,
where 0 < x; y; z < 1. The origin is taken where the pixel label is 1 (see Table 1 and Fig.
6

1). V (x; y; z) is determined by linear interpolation of the nodal voltages, such that

V (x; y; z) = Nr ur

(2)

and Nr = Nr (x; y; z). The actual functional forms of Nr in 3-D are:

N1 = (1 , x)(1 , y)(1 , z)
N2 = x(1 , y)(1 , z)
N3 = xy(1 , z)
N4 = (1 , x)y(1 , z)
N5 = (1 , x)(1 , y)z
N6 = x(1 , y)z
N7 = xyz
N8 = (1 , x)yz
Now the local eld in the pixel, ~e, is given in terms of derivatives of this local voltage,
x; y; z)
(3)
ep(x; y; z) = , @V (@x
p
In terms of the nodal voltages, this expression for ep (the p'th component of the local eld
~e) then becomes
"
#
,
@
,
@N
r
ep(x; y; z) = @x [Nr ur ] = @x ur
(4)
via

p

p

where ,@Nr =@xp  npr is a 3  8 matrix linking the 8-vector of nodal voltages to the 3-vector
of local elds, and ur has no explicit x dependence. Table 2 gives the components of this
matrix.

r ,@Nr =@x  n1r ,@Nr =@y  n2r ,@Nr =@z  n3r
1
2
3
4
5
6
7
8

(1 , y)(1 , z)
,(1 , y)(1 , z)
,y(1 , z)
y(1 , z)
(1 , y)z
,(1 , y)z
,yz
(1 , y)z

(1 , x)(1 , z)
x(1 , z)
,x(1 , z)
,(1 , x)(1 , z)
(1 , x)z
xz
,xz
,(1 , x)z

(1 , x)(1 , y)
x(1 , y)
xy
(1 , x)y
,(1 , x)(1 , y)
,x(1 , y)
,xy
,(1 , x)y

Table 2: Formulas for components of [@Nr =@xi ]  nir matrix
7

Now, the total energy dissipated in the pixel is



(5)
dx dy dz 21 eppq eq
0 0 0
In terms of the nodal voltages, this equation becomes
Z 1 Z 1 Z 1
h
i
1
T
En = 2 ur
dx dy dz npr pq nqs us
(6)
0 0 0
The indicated integrations can be easily carried out exactly using Simpson's rule, since
they are at most quadratic in x, y, or z, and Simpson's rule is exact for quadratic expressions.
These expressions could also be evaluated analytically, and then put into the program. When
these programs were developed, it was decided to use Simpson's rule for simplicity, and to
avoid doing all the derivations necessary to get the "sti ness" matrices in analytical form.
The subroutine FEMAT in the nite element programs carries out this integration. The
energy per pixel is then
(7)
En = 21 ur Drsus
where the matrix Drs, an 8  8 matrix in 3-D, is de ned by comparing eq. (7) to eq. (6).
This matrix is known as a "sti ness" matrix, a term originating in nite element treatments
of linear elasticity problems. The above term is added up for all pixels to give the global
energy, each pixel using its own value of the conductivity tensor. This sum is then the
total energy dissipated of the system under the applied eld, which must be minimized with
respect to the nodal voltages to nd the solution to this discretized version of the original
problem.
In order to be able to link up local energy with the global energy, it is important to
see how the local numbering scheme for the pixel nodes links up with the global numbering
scheme. The digital images are in the form pix(m), where pix is a 2-byte integer vector
(de ned INTEGER*2 in FORTRAN 77), and m is a one dimensional label. The digital
image in 3-D is de ned by (i; j; k), where i = 1; nx, j = 1; ny, and k = 1; nz, with the origin
of the image being the (1; 1; 1) node. The one dimensional labelling scheme is de ned by:

En =

Z 1Z 1Z 1

m = nx 
 ny 
 (k , 1) + nx 
 (j , 1) + i

(8)

The label m refers to the m'th node, and the m'th pixel. The m'th pixel has the m'th
node at its corner that is labelled "1" in the 1 , 8 local nite element numbering scheme. In
3-D , every node is part of 8 pixels. All the nodes in these pixels will need to be known with
respect to the node, since they are connected to it in the global energy via the local sti ness
matrices, and so their labels are encoded in the array ib. If the m'th node corresponds
to (i; j; k), then the 27 nodes that are part of these 8 pixels are given by the m labels
corresponding to adding or subtracting 0 or 1 from the i,j , k labels, so that ib = ib(m,27)
has 27 entries for each value of m. Table 3 gives this neighbor labelling scheme. For example,
ib(m,15) would be the m label of the node located at (,1,0,,1) with respect to the (i,j ,k)
position of the m'th node. Table 3 also gives the neighbor labelling scheme for 2-D, where

k = 0, since there is no third dimension. Note that in the interior of an image, the m labels
of the neighbors would always be simply known anyway, because of their xed relationship in
8


i 
j 
k 3-D neighbor # 2-D neighbor #
0
1
1
1
0
-1
-1
-1
0
0
1
1
1
0
-1
-1
-1
0
0
1
1
1
0
-1
-1
-1
0

1
1
0
-1
-1
-1
0
1
0
1
1
0
-1
-1
-1
0
1
0
1
1
0
-1
-1
-1
0
1
0

0
0
0
0
0
0
0
0
0
-1
-1
-1
-1
-1
-1
-1
-1
-1
1
1
1
1
1
1
1
1
1

1
2
3
4
5
6
7
8
27
9
10
11
12
13
14
15
16
25
17
18
19
20
21
22
23
24
26

1
2
3
4
5
6
7
8
9

Table 3: Neighbor labelling in 2-D and 3-D

9

fem label (3-D) neighbor label (3-D) fem label (2-D) neighbor label (2-D)
1
2
3
4
5
6
7
8

27
3
2
1
26
19
18
17

1
2
3
4

9
3
2
1

Table 4: Relation between nite element labelling and neighbor labeling
terms of i, j , and k, but with periodic boundary conditions (discussed below) a "neighbor"
could be at the other side of the image. The relationship between the local (1 , 8) labelling,
and the global (1 , 27) neighbor labelling used in array ib, is given in Table 4.
Periodic boundary conditions mean that if a neighbor is outside the digital image, it is
periodically continued on the opposite side of the system. For example, consider the node
at i = 10, j = 12, k = 20, in an nx = ny = nz = 20 digital image. The m label of this node
(pixel) would be 7830, via eq. (8). Neighbor number 18 would be located at i = 11, j = 13,
and k = 21. However, k = 21 > nz, so the k label is changed to k , nz = 21 , 20 = 1.
The m label of i = 11, j = 13, k = 1 is m = 251, so that ib(7830; 18) = 251. The process
is similar for the other faces (i = nx; j = ny), the edges (i = nx and j = ny, i = nx and
k = nz, j = ny and k = nz) and the corner point (i = nx; j = ny; k = nz).
Before going into the solution technique, one must rst consider the boundary conditions
of the problem. Equation (7) is positive de nite, and so has its minimum value when all the
voltages are zero, an uninteresting result. We use periodic boundary conditions to apply an
electric eld, so that the energy is minimized with respect to an applied eld, a more useful
case.
Consider a pixel as before, with i = nx; j < ny; k < nz, and label m. When we go to
evaluate the energy expression above, we nd no voltages at the 2,3,6, and 7 nodes, because
these would have i labels of (nx +1), which is unde ned. We use the voltages from across the
system, numbers 1,4,5, and 8 from the pixel with the same j and k label but with i = 1, and
label M = m , nx + 1. However, there is a jump of nx in length from where these voltages
are de ned. If the applied eld E~ = (Ex; Ey ; Ez ) is present, then on average there is a drop
in voltage of ,Ex nx between these nodes. Therefore, the voltages to be used in the energy
expression for the m'th pixel are: u1 = u1(m), u2 = u1(M ) , Ex nx, u3 = u4(M ) , Ex nx,
u4 = u4(m), u5 = u5(m), u6 = u5(M ) , Ex nx, u7 = u8(M ) , Ex nx, and u8 = u8(m).
We can write this, in general for a pixel at a boundary, as ur = Ur + r , where Ur is
an 8-vector of the voltages that are given by the ib(m; n) labels, and r is an 8-vector that
corrects them to what they should be in the pixel in question. The neighbor array ib is
10

designed to pick up the appropriate nodal voltages from across the image when evaluating
the energy of pixels on faces, edges, and corner.
Inserting this choice of voltages into the expression for the energy involving the sti ness
matrix, eq. (7), we get for the given pixel
En = 21 [ur Drsus + 2r Drsus + r Drss]
(9)
which gives a term quadratic in the nodal voltages, a term linear in the nodal voltages, and
a term constant with respect to the nodal voltages. In the particular case discussed, i = nx,
j < ny, k < nz, the components of the 8-vector r are: 1 = 0, 2 = ,Ex nx, 3 = ,Ex nx,
4 = 0, 5 = 0, 6 = ,Ex nx, 7 = ,Ex nx, 8 = 0.
We can rewrite eq. (9) as
(10)
En = 21 ur Drsus + br ur + C
where
bs = r Drs ; C = 21 r Drss
(11)
Adding these up over every pixel is done in the subroutine FEMAT, giving a global array
b that gives the term in the energy that is linear in the voltages, and a global constant C . It
is important to remember that the only contributions to b and C come from pixels having
nodes at the unit cell boundaries and having a non-zero sti ness matrix. That makes it easy
to make non-periodic boundary conditions by surrounding the system of interest by a layer
of insulating material, one pixel thick, which e ectively gets rid of the periodic boundary
conditions and makes b and C equal to zero (see Sec. 6.2). Table 5 shows the values of the
r variable for the various faces and edges of the digital image.
Once the energy equation is set up, all that remains is to nd the set of voltages that
minimize the electrical energy dissipated. This is done by a conjugate gradient method,
similar to that described in Ref. [5]. The subroutine DEMBX contains this routine, and is
the same for both electric and elastic problems, aside from some minor labelling di erences
of indices, due to the fact that voltage is a scalar while elastic displacement is a 3-D vector.
(see Section 4.2). No preconditioning of the matrix to be solved is done. It is not clear,
since the full Hessian matrix is never stored, whether it would even be possible to carry out
pre-conditioning in this case.
In the subroutine ENERGY, the gradient of the energy is computed, as this must become
very small in order to solve the problem, which is exactly solved when the gradient is zero.
Remember that the gradient is a vector containing all the partial derivatives of the energy
with respect to all the nodal voltages. Using the above expression, the gradient of the energy
will then be
@En = A u + b
(12)
mn n
m
@um
where now Amn and bm should be thought of as global quantities, so that all the terms
connected to um are included. The global matrix Amn is of course built up from the individual Drs matrices of the eight pixels that touch the node labelled m. The matrix Amn
is in principle large, but sparse. The way the nite element programs save memory is by
11

r i = nx

j = ny

i = nx
j = ny

i = nx j = ny
k = nz k = nz

0
,Ex nx
,Ex nx
,Ey ny
4
0
,Ey ny
0
,Ey ny
5
0
0
,Ez nz
0
6 ,Ex nx
0
,Ez nz ,Ex nx

0
0
,Ex nx
0
,Ex nx ,Ey ny

1
0
0
2 ,Ex nx
0
3 ,Ex nx ,Ey ny

k = nz
0
0
0

0

,Ez nz
,Ex nx
,Ez nz
7 ,Ex nx ,Ey ny ,Ez nz ,Ex nx ,Ex nx
,Ey ny ,Ez nz
8

0

,Ey ny
,Ez nz
,Ez nz
,Ey ny
,Ez nz

,Ey ny ,Ez nz ,Ey ny ,Ez nz ,Ey ny
,Ez nz

i = nx
j = ny
k = nz
0
,Ex nx
,Ex nx
,Ey ny
,Ey ny
,Ez nz
,Ex nx
,Ez nz
,Ex nx
,Ey ny
,Ez nz
,Ey ny
,Ez nz

Table 5: Values of r variable for system
computing Amn un using only the small single-pixel sti ness matrices and the appropriate
labelling scheme. This algorithm can be seen in subroutines ENERGY and DEMBX of the
nite element programs. The correct Drs term is used in place of the Amn term that would
be needed. In essence, the matrix A is re-built every time it is needed, without being stored.

2.3 Elastic moduli

The elasticity problem is set up similarly to the electrical conductivity problem, because the
elastic energy stored also obeys a variational principle. The elastic energy stored is given by
Z
1
En = 2 d3r pq Cpqrs rs
(13)
where the strain tensor pq and elastic moduli tensor Cpqrs are in full tensorial form, p,q,r,s =
1, 2, or 3, and the integral is over the volume of a single pixel. The total energy is obtained
by summing over all pixels. Since the strain tensor is symmetric, a simpler notation is
usually used, the Voigt notation, where the strain is taken to be a 6-vector containing the
six independent strains(xx; yy ; zz ; xz ; yz ; xy ), and then Cpqrs is written as C , where
@up
pp = @x
p
p @uq
pq = @u
@x + @x
q

12

p

We will use and exclusively as labels running over the six components of the strain
6-vectors. The energy equation becomes
Z
1
En = 2 d3r  C 
(14)
The idea of the nite element scheme, as for the electrical case, is to reduce the energy
equation to a quadratic form containing the components of the elastic displacement vector
de ned at the nodes of the pixels. There is an elastic displacement de ned at each node,
which has three components in 3-D. We denote this by u(m; 3). The m and ib integer labelling system is the same as before, but now all real variables have an extra label for the
Cartesian component being considered:

ump = p'th component of displacement at m'th node
C = elastic moduli tensor (Voigt notation) for a pixel
E~ = (Exx; Eyy ; Ezz ; Exz ; Eyz ; Exy ) = overall elastic strains applied to system
~ = (xx; yy ; zz ; xz ; yz ; xy ) = local strains at a point (x,y,z) in a pixel
Drp;sq = sti ness matrix in a pixel
Np;rq = shape matrix for cubic pixel
Each component of the displacement is linearly interpolated across the pixel in exactly
the same way as the voltage was done in the electric problem. The p'th component of the
three-vector ~u(x; y; z) at a point (x; y; z) in the pixel is then de ned as

up(x; y; z) = Np;rq (x; y; z)urq

(15)

where N1;r1 = N2;r2 = N3;r3 = Nr given for the electrical problem, and urq is the q'th
component of the displacement on the r'th node, r = 1; 8. Clearly, in this structure, we have
N1;r2 = N1;r3 = N2;r1 = N2;r3 = N3;r1 = N3;r2 = 0. N is a 3  (8; 3) matrix. To construct the
6-vector strain from this, we need to multiply by (or operate with) a matrix of derivatives,
Lpq , that is 6  3. The components Lpq are given in Table 6 below. This results in

 (x; y; z) = [L pNp;rq (x; y; z)] urq
or

(16)

 (x; y; z) = S ;rq (x; y; z)urq
(17)
where the components of S ;rq can be found in ELAS3D.F, in the subroutine FEMAT, where
the matrix es(n1; n2; n3) is equivalent to S ;rq . If eq. (17) is integrated over a pixel, it gives
the average strain in the pixel, in terms of the nodal displacements. When rst multiplied
by C and then integrated, the average stress in the pixel results. When the solution
displacements to the problem are obtained, these can be used to compute the average stress
and strain in the system, which de ne the e ective quantities and can also give insight into
local stress and strain elds around microstructural features.
Substituting into the energy expression above, in the Voigt notation, results in
Z
En = 21 d3r [S ;rpurp]T C [S ;sq usq ]
(18)
13

p (component of strain vector) q = 1 (x) q = 2 (y) q = 3 (z)

@=@x
0
0
@=@z
0
@=@y

1
2
3
4
5
6

0
@=@y
0
0
@=@z
@=@x

0
0
@=@z
@=@x
@=@y
0

Table 6: Components of Lpq
Grouping the S and C matrices together, and performing the integral over the pixel using
the part that has (x; y; z) dependence, results in
En = 21 uTrpDrp;sq usq
(19)
where
Z
Drp;sq = d3r [S ;rp]T C [S ;sq ]
(20)
is the sti ness matrix. It is also the same as the dynamical matrix arising in lattice models of
elastic phenomena found in theoretical physics, where displacements at nodes are connected
by various forces [6]. The rst part of subroutine FEMAT computes the sti ness matrix (dk
in the nite element programs) using Simpson's rule to perform the integration. As in the
electrical case, the integration is exact, as there is no term to be integrated that is higher
order than quadratic, and Simpson's rule is exact for quadratic functions.
Periodic boundary conditions result in exactly the same structure as in the electrical
case, but with now the extra index for the Cartesian coordinates, and a di erent form for
the vector rp , expressed in terms of the six independent applied strains. The components
of rp are given in Table 7. The periodic boundary conditions result in a term linear in
the displacements, denoted b 
 u, as well as a constant term C that is quadratic in the
applied strains. The gradient of the energy is the same as in the electric case, but with the
extra index for the Cartesian coordinates of the displacement. Note that the terms of b, by
de nition, are linear in the applied strains (see 2.2). This is important for the development
of the eigenstrain or thermal strain case, discussed next.

2.4 Thermal strains (eigenstrains)

In the case of thermal strains, sometimes called eigenstrains (terms used interchangeably in
this manual) [7], each phase can have a stress-free strain that comes about by thermal or
moisture expansion/shrinkage, or other causes. We denote this strain as e , where =1,6
as usual in the Voigt notation. The stress then becomes  = C ( , e ), where  is the
14

r,p i = nx j = ny k = nz
1,1
1,2
1,3
2,1
2,2
2,3
3,1

0
0
0

0
0
0
0
0
0

i = nx
j = ny

i = nx
k = nz

j = ny
k = nz

0
0
0

0
0
0

0
0
0
0
0
0

Exx nx
Exy nx
Exz nx
Exx nx Exy ny

0
0
0
0
0
0
0

Exx nx
Exy nx
Exz nx
Exx nx

Exx nx
Exy nx
Exz nx
Exx nx

Exy ny

3,2 Exy nx Eyy ny

0

Exy nx

Exy nx

Eyy ny

3,3 Exz nx Eyz ny

0

Exz nx

Exz nx

Eyz ny

4,1
0
Exy ny
0
4,2
0
Eyy ny
0
4,3
0
Eyz ny
0
5,1
0
0
Exz nz
5,2
0
0
Eyz nz
5,3
0
0
Ezz nz
6,1 Exx nx
0
Exz nz

Exy ny
Eyy ny
Eyz ny
0
0
0
Exx nx

0
0
0

Exy ny
Eyy ny
Eyz ny
Exz nz
Eyz nz
Ezz nz
Exz nz

6,2 Exy nx

0

Eyz nz

Exy nx

6,3 Exz nx

0

Ezz nz

Exz nx

7,1 Exx nx Exy ny Exz nz

Exx nx
+Exy ny

7,2 Exy nx Eyy ny Eyz nz

Exy nx Exy nx Eyy ny
+Eyy ny +Eyz nz +Eyz nz

7,3 Exz nx Eyz ny Ezz nz

Exz nx Exz nx Eyz ny
+Eyz ny +Ezz nz +Ezz nz

Exz nz
Eyz nz
Ezz nz
Exx nx
+Exz nz
Exy nx Eyz nz
+Eyz nz
Exz nx Ezz nz
+Ezz nz
Exx nx Exy ny
+Exz nz +Exz nz

8,1

0

Exy ny Exz nz

Exy ny

Exz nz

8,2

0

Eyy ny Eyz nz

Eyy ny

Eyz nz

8,3

0

Eyz ny Ezz nz

Eyz ny

Ezz nz

Exy ny
+Exz nz
Eyy ny
+Eyz nz
Eyz ny
+Ezz nz

Table 7: rp vectors for the 8 pixel corners
15

i = nx
j = ny
k = nz
0
0
0
Exx nx
Exy nx
Exz nx
Exx nx
+Exy ny
Exy nx
+Eyy ny
Exz nx
+Eyz ny
Exy ny
Eyy ny
Eyz ny
Exz nz
Eyz nz
Ezz nz
Exx nx
+Exz nz
Exy nx
+Eyz nz
Exz nx
+Ezz nz
Exx nx
+Exy ny
+Exz nz
Exy nx
+Eyy ny
+Eyz nz
Exz nx
+Eyz ny
+Ezz nz
Exy ny
+Exz nz
Eyy ny
+Eyz nz
Eyz ny
+Ezz nz

usual strain de ned by the elastic displacements. The elastic energy then becomes
Z
En = 21 d3r ( , e ) C ( , e )
(21)
Substituting for the strain using the linear interpolation scheme described earlier, we can
then perform the integration over a single pixel, keeping in mind that the thermal strains
are constants over the pixel and are not linearly interpolated. The resulting equation in the
nodal displacements is
En = 21 urpDrp;sq usq + 21 e C e + Trpurp
(22)
where the rst term is identical to the case without thermal strains, the second term is a
constant quadratic in the thermal strains, and the third term is linear in nodal displacements,
with Trp given by
Z
Trp = ,e C d3rS ;rp(x; y; z)
(23)
So even without periodic boundary conditions, there are terms linear and constant in
the displacements. Periodic boundary conditions are very similar to the elastic case studied
previously, except for an extra term picked up via the Trp term. In a pixel at the face, edge,
or corner of the image, the energy becomes
En = 21 urpDrp;sq usq + brpurp + 21 rpDrp;sqsq + 12 e C e + Trpurp + Trprp (24)
where the components of rp were given in Table 7, and the u's are the real displacements,
brought over from the opposing side or edge or corner of the image in the same way as in
the electrical or the no-eigenstrain elastic problems. It is important to note that brp and rp
are linear in the applied strains, and Trp is independent of the applied strains.
In the programs THERMAL3D.F or THERMAL2D.F, the whole system size is allowed
to change in order nd the overall thermal expansion that minimizes the energy of the
system. In this case, the applied strains, E , are now called the macrostrains, and become
dynamic variables that de ne the size and shape of the periodic unit cell. Their values are
determined in the conjugate gradient relaxation process, on an equal footing with the elastic
displacements. In the programs, the length of the u vector changes from ns = nx  ny  nz
to nss = ns +2, where u(ns +1,1)=Exx, u(ns +1,2)=Eyy , u(ns +1,3)=Ezz , u(ns +2,1)=Exz ,
u(ns + 2,2)=Eyz , and u(ns + 2,3)=Exy .
In subroutine DEMBX, the matrix of second derivatives, or the Hessian matrix, is used
to update the gradient and conjugate gradient direction in the relaxation process. For the
regular nodal displacements, the sti ness matrices make up the Hessian matrix. However,
when the six macrostrains are considered to be variables as well, the Hessian matrix goes
from being (ns; 3)  (ns; 3) in size to (ns + 2; 3)  (ns + 2; 3) in size. The extra partial
derivatives that have to be evaluated are (1) the mixed second derivatives, or the second
derivative of the energy with respect to a nodal displacement and a macrostrain, and (2) the
second derivative with respect to a macrostrain squared or two di erent macrostrains.
Figure 2 illustrates schematically what the full Hessian matrix looks like. The upper
left-hand corner is the main section, where the second derivatives of the energy are taken
16

with respect to the nodal displacements. Multiplication of a vector by this piece is taken
care of in the large DO loop involving the sti ness matrices in subroutine DEMBX. The
upper right and lower left-hand parts of the Hessian involve mixed second derivatives with
respect to a nodal displacement and a macrostrain. Examining eq. (24), which is the full
elastic energy with periodic boundary conditions and thermal energy terms, one can see that
the only term that has dependence on both the nodal displacements and the macrostrains
is the term brpurp . The kind of second derivative that is needed is

@En
(25)
@urp@E
In the term mentioned, the only urp dependence is the explicit one. The parameter brp
depends linearly on E , as can be seen in the elastic equivalent of eq. (11). In a linear
dependence, the partial derivative with respect to a certain variable is the same as the
function evaluated when that variable is one and all other variables are zero. Therefore,
@En = b (E = 1)
(26)
rp
@urp@E
where all the other macrostrains have been set to zero. This is the technique used in the
subroutine BGRAD, which computes the terms needed for these parts of the Hessian matrix.
The bottom right-hand part of the Hessian matrix involves second derivatives with respect
to the macrostrains. Again looking at eq. (24), the only term that has quadratic dependence
on the macrostrains is what was called the constant term before, that comes from the periodic
boundary conditions. In all the nite element programs with xed macrostrains, this term
is designated C . In the two thermal strain programs, this term is expressed by a matrix
called zcon. The second derivative terms are given simply by the various components of this
matrix. The subroutine CONST computes the elements of this matrix, which are constant
with respect to nodal displacements and macrostrains.
In subroutine DEMBX, the main part of the multiplication by the Hessian matrix is
carried out, and then the other three parts of the matrix multiplication are lled in using
subroutine BGRAD and the matrix zcon.

17

(ns,3)

(2,3)

(ns,3)

∂2En/∂u2

∂2En/(∂u∂ε)

(2,3)

∂2En/(∂ε∂u)

∂2En/∂ε2

Figure 2: Schematic view of the structure of the Hessian matrix when thermal strains are
used (u stands for elastic displacements,  represents the macrostrains).

18

3 Finite di erence theory

3.1 General description, comparison to nite element method

Included in this package of programs are nite di erence programs for solving linear conductivity problems on general digital images in 2-D and 3-D, for D.C. and A.C. problems.
There are some advantages and disadvantages when comparing nite di erence and nite
element computations. Regarding the programs described in this manual, the nite di erence programs take less memory, as they do not need the integer variable ib, which stores
neighbor information and is especially needed at the boundaries. This is because the nite
di erence programs handle the periodic boundary conditions in a di erent way, as will be
discussed below. The nite di erence programs in this manual could be written in an energy
form, and make use of the second derivative matrix and the variational principle approach
in the same way as the nite element formulation. They are not so written in this manual,
however.
The way the nite di erence algorithm handles boundary conditions between phases is
also di erent from the nite element algorithm. Electrical problems can be handled quite
readily with nite di erences, but there are some diculties with elastic problems. At
boundaries at which a displacement or stress component takes on a certain value, it is
straightforward to write nite di erence equations for elastic problems that handle this kind
of boundary, using non-centered di erence equations. For example, a nite di erence elastic
code can be written for a porous material, where there is a single kind of solid phase, and
so all solid-pore boundaries are zero stress boundaries [8]. Also, the case when the second
phase is in nitely sti , so that the displacement must be the same thoughout connected
parts of the phase, can also be handled by nite di erences. For a boundary where the
displacement and normal force are continuous, however, it is dicult to see how to write a
nite di erence program that accurately takes both these conditions properly into account.
However, two-phase elastic boundaries where both phases have nite moduli can be easily
be handled by nite elements.
As for material properties, nite di erence electrical conductivity programs can handle
arbitrary diagonal conductivity tensors. It is possible to extend them to tensors having
non-zero o -diagonal elements, although it makes phase-phase boundaries somewhat more
trickier to handle, and requires 2nd and 3rd neighbor information, instead of just nearest
neighbor information. Therefore the nite di erence programs described in this manual have
only diagonal conductivity tensors.
A major advantage of the nite di erence approach that we have found is in percolation
cases, where a conducting phase becomes sparsely connected. We have found that in simulating continuum percolation cases with digital images, assessing connectivity only through the
nearest neighbors agrees rather well with the equivalent continuum calculations [9]. When
using nite di erences, the only current ow possible is through nearest neighbor connections. In nite elements, however, a node is connected electrically with its rst, second,
and third nearest neighbors. Therefore a phase could become disconnected in terms of rst
neighbors, and so disconnected acording to a percolation algorithm, and yet stay connected
electrically, if a nite element scheme was being used. So in this case, the nite di erence
scheme is more physically realistic. If the resolution were in nite, then the result obtained
19

would be the same for the nite di erence and nite element methods. Since in a typical
problem, and especially so for a sparsely connected problem, the resolution is not in nite,
one needs to choose the method that will give more accurate results.
The nite di erence programs in this package are: AC3D.F and AC2D.F, for complex
( nite frequency or a.c.) problems in 3-D and 2-D, respectively, and DC3D.F and DC2D.F,
for d.c. problems in 3-D and 2-D. They are all based on a general digital image, where each
phase has a di erent (and frequency-dependent in the a.c. cases) conductivity. They are
not based on a true energy formulation, but rather on a solution of Kirchko 's laws for the
resulting conductance network, which is equivalent to solving Laplace's equation, as is shown
next.

3.2 Derivation of nite di erence form of Laplace's equation

For a steady state conductivity problem, where the currents are steady in time, the charge
conservation equation,
r 
 ~j + @
(27)
@t = 0
becomes simply
 
r 
 ~j = r 
 E~ = 0
(28)
or, in regions of constant conductivity,
 r2 V = 0
(29)
where V is the position dependent potential or voltage. Between phases having di erent
conductivities, the boundary conditions are that the current normal to the interface and the
potential are continuous.
Consider a region of uniform conductivity, and a point (i; j; k) in this region, in the middle
of a pixel. If the voltage at (i; j; k) is u(i; j; k), then, to second order in the pixel dimensions,
the potential at i  1, j  1, or k  1 is just
" 2 #
" #
1
@u
2
+ 2 a @@ 2 xu
(30)
u(i  1; j; k) = u(i; j; k)  a @x
i;j;k
i;j;k
"

#

"

#

2
u(i; j  1; k) = u(i; j; k)  b @u
+ 1 b2 @ 2u
(31)
@y i;j;k 2 @ y i;j;k
" 2 #
" #
@u
1
2
u(i; j; k  1) = u(i; j; k)  c @z
+ c @ 2u
(32)
2
@
z
i;j;k
i;j;k
where 
x = a, 
y = b, 
z = c, and a, b, and c are the dimensions of the pixel, in
some units. Then the nite di erence form of Laplace's equation, in terms of the 1-d label m
and the six nearest neighbors, can be obtained by adding together the above six equations in
pairs and solving for the 2nd derivative terms. The rst derivatives cancel out, the gradient
squared is formed from the sum of the second derivatives, and Laplace's equation in nite
di erence form, at each node m, becomes
X
m;n [u(n) , u(m)] = 0
(33)
n

20

where m;n is the conductance of the bond connecting node m to its nearest neighbor with
nearest neighbor label n, the sum is over the n = 1; 6 nearest neighbors, and for convenience,
both sides of eq. (33) have been multiplied by the volume of the pixel, abc. The values
of mn are m1 = m4 =  bca , m2 = m5 =  acb , and m3 = m6 =  abc , with neighbors
(1,2,3) in the (x,y,z) directions, and neighbors (4,5,6) in the (,x,,y,,z) directions. As has
been pointed out previously [10], this equation is formally identical to solving Kircho 's laws
of zero net current into each node for this conductance lattice, where the conductances are
those of a pixel of material with a given conductivity tensor in a given direction. Therefore
formally the length of the pixel appears in the conductance. Note that the nite di erence
programs are set up for cubic pixels, with a = b = c = 1. They can be easily modi ed
for a rectangular parallelipiped pixel, with three arbitrary dimensions. When setting up the
conductances, these lengths matter, as the conductances of a given conductivity pixel are
di erent in directions that have a di erent dimension.
There is an equation like eq. (33) for each node m in the system. Putting all these
equations together, a global equation can be written: Amn un = 0. If we build a quadratic
form out of this matrix A, 21 uAu, then this form is extremized when its gradient with
respect to u is zero: Amn un = 0, which is the same as the set of equations we are trying
to solve. The formal problem being solved is then the minimization of the quadratic form
1
2 uAu where the vector u is the voltage vector, and A is a sparse matrix composed of the
conductances of all the bonds in the problem. In the nite di erence programs, the matrix
A is implicitly stored in the vectors gx, gy, and gz. The storage requirements for the nite
di erence electric programs is much less than those for the nite element programs, and so
in this case, the global matrix A is stored. The result of A multiplied times an arbitrary
vector can be generated using the vectors gx, gy, and gz, which store the conductances of
the problem in the x, y, and z directions respectively. The subroutine PROD does this
matrix multiplication, while maintaining periodic boundary conditions and the applied eld
(see Section 3.3). If there were not the constraint of an applied eld, then of course the
minimization of a quadratic form such as 21 uAu would be trivially given by u = 0.

3.3 Boundary conditions

When two cubic pixels share a face, and are of the same material, the conductance of the
bond connecting them is just the conductance of the pixel,
m = m 
 (pixel length)

(34)

concentrated into a conducting bond (for a cubic pixel). The situation when two pixels share
a face, and are of two di erent materials A and B, is shown in Fig. 3. Node and pixel j is
on the left of the interface, and node and pixel j + 1 is on the right side of the interface. In
the gure, the conductivity of material A is shown as being di erent than that of material
B. If one were to think of a bond connecting j and j + 1, the conductance of that bond is
just a series combination of one half the conductance of each pixel,

,1
 = 21 + 21
(pixel length)
A
B

21

(35)

Figure 3: Illustration of nodes near a boundary in the nite di erence method.

22

This construction approximates boundaries. For a pure series situation, this construction
is perfectly accurate, and enables the program to give the correct answer. For a curved
boundary, like across a spherical inclusion, the answer will not be perfectly accurate. For
two adjoining pixels, this construction is, however, the best possible approximation.
The derivation of this condition is simple. Choose the pixel length to be unity, as is
done in all the programs. Consider the point P that is exactly on the boundary and halfway
between nodes j and j + 1. The voltage at P is VP . The boundary condition at the interface
is that the normal current across the interface must be continuous. In nite di erence terms,
that means that
A (VP , Vj ) = B (Vj+1 , VP )
(36)
which gives an equation for VP in terms of Vj and Vj+1 . Now one must build up the
appropriate nite di erence equations, but being careful of the boundary. One uses noncentered di erence equations for j and j , 1. The two expansions hold:
"

#

"

#

"

#

"

#

2
VP = Vj + 21 @V
+ 1 @ V2 + :::
@x j 8 @x j

(37)

1 @ 2 V + :::
+
(38)
Vj,1 = Vj , @V
@x j 2 @x2 j
Eliminating the rst derivative from this pair of equations, and using the boundary
condition equation to eliminate VP , results in the same kind of equation as eq. (33) for
node j , but with eq. (35) for the conductivity across the interface. Since in a digital image,
all boundaries are locally oriented in the x, y, or z directions, and are essentially 1-D, the
formalism in eqs. (35)-(38) above holds generally and is incorporated into the nite di erence
programs in subroutine BOND.
In all the nite di erence programs, periodic boundary conditions are used with the
application of an electric eld, with its appropriate voltage gradient. Periodic boundary
conditions are maintained using a shell of imaginary sites around the main system, as shown
in Fig. 4. In Fig. 4, there are nine real sites in this 3  3 pixel system, labelled 1 , 9, and
12 imaginary sites, labelled in italics according to which real site they correspond. Suppose
that a eld is applied from left to right in Fig. 4. The strength of the eld is 31 , so that
there is a voltage step of one placed across the sample. Therefore, the node with italic label
6 has a voltage equal to the voltage at real node 6, but exactly one unit higher. The node
with italic label 4 has the same voltage as that at real node 4, but one unit lower. Vertically,
since there is no applied eld, the nodes with italic labels 2 and 8 have the same voltages
as real nodes 2 and 8, respectively. Any update vector coming from the conjugate gradient
relaxation routine that is added to the voltage must obey pure periodic boundary conditions
(no applied eld), so that the inital applied voltage gradient (applied eld) is maintained.
In subroutine PROD in the nite di erence programs, there are DO loops that explicitly
maintain this periodicity after every multiplication of the main matrix.

23

Figure 4: Illustration of how periodic boundary conditions are implemented in the nite
di erence programs.

24

4 Explanation of key subroutines and program usage
Key subroutines found in the programs are rst explained, before describing how to use the
programs themselves. The user who prefers to treat the various subroutines as "black boxes"
can skip to Sec. 4.5, which describes the details of actually using the programs.

4.1 Subroutine ENERGY

The subroutine ENERGY is found in all of the nite element programs. It computes the total
energy of the system, which is a quadratic functional of the nodal voltages, in the electrical
case, and the nodal displacements, in the elastic case. It also computes the gradient of the
energy, to be fed into the subroutine DEMBX, which uses the gradient as a starting point for
the conjugate gradient routine. The heart of the subroutine is a matrix multiply of the global
sti ness matrix times the vector of displacements or voltages. This is accomplished using
only the small, local sti ness matrices dk. Essentially, each matrix multiply goes through
the same steps that would be necessary for actually building the global sti ness matrix, but
does not store the results, which saves tremendously on storage.

4.2 Subroutine DEMBX

Subroutine DEMBX is the conjugage gradient routine in both the nite element and nite
di erence programs. In both cases, DEMBX is essentially the same, with some details
di erent between the two kinds of programs. In the nite element elastic programs, the only
di erence from the electric case is the number of components of the vector quantities like
displacement. The quadratic energy functional minimized in the nite element programs
has explicit linear and constant terms, which come from the periodic boundary conditions.
The energy functional in the nite di erence programs has only an explicit quadratic term,
but the linear and constant terms are implicit in the periodic boundary conditions. A
standard conjugate gradient routine is used, which makes use of the Hessian matrix (matrix
of partial second derivatives of the energy function with respect to nodal values). The form
of the conjugate gradient routine can be found in Section 10.6 of Ref. [5]. Variables in the
programs have the same names as in Ref. [5].

4.3 Subroutine CURRENT/STRESS

This subroutine, called CURRENT for the electrical case, and STRESS for the elastic case,
computes, respectively, the total current tensor or the total stress tensor in the system. In
the nite element case, the averaging is done over each pixel ( nite element), using the nite
element average current and average stress matrices, and then added up over all the pixels.
In the nite di erence programs, the total current is added up over all the bonds connecting
nodes (see Section 7).

25

4.4 3-D : 2-D relations for programs

In terms of computer code, the 2-D programs, both nite element and nite di erence,
were created directly from the 3-D programs. This was done essentially by leaving out the
third (z) dimension everywhere. For the nite di erence programs, only two conductance
vectors, gx and gy, are needed to store the conductance information. For the nite element
programs, the sti ness matrices dk are only dimensioned for four nodes, not eight nodes, and
the elastic programs have only two degrees of freedom per node, not three as in 3-D. The
elastic modulus tensor is 6  6 in 3-D, but only 3  3 in 2-D. For the thermal strain programs,
in 3-D the last two entries of the displacement vector contain the six macrostrains, three
in each entry. In 2-D, this is also the case, but as there are only three macrostrains in all,
the second component of the last entry is not used in the program. The rest of the changes
between 2-D and 3-D are obvious, and are in the programs.
Analytically relating 2-D to 3-D elasticity can be a problem. Engineers usually always
think in 3-D, and to go to 2-D requires them to think in terms of plane strain (no z strains)
or plane stress (no z stresses). However, it is possible to set up the equations of elasticity in
2-D, independent of but analogous to 3-D. This is the approach used in the nite element
programs. The result looks like a plane strain approach. If one wishes to use the programs
as a plane strain or stress limit from three dimensions, then simply substitute the correct
moduli for the 2-D moduli in the programs.
The following formulas are used in the beginning of the nite element elastic programs,
and are worth repeating here, for isotropic phases. The subscripts indicate the dimensionality
of the modulus. K is the shear modulus and G is the bulk modulus. For isotropic elasticity,
the full elastic moduli tensor, Cij , can be expressed in terms of two independent constants.
These are usually taken to be either K and G, or E and  , the Young's modulus and Poisson's
ratio, respectively. The 2-D and 3-D relations interrelate the two pairs, as only two of them
are independent for isotropic elasticity. In 3-D,
E3
K3 = 3(1 ,
23 )
E3
G3 = (1 +
(39)
)
3

In 2-D,

K2 = 2(1E,2  )
2
E2
G2 = (1 +
2)
The inverse of these relations is also useful. In 3-D,
9 = 1 + 3
E3 K3 G3
K3 , 2G3)
3 = (3
2(3K + G )
3

26

3

(40)

(41)

and in 2-D,
4 = 1 + 1
E2 K2 G2
2 , G2 )
2 = ((K
K +G )
2

2

(42)

4.5 Actual program operation

This subsection gives details of how to actually use the various programs. Further details
are in the extensive comments in the programs themselves. In all the programs, the string
(USER) indicates a place where the user might have to modify the program to t his
particular problem. Possible changes include the value of conductivity or elastic moduli in
a phase, the number of phases, and the system size. The Gaussan quadrature program is
simple and so explained only in its commments. Sec. 7.7 gives a brief description of the two
percolation programs, with operational details left to the program listings, as these programs
are simple, too.

4.5.1 Electric and elastic nite element programs
The following gives some actual details of use for the nite element programs ELAS2D.F,
ELAS3D.F, ELECFEM2D.F, and ELECFEM3D.F. The eigenstrain programs have some
di erences in them, and will be discussed separately. The 2-D and 3-D programs are very
similar to each other, for both the elastic and electric cases, so that the discussion below
assumes 3-D.
The rst task is to assign the system size, by choosing the values of nx, ny, and nz.
These can be di erent from each other. The total number of nodes is, in the program,
ns = nx  ny  nz. The comments in the nite element programs show clearly which arrays
need to be dimensioned by the system size = total number of nodes. These variables can
be dimensioned at the same time with a global replacement, since the dimensions must be
changed in each subroutine as well as the main program.
The next task is to generate a digital-image structure in a separate program, or in the
subroutine PPIXEL. De ne i, j , and k as the indices running in the x, y, and z directions
in a right-handed Cartesian coordinate system. Once the structure has been generated, for
an image that is nx  ny  nz, the nite element programs assume it is stored in a le that
has been generated by the following piece of code:
DO 10 K = 1, NZ
DO 10 J = 1, NY
DO 10 I = 1, NZ
M = NX*NY*(K - 1) + NX*(J - 1) + I

1

WRITE(8,1) PIX(M)
FORMAT(i1)

27

10

CONTINUE

and reads the le in the same way in subroutine PPIXEL. In the programs, this le is called
MICROSTRUCTURE.DAT (the name of which can be set by the user). The two OPEN
statements in the beginning of the programs give the le names for the microstructure input
le (unit=9) and the results output le (unit=7). The values in pix(m) are phase labels,
starting at the value one, i.e., phase one has label 1, phase two has label 2, and so on, up to
the maximum value of nphase, which must be set by the user. If the user wishes to generate
a simple structure within the program, the place to do it is in subroutine PPIXEL.
The value of gtest is also assigned, and is used for determining when the gradient of the
elastic energy is small enough. One usually determines this using solutions of exactly known
composite problems (see Section 5), or operating directly on the problem of interest and
checking how the various calculated quantites, like e ective elastic moduli (conductivity),
local stresses (currents), etc. change with relaxation. The value of gtest should be determined
by numerical experimentation, for the speci c problem of interest. The only way to be sure
that gtest is small enough is to print out the values of the quantity desired and see how
many conjugate gradient cycles are enough so that this quantity is no longer changing. If
gtest is picked to be too small, much CPU time can be wasted in relaxing a system beyond
the point at which the desired answer no longer changes signi cantly.
Material properties have to be assigned to each distinct phase. For the elastic programs,
in the isotropic case bulk and shear moduli are assigned in the main program. The variable
containing the bulk and shear moduli for each phase is phasemod. In the anisotropic case,
the entire elastic moduli tensors must be inputted in subroutine FEMAT. In the electric
programs, the entire conductivity tensor must be supplied in the main program, for isotropic
or anisotropic cases. In ELAS3D.F, the values of the applied strain are assigned (exx, eyy,
ezz, exz, eyz, and exy) by the user, and in ELECFEM3D.F, the components of the applied
electric eld are assigned (ex, ey, ez) (similarly in 2-D).
Sometimes more than one microstructure will be considered in the same program, or the
same microstructure but with several di erent sets of properties will be considered. In this
case, the parameter npoints can be set to the appropriate value of iterations (micro =
1; npoints). One then gives an initial set of displacements (voltages). In ELAS3D.F
(ELECFEM3D.F), the initial displacements (voltages) are obtained by assuming that the
applied strain ( eld) is the same everywhere. It is sometimes advantageous, in the case
when several sets of phase properties are used on the same microstructure, for the initial
displacements (voltages) of one computation to be the nal values of the last computation.
This can save CPU time, and is easy to implement, by only going through the displacement
(voltage) initialization if micro = some xed value.
The subroutine FEMAT is then run, and sets up: the sti ness matrix dk for each kind
of pixel, the elastic moduli tensors cmod, and the various linear and constant terms needed
because of the periodic boundary conditions. Subroutine DEMBX then runs the conjugate
gradient algorithm to nd the answer to the problem that satis es the gtest criterion. The
user must supply the values (or use the default values) of two parameters, kmax and ldemb,
that control how dembx works. The value of kmax controls how many times DEMBX can
be called in a given computation, and ldemb tells how many conjugate gradient iterations
DEMBX will perform at each call. These two parameters can be adjusted so as to be
28

able to see how the relaxation is going. Each time DEMBX returns to the main program,
intermediate values of the average stress (current), the energy, and the value of gb  gb
are printed out, to serve as a check on how the relaxation is progressing. If the gtest
criterion has not been met, DEMBX is then re-entered to perform the next ldemb conjugate
gradient cycles. When relaxation has nally met the gtest criterion, subroutine STRESS
(CURRENT) is then used again to output the strain and stress ( eld and current) at every
pixel, in order to compute some kind of average or produce a map of local quantities.

4.5.2 Eigenstrain programs
The eigenstrain programs, THERMAL2D.F and THERMAL3D.F, are quite similar in many
ways to the elastic programs but have some signi cant di erences. The main di erence is
that in THERMAL3D.F, each phase can have an eigenstrain tensor, given by the variable
eigen. These are 6-vectors (Voigt notation), giving the eigenstrains in the x, y, and z and the
shear directions. Usually the shear terms are zero, as the processes that produce eigenstrains
do not usually produce shear strains. The programs relax the energy and treat the overall
size and shape of the system as being dynamic variables. The nal size (given macrostrains
in the x, y, and z directions) and shape (shear macrostrains) are given by the interaction
between the eigenstrains and elastic moduli of each of the phases. The initial displacements
of the problem, along with the macrostrains, can be all set to zero, or perhaps an initial
value of the macrostrains can be chosen, by guessing the nal answer, and the displacements
uniformly strained according to this chosen inital value. The conjugate gradient algorithm
is fairly robust, and should give the same answer for any reasonable initial values of the
variables.
Subroutine DEMBX, with the help of subroutines CONST and BGRAD for THERMAL3D.F and THERMAL2D.F, runs the conjugate gradient algorithm to nd the answer
to the problem that satis es the gtest criterion. Subroutine STRESS can then be used to
output the stresses and strains of the problem. The values of the macrostrains give the
overall expansion and shape change of the system. One should note that if all the shear
eigenstrains are initially zero, there will in general be a zero shear macrostrain found.

4.5.3 Finite di erence programs
The nite di erence electrical conductivity programs can be operated quite similarly to the
nite element electrical conductivity programs, with some small di erences given below.
(1) Since the nite di erence programs only handle diagonal conductivity tensors, the
local phase conductivity variable, sigma, is not a full tensor but only has space for the
diagonal elements.
(2) The subroutine DEMBX is not exited until the gtest criterion is met. The total number of conjugate gradient cycles possible within DEMBX is given by ncgsteps, a parameter
that is set by the user. Intermediate results are printed out every ncheck cycles, from within
DEMBX.
(3) The main variables, like the voltage vector u, are dimensioned ns2 = (nx +2)  (ny +
2)  (nz + 2), rather than ns = nx  ny  nz like in the nite element programs. Also, these
dimensions occur explicitly only in the main program, so only have to be changed there.
29

There are other di erences, but these are not in sections that must be changed by the
user. Sections 2 and 3 of this manual and the comments in the actual programs should be
consulted if the user wishes to know more about these di erences.

30

5 Exact solutions for testing programs
There are many exact solutions of composite problems that are quite useful for testing
programs like the ones in this package. Some of these are for special microstructures at
general volume fractions, some are for general microstructures with special choices of the
phase properties, while the others are for dilute microstructures, where a certain shaped
inclusion with di erent properties is introduced into an initially uniform matrix. A general
review of many of these dilute limits can be found in [11, 12, 13]. The relations most useful
for testing programs are given in this manual. These kinds of programs can also be useful in
exploring the dilute limit for inclusions with shapes that cannot be solved analytically [11].
The importance of having such exact solutions available for checking numerical computations should not be underestimated. It is usually easy to prepare numerical methods
for uniform regions. Even including a simple boundary, such as exists in series or parallel
problems, is not hard to do. But for random material problems, there are usually many
boundaries, of general shape. A proper test of a numerical method will include such boundaries. However, random problems usually have no analytical solution, leaving the problem of
how can a numerical result be assessed as to its accuracy? A numerical result for a random
system can seem perfectly reasonable, and yet be 100% wrong. The problems discussed in
this section give exact results for non-trivial boundaries and choices of phase moduli. These
exact solutions can then be used to rigorously test if a numerical method will give the correct
answer or not. These solutions can also be used to test things like the e ect of resolution
(number of pixels describing a microstructural feature) or nite size e ects, which result
when using periodic boundary conditions and a nite system size to simulate an in nite
system.

5.1 De nition of e ective properties

The main use of the programs described in this manual are to compute the e ective properties
of a multi-phase random composite. The e ective properties of a composite material are
de ned simply in terms of the averages of various quantities over the system. First of all,
for electric cases we have an applied eld, maintained via the periodic boundary conditions,
and in elastic cases, we have an applied strain, also maintained by the periodic boundary
conditions. Theorems from composite theory assure us that the electric eld average and the
strain average must equal the applied quantities. In other words, if h:::i indicates an average
over the entire system,
Z
(43)
hf (~r)i = V1 d3rf (~r)
then if, in the electrical case, E~ is the applied eld, the average of the microscopic eld,
which varies from pixel to pixel, is
hep(~r)i = Ep
(44)
while in the elastic case, a similar statement holds for the strain, where the applied strain is
E:
h (~r)i = E
(45)
31

In addition, an average can be performed over a single phase, denoted by
Z
1
hf (~r)in = V n d3rf (~r) = c1 hf (~r)(~r)i
(46)
n
n
where Vn is the volume of phase n, and cn is the volume fraction of phase n. The rst integral
is taken only over the volume of phase n, while the second integral is over the total volume,
with (~r) = 1 for phase n, and zero elsewhere.
There are two equivalent ways to de ne the e ective quantities: an energy method or an
average current/stress method. In the energy method, the total energy per unit volume is
equated to the energy per unit volume of a uniform medium. The equation used is:
(47)
En  21 E C eff E
in the elastic case, and
En = 12 Eppqeff Eq
(48)
in the electric case. Various applied elds and strains can be used to pick o various components of the e ective quantities.
Another way to de ne the e ective moduli or conductivity is through a stress average or
a current density average. For the electric case, the average current density is given by
hjpi = hpq eq i  pqeff Eq
(49)
which then de nes the e ective conductivity, as the average eld is just the applied eld.
For elastic cases, the equivalent expression for the average stress is
h i = hC  i  C eff E
(50)
which then de nes the e ective moduli tensor. The average stress and current formulations
are used in the programs, as one run can determine more of the e ective properties than in
the energy case, where there is only one number, the average energy, produced. The nite
element programs do output the energy, however, so that this method can be used. Good
review articles of composite theory include Refs. [14, 15]. For the rest of this manual, the
superscript eff is dropped, so that properties like conductivity and elastic tensors without
a phase label indicate e ective properties.

5.2 Series and parallel

The simplest exact composite relations that are sometimes useful are the well-known series
and parallel results for two di erent phases with volume fractions of x for phase 1 and 1 , x
for phase 2. If 1 and 2 are the isotropic conductivities of the two phases, then if the phases
are arranged in parallel , the e ective conductivity will be given by  = x1 + (1 , x)2 .
The eld in both phases will be uniform and equal to the applied eld. If the two phases
are arranged in series, then the e ective conductivity will be
"

 = x + (1 , x)
1
2
32

#,1

(51)

In this case, the eld in each phase will be uniform, but di erent. The current, however,
must be the same in both phases, as the direction of current ow is normal to the interface,
so that 1 hE~ i1 = 2 hE~ i2. With xhE~ i1 +(1 , x)hE~ i2 = E~ , it is simple to work out these elds,
which can then be used to check the phase average of the eld result for the program. It is
important that any numerical scheme be checked on more that just the series and parallel
problems, as these do not give realistic phase boundaries.

5.3 Small contrast of properties

When the di erence between the properties of a two phase material is small, a power series
expansion that can be made for the e ective properties in terms of this di erence becomes
useful. A conductivity result which is true for general two phase isotropic microstructures
has been derived by Brown [16] and extended by Torquato [14]. The e ective conductivity,
to second order in the di erence (2 , 1 ), is given by:
2
(52)
 = 1 + c2 (2 , 1 ) , d1 c1 c2 (2 , 1 ) + O(2 , 1 )3 + :::
1
where d is the dimensionality and ci is the volume fraction of phase i. The coecients for
the O(2 , 1 )3 and higher order terms involve details of the microstructure, and are given
in terms of various correlation functions over the phase geometry [14, 16].
A 22  22  22 pixel cube (phase 2) centered in a 30  30  30 unit cell was used to test
Brown's result for the case of small contrast between the two phase conductivities. For this
case, we have c1 = 0.39437, and c2 = 0.60563. We take 1 = 1 always, and vary 2 between
1 and 1:3. Figure 5 shows the result, plotted against 2 , where the quantity 1 + c2 (2 , 1 )
has been subtracted from both the numerical and theoretical results. This has been done
to show how the numerical results compare to the theoretical results in the quadratic term.
The numerical data follows the theoretical line very closely. As the value of 2 gets larger,
there should also be contributions from the cubic term, which could account for the small
di erences between the numerical and theoretical results. For cubic symmetry, the e ective
conductivity tensor is isotropic, so Brown's results can be used to analyze this system.
The same procedure can be carried out for the elastic moduli [17]. A simple way to
derive this result, at least up to second order in the modulus contrasts, is to take Hashin's
bounds [15], which bound the e ective properties between an upper and lower limit, and
expand them to second order in the modulus di erences. These bounds are known to be
exact to second order in the modulus di erences, and in fact the upper and lower bounds
agree exactly to this order. The 3-D results for the e ective bulk modulus K and shear
modulus G are:
K = K1 + c2(K2 , K1 ) , (K c+1c24 G ) (K2 , K1)2 + O(K2 , K1)3 + :::
(53)
1

3 1

G = G1 + c2(G2 , G1 ) , 25cG1c2((KK1++ 42GG1)) (G2 , G1)2 + O(G2 , G1)3 + :::

(54)

K = K1 + c2 (K2 , K1) , (K c1+c2G ) (K2 , K1)2 + O(K2 , K1 )3 + :::

(55)

In 2-D, the results are

1

1

1

3 1
1

33

3

Small contrast data on 30 system
3

Centered 22 cube
0.010

0.008

FEM
Brown (2nd order)

σeff

0.006

0.004

0.002

0.000
1.00

1.10

1.20

σ2

1.30

1.40

Figure 5: Showing  vs. 2 when 2 di ers only slightly from 1 . The points are nite
element data, and the straight line is Brown's exact expansion to second order in the contrast
(2 , 1).

34

G = G1 + c2 (G2 , G1) , c21Gc2((KK1 ++2GG1)) (G2 , G1 )2 + O(G2 , G1 )3 + :::
1

1

1

(56)

Similar results as those for the small contrast in conductivity case are obtained when
testing the accuracy of how well the nite element method computes the small contrast in
elastic moduli case. The bounds themselves [15] can also be useful for checking results, since
whatever e ective elastic moduli are found for an n-phase composite must lie between the
appropriate n-phase bounds. Just as in the electric case, the coecients for the O()3 and
higher order terms involve details of the microstructure, and are given in terms of various
correlation functions over the phase geometry [17].

5.4 Keller and Mendelsen 2-D result for conductivity

Another general conductivity result, due to Keller and Mendelson [18, 19], for a general 2-D
isotropic two phase microstructure, is that

(1 ; 2 )(2; 1 ) = 1 2

(57)

where (1 ; 2) means the e ective conductivity obtained when phase 1 has conductivity 1
and phase 2 has conductivity 2 , and (2; 1 ) means the phase conductivities have been
interchanged before determining the e ective conductivity.

5.5 Field uctuation result

For isotropic two component mixtures, in 2-D or 3-D, there are exact results connecting eld
uctuations and the e ective conductivity [20]. For elastic problems, one can also exactly
relate strain averages to the e ective bulk and shear moduli [20], though this topic is not
discussed in this manual.
For conductivity, one must rst formulate the average over a phase:
hf ij = c1 hj f i
(58)
j
where j is equal to 1 inside phase j , and zero elsewhere, and cj is the volume (area) fraction
of phase j; j = 1; 2. Plain brackets indicate an average over the whole system.
The exact relation can then be stated simply:

hE 2ij = 1 @
hE 2 i cj @j

(59)

where again  is the e ective conductivity. If  is known analytically as a function of j ,
then this di erentiation can be readily carried out. If not, this derivative can always be
evaluated numerically, by evaluating  for j  , with  being a small number.
Consider a system where  is analytically known. The checkerboard microstructure, as
shown in Fig. 6, can be evaluated by the Keller-Mendelson formula given above. Since the
microstructure is clearly invariant under inversion of 1 and 2 , the e ective conductivity
must then be  = (1 2 )1=2 .
35

Figure 6: Showing the checkerboard microstructure, with dark gray being phase 2 and light
gray phase 1.
The exact formula then gives hhEE2ii2 = (1 =2)1=2 . Fig. 7 shows the numerical results,
compared to the exact theoretical results, for a 128  128 system for a variety of conductivity
ratios. Excellent agreement is shown, with some systematic disagreement growing at larger
values of the conductivity di erence.
If we x the conductivity ratio between the two phases to be 10, we can then examine
how well the eld averages and e ective conductivity are computed as a function of system
size L  L. Table 8 gives the data obtained.
2

5.6 Equal shear modulus

A result that can be very useful in testing elastic programs is the equal shear modulus result,
true for any microstructure [21, 22]. If there are only two phases present, with equal shear
moduli G but di erent bulk moduli K1 and K2, then the e ective shear modulus of the entire
system is just G, and the e ective bulk modulus K of the system is, in 2-D,
c2K2 ) + K1K2
(60)
K = G(c1GK+1 +c K
1 2 + c2 K1
and in 3-D,
4 G(c K + c K ) + K K
K = 3 4 1G +1 c K2 +2 c K 1 2
(61)
3

1 2

36

2 1

Quantities

3.00

2.00
<σ>/σ1
2
2
2/
Exact
1.00

0.00
0.0

2.0

4.0

6.0

8.0

σ2/σ1

10.0

Figure 7: Showing the e ective conductivity and average of the electric eld magnitude
squared in phase 2 for the checkerboard, as a function of 2 . The points are numerical nite
element results, and the lines are the exact results discussed in the text. The system size
was 128  128.
L

=1(FEM) =1 (Exact) % Error

hE 2 i2
hE 2 i

16
32
64
128
256
512
1024

3.4442
3.3232
3.2550
3.2159
3.1934
3.1804
3.1728

0.4132
0.3791
0.3569
0.3424
0.3330
0.3269
0.3230

3.1623
3.1623
3.1623
3.1623
3.1623
3.1623
3.1623

8.9
5.1
2.9
1.7
1.0
0.6
0.3

2
(FEM) hhEE2ii2 (Exact) % Error

0.3162
0.3162
0.3162
0.3162
0.3162
0.3162
0.3162

30.7
19.9
12.9
8.3
5.3
3.4
2.1

Table 8: Size dependence (L  L system) of e ective conductivity and eld average for
checkerboard, as determined by nite element method. Each individual "check" of the
checkerboard is L=2  L=2. The conductivity ratio 2 =1 = 10.
37

where c1 and c2 are the volume fractions of the two phases. Note that the only di erence
between the two equations is the 4=3 factor in front of the shear modulus G. In 2-D, the
result is somewhat simpler if expressed in terms of the Young's modulus E and the Poisson's
ratio  , E = c1 E1 + c2E2 , and  = c11 + c2 2 [22], which actually looks like a parallel result
but is valid for general two phase microstructures.
We choose the same cubic image as was used for the small contrast in conductivity case
to test the equal shear moduli result. A 22  22  22 pixel cube (phase 2) is centered in
a 30  30  30 unit cell, so that the volume fractions are: c1 = 0:60563 and c2 = 0:39437.
The two shear moduli are taken equal to unity, G1 = G2 = 1, and the two bulk moduli are
K1 = 1, and K2 = 20. The exact answer, according to eq. (53) is K = 2:263250, while the
numerical answer is K = 2:269684, a di erence of only 0:3%.
In 2-D, a test of the equal shear moduli result can be combined with a test of the e ect
of digital resolution by considering the checkerboard problem. The shear moduli are both
equal to 2, and there is a ratio of 10 between the bulk moduli (K1 = 1, K2 = 10) (see Fig. 6
for a picture of the microstructure). The e ective bulk modulus is computed as a function
of system size L  L, where each "check" is L=2  L=2. The exact value of K is 2:8 using
eq. (60). Table 9 displays the data from this test. For L  16, the error in the e ective bulk
modulus is about 1% or less.
System size K

% di .

2
4
8
16
32
64
128
256
512

96.4
16.8
4.9
1.5
0.4
0.1
0.04
0.01
0.004

5.500
3.271
2.937
2.841
2.812
2.804
2.8012
2.8003
2.8001

Table 9: E ective bulk modulus for checkerboard{equal shear moduli case.

5.7 Intrinsic properties for spheres, circles, and cubes

A useful dilute limit is that of spherical (circular) inclusions randomly distributed in a matrix.
To be in the dilute limit, the volume fraction of the inclusion phase should probably be less
than or equal to 5%, although this limit can be higher or lower, depending on inclusion
shape and contrast of properties with the matrix. Call the inclusion phase 2. In general we
nd that, for some property being considered, say F , the property in the dilute limit has the
38

form

F = f1 (1 + [F ]c2 )
(62)
where [F] is called the intrinsic property [11], and F = f1 when c2 = 0. [F] is a function of
the shape of the particle, and the contrast between its properties (f2) and the properties of
the matrix (f1 ). This is true for any shape inclusion that has been averaged over orientation.
For the conductivity problem, for spherical inclusions in 3-D, the intrinsic conductivity []
is [11]
2 , 1 )
[] = 3(
(63)
(2 +  )
1

and the intrinsic elastic moduli are given by [23]


2





K1 + 43 G1 KK21 , 1
[K ] =
K2 + 43 G1

(64)





5 K1 + 34 G1 (G2 , G1)


(65)
[G] =
3G1 K1 + 89 G1 + 2G2 (K1 + 2G1)
For circular inclusions in 2-D, the corresponding equations for the conductivity and for the
elastic moduli are [24]
(66)
[] = 2((2+,1))
1



2



(K1 + G1) KK12 , 1
(67)
[K ] =
K2 + G1
[G] = 2 (K1 + G1 ) (G2 , G1 )
(68)
G1 (K1 + G1) + G2(K1 + G1 )
Note the similar structure between these sets of formulas in 2-D and 3-D.
To test a nite di erence or nite element program, one generally puts one particle in
the matrix, with a unit cell-size chosen so that the volume fraction is small enough to be in
the dilute limit. Even with periodic boundary conditions, the conductivity term or elastic
moduli term that is linear in the volume fraction of the inclusion phase is unchanged from
the in nite system limit. Refs. [11] and [12] contain electric, elastic, and viscosity dilute
limits for many other shape particles.
For the sphere, we use a 40  40  40 unit cell, and a 15 pixel diameter sphere, centered
on the middle of a pixel, so that the sphere volume fraction is 0:028. Figure 8 shows the
results for the intrinsic conductivity. The solid line is the exact solution, eq. (63). For x
close to zero, it appears that the nite element is more accurate, while for x >> 1, the nite
di erence seems to be more accurate. To test the e ect of resolution on the result, the results
at x = 10, the case with the worst accuracy, were re-run for a 31 pixel diameter sphere in
a 1003 unit cell. This made the volume fraction slightly smaller, which would also tend to
slightly improve the results. The numbers in Table 10 in parentheses are these results. For
a factor of 2 improvement in resolution, the absolute error in [] decreased by a factor of
three for the nite element method, and by a factor of eight for the nite di erence method.
39

Intrinsic conductivity for a sphere
3

d=15 sphere in 40 unit cell
3.0

2.0
FEM
FD
Exact

[σ]

1.0

0.0

−1.0

−2.0 −2
10

−1

10

0

10

1

10

2

10

x
Figure 8: Intrinsic conductivity for a 15 pixel diameter sphere embedded in a 403 unit cell, as
a function of the ratio of the sphere conducticity to the matrix conductivity. Finite element
and nite di erence data and the exact result are compared.

x

FEM

j % Error j FD

j % Error j Exact

0.1
0.2
0.5
2
5
10

-1.342
-1.121
-0.6029
0.7653
1.861
2.576 (2.353)

4.4
2.8
0.5
2.0
8.6
14.5 (4.6)

7.5
5.0
1.4
1.0
5.7
9.8 (1.2)

-1.382
-1.146
-0.6085
0.7577
1.812
2.471 (2.278)

-1.286
-1.091
-0.600
0.750
1.714
2.25

Table 10: Intrinsic conductivities for sphere, comparing the nite element and nite di erence
techniques.
40

The result for the nite di erence case only di ers by 1.2% from the exact value of 2:25 in
this case.
For the same system, we can also calculate the intrinsic bulk and shear moduli. Choosing
K1 = 1 and G1 = 0:75, and keeping, for simplicity, K2 =K1 = G2 =G1 = x (so that the same
Poisson's ratio is maintained in inclusion and matrix), it turns out that using eqs. (64) and
(65), both intrinsic moduli are the same and are equal to 2(x , 1)=(x + 1). Figure 9 shows
the numerical and exact results for both intrinsic moduli, for a 15 pixel diameter sphere in a
403 unit cell. The actual values are given in Table 11. The errors are qualitatively similar to
the intrinsic conductivity case. The x = 10 point was rerun for the 31 pixel diameter sphere
in a 1003 unit cell, and the result is shown in Table 11 in parentheses. For both the intrinsic
bulk and shear moduli, the absolute error decreased by almost exactly a factor of two for
the factor of two increase in resolution. There was little di erence in accuracy between the
computation of [K ] and [G].

Intrinsic elastic moduli for a sphere
3

d=15 sphere in 40 unit cell
2.5

[K], [G]

1.5

0.5

[K]
[G]
Exact

−0.5

−1.5

−2.5 −2
10

−1

10

0

10

1

10

2

10

x
Figure 9: Intrinsic elastic moduli for a 15 pixel diameter sphere embedded in a 403 unit cell,
as a function of the ratio of the sphere Young's modulus to the matrix Young's modulus.
The three sets of data show [K], [G], and the exact result, which is the same for both intrinsic
moduli.
In 2-D (to limit computational e ort), we can further analyze the resolution dependency
of the nite element calculation of the intrinsic elastic moduli. A circle is placed in a
41

x

[K]

j % Error j [G]

j % Error j Exact

0.1
0.2
0.5
2
5
10

-1.612
-1.31
-0.658
0.682
1.414
1.774 (1.706)

1.5
1.8
1.3
2.3
6.0
8.4 (4.3)

1.5
0.4
0.7
1.7
4.9
7.3 (3.7)

-1.612
-1.328
-0.662
0.678
1.398
1.755 (1.697)

-1.636
-1.333
-0.667
0.667
1.333
1.636

Table 11: Intrinsic elastic moduli for sphere, d = 15 in 403 system
computational cell, with a diameter of 1=10'th the cell size. The intrinsic moduli for the
bulk and shear moduli are then computed as a function of cell size. The circle moduli are
ten times larger than those of the matrix, with the same Poisson's ratio. Table 12 shows the
data for this computation.
System size [K]

% di . [G]

% di .

20
40
80
160
320
640

10.1
10.1
4.9
2.9
1.9
1.3

6.9
12.9
4.6
2.3
1.5
0.9

1.548
1.548
1.475
1.447
1.433
1.425

1.318
1.362
1.290
1.261
1.251
1.244

Table 12: Intrinsic elastic moduli for circle{e ect of digital resolution, using nite element
method
Another test of the accuracy of both the nite element and nite di erence electrical
programs is to run the example of the dilute limit of a cubical shaped inclusion. Since cubic
pixels can replicate the shape of a cubical inclusion exactly, then any e ect of a digitally
rough boundary (imperfect representation of the geometry) can be dispensed with. Eyges
[25] has done a very careful numerical treatment of this problem, which can be used as a
check on these programs. We use a 10  10  10 cube, centered in a 40  40  40 unit cell,
so that the cube volume fraction was 0:0156. The conductivity of the matrix, 1 , is taken
to be unity, with the conductivity of the inclusion, 2 , ranging between 0:1 and 10. For a
cubic symmetry system, the conductivity tensor is isotropic, so there is only one independent
diagonal component, and no o -diagonal components. Let x = 2 =1 . Figure 10 shows the
42

results. It is interesting to note that, for high values of x, the nite di erence result actually
tends to be a bit more accurate than does the nite element result. It is known that in the
limit of very high values of x, the nite element program tends to be about 8-9% high for
the dilute limit of cubes [11]. The formula used to plot Eyges' [25] data was

:486(x , 1)2 + (x , 1)
[(x)] = 1 + 00:820(
x , 1) + 0:143(x , 1)2

(69)

Intrinsic conductivity for a cube
3

3

10 cube in 40 unit cell
3.0

2.0

[σ]

1.0

0.0

−1.0

−2.0 −2
10

−1

10

0

10

1

10

2

10

x
Figure 10: Intrinsic conductivity for a 103 cube embedded in a 403 unit cell, as a function of
the ratio of the cube conductivity to the matrix conductivity. The three sets of data (circle,
square, line) compare the nite element method, the nite di erence method, and Eyges'
[25] data, respectively.

5.8 Vegard's law and Goodier result for thermal strains

When thermal strains (eigenstrains) are present, there are several other sets of exact results
available to check the output of the programs THERMAL3D.F and THERMAL2D.F. Suppose rst that for all n phases, each phase has the same elastic moduli but di erent isotropic
thermal strains, with the thermal strain in each direction being ei for the i'th phase. Then
the overall system strain in any direction will be simply  = ci ei , i = 1; n. This result is
43

known in the physics community as Vegard's law [26]. This result is true for any microstructure, in both two and three dimensions. Actually, the thermal strain tensor need not be
isotropic. The overall thermal strain tensor will be the sum of the volume fraction-weighted
phase thermal strain tensors.
A related result, not as well-known, by Goodier [27] applies to this same case where there
are only two isotropic phases. In this case, the trace of the strain in the inclusion phase (3D) is simply
(1 + 3 )
(70)
Tr  = e (1
, 3 )
where the thermal strains in any direction are ei = e, the trace does not include thermal
strains, and 3 is the 3-D Poisson's ratio. The strain trace is constant inside the inclusion.
In 2-D (plane strain) the result is
Tr  = e(1 + 2) = (1 ,e  )
(71)
3
where 2 is the 2-D Poisson's ratio. Equations (39)-(42) can be manipulated to show the
relationship between 2 and 3 .
The trace of the stress in the inclusion is then also constant, and is simply given by (3-D)
"
#
(1
+

)
3
Tr  = 3K3(Tr  , 3e) = 3K3e (1 ,  ) , 3
(72)
3
and in 2-D,
Tr  = 2K2(Tr  , 2e) = 2K2e[(1 + 2 ) , 2]
(73)
where K3 and K2 are the 3-D and 2-D (plane strain) bulk moduli, respectively.
It is known that if an elliptical (2-D) or ellipsoidal (3-D) inclusion is introduced into a
material, where the inclusion and matrix have the same elastic moduli but di erent thermal
strains, the individual components of the stress tensor are uniform inside the inclusion [7, 27].
Recently, a conjecture has been made that, in 2-D, if the inclusion is in the shape of a
equilateral 5-pointed star, the stress tensor components inside the star will also be uniform
[28]. The Goodier result [27] of course predicts that at least the trace of the stress tensor
will be uniform in any inclusion.
This conjecture about a 5-pointed star can be tested using the program THERMAL2D.F.
A digital image can be made of an inclusion in a matrix, where the inclusion has arbitrary
shape. With the properties assigned, the system can be relaxed and a stress picture can
be made, along with a picture of the trace of the stress tensor. In most case, it should be
obvious whether a quantity is uniform within the inclusion or not. In addition, the average
stress and its standard deviation can be computed within the inclusion. If a quantity is
indeed uniform, then the standard deviation should be much smaller than the average, and
re ect only things like nite size e ects, and using square pixels to approximate curved and
angular boundaries.
The top images in Fig. 11 show an elliptical inclusion, with an aspect ratio of three.
The stress map on the left is of ,xx , where the stress levels, from high to low, are redgreen-white-black (top to bottom in the accompanying color bar). This component of the
stress appears to be uniform inside the inclusion, as it is supposed to be. The 2- D moduli
44

Figure 11: Thermal stress for inclusions with the same elastic moduli as the matrix, but
di erent thermal eigenstrain. Stress is high (top) to low (bottom) in color bar). Left: ,xx ,
right: negative of the trace of the stress tensor. Images from top: Ellipse, 5-pointed star,
6-pointed star.
used were: E2 = 1, 2 = 0:3, K2 = 5=7, and G2 = 5=13, in arbitrary units. The inclusion
had a thermal strain of exx = eyy = 1, and the matrix had zero thermal strain. The right
top image shows the negative trace of the stress tensor, which is also clearly uniform inside
the inclusion. The stress in the inclusion is negative (compressive), since a positive thermal
strain, constrained by the outer matrix, results in a compressive stress inside the inclusion
(see Section 2.4).
The left image in the 2nd row from the top of Fig. 11 shows a map of ,xx , with the
same choice of elastic parameters, for a 5-pointed star-shaped inclusion. This component of
the stress tensor is clearly non-uniform in the inclusion. The right hand image shows the
negative trace of the stress tensor, which certainly appears uniform, according to Goodier's
result. With the choice of elastic parameters used, the trace of the stress tensor is predicted
to be, inside the inclusion, equal to ,1. For a 512  512 unit cell, and the area fraction of the
45

star being 7.6%, the trace of the average stress tensor inside the inclusion is ,0:93, in error
by only 7%. This di erence re ects the nite extent of the system, since the Goodier result is
for an in nite system, and errors at the star boundary, whose straight lines are approximated
by a digital boundary. A similar nite element computation was also carried out recently
[29], with similar results. The remaining row in Fig. 11 shows analogous computations for
a six-pointed star, with results similar to that of the 5-pointed star.

5.9 Hashin and Rosen thermal strain result

A result by Hashin and Rosen [30], which is true for both 2-D and 3-D, exactly relates, for
the case of two phases, the e ective bulk modulus K and the e ective thermal strain e when
the two phases do not have the same elastic moduli. The isotropic thermal strain for phase
i is ei, and the e ective isotropic thermal strain for the composite system is denoted e. The
result for the value of e is


c
c
1
(
e
,
e
)
1
2
2
1
(74)
e = c1 e1 + c2 e2 +  1 1  K , K , K
1
2
,
K2 K1
and is true for isotropic components and an isotropic composite. There are also versions of
this equation valid for anisotropic components and composites [30]. It is easy to combine
this equation with that for the dilute limit bulk modulus for spherical inclusions to get the
dilute limit thermal strain for a spherical inclusion that has di erent elastic properties than
does the matrix. This result, for an inclusion labelled "2" embedded in a matrix labelled
"1", is
K2 (K + 4 G )(e , e )
[e] = K1 1 3 14 2 1
(75)
(K2 + 3 G1)
To illustrate the accuracy with which the nite element programs obey the Hashin-Rosen
result, eq. (74), a 200  200 pixel system was set up. A total of 160 overlapping circles,
15 pixels in diameter, were thrown down randomly, such that the matrix phase had an area
fraction of c1 = 0:4779, and the inclusion phase had an area fraction c2 = 0:5221. The
matrix phase had E1 = 1:0, 1 = 0:2, and a thermal strain of 0:1 in the x and y directions
(e = 0:1). The inclusion phase had E2 = 0:3, 2 = 0:1, and a thermal strain of 0:4 in the x
and y directions (e = 0:4). Using the Hashin-Rosen equation, an overall expansion of 0:2109
is predicted. The program THERMAL2D.F gave a value of 0:2103 in the x- direction, 0:2112
in the y-direction, and an average expansion of 0:2108. This average value is within 0:05%
of the exact value. The small di erence between the x and y directions (< 0:5%) is due to
the fact that this fairly small numerical system is not exactly isotropic.

5.10 Mackenzie result for pressurized pore space

A useful exact result has been derived by Mackenzie [31], for an isotropic porous solid lled
with a gas or liquid at a pressure p. The expansion of the material is given exactly by an
equation involving only the bulk modulus of the solid frame, Ks , the e ective bulk modulus
of the porous solid, K , and the pressure p. This is analogous to the Hashin-Rosen result,
46

where only the e ective bulk modulus of the composite had to be known in order to be able
to compute the e ective thermal expansion. The strain  is exactly given by:


1
1
p
= 3 K ,K
(76)
s
In 2-D, simply replace the "3" by "2" in eq. (73), and use the 2-D bulk moduli (see Sec. 6.3
for suggestions on implementing this computation and a test using Mackenzie's result).

5.11 CLM Theorem

Recently, a new exact theorem has been proven for elastic composite problems in 2-D. There
can be general anisotropy, and any number of phases, with perfect bonding between phases
[22, 32, 33]. For this manual, we will consider only isotropic phases. Assume that the i'th
phase has bulk and shear moduli Ki and Gi, and the overall e ective moduli are K and
G. Now make a transformation in each phase of the following form: (1=Ki) = 1=Ki , C ,
(1=Gi) = 1=Gi + C , where C is an constant such that the new elastic moduli in each phase
are still positive. Then the theorem says that the new e ective moduli, K and G , are given
in terms of the old moduli by (1=K ) = 1=K , C , and (1=G) = 1=G + C . This theorem can
then be used to test elastic programs like ELAS2D.F.
The same microstructure can be used as in the Hashin-Rosen example, Sec. 5.9. The
phase moduli were originally K1 = 1, G1 = 0:5, K2 = 3, G2 = 1. The values of the
e ective moduli using these phase moduli are: K = 1:6627, G = 0:711, obtained using the
program ELAS2D.F. Using a value of C = 0:2, the transformed phase moduli are K1 = 1:25,
G1 = 5=11, K2 = 7:5, G2 = 5=6, and so the transformed e ective moduli are expected to
be: K = 2:492, G = 0:6225. The numerical values of the moduli, again obtained using
ELAS2D.F, are: K = 2:499, G = 0:6092, an error of only 0:3% in K, and 2:1% in G. This
error is a direct measure of how well pixels can resolve the true interface, since the CLM
theorem assumes a perfect interface.
0

0

0

0

0

0

0

0

0

47

0

6 Other possible uses of programs
This section describes some di erent ways that the nite element and nite di erence programs can be used. These uses are not, however, implemented in the programs discussed in
this manual, but are meant to give ideas to potential users of these programs.

6.1 Fixed voltages and displacements

It is relatively easy in these programs to run a problem with a xed displacement or a xed
voltage. To x a given nodal variable at a certain value, simply give it that value initially,
and then every time the gradient, gb, is calculated or changed, zero out the gradient with
respect to this variable, so that it can never change. In the nite element programs, this
needs to be done twice, once in ENERGY, after the gradient is rst computed, and then
also in DEMBX, after the gradient is updated using the Ah array, which results from the
matrix multiply in DEMBX. In the nite di erence programs, this zeroing out of gb is done
only in DEMBX, since there is no ENERGY subroutine in the nite di erence programs.
This technique can be used in THERMAL3D.F to be able to run an applied strain. One
simply sets the initial values of the macrostrains to their desired values, and then zeroes
out the gradient vector entries corresponding to these macrostrains. The macrostrains will
then stay constant, and the solution will be obtained for an applied strain problem where
eigenstrains can be speci ed.
For an electrical example of this process, see Ref. [11], where the intrinsic conductivity of
superconducting centro-symmetric inclusions was computed, using the nite element method,
by xing the voltage on the nodes belonging to the inclusion and its surface. This procedure
simulated a superconducting inclusion problem, in which the inclusion had no resistance and
was therefore an equipotential body. In this problem, the voltage of the inclusion was set to
zero at all the nodes composing the inclusion and its surface. A eld was applied via periodic
boundary conditions, and the rest of the voltages were then found that minimized the energy
of the problem. The e ective conductivity of the problem was found using the total energy,
not the average current, since the current through the superconducting inclusion was not
computed using this method. This was a case where the e ective conductivity could only be
de ned via the energy.
An amusing electrical example where this technique can be used successfully is from
an important problem in statistical physics, where it is desired to nd the total resistance
between two nodes in an in nite lattice. Each node is connected to its nearest neighbors
by unit resistors. This problem has been worked out completely for the in nite 2-D square
and the 3-D cubic lattice [34], analytically. Other papers that have looked at this problem
numerically and analytically include [35, 36], which include tables of results.
The method of solution is usually to introduce a current source at the origin, and then
an equal current sink at a lattice point of interest. The current is xed, and the voltage
between the two nodes is then calculated. The resistance between the two points is simply
the ratio of voltage to current introduced, according to Ohm's law. The same problem can be
done numerically, using the nite di erence method, by xing the voltages at the two nodes,
computing the current in at the one node and out at the other node (which will be equal and
opposite), and then computing the e ective resistance between the two points via the ratio
48

of xed voltage di erence to computed current. The conductivity chosen for the material
being used is arbitrary, since the e ective resistance will be proportional to the resistance
of a single pixel. We compare to the analytical result for the 2-D square lattice and the
3-D simple cubic lattice, for two nodes separated by 20 lattice spacings, where all results are
scaled by the resistance of a single pixel. In 2-D, the exact result for an in nite square lattice
is 1:468191. The numerical result, for a 200  200 resistor network with periodic boundary
conditions, is 1:473655, an error of only 0:4%. The exact result for an in nite 3-D simple
cubic lattice is 0:497499, and the numerical result is 0:497905, for a 200  200  60 lattice.
The error in this case is only 0:1%.
Formulas and tables of numerical results for checking other spacings can be found in
Refs. [34, 35, 36]. The integrals in Ref. [34] for the exact solutions can be easily evaluated,
with high accuracy, using Gaussian quadratures. A program for generating the Gaussian
points and weights for an N-point Gaussian quadrature evaluation is stored with the various
programs described in this manual, and is called GAUSS.F. Recall that Gaussian quadrature
is only for integrals between ,1 and 1. A general integral needs to be changed into this form
before using Gaussian quadrature [37]. A listing of GAUSS.F can be found in Sec. 9.3.6.
It is interesting to observe how the 2-D result for the resistance gradually becomes equal
to the 3-D result, as the lattice grows in size in the vertical dimension. Figure 12 shows the
resistance between nodes that are 20 lattice spacings apart, for lattices that were 200  200 in
the horizontal direction, and from 0 (2-D) to 60 spacings thick in the vertical direction. By
the time that the lattice is about as thick as the spacing between the two nodes of interest,
the e ective resistance has converged to within a few percent of the 3-D result.
Another application that arises in electrical problems is that of a piece of metal, usually of
some odd shape, buried in or attached to a piece of material with nite conductivity. A eld
is applied from electrodes at either end of the sample, and one wishes to know the resulting
voltage of the buried piece of metal. The electrodes have xed voltages. The voltage must
be the same throughout the metallic inclusion, but its value is free to oat according to its
shape, size, and position in order to solve the problem. One physical example would be an
impedance spectroscopy experiment, where the current is applied via a working and counter
electrode, and one wishes to know the voltage of a reference electrode that has been attached
to or buried in the sample [38]. One way to handle this situation is to make a list of which
nodes need to be at the same (but oating) voltage. The simplest starting point is to make
all these voltages equal to zero. The conductances between the nodes in the inclusion can
be made equal to one, for simplicity, though they will not really appear in the answer. Any
positive number will suce. Then, in every conjugate step where the gradient is computed,
the gradient for all this list of nodes should be made equal to the average of all the individual
gradients. In this way, the voltage of all these nodes will be forced to be always the same,
simulating a metallic inclusion, but can gradually change in order to match the elds outside
the inclusion. The correct solution will eventually be reached in the relaxation process.

6.2 Removing periodic boundary conditions

There will be times when one might not want to use the periodic boundary conditions that
are built into the programs. It is simple to get rid of them. Simply use a unit cell that is
(nx + 1)  (ny + 1)  (nz + 1), and set the i = nx + 1, j = ny + 1, and k = nz + 1 pixels
49

2.0

1.6

Reff/R

1.2

0.8

0.4

0.0

0

10

20

30

40

50

60

Thickness L
Figure 12: Showing the resistance between two nodes, separated by 20 lattice spacings, of a
simple cubic network, normalized by the resistance of one bond, as a function of the thickness
of the network.

50

to have zero conductivity or modulus, i.e. "air." In the nite element programs, this will
e ectively zero out the linear and quadratic term in the energy, which arise from the periodic
boundary conditions. Terms that connect variables across the system will be multiplied by
zero and thus contribute nothing. The program can then be run as is, but of course one must
allow for the layer of air on the outside in the eld averages. Also, one must put some kind
of constraint on the boundary in order to have a non- zero result, since the applied electric
eld or strain came from the boundary conditions. One easy way to apply a uniform strain
in an elastic problem is to use THERMAL3D.F, and allow every pixel to have the same
eigenstrain. The system will then expand, but there will be internal relaxations because of
the di erent moduli between phases. This is equivalent to applying an external strain [7].
One can also x the displacements of the outermost pixels, to mimic the desired applied
strain. In electrical problems, one can x the voltage at one or more places. Other ways
include putting a xed force or current somewhere in the system, which is discussed next.
Usually when applying these programs to a real 2-D image, it will be desirable to remove
periodic boundary conditions. Also, one will want to change an image, usually in gif or tif
or jpeg or some other graphics format, into an image of phase labels. The hardest part is
to convert the graphic image into an image of ASCII numbers. These number can then be
easily read and converted into the appropriate phase labels, either outside the programs or
within subroutine PPIXEL. The shareware program XV, easily obtainable on the Internet,
has an output le format called pgm or ppm. This is an ASCII format of either gray scales
from 0 to 255, or RGB triples, each ranging from 0 to 255. Simply read the image with XV,
then output an ASCII le, which is then easily converted to phase labels using a separate
small program.

6.3 Fixed currents and forces

Since the current in a resistor is just the conductance times the voltage di erence between the
two ends, it is somewhat easier to illustrate how to use xed currents in the nite di erence
algorithm. Consider the problem discussed in Section 6.1, that of the e ective resistance
between two nodes of a lattice, in 2-D or 3-D. There the problem was solved by xing the
voltage di erence between the two nodes, and then computing the current in and out the
two nodes. The same problem can also be done by xing the current in and out at the
two nodes. The gradient of the energy at a node is just the negative of the current at that
node. If we add 1 and ,1 to the gradients at the two nodes before relaxation occurs, this
will be like forcing a xed current at these nodes. The voltages will relax, so as to force the
total gradient to become zero, but with the current at the nodes xed at its original value,
since there are no other current or eld sources in the problem. The same solution to the
2-D problem, with some slight round-o di erence, was found as in Section 6.1. The same
technique could then be used if xed current boundary conditions, instead of periodic xed
voltage conditions, were desired. The currents at the boundary could be xed using this
method.
In the nite element programs, the negative gradient of the elastic energy with respect
to a nodal displacement is just the force at that node. In the same way, in the subroutine
ENERGY, after the gradient is rst computed, a xed value can be added to the gradient
at a particular node, to simulate a xed force at that node. If the subroutine ENERGY is
51

then run again, after LDEMB conjugate gradient cycles in DEMBX, this addition should be
renewed before going back into the conjugate gradient routine.
Usually the e ect of xed stresses is considered, not the e ect of xed forces. In this case,
if for example a pressure p is to be put on a surface, the way one does this is to consider
each pixel face, and put a force of p=4 on each node of each face, in 3-D. The edge length
of the pixels is one. That way, nodes that are shared by four pixel faces will end up with
p, nodes that are only shared by two pixel faces will have p=2, and nodes that are only on
one pixel face, will have p=4. This takes care of overcounting. Consider four (2  2) pixel
faces in a plane, containing 9 nodes. The total area is four, if the pixels have unit length.
If a force p was put on each node, then the total pressure on the face would be 9p=4, not
p. However, if the method suggested is used, then the middle node will have a force p, the
four side nodes will each have a force p=2, and the four corner nodes will each have a force
p=4, for a total pressure of 4p=4 = p. If the pixels are in a 1  4 arrangement, then the four
corner nodes will have a force p=4, and the six middle nodes will each have a force of p=2,
for a total pressure of 4p=4 = p.
This use of xed pressures can be tested using eq. (76), Mackenzie's exact result for an
isotropic porous solid lled with a uid at pressure p. If we consider a porous material with
a dilute concentration of spherical holes, we can combine eq. (64) for the bulk modulus of
such a system with Mackenzie's result to get an exact prediction for the strain expected at
a pressure p. Alternatively, we can directly compute, using ELAS3D.F, the e ective bulk
modulus of the porous solid. For a 253 unit cell, containing a single spherical hole of diameter
11 pixels, when the solid frame bulk modulus was 30:0825, the e ective bulk modulus of the
porous material was 27:3909 as computed by ELAS3D.F. When p = 0:1, the prediction from
Mackenzie's result is  = 1:08882  10,4. The numerical result, using the algorithm described
above, was  = 1:08888  10,4, a di erence of only 0:006%. So this algorithm works very
well.
One warning: often pores will go across the boundary when periodic boundary conditions
are used. In this case, this algorithm does not seem to work very well [41]. An alternate
method can be to enclose the porous solid with a "skin" of e ective material with elastic
properties such that the bulk modulus of the composite system is unchanged from that of
the porous solid alone. This algorithm does work on interior pores. When a skin is placed
around the material, all pores become interior pores.

6.4 Surface energies

If there is a surface energy associated with a material, and the material has a very large
surface area, a change in that surface energy can cause a measurable expansion or shrinkage
of the material. This has been seen in Vycor glass, at very low relative partial pressures of
an absorbed gas [39, 40]. As the partial pressure of the absorbed gas increases, the layer of
gas molecules on the internal surface of the Vycor increases in thickness, and thus lowers
the speci c surface free energy. This lowering in surface energy allows an expansion to take
place, which trades volume elastic energy for surface free energy. This expansion is linear in
the change in surface free energy [41, 39, 40].
If a 3-D model of the material is available in a digital form, then a surface energy can
be applied in nite element form. All surfaces are pixel faces, and are initially at. One
52

then must write, for general small strains of the pixel face, what the new area is in terms
of the nodal displacements. One can then linearize this when the surface distortion is small
(usually the overall strains are a percent or much less), to give an energy that can be added
to the elastic energy, which is linear in the displacements.
The key to deriving a surface energy, SA, where SA is the surface area of the digital
image found by counting pixel faces and is the speci c surface free energy, is knowing that
each pixel surface is originally at, before any strains are generated. Consider the z = 0 face
(1 , 2 , 3 , 4 face) of a pixel labelled like that shown in Fig. 1. The coordinates of the
number 1 node are (x1 + u1; y1 + v1 ; z1 + w1), where (x1 ; y1; z1 ) are the coordinates before
any strain has occurred, and(u1; v1 ; w1) are the x, y, and z displacements. There are similar
formulas for the other nodes. No matter how the pixel face has been strained (assuming
small distortions), the area of the face can be de ned as the area of the triangle 1 , 2 , 3,
and the triangle 2 , 3 , 4, since any three points are co-planar. The area of these two
triangles can be written as one half the sum of the magnitude of the cross-products of the
vectors making up their sides. The resulting formula contains the square root of the squares
of various combinations of the di erences of the nodal coordinates. The formula is somewhat
simpli ed for this face by remembering that z1 = z2 = z3 = z4 = 0. Making the assumption
of small di erences between nodal displacements (small strains), the area of this face can be
reduced to a linear form,
SA = r2[1 + 21r (u2 + u3 + v3 + v4 , u1 , u4 , v1 , v2)]
(77)
where the pixel was originally an r  r  r cube. For the 1 , 4 , 5 , 8 face (x=0), the formula
becomes
(78)
SA = r2[1 + 21r (v4 + v8 + w5 + w8 , v1 , v5 , w1 , w4 )]
and for the 1 , 2 , 6 , 5 face (y = 0) the linearized formula is
SA = r2[1 + 21r (u2 + u6 + w5 + w6 , u1 , u5 , w1 , w2)]
(79)
Consider a strain of 0:1 in the x-direction only. If the pixel has node 1 at the origin,
then u1 = u4 = u5 = u8 = 0, u2 = u3 = u6 = u7 = 0:1r, and all the other displacement
components are zero. Computing the above equations with these displacements gives SA =
1:1r2 for the z = 0 face, 1:1r2 for the y = 0 face, and SA = r2 for the x = 0 face, as expected.
A global vector can be built out of these displacements, which will also contain terms
for the macrostrains, via the periodic boundary conditions, which are picked up at the
boundaries. This vector is constant with respect to the macrostrains, and so will simply add
to the gradient vector, and will not come into the subroutine DEMBX at all, since only the
second derivative of the energy is used in that subroutine.
A simple example will suce to show the accuracy of this technique [41]. Consider a
solid block of material, with a dilute volume fraction c of spherical holes of radius R. In the
dilute limit the holes can be treated separately, so the analytical problem can be carried out
for a single hole. The solid material has bulk modulus K and shear modulus G, and is at
zero strain. Now suddenly add a surface energy per unit area to the surface of the hole.
The material will shrink in order to reduce the surface area of the hole and thus the surface
53

energy stored. The reduction in surface energy from the shrinkage will be exactly balanced
by the increase in volume stored elastic energy, giving the result:


3
4
,
 = 6R G + K c
(80)
For a 15 pixel diameter spherial hole in a 403 unit cell, the agreement with the exact
result for the coecient of c in eq. (78) was 4:7% [41]. Better agreement would be obtained
using a larger system.

54

7 Making and analyzing images and histograms
7.1 General features

It is often useful to make a current or stress map of the system being analyzed. These
are helpful to check against exact solutions, and can often give insight into systems not
capable of being understood analytically. The sections below will discuss current maps for
the electrical programs, both nite di erence and nite element, and stress maps for the
nite element programs. Histograms are a way of making plots of the stress or current
distribution functions. These distribution functions and their various moments are way of
going one step beyond just analyzing the e ective properties, and can give additional insight
into the random system being studied. Also, the programs BURN2D.F and BURN3D.F will
be discussed, which check an image for phase percolation.

7.2 Finite element electrical problems

It is simple to make a current map using the nite element programs ELECFEM2D.F or
ELECFEM3D.F. In the subroutine CURRENT, the total average current is computed. In
a small DO loop in the middle of the main DO loop, called DO 465, the average current
for a given pixel is computed. The components of the average current in a pixel are called
cur1, cur2, and cur3, for the x, y, and z directions. These variables can be written to a le,
or perhaps only the magnitude of the current need be stored. Then an image, using some
suitable scaling system (0-255 for gray scale, or a color system) can created using almost
any kind of imaging software. The shareware program XV can also be used to convert the
ASCII gray scale values back to a graphics format image.
In the nite element programs, the average current in a pixel is given in terms of the
nodal voltages, using the average over the pixel of the conductivity tensor times the matrix
in eq. (4). This matrix gives the local eld in the pixel in terms of the nodal voltages, so
the conductivity tensor times the local eld gives the local current.
Figure 13 shows (top images), for the applied eld in the horizontal direction, the horizontal current density, for two choices of the inclusion conductivity. The matrix always had
a conductivity of 1, and the inclusion had either a conductivity of 10 (left) or 0:1 (right).
The actual current densities have been scaled from 0 to 255 by assigning the value of 200 to
the average current density. Any current density that scaled to a value over 255 was simply
set to 255. Then a color scale was assigned according to this order: red, green, gray, and
black (top to bottom, high to low current density in accompanying color bar). In the left top
image of Fig. 13, the current tends to bunch up at the left and right of the inclusion, due
to the normal current being continuous at the boundary. The current density is depleted at
the top and bottom of the inclusion, as it is energetically favorable for the current to curve
and go through the high conductivity inclusion.
The right top image of Fig. 13 shows a similar picture for an inclusion conductivity of
0:1. Notice now the current tends to bunch up at the top and bottom of the inclusion, as it
is energetically more favorable to go around the low conductivity inclusion. Also, at the left
and right of the inclusion, the current is low, because at the boundary the normal current
must match the low current inside the inclusion.
55

Figure 13: Image of the horizontal current magnitudes, with the applied electric eld in the
x-direction in all images. The inclusion is phase 2. Left: 2 = 10, right: 2 = 0:1, and
both images had 1 = 1:0. Top: Finite element solution. Middle: nite di erence solution.
Bottom: exact solution, no periodic boundary conditions. Color bar shows high (red) to low
(black) current scale.
In the top images of Fig. 13, one can notice a ring of miscolored pixels around the
inclusion boundary. This is an artifact of having a digitally-rough approximation to a smooth
boundary. Locally averaging the pixel intensities around the boundary would make the
boundary much smoother, and give a better approximation to reality.

7.3 Finite di erence electrical problems

In the nite di erence electrical programs, one must be careful in de ning the average current
in a pixel. In d dimensions, there are 2d bonds coming into a node, with 2d currents to
consider. The most obvious way to de ne the average current in a pixel is to average the
current in the two x-bonds, the two y-bonds, and the two z-bonds, and thus obtain the three
components of the average current vector in the pixel. Subroutine CURRENT in the nite
di erence programs computes the total current for the whole image by summing over all the
56

pixel currents. Variables cur1, cur2, and cur3 are the local average currents in a pixel.
The middle images of Figure 13 shows the same problems as described in the previous
section but now for a nite di erence solution. The current maps are very similar, with
similar small anomalies at the inclusion boundary. To the eye, there is very little di erence
between the nite di erence and the nite element current maps. Recall from Fig. 8 and
Table 10 that in this range of inclusion to matrix conductivity ratios, 0:1 to 10, the nite
element and nite di erence intrinsic conductivities agreed rather well.
The bottom images of Fig. 13 show the equivalent maps for an exact solution to the same
problem, without periodic boundary conditions. The exact solution is described below, in
Section 7.6. Most of the variation of the current is near to the inclusion, so that it is
reasonable to compare in nite matrix analytical and periodic boundary numerical solutions.
The nite di erence and nite element solutions are seen, by comparing the di erent parts
of Fig. 13, to at least qualitatively give accurate solutions of the conductivity problem.
The analysis of the elds via their distribution in histogram form will show better the small
di erences between the exact and numerical solutions.

7.4 Finite element elastic problems

The subroutine STRESS, in the DO 465 loop, for all the elastic programs, computes the
stress per pixel and adds these up to get the overall average stress. Simply output the
variables str11, str22, str33, str13, str23, and str12, which are the local average stresses
in a given pixel, to generate any kind of stress map. The local strains are also computed,
s11, s22, s33, s13, s23, and s12, and can be output as well. In the thermal strain programs,
THERMAL2D.F and THERMAL3D.F, the local strain and the local stress can be similarly
output with or without the thermal strain included, if that is of interest in a given problem
solution.

7.5 General features of histograms

When a microstructure is analyzed, then all the local currents/stresses/strains are available,
since the full problem has been solved for every pixel. This information can be used to
generate pictures or maps of what a quantity looks like throughout a microstructure. This
is useful qualitative information. The same information can be used to generate histograms,
or distribution functions, of a quantity of interest, which is a more rigorous way of looking
at the same data. A histogram can give information on current or stress distributions and
moments of these distributions. The basic kind of graph has area or volume fraction as the
ordinate, and current or stress as the abscissa. A point with coordinates (g; p) means that a
fraction p of the system has property g.
It should be noted that system averages of current or stress, that de ne the e ective
conductivity or elastic moduli, are generally quite reliable and accurate, as has been seen in
the examples in Section 5. However, the actual pixel-by-pixel values of current or stress can
be somewhat o , due mostly to discretization errors and digital boundaries. At a boundary
which should be curved, locally one cannot have anything but at pixel boundaries. This
can throw o some of the values near the boundary.
57

7.6 Examples of histograms

Consider the same 200  200 system, with a 41-pixel diameter circular inclusion (phase 2)
in the middle of the matrix (phase 1), as was used in Fig. 13. Histograms for the average
current per pixel will be generated using both nite di erence and nite element methods,
and will be compared to the exact result, which uses the analytical solution of the problem.
This analytical solution is simple to generate. Using polar coordinates, (r; ), the potential outside the circle is ,Er cos  + B cos =r, and the potential inside the circle is ,Ar cos ,
where A and B are unknown constants and the applied eld E is in the x-direction only. By
requiring that the potential and the normal current be continuous at the r = R boundary
(the exterior of the circle), the coecients A and B are determined via these two equations.
The x component of the current is then:
"

#

2
2
2
Outside jx = 1 E 1 + R((+2 , )(1x)(2x+,y2y)2 )
(81)
1
2
Inside jx = (21+2E)
(82)
1
2
One problem in comparing the exact solution to the numerical solutions is that the exact
solution assumes an in nite matrix, while the numerical solutions use periodic boundary
conditions. For purposes of comparison, the currents from the exact solution at the pixel
centers, constricted to a 200  200 area around the inclusion, will be used to create the exact
solution histogram. An electric eld of magnitude unity was applied in the x-direction. For
all three computations, there will be 40; 000 numbers, the magnitude of the current in the
x- direction for each pixel. The same range of bins, and bin sizes, were chosen for all three
( nite element, nite di erence, exact) so as to make the histograms more comparable. The
exact solution could of course have been used to generate a histogram with in nite resolution.
In the rst example, the isotropic conductivities were: 1 = 1, 2 = 10. Figure 14 shows
all three histograms plotted from a minimum current of 0:2 to a maximum current of 2:0. The
nite element and nite di erence solutions actually had a few (10,20) pixels that had higher
currents than this, but as they would not show up on a graph, they have not been plotted.
These come from digital boundaries, and are not relevant to such a comparison. The nite
di erence and nite element histograms are essentially the same, and are slightly di erent
from the exact result. Part of this is due to the digital boundary of the circular inclusion,
which leads to inaccuracies in the currents near the surface. Part of the disagreement is also
due to the fact that the computations used periodic boundary conditions, while the exact
solution is for an in nite matrix.
In the second example, the isotropic conductivities were 1 = 1 and 2 = 0:1. Figure 15
shows all three histograms plotted from a minimum current of 0:1 to a maximum current
of 2:0. All currents were within these bounds. The nite di erence and nite element
histograms are essentially the same, and are slightly di erent from the exact result. Again,
part of this is due to the digital boundary of the circular inclusion, which leads to inaccuracies
in the currents near the surface. Similarly to Fig. 14, part of the disagreement between the
analytical and numerical histograms is also due to the fact that the computations used
periodic boundary conditions, while the exact solution was for an in nite matrix.

58

Area fraction

0.30

0.20

Exact
FE
FD
0.10

0.00
0.0

0.5

1.0

1.5

2.0

Current
Figure 14: Current distribution for the same circular inclusion problem as in Fig. 13,
calculated by all three methods ( nite element, nite di erence, and exact calculation, 1 =
1:0, 2 = 10.

59

Area fraction

0.30

0.20

Exact
FE
FD

0.10

0.00
0.0

0.5

1.0

1.5

2.0

Current
Figure 15: Current distribution for the same circular inclusion problem as in Fig. 13,
calculated by all three methods ( nite element, nite di erence, and exact calculation, 1 =
1:0, 2 = 0:1

60

The results for Figs. 14 and 15 make it clear that current distributions can di er, but
in such a way that averages ( rst order moments) are nearly the same. However, higher
moments will di er more. This is why the error on the e ective conductivity (1st moment)
was always less than the error in the average eld squared (2nd moment) in the checkerboard
example given in Table 8.

7.7 Phase percolation in images

If a phase percolates in a given direction in a microstructure, that means that the phase
is continuous from one side of the microstructure to the other in that direction. This is
important topological information and a ects greatly the ability of the phase to a ect the
overall properties [9, 10, 24, 42]. There are two auxiliary programs included with the nite
element and nite di erence programs, BURN2D.F and BURN3D.F, for computing phase
percolation in 2-D and 3-D using the burning algorithm [42]. Given a phase label, these
programs check for continuity of this phase in each of the principal directions.
The burning algorithm in principle "lights" a re at one end of the microstructure, in
the chosen phase, and lets the re burn in that phase until there are no more pixels of that
phase left unburned, at least ones that the re can get to via nearest-neighbor connections.
The other side of the microstructure is then checked to see if the re reached there or not.
If it did, then the chosen phase must be connected from one side to the other. If it did not,
then the phase does not percolate.
The microstructure is read into the programs. As these are simple programs, the comments contained in the programs should suce to describe their operation. These programs
can be used to analyze an original microstructure. They can also be used to analyze a current
or stress map, in the following way. Suppose we would like to know if the high currents in
a current map are limited to a few hot spots, or are continuous across the microstructure.
One could create an image where all pixels with currents above a certain value are given a
phase label of one, and all other pixels a phase label of two. Then the burning programs
could be used to see if the high current phase percolated or not.
The listing for BURN3D.F is given in Sec. 9.3.7. BURN2D.F is similar to the 3-D
version.

61

8 References
[1] J.C. Maxwell, A Treatise on Electricity and Magnetism (Dover, New York, 1954).
[2] B.P. Flannery, H.W. Deckman, W.G. Roberge, and K.L. D'Amico, Science 235, 1439
(1987).
[3] D.P. Bentz, N.S. Martys, P. Stutzman, M.S. Levenson, E.J. Garboczi, and L.M.
Schwartz, "X-ray microtomography of an ASTM C-109 mortar exposed to sulfate attack," in Microstructure of Cement-Based Systems/Bonding and Interfaces in Cementitious Materials, edited by S. Diamond et al. (Materials Research Society Vol. 370,
Pittsburgh, 1995), pp. 77-82.
[4] E.J. Garboczi and D.P. Bentz, "Fundamental Computer Simulation Models for CementBased Materials," in Materials Science of Concrete Vol. 2, edited by J. Skalny (American
Ceramic Society, Westerville, Ohio, 1991).
[5] William H. Press, Brian P. Flannery, Saul A. Teukolsky, and William T. Vetterling,
Numerical Recipes (Cambridge University Press, Cambridge, 1990).
[6] S. Feng, M.F. Thorpe, and E.J. Garboczi, Physical Review B 31, 276-283 (1985).
[7] T. Mura, Micromechanics of defects in solids: 2nd edition (Martinus Nijho , Hingham,
MA, 1987).
[8] J. Poutet, D. Manzoni, F. Hage-Chehade, C.J. Jacquin, M.J. Bouoteca, J.-F. Thovert,
and P.M. Adler, "The e ective mechanical properties of reconstructed porous media,"
Int. J. Rock Mech. Min. Sci. and Geomech. Abstr. 33, 409-415 (1996); ibid, "The
e ective mechanical properties of random porous media," J. Mech. Phys. Solids 44,
1587 (1996).
[9] E.J. Garboczi, M.F. Thorpe, M.S. DeVries, and A.R. Day, "Universal conductivity
curve," Phys. Rev. A 43, 6473 (1991).
[10] S. Kirkpatrick, "Percolation and conduction," Reviews of Modern Physics 45, 574-588
(1971).
[11] J.F. Douglas and E.J. Garboczi, "Intrinsic viscosity and the polarizability of particles
having a wide range of shapes," in Advances in Chemical Physics 16, edited by I.
Prigogine and S.A. Rice (John Wiley and Sons, New York, 1995).
[12] E.J. Garboczi and J.F. Douglas, "Intrinsic conductivity of objects having arbitrary
shape and conductivity," Physical Review E 53, 6169-6180 (1996).
[13] E.J. Garboczi and D.P. Bentz, "Analytical formulas for interfacial transition zone properties," Journal of Advanced Cement-Based Materials 6, 99-108 (1997).
[14] S. Torquato, Applied Mechanics Review 44, 37-76 (1991).
62

[15] Z. Hashin, "Analysis of composite materials: A survey," Journal of Applied Mechanics
50, 481-505 (1983).
[16] W.F. Brown, Journal of Chemical Physics 23, 1514-1517 (1955).
[17] S. Torquato, "E ective sti ness tensor of composite media-I. Exact series expansions,"
J. Mech. Phys. Solids 45, 1421-1448 (1997).
[18] J.B. Keller, J. Math. Phys. 5, 548 (1964).
[19] K.S. Mendelson, J. Appl. Phys. 46, 917 (1975).
[20] M. Bobeth and G. Diener, "Field uctuations in multi-component mixtures," J. Mech.
Phys. Solids 34, 1-17 (1986).
[21] R. Hill, Journal of the Mechanics and Physics of Solids 11, 357-372 (1963).
[22] M.F. Thorpe and I. Jasiuk, Proc. Roy. Soc. London A 438, 531-544 (1994).
[23] R.M. Christensen, Mechanics of Composite Materials (Krieger Publishing Co., Malabar,
FL, 1991).
[24] K.A. Snyder, E.J. Garboczi, and A.R. Day, "The elastic moduli of simple twodimensional isotropic composites: Computer simulation and e ective medium theory,"
J. Appl. Phys. 72, 5948-5955 (1992).
[25] L. Eyges and P. Gianino, IEEE Transactions on Antennas and Prop. 27, 557 (1979).
[26] L. Vegard, Z. Phys. 5, 17 (1921); M.F. Thorpe and E.J. Garboczi, Phys. Rev. B 42,
8405-8417 (1990).
[27] J.N. Goodier, Phil. Mag. Series 7, 23, 1017-1032 (1937).
[28] T. Mura, H.M. Shodja, T.Y. Lin, A. Safadi, and K. Hirashima, "The determination
of the elastic eld of a star shaped inclusion," presented at the Twelfth U.S. National
Congress of Applied Mechanics, University of Washington, Seattle, 1994.
[29] J. Wulf, P. Lipetzky, G.E. Beltz, and T. Steinkop , "A nite element model of the stress
eld in a star-shaped inclusion," Comp. Mater. Sci. 3, 423-429 (1995).
[30] B.W. Rosen and Z. Hashin, "E ective thermal expansion coecient and speci c heats
of composite materials," International Journal of Engineering Science 8, 157-173 (1970).
[31] J.K. Mackenzie, "The elastic constants of a solid containing spherical holes," Proc. Roy.
Soc. 683, 2-11 (1950).
[32] A. Cherkaev, K. Lurie, and G.W. Milton, "Invariant properties of the stress in plane
elasticity and equivalence classes of composites," Proc. Roy. Soc. Lond. A 438, 519-529
(1992).
63

[33] K.A. Lurie and A.V. Cherkaev, "The e ective properties of composite materials and
problems of optimal designs of constructions," Usp. Mekaniki (Adv. Mech.) 9, 3-81
(1986).
[34] H. Davies, "Poisson's partial di erence equation," Quart. J. Math. Oxford 6, 232-40
(1955).
[35] G. Venezian, "On the resistance between two points on a grid," Am. J. Phys. 62, 10001004 (1994). Note that in eq. (21), there is a typographical error. The constant "0.51469"
should be "3.23386."
[36] G. Montet, "The e ective resistance of passive networks," J. Math. Phys. 5, 1555-1559
(1964).
[37] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions (Dover, New
York, 1965), pp. 887-888.
[38] G. Hsieh, T.O. Mason, E.J. Garboczi, and L.R. Pederson, "Experimental limitations
in impedance spectroscopy: Part III{E ect of reference electrode geometry/position,"
Solid State Ionics 96, 153-172 (1997).
[39] D.J.C. Yates, "The expansion of porous glass on the adsorption of non-polar gases,"
Proc. Roy. Soc. London A 224, 526-544 (1954).
[40] G.W. Scherer, "Dilatation of porous glass," Journal of the American Ceramic Society
69, 473-480 (1986).
[41] D.P. Bentz, E.J. Garboczi, and D.A. Quenard, "Modelling drying shrinkage in reconstructed porous materials: Application to porous Vycor glass," Mod. Sim. Mater. Sci.
Eng. 6, 211-236 (1998).
[42] D. Stau er and A. Aharony, Introduction to percolation theory 2nd Ed. (Taylor and
Francis, London, 1992).

64

9 Listing of programs and computer requirements
9.1 Computer details, access information

The programs discussed in this manual were all written in FORTRAN 77, and use no special
statements that are machine speci c. The programs themselves are straightforward computational algorithms. In fact, these programs could quite easily be translated into another
language, or used by someone who had no previous knowledge of FORTRAN, just by learning the meaning of a few statements that are analogous to those in C, for example. All
programs should be run in double precision (8 bytes or 64 bits per real variable). Running
the nite di erence routines in single precision may give reasonable answers, but double
precision is more trustworthy. Running the nite element routines in single precision will
not work, except perhaps for very small systems, on the order of 103 = 1000 nodes.
In the nite element programs, all the main arrays are passed to subroutines via simple
common statements. Small variables and parameters are passed in the subroutine calls. In
contrast to this, in the nite di erence programs, all variables, including the main arrays, are
passed via subroutine call parameter lists. There are no common blocks used. In the nite
element programs, the dimensions of the system can be easily changed by doing a global
replace on the dimensions shown in the array declarations in the main program, since the
dimensions of the main arrays in the subroutines must be changed as well. The values of
nx, ny, and nz are then changed to match. The dimensions of the main arrays in the nite
di erence programs only need to be changed in the main program. There is no reason why
the system studied has to be a cube. All the programs will handle a rectangular parallelipiped
with nx 6= ny 6= nz. There are no input les except that for the system digital image. If
the user desires to have the various phase conductivities, elastic moduli, etc. given in input
les, that is easy to introduce into the programs.
The programs discussed in this manual are available via anonymous ftp from the directory edsel=ftp=pub=FDFEMANUAL, on the computer edsel:cbt:nist:gov. This manual is
also available online, as in Chapter 2 in an HTML electronic monograph on the computer
modelling of cement-based materials, at:

http://ciks.cbt.nist.gov/garboczi
See this address for instructions of how to down-load postscript and pdf versions of this
manual. The manual may also be requested from the author in paper form:

edward.garboczi@nist.gov

9.2 Memory requirements

The memory requirements of the various programs can be expressed in terms of the bytes
of memory required per pixel of the microstructure. The following numbers assume that
double precision (REAL*8 and COMPLEX*16) is used. Only the memory requirements for
arrays that are of the system size are counted. Variables like dk, phasemod, and sigma are
ignored in this count, since their memory requirements scale with the number of phases, and
this number is usually of order one.
65

Program
Bytes per pixel
DC2D.F
54
DC3D.F
62
AC2D.F
102
AC3D.F
118
ELECFEM2D.F
78
ELECFEM3D.F
150
ELAS2D.F
118
ELAS3D.F
230
THERMAL2D.F
134
THERMAL3D.F
254
Table 13: Memory requirements in terms of bytes per pixel for programs discussed in this
manual.

9.3 Listing of key programs

The FORTRAN 77 code is listed below for the following main programs: ELECFEM3D.F
(example of nite element electrical conductivity program), ELAS3D.F (example of nite
element elastic program), THERMAL3D.F (example of nite element elastic program with
thermal strains and variable system shape and size), DC3D.F (example of real nite difference electrical program), and AC3D.F (example of complex nite di erence electrical
program). The other programs are 2-D equivalents of these programs. Listings for two
auxiliary programs, GAUSS.F (Gaussian quadrature weights and points) and BURN3D.F
(phase percolation using burning algorithm), are also given.

66

9.3.1 ELECFEM3D.F
c
c

************************
BACKGROUND

elecfem3d.f

****************************

c
c
c
c
c
c
c

This program solves Laplace's equation in a random conducting
material using the finite element method. Each pixel in the 3-D digital
image is a cubic tri-linear finite element, having its own conductivity.
Periodic boundary conditions are maintained. In the comments below,
(USER) means that this is a section of code that the user might
have to change for his particular problem. Therefore the user is
encouraged to search for this string.

c

PROBLEM AND VARIABLE DEFINITION

c
c
c
c
c
c
c
c
c
c
c
c
c
c

The problem being solved is the minimization of the energy
1/2 uAu + b u + C, where A is the Hessian matrix composed of the
stiffness matrices (dk) for each pixel/element, b is a constant vector
and C is a constant that are determined by the applied field and
the periodic boundary conditions, and u is a vector of all the voltages.
The method used is the conjugate gradient relaxation algorithm.
Other variables are: gb is the gradient = Au+b, h and Ah are
auxiliary variables used in the conjugate gradient algorithm (in dembx),
dk(n,i,j) is the stiffness matrix of the n'th phase, sigma(n,i,j) is
the conductivity tensor of the n'th phase, pix is a vector that gives
the phase label of each pixel, ib is a matrix that gives the labels of
the 27 (counting itself) neighbors of a given node, prob is the volume
fractions of the various phases, and currx, curry, currz are the
volume averaged total currents in the x, y, and z directions.

c DIMENSIONS
c
c
c
c
c
c
c
c
c
c

The vectors u,gb,b,h, and Ah are dimensioned to be the system size,
ns=nx*ny*nz, where the digital image of the microstructure considered
is a rectangular parallelipiped ( nx x ny x nz) in size.
The arrays pix and ib are also dimensioned to the system size.
The array ib has 27 components, for the 27 neighbors of a node.
Note that the program is set up at present to have at most 100
different phases. This can easily be changed, simply by changing
the dimensions of dk, prob, and sigma. Nphase gives the number of
phases being considered.
All arrays are passed between subroutines using simple common statements.

c

STRONGLY SUGGESTED:

READ THE MANUAL BEFORE USING PROGRAM!!

67

c (USER) Change these dimensions and in other subroutines at the same time.
c For example, search and replace all occurrences throughout the program
c of "(8000" by "(64000", to go from a 20 x 20 x 20 system to a
c 40 x 40 x 40 system.
real u(8000),gb(8000),b(8000),dk(100,8,8)
real h(8000),Ah(8000)
real sigma(100,3,3),prob(100)
integer in(27),jn(27),kn(27)
integer*4 ib(8000,27)
integer*2 pix(8000)
common/list1/currx,curry,currz
common/list2/ex,ey,ez
common/list3/ib
common/list4/pix
common/list5/dk,b,C
common/list6/u
common/list7/gb
common/list8/sigma
common/list9/h,Ah
c
c

(USER) Unit 9 is the microstructure input file, unit 7 is
the results output file.
open(9,file='microstructure.dat')
open(7,file='outputfile.out')

c

(USER) nx,ny,nz give the size of the lattice
nx=20
ny=20
nz=20
c ns=total number of sites
ns=nx*ny*nz
write(7,9010) nx,ny,nz,ns
9010 format('nx= ',i4,' ny= ',i4,' nz= ',i4,' ns = ',i8)
c
c

(USER) nphase is the number of phases being considered in the problem.
The values of pix(m) will run from 1 to nphase.
nphase=2

c
c
c

(USER) gtest is the stopping criterion, compared to gg=gb*gb.
gtest=abc*ns, so that when gg < gtest, that average value per pixel
of gb is less than sqrt(abc).
gtest=1.e-16*ns

c

Construct the neighbor table, ib(m,n)

68

c
c

First construct 27 neighbor table in terms of delta i, delta j, delta k
(See Table 3 in manual)
in(1)=0
in(2)=1
in(3)=1
in(4)=1
in(5)=0
in(6)=-1
in(7)=-1
in(8)=-1
jn(1)=1
jn(2)=1
jn(3)=0
jn(4)=-1
jn(5)=-1
jn(6)=-1
jn(7)=0
jn(8)=1

555

c
c
c

do 555 n=1,8
kn(n)=0
kn(n+8)=-1
kn(n+16)=1
in(n+8)=in(n)
in(n+16)=in(n)
jn(n+8)=jn(n)
jn(n+16)=jn(n)
continue
in(25)=0
in(26)=0
in(27)=0
jn(25)=0
jn(26)=0
jn(27)=0
kn(25)=-1
kn(26)=1
kn(27)=0
Now construct neighbor table according to 1-d labels
Matrix ib(m,n) gives the 1-d label of the n'th neighbor (n=1,27) of
the node labelled m.
nxy=nx*ny
do 1020 k=1,nz

69

1004
1020
c
c

do 1020 j=1,ny
do 1020 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
do 1004 n=1,27
i1=i+in(n)
j1=j+jn(n)
k1=k+kn(n)
if(i1.lt.1) i1=i1+nx
if(i1.gt.nx) i1=i1-nx
if(j1.lt.1) j1=j1+ny
if(j1.gt.ny) j1=j1-ny
if(k1.lt.1) k1=k1+nz
if(k1.gt.nz) k1=k1-nz
m1=nxy*(k1-1)+nx*(j1-1)+i1
ib(m,n)=m1
continue
continue

Compute the electrical conductivity of each microstructure.
(USER) npoints is the number of microstructures to use.
npoints=1
do 8000 micro=1,npoints

c
c

Read in a microstructure in subroutine ppixel, and set up pix(m)
with the appropriate phase assignments.
call ppixel(nx,ny,nz,ns,nphase)
c Count and output the volume fractions of the different phases
call assig(ns,nphase,prob)
do 805 i=1,nphase
write(7,*) 'Volume fraction of phase ',i,' = ',prob(i)
805
continue
c
c
c
c

(USER) sigma(100,3,3) is the electrical conductivity tensor of each phase
The user can make the value of sigma to be different for each
phase of the microstructure if so desired (up to 100 phases as currently
dimensioned).
sigma(1,1,1)=1.0
sigma(1,1,2)=0.0
sigma(1,1,3)=0.0
sigma(1,2,2)=1.0
sigma(1,2,3)=0.0
sigma(1,3,3)=1.0
sigma(1,2,1)=sigma(1,1,2)
sigma(1,3,1)=sigma(1,1,3)
sigma(1,3,2)=sigma(1,2,3)

70

sigma(2,1,1)=0.5
sigma(2,1,2)=0.0
sigma(2,1,3)=0.0
sigma(2,2,2)=0.5
sigma(2,2,3)=0.0
sigma(2,3,3)=0.5
sigma(2,2,1)=sigma(2,1,2)
sigma(2,3,1)=sigma(2,1,3)
sigma(2,3,2)=sigma(2,2,3)
c

write out the phase electrical conductivity tensors
do 11 i=1,nphase
write(7,*) 'Phase ',i,' conductivity tensor is:'
write(7,*) sigma(i,1,1),sigma(i,1,2),sigma(i,1,3)
write(7,*) sigma(i,2,1),sigma(i,2,2),sigma(i,2,3)
write(7,*) sigma(i,3,1),sigma(i,3,2),sigma(i,3,3)
11
continue
c

(USER) Set applied electric field.
ex=1.0
ey=1.0
ez=1.0
write(7,*) 'Applied field components:'
write(7,*) 'ex = ',ex,' ey = ',ey,' ez = ',ez

c
c

Set up the finite element "stiffness" matrices and the Constant and
vector required for the energy

call femat(nx,ny,nz,ns,nphase)
c Apply a homogeneous macroscopic electric field as the initial condition
do 1050 k=1,nz
do 1050 j=1,ny
do 1050 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
x=float(i-1)
y=float(j-1)
z=float(k-1)
u(m)=-x*ex-y*ey-z*ez
1050 continue

c
c

Relaxation Loop
(USER) kmax is the maximum number of times dembx will be called, with

71

c
c
c

ldemb conjugate gradient steps done during each call. The total
number of conjugate gradient cycles allowed for a given conductivity
computation is kmax*ldemb.
kmax=40
ldemb=50
ltot=0

c

Call energy to get initial energy and initial gradient
call energy(nx,ny,nz,ns,utot)
c gg is the norm squared of the gradient (gg=gb*gb)
gg=0.0
do 100 m=1,ns
gg=gg+gb(m)*gb(m)
100
continue
write(7,*) 'Initial energy = ',utot,'gg = ',gg
call flush(7)
do 5000 kkk=1,kmax
Call dembx to go into conjugate gradient solver
call dembx(ns,Lstep,gg,dk,gtest,ldemb,kkk)
ltot=ltot+Lstep
c Call energy to compute energy after dembx call. If gg < gtest, this
c will be the final energy. If gg is still larger than gtest, then this
c will give an intermediate energy with which to check how the relaxation
c process is coming along.
call energy(nx,ny,nz,ns,utot)
write(7,*) 'Energy = ',utot,'gg = ',gg
write(7,*) ltot, ' conj. grad. steps'
if(gg.lt.gtest) goto 444
c

c
c

If relaxation process will continue, compute and output currents as an
additional aid to judge how the relaxation procedure if progressing.
call current(nx,ny,nz,ns)
c Output intermediate currents
write(7,*)
write(7, *) ' Current in x direction = ',currx
write(7, *) ' Current in y direction = ',curry
write(7, *) ' Current in z direction = ',currz
call flush(7)
5000

continue

444

call current(nx,ny,nz,ns)

c Output final currents

72

write(7,*)
write(7, *) ' Current in x direction = ',currx
write(7, *) ' Current in y direction = ',curry
write(7, *) ' Current in z direction = ',currz
call flush(7)
8000

continue
end

c
c
c

Subroutine that sets up the stiffness matrices, linear term in
voltages, and constant term C that appear in the total energy due
to the periodic boundary conditions.
subroutine femat(nx,ny,nz,ns,nphase)
real dk(100,8,8),xn(8),b(8000),C
real dndx(8),dndy(8),dndz(8)
real g(3,3,3),sigma(100,3,3)
real es(3,8)
integer is(8)
integer*4 ib(8000,27)
integer*2 pix(8000)

common/list2/ex,ey,ez
common/list3/ib
common/list4/pix
common/list5/dk,b,C
common/list8/sigma
nxy=nx*ny
c

initialize stiffness matrices
do 40 m=1,nphase
do 40 j=1,8
do 40 i=1,8
dk(m,i,j)=0.0
40
continue
c

set up Simpson's rule integration weight vector
do 30 k=1,3
do 30 j=1,3
do 30 i=1,3
nm=0
if(i.eq.2) nm=nm+1
if(j.eq.2) nm=nm+1

73

30
c
c
c
c

c
c

c

if(k.eq.2) nm=nm+1
g(i,j,k)=4.0**nm
continue
loop over the nphase kinds of pixels and Simpson's rule quadrature
points in order to compute the stiffness matrices. Stiffness matrices
of trilinear finite elements are quadratic in x, y, and z, so that
Simpson's rule quadrature is exact.
do 4000 ijk=1,nphase
do 3000 k=1,3
do 3000 j=1,3
do 3000 i=1,3
x=float(i-1)/2.0
y=float(j-1)/2.0
z=float(k-1)/2.0
dndx means the negative derivative with respect to x of the shape
matrix N (see manual, Sec. 2.2), dndy, dndz are similar.
dndx(1)=-(1.0-y)*(1.0-z)
dndx(2)=(1.0-y)*(1.0-z)
dndx(3)=y*(1.0-z)
dndx(4)=-y*(1.0-z)
dndx(5)=-(1.0-y)*z
dndx(6)=(1.0-y)*z
dndx(7)=y*z
dndx(8)=-y*z
dndy(1)=-(1.0-x)*(1.0-z)
dndy(2)=-x*(1.0-z)
dndy(3)=x*(1.0-z)
dndy(4)=(1.0-x)*(1.0-z)
dndy(5)=-(1.0-x)*z
dndy(6)=-x*z
dndy(7)=x*z
dndy(8)=(1.0-x)*z
dndz(1)=-(1.0-x)*(1.0-y)
dndz(2)=-x*(1.0-y)
dndz(3)=-x*y
dndz(4)=-(1.0-x)*y
dndz(5)=(1.0-x)*(1.0-y)
dndz(6)=x*(1.0-y)
dndz(7)=x*y
dndz(8)=(1.0-x)*y
now build electric field matrix
do 2799 n1=1,3
do 2799 n2=1,8
es(n1,n2)=0.0

74

2799

continue
do 2797 n=1,8
es(1,n)=dndx(n)
es(2,n)=dndy(n)
es(3,n)=dndz(n)
2797 continue
c now do matrix multiply to determine value at (x,y,z), multiply by
c proper weight, and sum into dk, the stiffness matrix
do 900 ii=1,8
do 900 jj=1,8
c Define sum over field matrices and conductivity tensor that defines
c the stiffness matrix.
sum=0.0
do 890 kk=1,3
do 890 ll=1,3
sum=sum+es(kk,ii)*sigma(ijk,kk,ll)*es(ll,jj)
890
continue
dk(ijk,ii,jj)=dk(ijk,ii,jj)+g(i,j,k)*sum/216.
900
continue
3000
4000
c
c
c
c
c

Set up vector for linear term, b, and constant term, C,
in the electrical energy. This is done using the stiffness matrices,
and the periodic terms in the applied field that come in at the boundary
pixels via the periodic boundary conditions and the condition that
an applied macroscopic field exists (see Sec. 2.2 in manual).

5000
c
c
c
c

continue
continue

do 5000 m=1,ns
b(m)=0.0
continue

For all cases, correspondence between 1-8 finite element node labels
and 1-27 neighbor labels is: 1:ib(m,27),2:ib(m,3),3:ib(m,2),
4:ib(m,1),5:ib(m,26),6:ib(m,19),7:ib(m,18),8:ib(m,17)
(see Table 4 in manual)
is(1)=27
is(2)=3
is(3)=2
is(4)=1
is(5)=26
is(6)=19
is(7)=18
is(8)=17

75

C=0.0
c x=nx face
i=nx
do 2001 i8=1,8
xn(i8)=0.0
if(i8.eq.2.or.i8.eq.3.or.i8.eq.6.or.i8.eq.7) then
xn(i8)=-ex*nx
end if
2001 continue
do 2000 j=1,ny-1
do 2000 k=1,nz-1
m=nxy*(k-1)+j*nx
do 1900 mm=1,8
sum=0.0
do 1899 m8=1,8
sum=sum+xn(m8)*dk(pix(m),m8,mm)
C=C+0.5*xn(m8)*dk(pix(m),m8,mm)*xn(mm)
1899 continue
b(ib(m,is(mm)))=b(ib(m,is(mm)))+sum
1900 continue
2000 continue
c y=ny face
j=ny
do 2011 i8=1,8
xn(i8)=0.0
if(i8.eq.3.or.i8.eq.4.or.i8.eq.7.or.i8.eq.8) then
xn(i8)=-ey*ny
end if
2011 continue
do 2010 i=1,nx-1
do 2010 k=1,nz-1
m=nxy*(k-1)+nx*(ny-1)+i
do 1901 mm=1,8
sum=0.0
do 2099 m8=1,8
sum=sum+xn(m8)*dk(pix(m),m8,mm)
C=C+0.5*xn(m8)*dk(pix(m),m8,mm)*xn(mm)
2099 continue
b(ib(m,is(mm)))=b(ib(m,is(mm)))+sum
1901 continue
2010 continue
c z=nz face
k=nz
do 2021 i8=1,8

76

xn(i8)=0.0
if(i8.eq.5.or.i8.eq.6.or.i8.eq.7.or.i8.eq.8) then
xn(i8)=-ez*nz
end if
2021 continue
do 2020 i=1,nx-1
do 2020 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+i
do 1902 mm=1,8
sum=0.0
do 2019 m8=1,8
sum=sum+xn(m8)*dk(pix(m),m8,mm)
C=C+0.5*xn(m8)*dk(pix(m),m8,mm)*xn(mm)
2019 continue
b(ib(m,is(mm)))=b(ib(m,is(mm)))+sum
1902 continue
2020 continue
c x=nx y=ny edge
i=nx
y=ny
do 2031 i8=1,8
xn(i8)=0.0
if(i8.eq.2.or.i8.eq.6) then
xn(i8)=-ex*nx
end if
if(i8.eq.4.or.i8.eq.8) then
xn(i8)=-ey*ny
end if
if(i8.eq.3.or.i8.eq.7) then
xn(i8)=-ey*ny-ex*nx
end if
2031 continue
do 2030 k=1,nz-1
m=nxy*k
do 1903 mm=1,8
sum=0.0
do 2029 m8=1,8
sum=sum+xn(m8)*dk(pix(m),m8,mm)
C=C+0.5*xn(m8)*dk(pix(m),m8,mm)*xn(mm)
2029 continue
b(ib(m,is(mm)))=b(ib(m,is(mm)))+sum
1903 continue
2030 continue
c x=nx z=nz edge
i=nx

77

k=nz
do 2041 i8=1,8
xn(i8)=0.0
if(i8.eq.2.or.i8.eq.3) then
xn(i8)=-ex*nx
end if
if(i8.eq.5.or.i8.eq.8) then
xn(i8)=-ez*nz
end if
if(i8.eq.6.or.i8.eq.7) then
xn(i8)=-ez*nz-ex*nx
end if
2041 continue
do 2040 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+nx
do 1904 mm=1,8
sum=0.0
do 2039 m8=1,8
sum=sum+xn(m8)*dk(pix(m),m8,mm)
C=C+0.5*xn(m8)*dk(pix(m),m8,mm)*xn(mm)
2039 continue
b(ib(m,is(mm)))=b(ib(m,is(mm)))+sum
1904 continue
2040 continue
c y=ny z=nz edge
j=ny
k=nz
do 2051 i8=1,8
xn(i8)=0.0
if(i8.eq.5.or.i8.eq.6) then
xn(i8)=-ez*nz
end if
if(i8.eq.3.or.i8.eq.4) then
xn(i8)=-ey*ny
end if
if(i8.eq.7.or.i8.eq.8) then
xn(i8)=-ey*ny-ez*nz
end if
2051 continue
do 2050 i=1,nx-1
m=nxy*(nz-1)+nx*(ny-1)+i
do 1905 mm=1,8
sum=0.0
do 2049 m8=1,8
sum=sum+xn(m8)*dk(pix(m),m8,mm)

78

C=C+0.5*xn(m8)*dk(pix(m),m8,mm)*xn(mm)
continue
b(ib(m,is(mm)))=b(ib(m,is(mm)))+sum
1905 continue
2050 continue
c x=nx y=ny z=nz corner
i=nx
j=ny
k=nz
do 2061 i8=1,8
xn(i8)=0.0
if(i8.eq.2) then
xn(i8)=-ex*nx
end if
if(i8.eq.4) then
xn(i8)=-ey*ny
end if
if(i8.eq.5) then
xn(i8)=-ez*nz
end if
if(i8.eq.8) then
xn(i8)=-ey*ny-ez*nz
end if
if(i8.eq.6) then
xn(i8)=-ex*nx-ez*nz
end if
if(i8.eq.3) then
xn(i8)=-ex*nx-ey*ny
end if
if(i8.eq.7) then
xn(i8)=-ex*nx-ey*ny-ez*nz
end if
2061 continue
m=nx*ny*nz
do 1906 mm=1,8
sum=0.0
do 2059 m8=1,8
sum=sum+xn(m8)*dk(pix(m),m8,mm)
C=C+0.5*xn(m8)*dk(pix(m),m8,mm)*xn(mm)
2059 continue
b(ib(m,is(mm)))=b(ib(m,is(mm)))+sum
1906 continue
2049

return
end

79

c

Subroutine computes the total energy, utot, and gradient, gb

subroutine energy(nx,ny,nz,ns,utot)
real u(8000),gb(8000)
real b(8000),C
real dk(100,8,8)
real utot
integer*4 ib(8000,27)
integer*2 pix(8000)
common/list2/ex,ey,ez
common/list3/ib
common/list4/pix
common/list5/dk,b,C
common/list6/u
common/list7/gb
do 2090 m=1,ns
gb(m)=0.0
2090 continue
c Energy loop. Do global matrix multiply via small stiffness matrices,
c gb=Au + b. The long statement below correctly brings in all the
c terms from the global matrix A using only the small stiffness matrices.
do 3000 m=1,ns
gb(m)=gb(m)+u(ib(m,1))*( dk(pix(ib(m,27)),1,4)+
&dk(pix(ib(m,7)),2,3)+
&dk(pix(ib(m,25)),5,8)+dk(pix(ib(m,15)),6,7) )+
&u(ib(m,2))*( dk(pix(ib(m,27)),1,3)+dk(pix(ib(m,25)),5,7) )+
&u(ib(m,3))*( dk(pix(ib(m,27)),1,2)+dk(pix(ib(m,5)),4,3)+
&dk(pix(ib(m,13)),8,7)+dk(pix(ib(m,25)),5,6) )+
&u(ib(m,4))*( dk(pix(ib(m,5)),4,2)+dk(pix(ib(m,13)),8,6) )+
&u(ib(m,5))*( dk(pix(ib(m,6)),3,2)+dk(pix(ib(m,5)),4,1)+
&dk(pix(ib(m,14)),6,7)+dk(pix(ib(m,13)),8,5) )+
&u(ib(m,6))*( dk(pix(ib(m,6)),3,1)+dk(pix(ib(m,14)),7,5) )+
&u(ib(m,7))*( dk(pix(ib(m,6)),3,4)+dk(pix(ib(m,7)),2,1)+
&dk(pix(ib(m,14)),7,8)+dk(pix(ib(m,15)),6,5) )+
&u(ib(m,8))*( dk(pix(ib(m,7)),2,4)+dk(pix(ib(m,15)),6,8) )+
&u(ib(m,9))*( dk(pix(ib(m,25)),5,4)+dk(pix(ib(m,15)),6,3) )+
&u(ib(m,10))*( dk(pix(ib(m,25)),5,3) )+
&u(ib(m,11))*( dk(pix(ib(m,13)),8,3)+dk(pix(ib(m,25)),5,2) )+
&u(ib(m,12))*( dk(pix(ib(m,13)),8,2) )+
&u(ib(m,13))*( dk(pix(ib(m,13)),8,1)+dk(pix(ib(m,14)),7,2) )+

80

&u(ib(m,14))*( dk(pix(ib(m,14)),7,1) )+
&u(ib(m,15))*( dk(pix(ib(m,14)),7,4)+dk(pix(ib(m,15)),6,1) )+
&u(ib(m,16))*( dk(pix(ib(m,15)),6,4) )+
&u(ib(m,17))*( dk(pix(ib(m,27)),1,8)+dk(pix(ib(m,7)),2,7) )+
&u(ib(m,18))*( dk(pix(ib(m,27)),1,7) )+
&u(ib(m,19))*( dk(pix(ib(m,27)),1,6)+dk(pix(ib(m,5)),4,7) )+
&u(ib(m,20))*( dk(pix(ib(m,5)),4,6) )+
&u(ib(m,21))*( dk(pix(ib(m,5)),4,5)+dk(pix(ib(m,6)),3,6) )+
&u(ib(m,22))*( dk(pix(ib(m,6)),3,5) )+
&u(ib(m,23))*( dk(pix(ib(m,6)),3,8)+dk(pix(ib(m,7)),2,5) )+
&u(ib(m,24))*( dk(pix(ib(m,7)),2,8) )+
&u(ib(m,25))*( dk(pix(ib(m,14)),7,3)+dk(pix(ib(m,13)),8,4)+
&dk(pix(ib(m,15)),6,2)+dk(pix(ib(m,25)),5,1) )+
&u(ib(m,26))*( dk(pix(ib(m,6)),3,7)+dk(pix(ib(m,5)),4,8)+
&dk(pix(ib(m,27)),1,5)+dk(pix(ib(m,7)),2,6) )+
&u(ib(m,27))*( dk(pix(ib(m,27)),1,1)+dk(pix(ib(m,7)),2,2)+
&dk(pix(ib(m,6)),3,3)+dk(pix(ib(m,5)),4,4)+dk(pix(ib(m,25)),5,5)+
&dk(pix(ib(m,15)),6,6)+dk(pix(ib(m,14)),7,7)+
&dk(pix(ib(m,13)),8,8) )
3000 continue
utot=0.0
do 3100 m=1,ns
utot=utot+0.5*u(m)*gb(m)+b(m)*u(m)
gb(m)=gb(m)+b(m)
3100 continue
utot=utot+C
return
end
c

Subroutine that carries out the conjugate gradient relaxation process
subroutine dembx(ns,Lstep,gg,dk,gtest,ldemb,kkk)
real u(8000),gb(8000),dk(100,8,8)
real Ah(8000),h(8000),B,lambda,gamma
integer*4 ib(8000,27)
integer*2 pix(8000)

common/list3/ib
common/list4/pix
common/list6/u
common/list7/gb
common/list9/h,Ah

81

c
c
c
c
c
c

Initialize the conjugate direction vector on first call to dembx only.
For calls to dembx after the first, we want to continue using the
value fo h determined in the previous call. Of course, if npooints
is greater than 1, then this initialization step will be run every
a new microstructure is used, as kkk will be reset to 1 every time
the counter micro is increased.
if(kkk.eq.1) then
do 50 m=1,ns
h(m)=gb(m)
50
continue
end if
c Lstep counts the number of conjugate gradient steps taken in each call
c to dembx.
Lstep=0
c

Conjugate gradient loop
do 800 ijk=1,ldemb
Lstep=Lstep+1

290
c
c
c

do 290 m=1,ns
Ah(m)=0.0
continue
Do global matrix multiply via small stiffness matrices, Ah = A * h.
The long statement below correctly brings in all the terms from
the global matrix A using only the small stiffness matrices.
do 400 m=1,ns
Ah(m)=Ah(m)+h(ib(m,1))*( dk(pix(ib(m,27)),1,4)+
&dk(pix(ib(m,7)),2,3)+
&dk(pix(ib(m,25)),5,8)+dk(pix(ib(m,15)),6,7) )+
&h(ib(m,2))*( dk(pix(ib(m,27)),1,3)+dk(pix(ib(m,25)),5,7) )+
&h(ib(m,3))*( dk(pix(ib(m,27)),1,2)+dk(pix(ib(m,5)),4,3)+
&dk(pix(ib(m,13)),8,7)+dk(pix(ib(m,25)),5,6) )+
&h(ib(m,4))*( dk(pix(ib(m,5)),4,2)+dk(pix(ib(m,13)),8,6) )+
&h(ib(m,5))*( dk(pix(ib(m,6)),3,2)+dk(pix(ib(m,5)),4,1)+
&dk(pix(ib(m,14)),6,7)+dk(pix(ib(m,13)),8,5) )+
&h(ib(m,6))*( dk(pix(ib(m,6)),3,1)+dk(pix(ib(m,14)),7,5) )+
&h(ib(m,7))*( dk(pix(ib(m,6)),3,4)+dk(pix(ib(m,7)),2,1)+
&dk(pix(ib(m,14)),7,8)+dk(pix(ib(m,15)),6,5) )+
&h(ib(m,8))*( dk(pix(ib(m,7)),2,4)+dk(pix(ib(m,15)),6,8) )+
&h(ib(m,9))*( dk(pix(ib(m,25)),5,4)+dk(pix(ib(m,15)),6,3) )+
&h(ib(m,10))*( dk(pix(ib(m,25)),5,3) )+

82

&h(ib(m,11))*( dk(pix(ib(m,13)),8,3)+dk(pix(ib(m,25)),5,2) )+
&h(ib(m,12))*( dk(pix(ib(m,13)),8,2) )+
&h(ib(m,13))*( dk(pix(ib(m,13)),8,1)+dk(pix(ib(m,14)),7,2) )+
&h(ib(m,14))*( dk(pix(ib(m,14)),7,1) )+
&h(ib(m,15))*( dk(pix(ib(m,14)),7,4)+dk(pix(ib(m,15)),6,1) )+
&h(ib(m,16))*( dk(pix(ib(m,15)),6,4) )+
&h(ib(m,17))*( dk(pix(ib(m,27)),1,8)+dk(pix(ib(m,7)),2,7) )+
&h(ib(m,18))*( dk(pix(ib(m,27)),1,7) )+
&h(ib(m,19))*( dk(pix(ib(m,27)),1,6)+dk(pix(ib(m,5)),4,7) )+
&h(ib(m,20))*( dk(pix(ib(m,5)),4,6) )+
&h(ib(m,21))*( dk(pix(ib(m,5)),4,5)+dk(pix(ib(m,6)),3,6) )+
&h(ib(m,22))*( dk(pix(ib(m,6)),3,5) )+
&h(ib(m,23))*( dk(pix(ib(m,6)),3,8)+dk(pix(ib(m,7)),2,5) )+
&h(ib(m,24))*( dk(pix(ib(m,7)),2,8) )+
&h(ib(m,25))*( dk(pix(ib(m,14)),7,3)+dk(pix(ib(m,13)),8,4)+
&dk(pix(ib(m,15)),6,2)+dk(pix(ib(m,25)),5,1) )+
&h(ib(m,26))*( dk(pix(ib(m,6)),3,7)+dk(pix(ib(m,5)),4,8)+
&dk(pix(ib(m,27)),1,5)+dk(pix(ib(m,7)),2,6) )+
&h(ib(m,27))*( dk(pix(ib(m,27)),1,1)+dk(pix(ib(m,7)),2,2)+
&dk(pix(ib(m,6)),3,3)+dk(pix(ib(m,5)),4,4)+dk(pix(ib(m,25)),5,5)+
&dk(pix(ib(m,15)),6,6)+dk(pix(ib(m,14)),7,7)+
&dk(pix(ib(m,13)),8,8) )
400

continue

530

hAh=0.0
do 530 m=1,ns
hAh=hAh+h(m)*Ah(m)
continue

540

lambda=gg/hAh
do 540 m=1,ns
u(m)=u(m)-lambda*h(m)
gb(m)=gb(m)-lambda*Ah(m)
continue

550

gglast=gg
gg=0.0
do 550 m=1,ns
gg=gg+gb(m)*gb(m)
continue
if(gg.le.gtest) goto 1000
gamma=gg/gglast
do 570 m=1,ns

83

570

h(m)=gb(m)+gamma*h(m)
continue

800

continue

1000

continue
return
end

c Subroutine that computes average current in three directions
subroutine current(nx,ny,nz,ns)
real af(3,8)
real u(8000),uu(8)
real sigma(100,3,3)
integer*4 ib(8000,27)
integer*2 pix(8000)
common/list1/currx,curry,currz
common/list2/ex,ey,ez
common/list3/ib
common/list4/pix
common/list6/u
common/list8/sigma

c
c

nxy=nx*ny
af is the average field matrix, average field in a pixel is af*u(pixel).
The matrix af relates the nodal voltages to the average field in the pixel.

c Set up single element average field matrix
af(1,1)=0.25
af(1,2)=-0.25
af(1,3)=-0.25
af(1,4)=0.25
af(1,5)=0.25
af(1,6)=-0.25
af(1,7)=-0.25
af(1,8)=0.25
af(2,1)=0.25
af(2,2)=0.25
af(2,3)=-0.25
af(2,4)=-0.25

84

af(2,5)=0.25
af(2,6)=0.25
af(2,7)=-0.25
af(2,8)=-0.25
af(3,1)=0.25
af(3,2)=0.25
af(3,3)=0.25
af(3,4)=0.25
af(3,5)=-0.25
af(3,6)=-0.25
af(3,7)=-0.25
af(3,8)=-0.25
c

c

c

c
c
c

now compute current in each pixel
currx=0.0
curry=0.0
currz=0.0
compute average field in each pixel
do 470 k=1,nz
do 470 j=1,ny
do 470 i=1,nx
m=(k-1)*nxy+(j-1)*nx+i
load in elements of 8-vector using pd. bd. conds.
uu(1)=u(m)
uu(2)=u(ib(m,3))
uu(3)=u(ib(m,2))
uu(4)=u(ib(m,1))
uu(5)=u(ib(m,26))
uu(6)=u(ib(m,19))
uu(7)=u(ib(m,18))
uu(8)=u(ib(m,17))
Correct for periodic boundary conditions, some voltages are wrong
for a pixel on a periodic boundary. Since they come from an opposite
face, need to put in applied fields to correct them.
if(i.eq.nx) then
uu(2)=uu(2)-ex*nx
uu(3)=uu(3)-ex*nx
uu(6)=uu(6)-ex*nx
uu(7)=uu(7)-ex*nx
end if
if(j.eq.ny) then
uu(3)=uu(3)-ey*ny
uu(4)=uu(4)-ey*ny
uu(7)=uu(7)-ey*ny
uu(8)=uu(8)-ey*ny

85

end if
if(k.eq.nz) then
uu(5)=uu(5)-ez*nz
uu(6)=uu(6)-ez*nz
uu(7)=uu(7)-ez*nz
uu(8)=uu(8)-ez*nz
end if
c cur1, cur2, cur3 are the local currents averaged over the pixel
cur1=0.0
cur2=0.0
cur3=0.0
do 465 n=1,8
do 465 nn=1,3
cur1=cur1+sigma(pix(m),1,nn)*af(nn,n)*uu(n)
cur2=cur2+sigma(pix(m),2,nn)*af(nn,n)*uu(n)
cur3=cur3+sigma(pix(m),3,nn)*af(nn,n)*uu(n)
465
continue
c sum into the global average currents
currx=currx+cur1
curry=curry+cur2
currz=currz+cur3
470
continue
c Volume average currents
currx=currx/float(ns)
curry=curry/float(ns)
currz=currz/float(ns)
return
end
c

Subroutine that counts phase volume fractions
subroutine assig(ns,nphase,prob)
integer ns,nphase
integer*2 pix(8000)
real prob(100)
common/list4/pix

do 90 i=1,nphase
prob(i)=0.0
90
continue
do 100 m=1,ns

86

do 100 i=1,nphase
if(pix(m).eq.i) then
prob(i)=prob(i)+1
endif
100
continue
do 110 i=1,nphase
prob(i)=prob(i)/float(ns)
110
continue
return
end
c

Subroutine that sets up microstructural image
subroutine ppixel(nx,ny,nz,ns,nphase)
integer*2 pix(8000)
common/list4/pix

c
c

100

(USER) If you want to set up a test image inside the program, instead
of reading it in from a file, this should be done inside this subroutine.
do 100 k=1,nz
do 100 j=1,ny
do 100 i=1,nx
m=nx*ny*(k-1)+nx*(j-1)+i
read(9,*) pix(m)
continue

c

Check for wrong phase labels--less than 1 or greater than nphase
do 500 m=1,ns
if(pix(m).lt.1) then
write(7,*) 'Phase label in pix < 1--error at ',m
end if
if(pix(m).gt.nphase) then
write(7,*) 'Phase label in pix > nphase--error at ',m
end if
500
continue
return
end

87

9.3.2 ELAS3D.F
c
c

***********************
BACKGROUND

elas3d.f

***********************************

c
c
c
c
c
c
c
c

This program solves the linear elastic equations in a
random linear elastic material, subject to an applied macroscopic strain,
using the finite element method. Each pixel in the 3-D digital
image is a cubic tri-linear finite element, having its own
elastic moduli tensor. Periodic boundary conditions are maintained.
In the comments below, (USER) means that this is a section of code that
the user might have to change for his particular problem. Therefore the
user is encouraged to search for this string.

c

PROBLEM AND VARIABLE DEFINITION

c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c

The problem being solved is the minimization of the energy
1/2 uAu + b u + C, where A is the Hessian matrix composed of the
stiffness matrices (dk) for each pixel/element, b is a constant vector
and C is a constant that are determined by the applied strain and
the periodic boundary conditions, and u is a vector of
all the displacements. The solution
method used is the conjugate gradient relaxation algorithm.
Other variables are: gb is the gradient = Au+b, h and Ah are
auxiliary variables used in the conjugate gradient algorithm (in dembx),
dk(n,i,j) is the stiffness matrix of the n'th phase, cmod(n,i,j) is
the elastic moduli tensor of the n'th phase, pix is a vector that gives
the phase label of each pixel, ib is a matrix that gives the labels of
the 27 (counting itself) neighbors of a given node, prob is the volume
fractions of the various phases,
strxx, stryy, strzz, strxz, stryz, and strxy are the six Voigt
volume averaged total stresses, and
sxx, syy, szz, sxz, syz, and sxy are the six Voigt
volume averaged total strains.

c

DIMENSIONS

c
c
c
c
c
c
c
c

The vectors u,gb,b,h, and Ah are dimensioned to be the system size,
ns=nx*ny*nz, with three components, where the digital image of the
microstructure considered is a rectangular paralleliped, nx x ny x nz
in size. The arrays pix and ib are are also dimensioned to the system size.
The array ib has 27 components, for the 27 neighbors of a node.
Note that the program is set up at present to have at most 100
different phases. This can easily be changed, simply by changing
the dimensions of dk, prob, and cmod. The parameter nphase gives the

88

c
c

number of phases being considered in the problem.
All arrays are passed between subroutines using simple common statements.

c

STRONGLY SUGGESTED:

READ THE MANUAL BEFORE USING PROGRAM!!

c (USER) Change these dimensions and in other subroutines at same time.
c For example, search and replace all occurrences throughout the
c program of "(8000" by "(64000", to go from a
c 20 x 20 x 20 system to a 40 x 40 x 40 system.
real u(8000,3),gb(8000,3),b(8000,3)
real h(8000,3),Ah(8000,3)
real cmod(100,6,6),dk(100,8,3,8,3)
real phasemod(100,2),prob(100)
integer in(27),jn(27),kn(27)
integer*4 ib(8000,27)
integer*2 pix(8000)
common/list1/strxx,stryy,strzz,strxz,stryz,strxy
common/list2/exx,eyy,ezz,exz,eyz,exy
common/list3/ib
common/list4/pix
common/list5/dk,b,C
common/list6/u
common/list7/gb
common/list8/cmod
common/list9/h,Ah
common/list10/sxx,syy,szz,sxz,syz,sxy
c (USER) Unit 9 is the microstructure input file,
c unit 7 is the results output file.
open (unit=9,file='microstructure.dat')
open (unit=7,file='outputfile.out')
c (USER)

nx,ny,nz give the size of the lattice
nx=20
ny=20
nz=20
c ns=total number of sites
ns=nx*ny*nz
write(7,9010) nx,ny,nz,ns
9010 format('nx= ',i4,' ny= ',i4,' nz= ',i4,' ns= 'i8)
c (USER) nphase is the number of phases being considered in the problem.
c The values of pix(m) will run from 1 to nphase.
nphase=2

89

c
c
c
c

(USER) gtest is the stopping criterion, the number
to which the quantity gg=gb*gb is compared.
Usually gtest = abc*ns, so that when gg < gtest, the rms value
per pixel of gb is less than sqrt(abc).
gtest=1.e-12*ns

c (USER)
c The parameter phasemod(i,j) is the bulk (i,1) and shear (i,2) moduli of
c the i'th phase. These can be input in terms of Young's moduli E(i,1) and
c Poisson's ratio nu (i,2). The program, in do loop 1144, then changes them
c to bulk and shear moduli, using relations for isotropic elastic
c moduli. For anisotropic elastic material, one can directly input
c the elastic moduli tensor cmod in subroutine femat, and skip this part.
c If you wish to input in terms of bulk (1) and shear (2), then make sure
c to comment out the do 1144 loop.
phasemod(1,1)=1.0
phasemod(1,2)=0.2
phasemod(2,1)=0.5
phasemod(2,2)=0.2
c
c

(USER) Program uses bulk modulus (1) and shear modulus (2), so transform
Young's modulis (1) and Poisson's ratio (2).
do 1144 i=1,nphase
save=phasemod(i,1)
phasemod(i,1)=phasemod(i,1)/3./(1.-2.*phasemod(i,2))
phasemod(i,2)=save/2./(1.+phasemod(i,2))
1144
continue
c

Construct the neighbor table, ib(m,n)

c
c

First construct the 27 neighbor table in terms of delta i, delta j, and
delta k information (see Table 3 in manual)
in(1)=0
in(2)=1
in(3)=1
in(4)=1
in(5)=0
in(6)=-1
in(7)=-1
in(8)=-1
jn(1)=1
jn(2)=1
jn(3)=0

90

jn(4)=-1
jn(5)=-1
jn(6)=-1
jn(7)=0
jn(8)=1

555

c
c
c

do 555 n=1,8
kn(n)=0
kn(n+8)=-1
kn(n+16)=1
in(n+8)=in(n)
in(n+16)=in(n)
jn(n+8)=jn(n)
jn(n+16)=jn(n)
continue
in(25)=0
in(26)=0
in(27)=0
jn(25)=0
jn(26)=0
jn(27)=0
kn(25)=-1
kn(26)=1
kn(27)=0
Now construct neighbor table according to 1-d labels
Matrix ib(m,n) gives the 1-d label of the n'th neighbor (n=1,27) of
the node labelled m.
nxy=nx*ny
do 1020 k=1,nz
do 1020 j=1,ny
do 1020 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
do 1004 n=1,27
i1=i+in(n)
j1=j+jn(n)
k1=k+kn(n)
if(i1.lt.1) i1=i1+nx
if(i1.gt.nx) i1=i1-nx
if(j1.lt.1) j1=j1+ny
if(j1.gt.ny) j1=j1-ny
if(k1.lt.1) k1=k1+nz
if(k1.gt.nz) k1=k1-nz
m1=nxy*(k1-1)+nx*(j1-1)+i1
ib(m,n)=m1

91

1004
1020

continue
continue

c Compute the average stress and strain in each microstructure.
c (USER) npoints is the number of microstructures to use.
npoints=1
do 8000 micro=1,npoints
c Read in a microstructure in subroutine ppixel, and set up pix(m)
c with the appropriate phase assignments.
call ppixel(nx,ny,nz,ns,nphase)
c Count and output the volume fractions of the different phases
call assig(ns,nphase,prob)
do 111 i=1,nphase
write(7,9020) i,phasemod(i,1),phasemod(i,2)
9020
format(' Phase ',i3,' bulk = ',f12.6,' shear = ',f12.6)
111 continue
do 8050 i=1,nphase
write(7,9065) i,prob(i)
9065 format(' Volume fraction of phase ',i3,'
8050 continue
c
c
c
c

is ',f8.5)

(USER) Set applied strains
Actual shear strain applied in do 1050 loop is exy, exz, and eyz as
given in the statements below. The engineering shear strain, by which
the shear modulus is usually defined, is twice these values.
exx=0.1
eyy=0.1
ezz=0.1
exz=0.1/2.
eyz=0.2/2.
exy=0.3/2.
write(7,*) 'Applied engineering strains'
write(7,*) ' exx eyy ezz exz eyz exy'
write(7,*) exx,eyy,ezz,2.*exz,2.*eyz,2.*exy

c Set up the elastic modulus variables, finite element stiffness matrices,
c the constant, C, and vector, b, required for computing the energy.
c (USER) If anisotropic elastic moduli tensors are used, these need to be
c input in subroutine femat.
call femat(nx,ny,nz,ns,phasemod,nphase)
c Apply chosen strains as a homogeneous macroscopic strain

92

c as the initial condition.
do 1050 k=1,nz
do 1050 j=1,ny
do 1050 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
x=float(i-1)
y=float(j-1)
z=float(k-1)
u(m,1)=x*exx+y*exy+z*exz
u(m,2)=x*exy+y*eyy+z*eyz
u(m,3)=x*exz+y*eyz+z*ezz
1050 continue
c
c
c
c
c

RELAXATION LOOP
(USER) kmax is the maximum number of times dembx will be called, with
ldemb conjugate gradient steps performed during each call. The total
number of conjugate gradient steps allowed for a given elastic
computation is kmax*ldemb.
kmax=40
ldemb=50
ltot=0
c Call energy to get initial energy and initial gradient
call energy(nx,ny,nz,ns,utot)
c gg is the norm squared of the gradient (gg=gb*gb)
gg=0.0
do 100 m3=1,3
do 100 m=1,ns
gg=gg+gb(m,m3)*gb(m,m3)
100
continue
write(7,*) 'Initial energy = ',utot,' gg = ',gg
call flush(7)

do 5000 kkk=1,kmax
c call dembx to go into the conjugate gradient solver
call dembx(ns,Lstep,gg,dk,gtest,ldemb,kkk)
ltot=ltot+Lstep
c Call energy to compute energy after dembx call. If gg < gtest, this
c will be the final energy. If gg is still larger than gtest, then this
c will give an intermediate energy with which to check how the
c relaxation process is coming along.
call energy(nx,ny,nz,ns,utot)
write(7,*) 'Energy = ',utot,' gg = ',gg
write(7,*) 'Number of conjugate steps = ',ltot
call flush(7)
c If relaxation process is finished, jump out of loop

93

if(gg.le.gtest) goto 444
c If relaxation process will continue, compute and output stresses
c and strains as an additional aid to judge how the
c relaxation procedure is progressing.
call stress(nx,ny,nz,ns)
write(7,*) ' stresses: xx,yy,zz,xz,yz,xy'
write(7,*) strxx,stryy,strzz,strxz,stryz,strxy
write(7,*) ' strains: xx,yy,zz,xz,yz,xy'
write(7,*) sxx,syy,szz,sxz,syz,sxy
5000
continue
444

call stress(nx,ny,nz,ns)
write(7,*) ' stresses: xx,yy,zz,xz,yz,xy'
write(7,*) strxx,stryy,strzz,strxz,stryz,strxy
write(7,*) ' strains: xx,yy,zz,xz,yz,xy'
write(7,*) sxx,syy,szz,sxz,syz,sxy
8000

continue
end

c
c
c
c

Subroutine that sets up the elastic moduli variables,
the stiffness matrices,dk, the linear term in
displacements, b, and the constant term, C, that appear in the total energy
due to the periodic boundary conditions
subroutine femat(nx,ny,nz,ns,phasemod,nphase)
real dk(100,8,3,8,3),phasemod(100,2),dndx(8),dndy(8),dndz(8)
real b(8000,3),g(3,3,3),C,ck(6,6),cmu(6,6),cmod(100,6,6)
real es(6,8,3),delta(8,3)
integer is(8)
integer*4 ib(8000,27)
integer*2 pix(8000)

common/list2/exx,eyy,ezz,exz,eyz,exy
common/list3/ib
common/list4/pix
common/list5/dk,b,C
common/list8/cmod
nxy=nx*ny
c
c
c

(USER) NOTE: complete elastic modulus matrix is used, so an anisotropic
matrix could be directly input at any point, since program is written
to use a general elastic moduli tensor, but is only explicitly

94

c

implemented for isotropic materials.

c

initialize stiffness matrices
do 40 m=1,nphase
do 40 l=1,3
do 40 k=1,3
do 40 j=1,8
do 40 i=1,8
dk(m,i,k,j,l)=0.0
40
continue
c set up elastic moduli matrices for each kind of element
c ck and cmu are the bulk and shear modulus matrices, which need to be
c weighted by the actual bulk and shear moduli
ck(1,1)=1.0
ck(1,2)=1.0
ck(1,3)=1.0
ck(1,4)=0.0
ck(1,5)=0.0
ck(1,6)=0.0
ck(2,1)=1.0
ck(2,2)=1.0
ck(2,3)=1.0
ck(2,4)=0.0
ck(2,5)=0.0
ck(2,6)=0.0
ck(3,1)=1.0
ck(3,2)=1.0
ck(3,3)=1.0
ck(3,4)=0.0
ck(3,5)=0.0
ck(3,6)=0.0
ck(4,1)=0.0
ck(4,2)=0.0
ck(4,3)=0.0
ck(4,4)=0.0
ck(4,5)=0.0
ck(4,6)=0.0
ck(5,1)=0.0
ck(5,2)=0.0
ck(5,3)=0.0
ck(5,4)=0.0
ck(5,5)=0.0
ck(5,6)=0.0

95

ck(6,1)=0.0
ck(6,2)=0.0
ck(6,3)=0.0
ck(6,4)=0.0
ck(6,5)=0.0
ck(6,6)=0.0
cmu(1,1)=4.0/3.0
cmu(1,2)=-2.0/3.0
cmu(1,3)=-2.0/3.0
cmu(1,4)=0.0
cmu(1,5)=0.0
cmu(1,6)=0.0
cmu(2,1)=-2.0/3.0
cmu(2,2)=4.0/3.0
cmu(2,3)=-2.0/3.0
cmu(2,4)=0.0
cmu(2,5)=0.0
cmu(2,6)=0.0
cmu(3,1)=-2.0/3.0
cmu(3,2)=-2.0/3.0
cmu(3,3)=4.0/3.0
cmu(3,4)=0.0
cmu(3,5)=0.0
cmu(3,6)=0.0
cmu(4,1)=0.0
cmu(4,2)=0.0
cmu(4,3)=0.0
cmu(4,4)=1.0
cmu(4,5)=0.0
cmu(4,6)=0.0
cmu(5,1)=0.0
cmu(5,2)=0.0
cmu(5,3)=0.0
cmu(5,4)=0.0
cmu(5,5)=1.0
cmu(5,6)=0.0
cmu(6,1)=0.0
cmu(6,2)=0.0
cmu(6,3)=0.0
cmu(6,4)=0.0
cmu(6,5)=0.0
cmu(6,6)=1.0
do 31 k=1,nphase

96

11
21
31

do 21 j=1,6
do 11 i=1,6
cmod(k,i,j)=phasemod(k,1)*ck(i,j)+phasemod(k,2)*cmu(i,j)
continue
continue
continue

c

set up Simpson's integration rule weight vector
do 30 k=1,3
do 30 j=1,3
do 30 i=1,3
nm=0
if(i.eq.2) nm=nm+1
if(j.eq.2) nm=nm+1
if(k.eq.2) nm=nm+1
g(i,j,k)=4.0**nm
30
continue
c
c
c
c

c
c

loop over the nphase kinds of pixels and Simpson's rule quadrature
points in order to compute the stiffness matrices. Stiffness matrices
of trilinear finite elements are quadratic in x, y, and z, so that
Simpson's rule quadrature gives exact results.
do 4000 ijk=1,nphase
do 3000 k=1,3
do 3000 j=1,3
do 3000 i=1,3
x=float(i-1)/2.0
y=float(j-1)/2.0
z=float(k-1)/2.0
dndx means the negative derivative, with respect to x, of the shape
matrix N (see manual, Sec. 2.2), dndy, and dndz are similar.
dndx(1)=-(1.0-y)*(1.0-z)
dndx(2)=(1.0-y)*(1.0-z)
dndx(3)=y*(1.0-z)
dndx(4)=-y*(1.0-z)
dndx(5)=-(1.0-y)*z
dndx(6)=(1.0-y)*z
dndx(7)=y*z
dndx(8)=-y*z
dndy(1)=-(1.0-x)*(1.0-z)
dndy(2)=-x*(1.0-z)
dndy(3)=x*(1.0-z)
dndy(4)=(1.0-x)*(1.0-z)
dndy(5)=-(1.0-x)*z
dndy(6)=-x*z

97

dndy(7)=x*z
dndy(8)=(1.0-x)*z
dndz(1)=-(1.0-x)*(1.0-y)
dndz(2)=-x*(1.0-y)
dndz(3)=-x*y
dndz(4)=-(1.0-x)*y
dndz(5)=(1.0-x)*(1.0-y)
dndz(6)=x*(1.0-y)
dndz(7)=x*y
dndz(8)=(1.0-x)*y
c now build strain matrix
do 2799 n1=1,6
do 2799 n2=1,8
do 2799 n3=1,3
es(n1,n2,n3)=0.0
2799 continue
do 2797 n=1,8
es(1,n,1)=dndx(n)
es(2,n,2)=dndy(n)
es(3,n,3)=dndz(n)
es(4,n,1)=dndz(n)
es(4,n,3)=dndx(n)
es(5,n,2)=dndz(n)
es(5,n,3)=dndy(n)
es(6,n,1)=dndy(n)
es(6,n,2)=dndx(n)
2797 continue
c Matrix multiply to determine value at (x,y,z), multiply by
c proper weight, and sum into dk, the stiffness matrix
do 900 mm=1,3
do 900 nn=1,3
do 900 ii=1,8
do 900 jj=1,8
c Define sum over strain matrices and elastic moduli matrix for
c stiffness matrix
sum=0.0
do 890 kk=1,6
do 890 ll=1,6
sum=sum+es(kk,ii,mm)*cmod(ijk,kk,ll)*es(ll,jj,nn)
890
continue
dk(ijk,ii,mm,jj,nn)=dk(ijk,ii,mm,jj,nn)+g(i,j,k)*sum/216.
900
continue
3000 continue
4000 continue

98

c
c
c
c
c
c
c

Set up vector for linear term, b, and constant term, C,
in the elastic energy. This is done using the stiffness matrices,
and the periodic terms in the applied strain that come in at the
boundary pixels via the periodic boundary conditions and the
condition that an applied macroscopic strain exists (see Sec. 2.2
in the manual). It is easier to set b up this way than to analytically
write out all the terms involved.

c

Initialize b and C
do 5000 m3=1,3
do 5000 m=1,ns
b(m,m3)=0.0
5000 continue
C=0.0
c
c
c
c
c
c

For all cases, the correspondence between 1-8 finite element node
labels and 1-27 neighbor labels is (see Table 4 in manual):
1:ib(m,27), 2:ib(m,3),
3:ib(m,2),4:ib(m,1),
5:ib(m,26),6:ib(m,19)
7:ib(m,18),8:ib(m,17)
is(1)=27
is(2)=3
is(3)=2
is(4)=1
is(5)=26
is(6)=19
is(7)=18
is(8)=17

c

x=nx face
do 2001 i3=1,3
do 2001 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3.or.i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
2001 continue
do 2000 j=1,ny-1
do 2000 k=1,nz-1
m=nxy*(k-1)+j*nx
do 1900 nn=1,3
do 1900 mm=1,8

99

sum=0.0
do 1899 m3=1,3
do 1899 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
C=C+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
1899 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1900 continue
2000 continue
c y=ny face
do 2011 i3=1,3
do 2011 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.3.or.i8.eq.4.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
2011 continue
do 2010 i=1,nx-1
do 2010 k=1,nz-1
m=nxy*(k-1)+nx*(ny-1)+i
do 1901 nn=1,3
do 1901 mm=1,8
sum=0.0
do 2099 m3=1,3
do 2099 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
C=C+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2099 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1901 continue
2010 continue
c z=nz face
do 2021 i3=1,3
do 2021 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
2021 continue
do 2020 i=1,nx-1
do 2020 j=1,ny-1

100

m=nxy*(nz-1)+nx*(j-1)+i
do 1902 nn=1,3
do 1902 mm=1,8
sum=0.0
do 2019 m3=1,3
do 2019 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
C=C+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2019 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1902 continue
2020 continue
c x=nx y=ny edge
do 2031 i3=1,3
do 2031 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.6) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.3.or.i8.eq.7) then
delta(i8,1)=exy*ny+exx*nx
delta(i8,2)=eyy*ny+exy*nx
delta(i8,3)=eyz*ny+exz*nx
end if
2031 continue
do 2030 k=1,nz-1
m=nxy*k
do 1903 nn=1,3
do 1903 mm=1,8
sum=0.0
do 2029 m3=1,3
do 2029 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
C=C+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2029 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1903 continue
2030 continue

101

c

x=nx z=nz edge
do 2041 i3=1,3
do 2041 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.5.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exz*nz+exx*nx
delta(i8,2)=eyz*nz+exy*nx
delta(i8,3)=ezz*nz+exz*nx
end if
2041 continue
do 2040 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+nx
do 1904 nn=1,3
do 1904 mm=1,8
sum=0.0
do 2039 m3=1,3
do 2039 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
C=C+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2039 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1904 continue
2040 continue
c y=ny z=nz edge
do 2051 i3=1,3
do 2051 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.3.or.i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny

102

delta(i8,3)=eyz*ny
end if
if(i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz
end if
2051 continue
do 2050 i=1,nx-1
m=nxy*(nz-1)+nx*(ny-1)+i
do 1905 nn=1,3
do 1905 mm=1,8
sum=0.0
do 2049 m3=1,3
do 2049 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
C=C+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2049 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1905 continue
2050 continue
c x=nx y=ny z=nz corner
do 2061 i3=1,3
do 2061 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.5) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz
end if

103

2061

2059
1906

if(i8.eq.6) then
delta(i8,1)=exx*nx+exz*nz
delta(i8,2)=exy*nx+eyz*nz
delta(i8,3)=exz*nx+ezz*nz
end if
if(i8.eq.3) then
delta(i8,1)=exx*nx+exy*ny
delta(i8,2)=exy*nx+eyy*ny
delta(i8,3)=exz*nx+eyz*ny
end if
if(i8.eq.7) then
delta(i8,1)=exx*nx+exy*ny+exz*nz
delta(i8,2)=exy*nx+eyy*ny+eyz*nz
delta(i8,3)=exz*nx+eyz*ny+ezz*nz
end if
continue
m=nx*ny*nz
do 1906 nn=1,3
do 1906 mm=1,8
sum=0.0
do 2059 m3=1,3
do 2059 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
C=C+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
continue
return
end

c

Subroutine computes the total energy, utot, and the gradient, gb
subroutine energy(nx,ny,nz,ns,utot)

real u(8000,3),gb(8000,3)
real b(8000,3),C,utot
real dk(100,8,3,8,3)
integer*4 ib(8000,27)
integer*2 pix(8000)
common/list2/exx,eyy,ezz,exz,eyz,exy
common/list3/ib
common/list4/pix
common/list5/dk,b,C

104

common/list6/u
common/list7/gb
do 2090 m3=1,3
do 2090 m=1,ns
gb(m,m3)=0.0
2090 continue
c
c
c

Do global matrix multiply via small stiffness matrices, gb = A * u
The long statement below correctly brings in all the terms from
the global matrix A using only the small stiffness matrices.

do 3000 j=1,3
do 3000 n=1,3
do 3000 m=1,ns
gb(m,j)=gb(m,j)+u(ib(m,1),n)*( dk(pix(ib(m,27)),1,j,4,n)
&+dk(pix(ib(m,7)),2,j,3,n)
&+dk(pix(ib(m,25)),5,j,8,n)+dk(pix(ib(m,15)),6,j,7,n) )+
&u(ib(m,2),n)*( dk(pix(ib(m,27)),1,j,3,n)
&+dk(pix(ib(m,25)),5,j,7,n) )+
&u(ib(m,3),n)*( dk(pix(ib(m,27)),1,j,2,n)+dk(pix(ib(m,5)),4,j,3,n)+
&dk(pix(ib(m,13)),8,j,7,n)+dk(pix(ib(m,25)),5,j,6,n) )+
&u(ib(m,4),n)*( dk(pix(ib(m,5)),4,j,2,n)
&+dk(pix(ib(m,13)),8,j,6,n) )+
&u(ib(m,5),n)*( dk(pix(ib(m,6)),3,j,2,n)+dk(pix(ib(m,5)),4,j,1,n)+
&dk(pix(ib(m,14)),7,j,6,n)+dk(pix(ib(m,13)),8,j,5,n) )+
&u(ib(m,6),n)*( dk(pix(ib(m,6)),3,j,1,n)
&+dk(pix(ib(m,14)),7,j,5,n) )+
&u(ib(m,7),n)*( dk(pix(ib(m,6)),3,j,4,n)+dk(pix(ib(m,7)),2,j,1,n)+
&dk(pix(ib(m,14)),7,j,8,n)+dk(pix(ib(m,15)),6,j,5,n) )+
&u(ib(m,8),n)*( dk(pix(ib(m,7)),2,j,4,n)
&+dk(pix(ib(m,15)),6,j,8,n) )+
&u(ib(m,9),n)*( dk(pix(ib(m,25)),5,j,4,n)
&+dk(pix(ib(m,15)),6,j,3,n) )+
&u(ib(m,10),n)*( dk(pix(ib(m,25)),5,j,3,n) )+
&u(ib(m,11),n)*( dk(pix(ib(m,13)),8,j,3,n)
&+dk(pix(ib(m,25)),5,j,2,n) )+
&u(ib(m,12),n)*( dk(pix(ib(m,13)),8,j,2,n) )+
&u(ib(m,13),n)*( dk(pix(ib(m,13)),8,j,1,n)
&+dk(pix(ib(m,14)),7,j,2,n) )+
&u(ib(m,14),n)*( dk(pix(ib(m,14)),7,j,1,n) )+
&u(ib(m,15),n)*( dk(pix(ib(m,14)),7,j,4,n)
&+dk(pix(ib(m,15)),6,j,1,n) )+
&u(ib(m,16),n)*( dk(pix(ib(m,15)),6,j,4,n) )+
&u(ib(m,17),n)*( dk(pix(ib(m,27)),1,j,8,n)

105

&+dk(pix(ib(m,7)),2,j,7,n) )+
&u(ib(m,18),n)*( dk(pix(ib(m,27)),1,j,7,n) )+
&u(ib(m,19),n)*( dk(pix(ib(m,27)),1,j,6,n)
&+dk(pix(ib(m,5)),4,j,7,n) )+
&u(ib(m,20),n)*( dk(pix(ib(m,5)),4,j,6,n) )+
&u(ib(m,21),n)*( dk(pix(ib(m,5)),4,j,5,n)
&+dk(pix(ib(m,6)),3,j,6,n) )+
&u(ib(m,22),n)*( dk(pix(ib(m,6)),3,j,5,n) )+
&u(ib(m,23),n)*( dk(pix(ib(m,6)),3,j,8,n)
&+dk(pix(ib(m,7)),2,j,5,n) )+
&u(ib(m,24),n)*( dk(pix(ib(m,7)),2,j,8,n) )+
&u(ib(m,25),n)*( dk(pix(ib(m,14)),7,j,3,n)
&+dk(pix(ib(m,13)),8,j,4,n)+
&dk(pix(ib(m,15)),6,j,2,n)+dk(pix(ib(m,25)),5,j,1,n) )+
&u(ib(m,26),n)*( dk(pix(ib(m,6)),3,j,7,n)
&+dk(pix(ib(m,5)),4,j,8,n)+
&dk(pix(ib(m,27)),1,j,5,n)+dk(pix(ib(m,7)),2,j,6,n) )+
&u(ib(m,27),n)*( dk(pix(ib(m,27)),1,j,1,n)
&+dk(pix(ib(m,7)),2,j,2,n)+
&dk(pix(ib(m,6)),3,j,3,n)+dk(pix(ib(m,5)),4,j,4,n)
&+dk(pix(ib(m,25)),5,j,5,n)+
&dk(pix(ib(m,15)),6,j,6,n)+dk(pix(ib(m,14)),7,j,7,n)+
&dk(pix(ib(m,13)),8,j,8,n) )
3000 continue
utot=C
do 3100 m3=1,3
do 3100 m=1,ns
utot=utot+0.5*u(m,m3)*gb(m,m3)+b(m,m3)*u(m,m3)
gb(m,m3)=gb(m,m3)+b(m,m3)
3100 continue
return
end
c

Subroutine that carries out the conjugate gradient relaxation process
subroutine dembx(ns,Lstep,gg,dk,gtest,ldemb,kkk)
real gb(8000,3),u(8000,3),dk(100,8,3,8,3)
real h(8000,3),Ah(8000,3)
real lambda,gamma
integer*4 ib(8000,27)
integer*2 pix(8000)
common/list3/ib

106

common/list4/pix
common/list6/u
common/list7/gb
common/list9/h,Ah
c
c
c
c
c
c

Initialize the conjugate direction vector on first call to dembx only
For calls to dembx after the first, we want to continue using the
value of h determined in the previous call. Of course, if npoints is
greater than 1, this initialization step will be run for every new
microstructure used, as kkk is reset to 1 every time the counter micro
is increased.
if(kkk.eq.1) then
do 500 m3=1,3
do 500 m=1,ns
h(m,m3)=gb(m,m3)
500
continue
end if
c Lstep counts the number of conjugate gradient steps taken in
c each call to dembx
Lstep=0
do 800 ijk=1,ldemb
Lstep=Lstep+1
do 290 m3=1,3
do 290 m=1,ns
Ah(m,m3)=0.0
290
continue
c Do global matrix multiply via small stiffness matrices, Ah = A * h
c The long statement below correctly brings in all the terms from
c the global matrix A using only the small stiffness matrices dk.
do 400 j=1,3
do 400 n=1,3
do 400 m=1,ns
Ah(m,j)=Ah(m,j)+h(ib(m,1),n)*( dk(pix(ib(m,27)),1,j,4,n)
&+dk(pix(ib(m,7)),2,j,3,n)
&+dk(pix(ib(m,25)),5,j,8,n)+dk(pix(ib(m,15)),6,j,7,n) )+
&h(ib(m,2),n)*( dk(pix(ib(m,27)),1,j,3,n)
&+dk(pix(ib(m,25)),5,j,7,n) )+
&h(ib(m,3),n)*( dk(pix(ib(m,27)),1,j,2,n)+dk(pix(ib(m,5)),4,j,3,n)+
&dk(pix(ib(m,13)),8,j,7,n)+dk(pix(ib(m,25)),5,j,6,n) )+
&h(ib(m,4),n)*( dk(pix(ib(m,5)),4,j,2,n)
&+dk(pix(ib(m,13)),8,j,6,n) )+
&h(ib(m,5),n)*( dk(pix(ib(m,6)),3,j,2,n)+dk(pix(ib(m,5)),4,j,1,n)+
&dk(pix(ib(m,14)),7,j,6,n)+dk(pix(ib(m,13)),8,j,5,n) )+

107

400

&h(ib(m,6),n)*( dk(pix(ib(m,6)),3,j,1,n)
&+dk(pix(ib(m,14)),7,j,5,n) )+
&h(ib(m,7),n)*( dk(pix(ib(m,6)),3,j,4,n)+dk(pix(ib(m,7)),2,j,1,n)+
&dk(pix(ib(m,14)),7,j,8,n)+dk(pix(ib(m,15)),6,j,5,n) )+
&h(ib(m,8),n)*( dk(pix(ib(m,7)),2,j,4,n)
&+dk(pix(ib(m,15)),6,j,8,n) )+
&h(ib(m,9),n)*( dk(pix(ib(m,25)),5,j,4,n)
&+dk(pix(ib(m,15)),6,j,3,n) )+
&h(ib(m,10),n)*( dk(pix(ib(m,25)),5,j,3,n) )+
&h(ib(m,11),n)*( dk(pix(ib(m,13)),8,j,3,n)
&+dk(pix(ib(m,25)),5,j,2,n) )+
&h(ib(m,12),n)*( dk(pix(ib(m,13)),8,j,2,n) )+
&h(ib(m,13),n)*( dk(pix(ib(m,13)),8,j,1,n)
&+dk(pix(ib(m,14)),7,j,2,n) )+
&h(ib(m,14),n)*( dk(pix(ib(m,14)),7,j,1,n) )+
&h(ib(m,15),n)*( dk(pix(ib(m,14)),7,j,4,n)
&+dk(pix(ib(m,15)),6,j,1,n) )+
&h(ib(m,16),n)*( dk(pix(ib(m,15)),6,j,4,n) )+
&h(ib(m,17),n)*( dk(pix(ib(m,27)),1,j,8,n)
&+dk(pix(ib(m,7)),2,j,7,n) )+
&h(ib(m,18),n)*( dk(pix(ib(m,27)),1,j,7,n) )+
&h(ib(m,19),n)*( dk(pix(ib(m,27)),1,j,6,n)
&+dk(pix(ib(m,5)),4,j,7,n) )+
&h(ib(m,20),n)*( dk(pix(ib(m,5)),4,j,6,n) )+
&h(ib(m,21),n)*( dk(pix(ib(m,5)),4,j,5,n)
&+dk(pix(ib(m,6)),3,j,6,n) )+
&h(ib(m,22),n)*( dk(pix(ib(m,6)),3,j,5,n) )+
&h(ib(m,23),n)*( dk(pix(ib(m,6)),3,j,8,n)
&+dk(pix(ib(m,7)),2,j,5,n) )+
&h(ib(m,24),n)*( dk(pix(ib(m,7)),2,j,8,n) )+
&h(ib(m,25),n)*( dk(pix(ib(m,14)),7,j,3,n)
&+dk(pix(ib(m,13)),8,j,4,n)+
&dk(pix(ib(m,15)),6,j,2,n)+dk(pix(ib(m,25)),5,j,1,n) )+
&h(ib(m,26),n)*( dk(pix(ib(m,6)),3,j,7,n)
&+dk(pix(ib(m,5)),4,j,8,n)+
&dk(pix(ib(m,27)),1,j,5,n)+dk(pix(ib(m,7)),2,j,6,n) )+
&h(ib(m,27),n)*( dk(pix(ib(m,27)),1,j,1,n)
&+dk(pix(ib(m,7)),2,j,2,n)+
&dk(pix(ib(m,6)),3,j,3,n)+dk(pix(ib(m,5)),4,j,4,n)
&+dk(pix(ib(m,25)),5,j,5,n)+
&dk(pix(ib(m,15)),6,j,6,n)+dk(pix(ib(m,14)),7,j,7,n)+
&dk(pix(ib(m,13)),8,j,8,n) )
continue
hAh=0.0

108

530

do 530 m3=1,3
do 530 m=1,ns
hAh=hAh+h(m,m3)*Ah(m,m3)
continue

540

lambda=gg/hAh
do 540 m3=1,3
do 540 m=1,ns
u(m,m3)=u(m,m3)-lambda*h(m,m3)
gb(m,m3)=gb(m,m3)-lambda*Ah(m,m3)
continue
gglast=gg
gg=0.0
do 550 m3=1,3
do 550 m=1,ns
gg=gg+gb(m,m3)*gb(m,m3)
continue
if(gg.lt.gtest) goto 1000

550

570

gamma=gg/gglast
do 570 m3=1,3
do 570 m=1,ns
h(m,m3)=gb(m,m3)+gamma*h(m,m3)
continue

800

continue

1000

continue
return
end

c
c

Subroutine that computes the six average stresses and six
average strains.
subroutine stress(nx,ny,nz,ns)
real u(8000,3),gb(8000,3),uu(8,3)
real dndx(8),dndy(8),dndz(8),es(6,8,3),cmod(100,6,6)
integer*4 ib(8000,27)
integer*2 pix(8000)

common/list1/strxx,stryy,strzz,strxz,stryz,strxy
common/list2/exx,eyy,ezz,exz,eyz,exy
common/list3/ib

109

common/list4/pix
common/list6/u
common/list7/gb
common/list8/cmod
common/list10/sxx,syy,szz,sxz,syz,sxy
nxy=nx*ny
c
c
c

set up single element strain matrix
dndx, dndy, and dndz are the components of the average strain
matrix in a pixel

dndx(1)=-0.25
dndx(2)=0.25
dndx(3)=0.25
dndx(4)=-0.25
dndx(5)=-0.25
dndx(6)=0.25
dndx(7)=0.25
dndx(8)=-0.25
dndy(1)=-0.25
dndy(2)=-0.25
dndy(3)=0.25
dndy(4)=0.25
dndy(5)=-0.25
dndy(6)=-0.25
dndy(7)=0.25
dndy(8)=0.25
dndz(1)=-0.25
dndz(2)=-0.25
dndz(3)=-0.25
dndz(4)=-0.25
dndz(5)=0.25
dndz(6)=0.25
dndz(7)=0.25
dndz(8)=0.25
c Build averaged strain matrix, follows code in femat, but for average
c strain over the pixel, not the strain at a point.
do 2799 n1=1,6
do 2799 n2=1,8
do 2799 n3=1,3
es(n1,n2,n3)=0.0
2799 continue
do 2797 n=1,8
es(1,n,1)=dndx(n)

110

2797

es(2,n,2)=dndy(n)
es(3,n,3)=dndz(n)
es(4,n,1)=dndz(n)
es(4,n,3)=dndx(n)
es(5,n,2)=dndz(n)
es(5,n,3)=dndy(n)
es(6,n,1)=dndy(n)
es(6,n,2)=dndx(n)
continue

c

Compute components of the average stress and strain tensors in each pixel
strxx=0.0
stryy=0.0
strzz=0.0
strxz=0.0
stryz=0.0
strxy=0.0
sxx=0.0
syy=0.0
szz=0.0
sxz=0.0
syz=0.0
sxy=0.0
do 470 k=1,nz
do 470 j=1,ny
do 470 i=1,nx
m=(k-1)*nxy+(j-1)*nx+i
c load in elements of 8-vector using pd. bd. conds.
do 9898 mm=1,3
uu(1,mm)=u(m,mm)
uu(2,mm)=u(ib(m,3),mm)
uu(3,mm)=u(ib(m,2),mm)
uu(4,mm)=u(ib(m,1),mm)
uu(5,mm)=u(ib(m,26),mm)
uu(6,mm)=u(ib(m,19),mm)
uu(7,mm)=u(ib(m,18),mm)
uu(8,mm)=u(ib(m,17),mm)
9898 continue
c Correct for periodic boundary conditions, some displacements are wrong
c for a pixel on a periodic boundary. Since they come from an opposite
c face, need to put in applied strain to correct them.
if(i.eq.nx) then
uu(2,1)=uu(2,1)+exx*nx
uu(2,2)=uu(2,2)+exy*nx
uu(2,3)=uu(2,3)+exz*nx

111

uu(3,1)=uu(3,1)+exx*nx
uu(3,2)=uu(3,2)+exy*nx
uu(3,3)=uu(3,3)+exz*nx
uu(6,1)=uu(6,1)+exx*nx
uu(6,2)=uu(6,2)+exy*nx
uu(6,3)=uu(6,3)+exz*nx
uu(7,1)=uu(7,1)+exx*nx
uu(7,2)=uu(7,2)+exy*nx
uu(7,3)=uu(7,3)+exz*nx
end if
if(j.eq.ny) then
uu(3,1)=uu(3,1)+exy*ny
uu(3,2)=uu(3,2)+eyy*ny
uu(3,3)=uu(3,3)+eyz*ny
uu(4,1)=uu(4,1)+exy*ny
uu(4,2)=uu(4,2)+eyy*ny
uu(4,3)=uu(4,3)+eyz*ny
uu(7,1)=uu(7,1)+exy*ny
uu(7,2)=uu(7,2)+eyy*ny
uu(7,3)=uu(7,3)+eyz*ny
uu(8,1)=uu(8,1)+exy*ny
uu(8,2)=uu(8,2)+eyy*ny
uu(8,3)=uu(8,3)+eyz*ny
end if
if(k.eq.nz) then
uu(5,1)=uu(5,1)+exz*nz
uu(5,2)=uu(5,2)+eyz*nz
uu(5,3)=uu(5,3)+ezz*nz
uu(6,1)=uu(6,1)+exz*nz
uu(6,2)=uu(6,2)+eyz*nz
uu(6,3)=uu(6,3)+ezz*nz
uu(7,1)=uu(7,1)+exz*nz
uu(7,2)=uu(7,2)+eyz*nz
uu(7,3)=uu(7,3)+ezz*nz
uu(8,1)=uu(8,1)+exz*nz
uu(8,2)=uu(8,2)+eyz*nz
uu(8,3)=uu(8,3)+ezz*nz
end if
c

local stresses and strains in a pixel
str11=0.0
str22=0.0
str33=0.0
str13=0.0
str23=0.0

112

str12=0.0
s11=0.0
s22=0.0
s33=0.0
s13=0.0
s23=0.0
s12=0.0
do 465 n3=1,3
do 465 n8=1,8
s11=s11+es(1,n8,n3)*uu(n8,n3)
s22=s22+es(2,n8,n3)*uu(n8,n3)
s33=s33+es(3,n8,n3)*uu(n8,n3)
s13=s13+es(4,n8,n3)*uu(n8,n3)
s23=s23+es(5,n8,n3)*uu(n8,n3)
s12=s12+es(6,n8,n3)*uu(n8,n3)
do 465 n=1,6
str11=str11+cmod(pix(m),1,n)*es(n,n8,n3)*uu(n8,n3)
str22=str22+cmod(pix(m),2,n)*es(n,n8,n3)*uu(n8,n3)
str33=str33+cmod(pix(m),3,n)*es(n,n8,n3)*uu(n8,n3)
str13=str13+cmod(pix(m),4,n)*es(n,n8,n3)*uu(n8,n3)
str23=str23+cmod(pix(m),5,n)*es(n,n8,n3)*uu(n8,n3)
str12=str12+cmod(pix(m),6,n)*es(n,n8,n3)*uu(n8,n3)
465
continue
c sum local strains and stresses into global values
strxx=strxx+str11
stryy=stryy+str22
strzz=strzz+str33
strxz=strxz+str13
stryz=stryz+str23
strxy=strxy+str12
sxx=sxx+s11
syy=syy+s22
szz=szz+s33
sxz=sxz+s13
syz=syz+s23
sxy=sxy+s12
470
continue
c

Volume average of global stresses and strains
strxx=strxx/float(ns)
stryy=stryy/float(ns)
strzz=strzz/float(ns)
strxz=strxz/float(ns)
stryz=stryz/float(ns)
strxy=strxy/float(ns)

113

sxx=sxx/float(ns)
syy=syy/float(ns)
szz=szz/float(ns)
sxz=sxz/float(ns)
syz=syz/float(ns)
sxy=sxy/float(ns)
return
end
c

Subroutine that counts volume fractions
subroutine assig(ns,nphase,prob)
integer*2 pix(8000)
real prob(100)
common/list4/pix

do 90 i=1,nphase
prob(i)=0.0
90
continue
do 100 m=1,ns
do 100 i=1,nphase
if(pix(m).eq.i) then
prob(i)=prob(i)+1
end if
100
continue
do 110 i=1,nphase
prob(i)=prob(i)/float(ns)
110
continue
return
end
c

Subroutine that sets up microstructural image
subroutine ppixel(nx,ny,nz,ns,nphase)
integer*2 pix(8000)
common/list4/pix

c
c

(USER) If you want to set up a test image inside the program, instead of
reading it in from a file, this should be done inside this subroutine.

114

200

nxy=nx*ny
do 200 k=1,nz
do 200 j=1,ny
do 200 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
read(9,*) pix(m)
continue

c

Check for wrong phase labels--less than 1 or greater than nphase
do 500 m=1,ns
if(pix(m).lt.1) then
write(7,*) 'Phase label in pix < 1--error at ',m
end if
if(pix(m).gt.nphase) then
write(7,*) 'Phase label in pix > nphase--error at ',m
end if
500
continue
return
end

115

9.3.3 THERMAL3D.F
c
c

***********************
BACKGROUND

thermal3d.f

c
c
c
c
c
c
c
c
c
c

Program adjusts dimensions of unit cell,
[(1 + macrostrain) times dimension],
in response to phases that have a non-zero eigenstrain and arbitrary
elastic moduli tensors.
All six macrostrains can adjust their values (3-d program), and are
stored in the last two positions in the displacement vector u,
as listed below. Periodic boundaries are maintained.
In the comments below, (USER) means that this is a section of code
that the user might have to change for his particular problem.
Therefore the user is encouraged to search for this string.

c

PROBLEM AND VARIABLE DEFINITION

c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c

The problem being solved is the minimization of the elastic energy
1/2 uAu + bu + C + Tu + Y, where b and C are also functions of the
macrostrains.
The small array zcon computes the thermal strain energy associated
with macrostrains (C term), T is the thermal energy term linear in the
displacements (built from ss), b is the regular energy term linear in the
b is the regular energy term linear in the
displacements, u is the displacements including the macrostrains,
gb is the energy gradient vector, h,Ah are auxiliary vectors,
dk is the single pixel stiffness matrix, pix is the phase
identification vector, and ib is the
integer matrix for mapping labels from the 1-27 nearest neighbor
labelling to the 1-d system labelling.
The array prob(i) contains the volume fractions of the i'th phase,
strxx, etc. are the six independent (Voigt notation) volume
averaged stresses, sxx, etc. are the six independent (Voigt notation)
volume averaged strains (not counting the thermal strains).
The variable cmod(i,6,6) gives the elastic moduli tensor
of the i'th phase, eigen(i,6) gives the six independent elements
of the eigenstrain tensor for the i'th phase (Voigt notation)
and dk(i,8,3,8,3) is the stiffness matrix of the i'th
phase. The parameter nphase gives the number of phases being considered
in the problem, and is set by the user.

c

DIMENSIONS

c

The main arrays of the problem, u, gb, h, Ah, b, and T, are dimensioned

116

***************************

c
c
c
c
c
c
c
c

as (nx*ny*nz)+2, which is the number of nodal displacements plus two for
the macrostrains.
Currently the program assumes the number of different phases is
100, since phasemod and eigen (the moduli and eigenstrains for each phase)
and dk are dimensioned to have at most 100 different kinds. This
is easily changed, by changing the dimension of these three variables
throughout the program. The parameter nphase gives the number of phases
All major arrays are passed to subroutines via simple common statements.

c

NOTE ON USE OF PROGRAM

c
c
c
c
c
c
c
c
c
c
c
c

Program is set up to allow the macrostrains,
which control the overall size of the system, to be dynamic
variables, which are adjusted in order to minimize the overall
energy. That means that if there are no eigenstrains specified
for any of the phases, the overall strain will always relax to
zero. If it is desired to simply apply a fixed strain, with no
eigenstrains, then in subroutines Energy and Dembx, one must
zero out the elements of gb (in Energy and in Dembx) that
correspond to the macrostrains. This is easily done.
This will fix the gradients of the macrostrains to always to be
zero, so that they will not change, so the applied strain (initial
values of the macrostrains) will remain fixed.

c

STRONGLY SUGGESTED:

READ MANUAL BEFORE USING PROGRAM!!!

c
c
c
c

(USER) Change these dimensions and in other subroutines at same
time. For example, search and replace all occurrences throughout
the program of "(8002" by "(64000", to go from a 20 x 20 x 20
system to a 40 x 40 x 40 system.
real u(8002,3),gb(8002,3),b(8002,3)
real h(8002,3),Ah(8002,3),T(8002,3)
real C,dk(100,8,3,8,3)
real cmod(100,6,6),ss(100,8,3),eigen(100,6)
real zcon(2,3,2,3),pk(6,8,3)
real phasemod(100,2),prob(100)
integer in(27),jn(27),kn(27)
integer*4 ib(8002,27)
integer*2 pix(8002)
common/list1/strxx,stryy,strzz,strxz,stryz,strxy
common/list2/h,Ah
common/list3/ib
common/list4/pix
common/list5/dk,b,C,zcon,Y

117

common/list6/u
common/list7/gb
common/list8/cmod,T,eigen
common/list10/phasemod,nphase,ss
common/list11/sxx,syy,szz,sxz,syz,sxy
c (USER) Unit 9 is the microstructure input file, unit 7
c is the results output file.
open (9,file='microstructure.dat')
open (7,file='outputfile.out')
c (USER) nx,ny,nz are the size of the lattice
nx=20
ny=20
nz=20
c ns=total number of sites
ns=nx*ny*nz
write(7,9010) nx,ny,nz,ns
9010 format(' nx= ',i4,' ny= ',i4,' nz= ',i4,' ns = ',i8)
c
c
c

Add two more entries in the displacement vector for the 6 macrostrains,
u(ns+1,1) = exx,u(ns+1,2) = eyy, u(ns+1,3) = ezz,
u(nss,1) = exz, u(nss,2) = eyz, u(nss,3) = exy
nss=ns+2

c (USER) nphase is the number of phases being considered in the problem.
c The values of pix(m) will run from 1 to nphase.
nphase=2
c
c
c

(USER) gtest is the stopping criterion, compared to gg=gb*gb.
If gtest=abc*ns, when gg < gtest, the rms value per pixel
of gb is less than sqrt(abc)
gtest=1.e-20*(nx*ny*nz)
write(7,*) 'relaxation criterion gtest = ',gtest

c
c
c
c
c
c
c
c
c
c

(USER)
The parameter phasemod(i,j) is the bulk (i,1) and shear (i,2) moduli of
the i'th phase. These can be
input in terms of Young's modulus E (i,1) and Poisson's ratio nu (i,2).
The program, in the do 1144 loop, changes them to bulk and shear
moduli, using relations for isotropic elastic moduli.
For anisotropic moduli tensors, one can directly input the whole tensor
cmod in subroutine femat, and skip this part.
If you wish to input in terms of bulk (i,1) and shear (i,2) moduli,
then simply comment out do 1144 loop.

118

phasemod(1,1)=1.0
phasemod(1,2)=0.2
phasemod(2,1)=1.0
phasemod(2,2)=0.2

1144

do 1144 i=1,nphase
save=phasemod(i,1)
phasemod(i,1)=phasemod(i,1)/3./(1.-2.*phasemod(i,2))
phasemod(i,2)=save/2./(1.+phasemod(i,2))
continue

c
c

(USER) input eigenstrains for each phase
(1=xx, 2=yy, 3=zz, 4=xz, 5=yz, 6=xy).
eigen(1,1)=0.
eigen(1,2)=0.
eigen(1,3)=0.
eigen(1,4)=0.
eigen(1,5)=0.
eigen(1,6)=0.
eigen(2,1)=0.1
eigen(2,2)=0.
eigen(2,3)=0.
eigen(2,4)=0.
eigen(2,5)=0.
eigen(2,6)=0.

c

Construct the 27 neighbor table, ib(m,n)

c
c

First construct the 27 neighbor table in terms of delta i, delta j,
and delta k information (see Table 3 in manual)
in(1)=0
in(2)=1
in(3)=1
in(4)=1
in(5)=0
in(6)=-1
in(7)=-1
in(8)=-1
jn(1)=1
jn(2)=1
jn(3)=0
jn(4)=-1
jn(5)=-1
jn(6)=-1

119

jn(7)=0
jn(8)=1
do 555 n=1,8
kn(n)=0
kn(n+8)=-1
kn(n+16)=1
in(n+8)=in(n)
in(n+16)=in(n)
jn(n+8)=jn(n)
jn(n+16)=jn(n)
555
continue
in(25)=0
in(26)=0
in(27)=0
jn(25)=0
jn(26)=0
jn(27)=0
kn(25)=-1
kn(26)=1
kn(27)=0
c Now construct neighbor table according to 1-d labels
c Matrix ib(m,n) gives the 1-d label of the n'th neighbor (n=1,27) of
c the node labelled m.
nxy=nx*ny
do 1020 k=1,nz
do 1020 j=1,ny
do 1020 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
do 1004 n=1,27
i1=i+in(n)
j1=j+jn(n)
k1=k+kn(n)
if(i1.lt.1) i1=i1+nx
if(i1.gt.nx) i1=i1-nx
if(j1.lt.1) j1=j1+ny
if(j1.gt.ny) j1=j1-ny
if(k1.lt.1) k1=k1+nz
if(k1.gt.nz) k1=k1-nz
m1=nxy*(k1-1)+nx*(j1-1)+i1
ib(m,n)=m1
1004 continue
1020 continue
c

Compute the average stress and strain, as well as the macrostrains (overall

120

c
c

system size and shape) in each microstructure.
(USER) npoints is the number of microstructures to use.
npoints=1

do 8000 micro=1,npoints
c Read in a microstructure in subroutine ppixel, and set up pix(m)
c with the appropriate phase assignments.
call ppixel(nx,ny,nz,ns,nphase)
c Count and output the volume fractions of the different phases
call assig(ns,nphase,prob)
do 8050 i=1,nphase
write(7,9065) i,prob(i)
9065 format(' Volume fraction of phase ',i3,' is ',f10.8)
8050 continue
c output elastic moduli (bulk and shear) for each phase
write(7,*) ' Phase Moduli'
do 111 i=1,nphase
write(7,9020) i,phasemod(i,1),phasemod(i,2)
9020 format(' Phase ',i3,' bulk = ',f12.6,' shear = ',f12.6)
111 continue
c output thermal strains for each phase
write(7,*) ' Thermal Strains'
do 119 i=1,nphase
write(7,9029) i,eigen(i,1),eigen(i,2),eigen(i,3)
write(7,9029) i,eigen(i,4),eigen(i,5),eigen(i,6)
9029
format('Phase ',i3,' ',3f6.2)
119 continue
c (USER) Set inital macrostrains of computational cell
u(ns+1,1)=0.0
u(ns+1,2)=0.0
u(ns+1,3)=0.0
u(nss,1)=0.0
u(nss,2)=0.0
u(nss,3)=0.0
c Apply homogeneous macroscopic strain as the initial condition
c to displacement variables
do 1050 k=1,nz
do 1050 j=1,ny
do 1050 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
x=float(i-1)
y=float(j-1)
z=float(k-1)
u(m,1)=x*u(ns+1,1)+y*u(nss,3)+z*u(nss,1)
u(m,2)=x*u(nss,3)+y*u(ns+1,2)+z*u(nss,2)

121

u(m,3)=x*u(nss,1)+y*u(nss,2)+z*u(ns+1,3)
1050 continue
c
c
c
c
c
c
c
c

Set up the finite element stiffness matrices,the constant, C,
the vector, b, required for the energy. b and C depend on the macrostrains.
When they are updated, the values of b and C are updated too via
calling subroutine femat.
Only compute the thermal strain terms the first time femat is called,
(iskip=0) as they are unaffected by later changes (iskip=1) in
displacements and macrostrains.
Compute initial value of gradient gb and gg=gb*gb.
iskip=0
call femat(nx,ny,nz,ns,iskip)
call energy(nx,ny,nz,ns,utot)
gg=0.0
do 100 m3=1,3
do 100 m=1,nss
gg=gg+gb(m,m3)*gb(m,m3)
100
continue
write(7,9042) utot,gg
9042 format(' energy = ',e15.8,' gg= ',e15.8)
call flush(7)

c
c
c
c
c

Relaxation loop
(USER) kmax is the maximum number of times that dembx will be called,
with ldemb conjugate gradient steps performed during each call.
The total number of conjugate gradient steps allowed for a given elastic
computation is kmax*ldemb.
kmax=40
ldemb=50
ltot=0
do 5000 kkk=1,kmax

c

c
c
c
c
c
c
c

Call dembx to implement conjugate gradient routine
write(7,*) 'Going into dembx, call no. ',kkk
call dembx(nx,ny,nz,ns,Lstep,gg,gtest,ldemb,kkk)
ltot=ltot+Lstep
Call energy to compute energy after dembx call. If gg < gtest, this
will be the final energy. If gg is still larger than gtest, then this
will give an intermediate energy with which to check how the
relaxation process is coming along. The call to energy does not
change the gradient or the value of gg.
Need to first call femat to update the vector b, as the value of the
components of b depend on the macrostrains.

122

iskip=1
call femat(nx,ny,nz,ns,iskip)
call energy(nx,ny,nz,ns,utot)
write(7,9043) utot,gg,ltot
9043
format(' energy = ',e15.8,' gg=
call flush(7)
c
c
c

',e15.8,' ltot = ',i6)

If relaxation process is finished, jump out of loop
if(gg.lt.gtest) goto 444
Output stresses, strains, and macrostrains as an additional aid in judging
how well the relaxation process is proceeding.
call stress(nx,ny,nz,ns)
write(7,*) ' stresses: xx,yy,zz,xz,yz,xy'
write(7,*) strxx,stryy,strzz,strxz,stryz,strxy
write(7,*) ' strains: xx,yy,zz,xz,yz,xy'
write(7,*) sxx,syy,szz,sxz,syz,sxy
write(7,*)
write(7,*)
write(7,*)
write(7,*)

5000
444

' macrostrains in same order'
u(ns+1,1),u(ns+1,2),u(ns+1,3)
u(nss,1),u(nss,2),u(nss,3)
'avg = ',(u(ns+1,1)+u(ns+1,2)+u(ns+1,3))/3.

continue
call stress(nx,ny,nz,ns)
write(7,*) ' stresses: xx,yy,zz,xz,yz,xy'
write(7,*) strxx,stryy,strzz,strxz,stryz,strxy
write(7,*) ' strains: xx,yy,zz,xz,yz,xy'
write(7,*) sxx,syy,szz,sxz,syz,sxy
write(7,*)
write(7,*)
write(7,*)
write(7,*)

8000

' macrostrains in same order'
u(ns+1,1),u(ns+1,2),u(ns+1,3)
u(nss,1),u(nss,2),u(nss,3)
'avg = ',(u(ns+1,1)+u(ns+1,2)+u(ns+1,3))/3.

continue
end

c
c
c
c

Subroutine sets up the stiffness matrices, the linear term in the
regular displacements, b, and the constant term, C, which come from
the periodic boundary conditions, the term linear in the displacments,
T, that comes from the thermal strains, and the constant term Y.
subroutine femat(nx,ny,nz,ns,iskip)

123

real u(8002,3),b(8002,3),T(8002,3)
real dk(100,8,3,8,3),phasemod(100,2),dndx(8),dndy(8),dndz(8)
real g(3,3,3),econ,ck(6,6),cmu(6,6),cmod(100,6,6)
real es(6,8,3),zcon(2,3,2,3),ss(100,8,3)
real eigen(100,6),delta(8,3)
integer is(8),iskip
integer*4 ib(8002,27)
integer*2 pix(8002)
common/list3/ib
common/list4/pix
common/list5/dk,b,C,zcon,Y
common/list6/u
common/list8/cmod,T,eigen
common/list10/phasemod,nphase,ss
nxy=nx*ny
c
c
c
c

Generate dk, zcon, T, and Y on first pass. After that they are
constant, snce they are independent of the macrostrains. Only b gets
upgraded as the macrostrains change.
Line number 1221 is the routine for b.
if(iskip.eq.1) goto 1221

c

initialize stiffness matrices
do 40 m=1,nphase
do 40 l=1,3
do 40 k=1,3
do 40 j=1,8
do 40 i=1,8
dk(m,i,k,j,l)=0.0
40
continue
c initialize zcon matrix (gives C term for arbitrary macrostrains)
do 42 i=1,2
do 42 j=1,2
do 42 mi=1,3
do 42 mj=1,3
zcon(i,mi,j,mj)=0.0
42
continue
c (USER) An anisotropic elastic moduli tensor could be input at this point,
c bypassing this part, which assumes isotropic elasticity, so that there
c are only two independent numbers making up the elastic moduli tensor,
c the bulk modulus K and the shear modulus G.
c

Set up elastic moduli matrices for each kind of element

124

c
c

ck and cmu are the bulk modulus and shear modulus matrices, which
need to multiplied by the actual bulk and shear moduli in each phase.
ck(1,1)=1.0
ck(1,2)=1.0
ck(1,3)=1.0
ck(1,4)=0.0
ck(1,5)=0.0
ck(1,6)=0.0
ck(2,1)=1.0
ck(2,2)=1.0
ck(2,3)=1.0
ck(2,4)=0.0
ck(2,5)=0.0
ck(2,6)=0.0
ck(3,1)=1.0
ck(3,2)=1.0
ck(3,3)=1.0
ck(3,4)=0.0
ck(3,5)=0.0
ck(3,6)=0.0
ck(4,1)=0.0
ck(4,2)=0.0
ck(4,3)=0.0
ck(4,4)=0.0
ck(4,5)=0.0
ck(4,6)=0.0
ck(5,1)=0.0
ck(5,2)=0.0
ck(5,3)=0.0
ck(5,4)=0.0
ck(5,5)=0.0
ck(5,6)=0.0
ck(6,1)=0.0
ck(6,2)=0.0
ck(6,3)=0.0
ck(6,4)=0.0
ck(6,5)=0.0
ck(6,6)=0.0
cmu(1,1)=4.0/3.0
cmu(1,2)=-2.0/3.0
cmu(1,3)=-2.0/3.0
cmu(1,4)=0.0
cmu(1,5)=0.0

125

cmu(1,6)=0.0
cmu(2,1)=-2.0/3.0
cmu(2,2)=4.0/3.0
cmu(2,3)=-2.0/3.0
cmu(2,4)=0.0
cmu(2,5)=0.0
cmu(2,6)=0.0
cmu(3,1)=-2.0/3.0
cmu(3,2)=-2.0/3.0
cmu(3,3)=4.0/3.0
cmu(3,4)=0.0
cmu(3,5)=0.0
cmu(3,6)=0.0
cmu(4,1)=0.0
cmu(4,2)=0.0
cmu(4,3)=0.0
cmu(4,4)=1.0
cmu(4,5)=0.0
cmu(4,6)=0.0
cmu(5,1)=0.0
cmu(5,2)=0.0
cmu(5,3)=0.0
cmu(5,4)=0.0
cmu(5,5)=1.0
cmu(5,6)=0.0
cmu(6,1)=0.0
cmu(6,2)=0.0
cmu(6,3)=0.0
cmu(6,4)=0.0
cmu(6,5)=0.0
cmu(6,6)=1.0
do 31 k=1,nphase
do 21 j=1,6
do 11 i=1,6
cmod(k,i,j)=phasemod(k,1)*ck(i,j)+phasemod(k,2)*cmu(i,j)
11
continue
21
continue
31
continue
c Set up Simpson's integration rule weight vector
do 30 k=1,3
do 30 j=1,3
do 30 i=1,3
nm=0
if(i.eq.2) nm=nm+1

126

30
c
c
c
c

c
c

c

if(j.eq.2) nm=nm+1
if(k.eq.2) nm=nm+1
g(i,j,k)=4.0**nm
continue
Loop over the nphase kinds of pixels and
Simpson's rule quadrature points in order to compute the stiffness
matrices. Stiffness matrices of trilinear finite elements are quadratic
in x, y, and z, so that Simpson's rule quadrature gives exact results.
do 4000 ijk=1,nphase
do 3000 k=1,3
do 3000 j=1,3
do 3000 i=1,3
x=float(i-1)/2.0
y=float(j-1)/2.0
z=float(k-1)/2.0
dndx means the negative derivative with respect to x, of the shape
matrix N (see manual, Sec. 2.2), dndy and dndz are similar.
dndx(1)=-(1.0-y)*(1.0-z)
dndx(2)=(1.0-y)*(1.0-z)
dndx(3)=y*(1.0-z)
dndx(4)=-y*(1.0-z)
dndx(5)=-(1.0-y)*z
dndx(6)=(1.0-y)*z
dndx(7)=y*z
dndx(8)=-y*z
dndy(1)=-(1.0-x)*(1.0-z)
dndy(2)=-x*(1.0-z)
dndy(3)=x*(1.0-z)
dndy(4)=(1.0-x)*(1.0-z)
dndy(5)=-(1.0-x)*z
dndy(6)=-x*z
dndy(7)=x*z
dndy(8)=(1.0-x)*z
dndz(1)=-(1.0-x)*(1.0-y)
dndz(2)=-x*(1.0-y)
dndz(3)=-x*y
dndz(4)=-(1.0-x)*y
dndz(5)=(1.0-x)*(1.0-y)
dndz(6)=x*(1.0-y)
dndz(7)=x*y
dndz(8)=(1.0-x)*y
now build strain matrix
do 2799 n1=1,6
do 2799 n2=1,8

127

do 2799 n3=1,3
es(n1,n2,n3)=0.0
2799 continue
do 2797 n=1,8
es(1,n,1)=dndx(n)
es(2,n,2)=dndy(n)
es(3,n,3)=dndz(n)
es(4,n,1)=dndz(n)
es(4,n,3)=dndx(n)
es(5,n,2)=dndz(n)
es(5,n,3)=dndy(n)
es(6,n,1)=dndy(n)
es(6,n,2)=dndx(n)
2797 continue
c now do matrix multiply to determine value at (x,y,z), multiply by
c proper weight, and sum into dk, the stiffness matrix
do 900 mm=1,3
do 900 nn=1,3
do 900 ii=1,8
do 900 jj=1,8
c define sum over strain matrices and elastic moduli matrix for
c stiffness matrix
sum=0.0
do 890 kk=1,6
do 890 ll=1,6
sum=sum+es(kk,ii,mm)*cmod(ijk,kk,ll)*es(ll,jj,nn)
890
continue
dk(ijk,ii,mm,jj,nn)=dk(ijk,ii,mm,jj,nn)+g(i,j,k)*sum/216.
900
continue
3000 continue
4000 continue
c
c

Now compute the ss matrices, which give the thermal strain terms
for the i'th phase, single pixel.
dndx(1)=-0.25
dndx(2)=0.25
dndx(3)=0.25
dndx(4)=-0.25
dndx(5)=-0.25
dndx(6)=0.25
dndx(7)=0.25
dndx(8)=-0.25
dndy(1)=-0.25
dndy(2)=-0.25

128

dndy(3)=0.25
dndy(4)=0.25
dndy(5)=-0.25
dndy(6)=-0.25
dndy(7)=0.25
dndy(8)=0.25
dndz(1)=-0.25
dndz(2)=-0.25
dndz(3)=-0.25
dndz(4)=-0.25
dndz(5)=0.25
dndz(6)=0.25
dndz(7)=0.25
dndz(8)=0.25
c now build average strain matrix
do 3799 n1=1,6
do 3799 n2=1,8
do 3799 n3=1,3
es(n1,n2,n3)=0.0
3799 continue
do 3797 n=1,8
es(1,n,1)=dndx(n)
es(2,n,2)=dndy(n)
es(3,n,3)=dndz(n)
es(4,n,1)=dndz(n)
es(4,n,3)=dndx(n)
es(5,n,2)=dndz(n)
es(5,n,3)=dndy(n)
es(6,n,1)=dndy(n)
es(6,n,2)=dndx(n)
3797 continue
do 3598 mmm=1,nphase
do 3798 nn=1,3
do 3798 mm=1,8
sum=0.0
do 3698 nm=1,6
do 3698 n=1,6
sum=sum+cmod(mmm,n,nm)*es(n,mm,nn)*eigen(mmm,nm)
3698 continue
ss(mmm,mm,nn)=sum
3798 continue
3598 continue
c
c

now call subroutine const to generate zcon
zcon is a (2,3) x (2,3) matrix

129

call const(dk,ns,zcon,nx,ny,nz)
c
c
c

Now set up linear term, T, for thermal energy. It does not depend
on the macrostrains or displacements, so there is no need to update it
as the macrostrains change. T is built up out of the ss matrices.

6066

nss=ns+2
do 6066 m3=1,3
do 6066 m=1,nss
T(m,m3)=0.0
continue

c
c
c
c

For all cases, the correspondence between 1-8 finite element node
labels and the 1-27 neighbor labels is (see Table 4 in manual):
1:ib(m,27), 2:ib(m,3),3:ib(m,2),4:ib(m,1)
5:ib(m,26),6:ib(m,19),7:ib(m,18),8:ib(m,17)
is(1)=27
is(2)=3
is(3)=2
is(4)=1
is(5)=26
is(6)=19
is(7)=18
is(8)=17
c Do all points, but no macrostrain terms
c note: factor of 2 on linear thermal term is cancelled
c by factor of 1/2 out in front of total energy term
do 6601 k=1,nz
do 6601 j=1,ny
do 6601 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
do 6600 mm=1,8
do 6600 nn=1,3
T(ib(m,is(mm)),nn)=T(ib(m,is(mm)),nn)-ss(pix(m),mm,nn)
6600 continue
6601 continue
c

now need to pick up and sum in all terms multiplying macrostrains
do 7788 ipp=1,2
do 7788 jpp=1,3
exx=0.0
eyy=0.0
ezz=0.0
exz=0.0
eyz=0.0

130

exy=0.0
if(ipp.eq.1.and.jpp.eq.1)
if(ipp.eq.1.and.jpp.eq.2)
if(ipp.eq.1.and.jpp.eq.3)
if(ipp.eq.2.and.jpp.eq.1)
if(ipp.eq.2.and.jpp.eq.2)
if(ipp.eq.2.and.jpp.eq.3)

exx=1.0
eyy=1.0
ezz=1.0
exz=1.0
eyz=1.0
exy=1.0

c

x=nx face
do 6001 i3=1,3
do 6001 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3.or.i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
6001 continue

6900
6000

do 6000 j=1,ny-1
do 6000 k=1,nz-1
m=nxy*(k-1)+j*nx
do 6900 nn=1,3
do 6900 mm=1,8
T(ns+ipp,jpp)=T(ns+ipp,jpp)-ss(pix(m),mm,nn)*delta(mm,nn)
continue
continue

c

y=ny face
do 6011 i3=1,3
do 6011 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.3.or.i8.eq.4.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
6011 continue
do 6010 i=1,nx-1
do 6010 k=1,nz-1
m=nxy*(k-1)+nx*(ny-1)+i
do 6901 nn=1,3
do 6901 mm=1,8
T(ns+ipp,jpp)=T(ns+ipp,jpp)-ss(pix(m),mm,nn)*delta(mm,nn)
6901 continue

131

6010 continue
c z=nz face
do 6021 i3=1,3
do 6021 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
6021 continue
do 6020 i=1,nx-1
do 6020 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+i
do 6902 nn=1,3
do 6902 mm=1,8
T(ns+ipp,jpp)=T(ns+ipp,jpp)-ss(pix(m),mm,nn)*delta(mm,nn)
6902 continue
6020 continue
c x=nx y=ny edge
do 6031 i3=1,3
do 6031 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.6) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.3.or.i8.eq.7) then
delta(i8,1)=exy*ny+exx*nx
delta(i8,2)=eyy*ny+exy*nx
delta(i8,3)=eyz*ny+exz*nx
end if
6031 continue
do 6030 k=1,nz-1
m=nxy*k
do 6903 nn=1,3
do 6903 mm=1,8
T(ns+ipp,jpp)=T(ns+ipp,jpp)-ss(pix(m),mm,nn)*delta(mm,nn)
6903 continue

132

6030 continue
c x=nx z=nz edge
do 6041 i3=1,3
do 6041 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.5.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exz*nz+exx*nx
delta(i8,2)=eyz*nz+exy*nx
delta(i8,3)=ezz*nz+exz*nx
end if
6041 continue
do 6040 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+nx
do 6904 nn=1,3
do 6904 mm=1,8
T(ns+ipp,jpp)=T(ns+ipp,jpp)-ss(pix(m),mm,nn)*delta(mm,nn)
6904 continue
6040 continue
c y=ny z=nz edge
do 6051 i3=1,3
do 6051 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.3.or.i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz

133

delta(i8,3)=eyz*ny+ezz*nz
end if
6051 continue
do 6050 i=1,nx-1
m=nxy*(nz-1)+nx*(ny-1)+i
do 6905 nn=1,3
do 6905 mm=1,8
T(ns+ipp,jpp)=T(ns+ipp,jpp)-ss(pix(m),mm,nn)*delta(mm,nn)
6905 continue
6050 continue
c x=nx y=ny z=nz corner
do 6061 i3=1,3
do 6061 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.5) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz
end if
if(i8.eq.6) then
delta(i8,1)=exx*nx+exz*nz
delta(i8,2)=exy*nx+eyz*nz
delta(i8,3)=exz*nx+ezz*nz
end if
if(i8.eq.3) then
delta(i8,1)=exx*nx+exy*ny
delta(i8,2)=exy*nx+eyy*ny
delta(i8,3)=exz*nx+eyz*ny
end if
if(i8.eq.7) then

134

delta(i8,1)=exx*nx+exy*ny+exz*nz
delta(i8,2)=exy*nx+eyy*ny+eyz*nz
delta(i8,3)=exz*nx+eyz*ny+ezz*nz
end if
6061 continue
m=nx*ny*nz
do 6906 mm=1,8
do 6906 nn=1,3
T(ns+ipp,jpp)=T(ns+ipp,jpp)-ss(pix(m),mm,nn)*delta(mm,nn)
6906 continue
7788 continue
c now compute Y, the 0.5(eigen)Cij(eigen) energy, doesn't ever change
c with macrostrain or displacements
Y=0.0
do 8811 m=1,ns
do 8811 n=1,6
do 8811 nn=1,6
Y=Y+0.5*eigen(pix(m),n)*cmod(pix(m),n,nn)*eigen(pix(m),nn)
8811 continue
c
c

Following needs to be run after every change in macrostrain
when energy is recomputed.

1221 continue
c Use auxiliary variables (exx, etc.) instead of u() variable, for
c convenience, and to make the following code easier to read.
exx=u(ns+1,1)
eyy=u(ns+1,2)
ezz=u(ns+1,3)
exz=u(nss,1)
eyz=u(nss,2)
exy=u(nss,3)
c Now set up vector for linear term that comes from periodic boundary
c conditions. Notation and conventions same as for T term.
c This is done using the stiffness matrices, and the periodic terms
c in the macrostrains. It is easier to set up b this way than to
c analytically write out all the terms involved.
do 5000 m3=1,3
do 5000 m=1,ns
b(m,m3)=0.0
5000 continue
c For all cases, the correspondence between 1-8 finite element node
c labels and the 1-27 neighbor labels is (see Table 4 in manual):
c 1:ib(m,27), 2:ib(m,3),3:ib(m,2),4:ib(m,1)

135

c

5:ib(m,26),6:ib(m,19),7:ib(m,18),8:ib(m,17)
is(1)=27
is(2)=3
is(3)=2
is(4)=1
is(5)=26
is(6)=19
is(7)=18
is(8)=17

C=0.0
x=nx face
do 2001 i3=1,3
do 2001 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3.or.i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
2001 continue
c

do 2000 j=1,ny-1
do 2000 k=1,nz-1
m=nxy*(k-1)+j*nx
do 1900 nn=1,3
do 1900 mm=1,8
sum=0.0
do 1899 m3=1,3
do 1899 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
1899 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1900 continue
2000 continue
c y=ny face
do 2011 i3=1,3
do 2011 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.3.or.i8.eq.4.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
2011 continue

136

do 2010 i=1,nx-1
do 2010 k=1,nz-1
m=nxy*(k-1)+nx*(ny-1)+i
do 1901 nn=1,3
do 1901 mm=1,8
sum=0.0
do 2099 m3=1,3
do 2099 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2099 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1901 continue
2010 continue
c z=nz face
do 2021 i3=1,3
do 2021 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
2021 continue
do 2020 i=1,nx-1
do 2020 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+i
do 1902 nn=1,3
do 1902 mm=1,8
sum=0.0
do 2019 m3=1,3
do 2019 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2019 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1902 continue
2020 continue
c x=nx y=ny edge
do 2031 i3=1,3
do 2031 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.6) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if

137

if(i8.eq.4.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.3.or.i8.eq.7) then
delta(i8,1)=exy*ny+exx*nx
delta(i8,2)=eyy*ny+exy*nx
delta(i8,3)=eyz*ny+exz*nx
end if
2031 continue
do 2030 k=1,nz-1
m=nxy*k
do 1903 nn=1,3
do 1903 mm=1,8
sum=0.0
do 2029 m3=1,3
do 2029 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2029 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1903 continue
2030 continue
c x=nx z=nz edge
do 2041 i3=1,3
do 2041 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.5.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exz*nz+exx*nx
delta(i8,2)=eyz*nz+exy*nx
delta(i8,3)=ezz*nz+exz*nx
end if
2041 continue
do 2040 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+nx

138

do 1904 nn=1,3
do 1904 mm=1,8
sum=0.0
do 2039 m3=1,3
do 2039 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2039 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1904 continue
2040 continue
c y=ny z=nz edge
do 2051 i3=1,3
do 2051 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.3.or.i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz
end if
2051 continue
do 2050 i=1,nx-1
m=nxy*(nz-1)+nx*(ny-1)+i
do 1905 nn=1,3
do 1905 mm=1,8
sum=0.0
do 2049 m3=1,3
do 2049 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2049 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1905 continue
2050 continue
c x=nx y=ny z=nz corner
do 2061 i3=1,3
do 2061 i8=1,8

139

2061

2059

delta(i8,i3)=0.0
if(i8.eq.2) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.5) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz
end if
if(i8.eq.6) then
delta(i8,1)=exx*nx+exz*nz
delta(i8,2)=exy*nx+eyz*nz
delta(i8,3)=exz*nx+ezz*nz
end if
if(i8.eq.3) then
delta(i8,1)=exx*nx+exy*ny
delta(i8,2)=exy*nx+eyy*ny
delta(i8,3)=exz*nx+eyz*ny
end if
if(i8.eq.7) then
delta(i8,1)=exx*nx+exy*ny+exz*nz
delta(i8,2)=exy*nx+eyy*ny+eyz*nz
delta(i8,3)=exz*nx+eyz*ny+ezz*nz
end if
continue
m=nx*ny*nz
do 1906 nn=1,3
do 1906 mm=1,8
sum=0.0
do 2059 m3=1,3
do 2059 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
continue

140

1906

b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
continue
return
end

c
c
c
c
c
c

Subroutine that computes derivatives of the b-vector with respect
to the macrostrains. Since b is linear in the macrostrains, the
derivative with respect to any one of them can be computed simply
by letting that macrostrain, within the subroutine, be equal to one,
and all the other macrostrains to be zero.
Very similar to 1221 loop in femat for b.
subroutine bgrad(nx,ny,nz,ns,exx,eyy,ezz,exz,eyz,exy)
real b(8002,3)
real dk(100,8,3,8,3),delta(8,3),zcon(2,3,2,3)
integer is(8)
integer*4 ib(8002,27)
integer*2 pix(8002)
common/list3/ib
common/list4/pix
common/list5/dk,b,C,zcon,Y

c
c

nxy=nx*ny
exx, eyy, ezz, exz, eyz, exy are the artificial macrostrains used
to get the gradient terms (appropriate combinations of 1's and 0's).

c

Set up vector for linear term

5000

c

do 5000 m3=1,3
do 5000 m=1,ns
b(m,m3)=0.0
continue
is(1)=27
is(2)=3
is(3)=2
is(4)=1
is(5)=26
is(6)=19
is(7)=18
is(8)=17

x=nx face
do 2001 i3=1,3

141

2001

do 2001 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3.or.i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
continue

do 2000 j=1,ny-1
do 2000 k=1,nz-1
m=nxy*(k-1)+j*nx
do 1900 nn=1,3
do 1900 mm=1,8
sum=0.0
do 1899 m3=1,3
do 1899 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
1899 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1900 continue
2000 continue
c y=ny face
do 2011 i3=1,3
do 2011 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.3.or.i8.eq.4.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
2011 continue
do 2010 i=1,nx-1
do 2010 k=1,nz-1
m=nxy*(k-1)+nx*(ny-1)+i
do 1901 nn=1,3
do 1901 mm=1,8
sum=0.0
do 2099 m3=1,3
do 2099 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2099 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1901 continue
2010 continue

142

c

z=nz face
do 2021 i3=1,3
do 2021 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
2021 continue
do 2020 i=1,nx-1
do 2020 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+i
do 1902 nn=1,3
do 1902 mm=1,8
sum=0.0
do 2019 m3=1,3
do 2019 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2019 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1902 continue
2020 continue
c x=nx y=ny edge
do 2031 i3=1,3
do 2031 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.6) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.3.or.i8.eq.7) then
delta(i8,1)=exy*ny+exx*nx
delta(i8,2)=eyy*ny+exy*nx
delta(i8,3)=eyz*ny+exz*nx
end if
2031 continue
do 2030 k=1,nz-1
m=nxy*k

143

do 1903 nn=1,3
do 1903 mm=1,8
sum=0.0
do 2029 m3=1,3
do 2029 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2029 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1903 continue
2030 continue
c x=nx z=nz edge
do 2041 i3=1,3
do 2041 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.5.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exz*nz+exx*nx
delta(i8,2)=eyz*nz+exy*nx
delta(i8,3)=ezz*nz+exz*nx
end if
2041 continue
do 2040 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+nx
do 1904 nn=1,3
do 1904 mm=1,8
sum=0.0
do 2039 m3=1,3
do 2039 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2039 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1904 continue
2040 continue
c y=ny z=nz edge
do 2051 i3=1,3
do 2051 i8=1,8

144

delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.3.or.i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz
end if
2051 continue
do 2050 i=1,nx-1
m=nxy*(nz-1)+nx*(ny-1)+i
do 1905 nn=1,3
do 1905 mm=1,8
sum=0.0
do 2049 m3=1,3
do 2049 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
2049 continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
1905 continue
2050 continue
c x=nx y=ny z=nz corner
do 2061 i3=1,3
do 2061 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.5) then
delta(i8,1)=exz*nz

145

2061

2059
1906

c
c
c

delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz
end if
if(i8.eq.6) then
delta(i8,1)=exx*nx+exz*nz
delta(i8,2)=exy*nx+eyz*nz
delta(i8,3)=exz*nx+ezz*nz
end if
if(i8.eq.3) then
delta(i8,1)=exx*nx+exy*ny
delta(i8,2)=exy*nx+eyy*ny
delta(i8,3)=exz*nx+eyz*ny
end if
if(i8.eq.7) then
delta(i8,1)=exx*nx+exy*ny+exz*nz
delta(i8,2)=exy*nx+eyy*ny+eyz*nz
delta(i8,3)=exz*nx+eyz*ny+ezz*nz
end if
continue
m=nx*ny*nz
do 1906 nn=1,3
do 1906 mm=1,8
sum=0.0
do 2059 m3=1,3
do 2059 m8=1,8
sum=sum+delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)
continue
b(ib(m,is(mm)),nn)=b(ib(m,is(mm)),nn)+sum
continue
return
end

Subroutine computes the quadratic term in the macrostrains, that comes
from the periodic boundary conditions, and sets it up as a
(2,3) x (2,3) matrix that couples to the six macrostrains
subroutine const(dk,ns,zcon,nx,ny,nz)
real dk(100,8,3,8,3),zcon(2,3,2,3),delta(8,3)
real pp(6,6),s(6,6)
integer*2 pix(8002)

146

common/list4/pix
c
c
c
c
c
c

routine to set up 6 x 6 matrix for energy term involving macro-strains
only, pulled out of femat
Idea is to evaluate the quadratic term in the macrostrains repeatedly
for choices of strain like
exx=1, exy=1, all others = 0, build up 21 choices, then recombine
to get matrix elements by themselves
nxy=nx*ny
nss=ns+2

1111

do 1111 i=1,6
do 1111 j=1,6
s(i,j)=0.0
pp(i,j)=0.0
continue
do 5000 ii=1,6
do 5000 jj=ii,6
econ=0.0
exx=0.0
eyy=0.0
ezz=0.0
exz=0.0
eyz=0.0
exy=0.0
if(ii.eq.1.and.jj.eq.1)
if(ii.eq.2.and.jj.eq.2)
if(ii.eq.3.and.jj.eq.3)
if(ii.eq.4.and.jj.eq.4)
if(ii.eq.5.and.jj.eq.5)
if(ii.eq.6.and.jj.eq.6)
if(ii.eq.1.and.jj.eq.2)
exx=1.0
eyy=1.0
end if
if(ii.eq.1.and.jj.eq.3)
exx=1.0
ezz=1.0
end if
if(ii.eq.1.and.jj.eq.4)
exx=1.0
exz=1.0
end if

exx=1.0
eyy=1.0
ezz=1.0
exz=1.0
eyz=1.0
exy=1.0
then

then

then

147

if(ii.eq.1.and.jj.eq.5)
exx=1.0
eyz=1.0
end if
if(ii.eq.1.and.jj.eq.6)
exx=1.0
exy=1.0
end if
if(ii.eq.2.and.jj.eq.3)
eyy=1.0
ezz=1.0
end if
if(ii.eq.2.and.jj.eq.4)
eyy=1.0
exz=1.0
end if
if(ii.eq.2.and.jj.eq.5)
eyy=1.0
eyz=1.0
end if
if(ii.eq.2.and.jj.eq.6)
eyy=1.0
exy=1.0
end if
if(ii.eq.3.and.jj.eq.4)
ezz=1.0
exz=1.0
end if
if(ii.eq.3.and.jj.eq.5)
ezz=1.0
eyz=1.0
end if
if(ii.eq.3.and.jj.eq.6)
ezz=1.0
exy=1.0
end if
if(ii.eq.4.and.jj.eq.5)
exz=1.0
eyz=1.0
end if
if(ii.eq.4.and.jj.eq.6)
exz=1.0
exy=1.0
end if
if(ii.eq.5.and.jj.eq.6)

then

then

then

then

then

then

then

then

then

then

then

then

148

eyz=1.0
exy=1.0
end if
c x=nx face
do 2001 i3=1,3
do 2001 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3.or.i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
2001 continue
do 2000 j=1,ny-1
do 2000 k=1,nz-1
m=nxy*(k-1)+j*nx
do 1900 nn=1,3
do 1900 mm=1,8
do 1899 m3=1,3
do 1899 m8=1,8
econ=econ+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
1899 continue
1900 continue
2000 continue
c y=ny face
do 2011 i3=1,3
do 2011 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.3.or.i8.eq.4.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
2011 continue
do 2010 i=1,nx-1
do 2010 k=1,nz-1
m=nxy*(k-1)+nx*(ny-1)+i
do 1901 nn=1,3
do 1901 mm=1,8
do 2099 m3=1,3
do 2099 m8=1,8
econ=econ+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2099 continue
1901 continue

149

2010 continue
c z=nz face
do 2021 i3=1,3
do 2021 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6.or.i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
2021 continue
do 2020 i=1,nx-1
do 2020 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+i
do 1902 nn=1,3
do 1902 mm=1,8
do 2019 m3=1,3
do 2019 m8=1,8
econ=econ+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2019 continue
1902 continue
2020 continue
c x=nx y=ny edge
do 2031 i3=1,3
do 2031 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.6) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4.or.i8.eq.8) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.3.or.i8.eq.7) then
delta(i8,1)=exy*ny+exx*nx
delta(i8,2)=eyy*ny+exy*nx
delta(i8,3)=eyz*ny+exz*nx
end if
2031 continue
do 2030 k=1,nz-1
m=nxy*k
do 1903 nn=1,3

150

do 1903 mm=1,8
do 2029 m3=1,3
do 2029 m8=1,8
econ=econ+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2029 continue
1903 continue
2030 continue
c x=nx z=nz edge
do 2041 i3=1,3
do 2041 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2.or.i8.eq.3) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.5.or.i8.eq.8) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.6.or.i8.eq.7) then
delta(i8,1)=exz*nz+exx*nx
delta(i8,2)=eyz*nz+exy*nx
delta(i8,3)=ezz*nz+exz*nx
end if
2041 continue
do 2040 j=1,ny-1
m=nxy*(nz-1)+nx*(j-1)+nx
do 1904 nn=1,3
do 1904 mm=1,8
do 2039 m3=1,3
do 2039 m8=1,8
econ=econ+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2039 continue
1904 continue
2040 continue
c y=ny z=nz edge
do 2051 i3=1,3
do 2051 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.5.or.i8.eq.6) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz

151

end if
if(i8.eq.3.or.i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.7.or.i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz
end if
2051 continue
do 2050 i=1,nx-1
m=nxy*(nz-1)+nx*(ny-1)+i
do 1905 nn=1,3
do 1905 mm=1,8
do 2049 m3=1,3
do 2049 m8=1,8
econ=econ+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
2049 continue
1905 continue
2050 continue
c x=nx y=ny z=nz corner
do 2061 i3=1,3
do 2061 i8=1,8
delta(i8,i3)=0.0
if(i8.eq.2) then
delta(i8,1)=exx*nx
delta(i8,2)=exy*nx
delta(i8,3)=exz*nx
end if
if(i8.eq.4) then
delta(i8,1)=exy*ny
delta(i8,2)=eyy*ny
delta(i8,3)=eyz*ny
end if
if(i8.eq.5) then
delta(i8,1)=exz*nz
delta(i8,2)=eyz*nz
delta(i8,3)=ezz*nz
end if
if(i8.eq.8) then
delta(i8,1)=exy*ny+exz*nz
delta(i8,2)=eyy*ny+eyz*nz
delta(i8,3)=eyz*ny+ezz*nz

152

2061

2059
1906

end if
if(i8.eq.6) then
delta(i8,1)=exx*nx+exz*nz
delta(i8,2)=exy*nx+eyz*nz
delta(i8,3)=exz*nx+ezz*nz
end if
if(i8.eq.3) then
delta(i8,1)=exx*nx+exy*ny
delta(i8,2)=exy*nx+eyy*ny
delta(i8,3)=exz*nx+eyz*ny
end if
if(i8.eq.7) then
delta(i8,1)=exx*nx+exy*ny+exz*nz
delta(i8,2)=exy*nx+eyy*ny+eyz*nz
delta(i8,3)=exz*nx+eyz*ny+ezz*nz
end if
continue
m=nx*ny*nz
do 1906 nn=1,3
do 1906 mm=1,8
do 2059 m3=1,3
do 2059 m8=1,8
econ=econ+0.5*delta(m8,m3)*dk(pix(m),m8,m3,mm,nn)*delta(mm,nn)
continue
continue
pp(ii,jj)=econ*2.

5000

continue
do 6000 i=1,6
do 6000 j=i,6
if(i.eq.j) s(i,j)=pp(i,j)
if(i.ne.j) then
s(i,j)=pp(i,j)-pp(i,i)-pp(j,j)
end if
6000 continue
do 7000 i=1,6
do 7000 j=1,6
pp(i,j)=0.5*(s(i,j)+s(j,i))
7000 continue
c now map pp(i,j) into zcon(2,3,2,3)
do 7200 i=1,2
do 7200 j=1,2
do 7200 mi=1,3
do 7200 mj=1,3
if(i.eq.1) ii=i+mi-1

153

1
7200

if(i.eq.2) ii=i+mi+1
if(j.eq.1) jj=j+mj-1
if(j.eq.2) jj=j+mj+1
zcon(i,mi,j,mj)=pp(ii,jj)
write(7,1) i,mi,j,mj,zcon(i,mi,j,mj)
format(4i5,f20.10)
continue
return
end

c
c

Subroutine computes the total energy, utot, and the gradient, gb,
for the regular displacements as well as for the macrostrains
subroutine energy(nx,ny,nz,ns,utot)

real u(8002,3),gb(8002,3)
real b(8002,3),T(8002,3)
real cmod(100,6,6),pk(6,8,3),eigen(100,6)
real dk(100,8,3,8,3),zcon(2,3,2,3),C
integer*4 ib(8002,27)
integer*2 pix(8002)
common/list3/ib
common/list4/pix
common/list5/dk,b,C,zcon,Y
common/list6/u
common/list7/gb
common/list8/cmod,T,eigen
nss=ns+2
c
c
c

Do global matrix multiply via small stiffness matrices, Ah = A * h
The long statement below correctly brings in all the terms from
the global matrix A using only the small stiffness matrices.

do 2090 m3=1,3
do 2090 m=1,nss
gb(m,m3)=0.0
2090 continue
do 3000 j=1,3
do 3000 n=1,3
do 3000 m=1,ns
gb(m,j)=gb(m,j)+u(ib(m,1),n)*( dk(pix(ib(m,27)),1,j,4,n)

154

&+dk(pix(ib(m,7)),2,j,3,n)
&+dk(pix(ib(m,25)),5,j,8,n)+dk(pix(ib(m,15)),6,j,7,n) )+
&u(ib(m,2),n)*( dk(pix(ib(m,27)),1,j,3,n)
&+dk(pix(ib(m,25)),5,j,7,n) )+
&u(ib(m,3),n)*( dk(pix(ib(m,27)),1,j,2,n)+dk(pix(ib(m,5)),4,j,3,n)+
&dk(pix(ib(m,13)),8,j,7,n)+dk(pix(ib(m,25)),5,j,6,n) )+
&u(ib(m,4),n)*( dk(pix(ib(m,5)),4,j,2,n)
&+dk(pix(ib(m,13)),8,j,6,n) )+
&u(ib(m,5),n)*( dk(pix(ib(m,6)),3,j,2,n)+dk(pix(ib(m,5)),4,j,1,n)+
&dk(pix(ib(m,14)),7,j,6,n)+dk(pix(ib(m,13)),8,j,5,n) )+
&u(ib(m,6),n)*( dk(pix(ib(m,6)),3,j,1,n)
&+dk(pix(ib(m,14)),7,j,5,n) )+
&u(ib(m,7),n)*( dk(pix(ib(m,6)),3,j,4,n)+dk(pix(ib(m,7)),2,j,1,n)+
&dk(pix(ib(m,14)),7,j,8,n)+dk(pix(ib(m,15)),6,j,5,n) )+
&u(ib(m,8),n)*( dk(pix(ib(m,7)),2,j,4,n)
&+dk(pix(ib(m,15)),6,j,8,n) )+
&u(ib(m,9),n)*( dk(pix(ib(m,25)),5,j,4,n)
&+dk(pix(ib(m,15)),6,j,3,n) )+
&u(ib(m,10),n)*( dk(pix(ib(m,25)),5,j,3,n) )+
&u(ib(m,11),n)*( dk(pix(ib(m,13)),8,j,3,n)
&+dk(pix(ib(m,25)),5,j,2,n) )+
&u(ib(m,12),n)*( dk(pix(ib(m,13)),8,j,2,n) )+
&u(ib(m,13),n)*( dk(pix(ib(m,13)),8,j,1,n)
&+dk(pix(ib(m,14)),7,j,2,n) )+
&u(ib(m,14),n)*( dk(pix(ib(m,14)),7,j,1,n) )+
&u(ib(m,15),n)*( dk(pix(ib(m,14)),7,j,4,n)
&+dk(pix(ib(m,15)),6,j,1,n) )+
&u(ib(m,16),n)*( dk(pix(ib(m,15)),6,j,4,n) )+
&u(ib(m,17),n)*( dk(pix(ib(m,27)),1,j,8,n)
&+dk(pix(ib(m,7)),2,j,7,n) )+
&u(ib(m,18),n)*( dk(pix(ib(m,27)),1,j,7,n) )+
&u(ib(m,19),n)*( dk(pix(ib(m,27)),1,j,6,n)
&+dk(pix(ib(m,5)),4,j,7,n) )+
&u(ib(m,20),n)*( dk(pix(ib(m,5)),4,j,6,n) )+
&u(ib(m,21),n)*( dk(pix(ib(m,5)),4,j,5,n)
&+dk(pix(ib(m,6)),3,j,6,n) )+
&u(ib(m,22),n)*( dk(pix(ib(m,6)),3,j,5,n) )+
&u(ib(m,23),n)*( dk(pix(ib(m,6)),3,j,8,n)
&+dk(pix(ib(m,7)),2,j,5,n) )+
&u(ib(m,24),n)*( dk(pix(ib(m,7)),2,j,8,n) )+
&u(ib(m,25),n)*( dk(pix(ib(m,14)),7,j,3,n)
&+dk(pix(ib(m,13)),8,j,4,n)+
&dk(pix(ib(m,15)),6,j,2,n)+dk(pix(ib(m,25)),5,j,1,n) )+
&u(ib(m,26),n)*( dk(pix(ib(m,6)),3,j,7,n)
&+dk(pix(ib(m,5)),4,j,8,n)+

155

&dk(pix(ib(m,27)),1,j,5,n)+dk(pix(ib(m,7)),2,j,6,n) )+
&u(ib(m,27),n)*( dk(pix(ib(m,27)),1,j,1,n)
&+dk(pix(ib(m,7)),2,j,2,n)+
&dk(pix(ib(m,6)),3,j,3,n)+dk(pix(ib(m,5)),4,j,4,n)
&+dk(pix(ib(m,25)),5,j,5,n)+
&dk(pix(ib(m,15)),6,j,6,n)+dk(pix(ib(m,14)),7,j,7,n)+
&dk(pix(ib(m,13)),8,j,8,n) )
3000 continue
utot=0.0
gtot=0.0
g2=0.0
g1=0.0
do 3100 m3=1,3
do 3100 m=1,ns
utot=utot+0.5*u(m,m3)*gb(m,m3)+b(m,m3)*u(m,m3)
c this is gradient of energy with respect to normal displacements
gb(m,m3)=gb(m,m3)+b(m,m3)
3100 continue
c

compute "constant" macrostrain energy term
C=0.0
do 7200 i=1,2
do 7200 j=1,2
do 7200 mi=1,3
do 7200 mj=1,3
C=C+0.5*u(ns+i,mi)*zcon(i,mi,j,mj)*u(ns+j,mj)
7200 continue
utot=utot+C
c now add in constant term from thermal energy, Y
utot=utot+Y
c now add in linear term in thermal energy
do 7171 m3=1,3
do 7171 m=1,nss
utot=utot+T(m,m3)*u(m,m3)
7171 continue
c
c

now compute gradient with respect to macrostrains
put in piece from first derivative of zcon quadratic term
do 7300 i=1,2
do 7300 mi=1,3
sum=0.0
do 7250 j=1,2
do 7250 mj=1,3

156

sum=sum+zcon(i,mi,j,mj)*u(ns+j,mj)
continue
gb(ns+i,mi)=sum
7300 continue
c add in piece of gradient, for displacements as well as macrostrains,
c that come from linear term in thermal energy
do 3150 m3=1,3
do 3150 m=1,nss
gb(m,m3)=gb(m,m3)+T(m,m3)
3150 continue
c now generate part that comes from b . u term
c do by calling b generation with appropriate macrostrain set to 1 to
c get that partial derivative, just use bgrad (taken from femat),
c skip dk and zcon part
do 8100 ii=1,6
exx=0.0
eyy=0.0
ezz=0.0
exz=0.0
eyz=0.0
exy=0.0
if(ii.eq.1) exx=1.0
if(ii.eq.2) eyy=1.0
if(ii.eq.3) ezz=1.0
if(ii.eq.4) exz=1.0
if(ii.eq.5) eyz=1.0
if(ii.eq.6) exy=1.0
call bgrad(nx,ny,nz,ns,exx,eyy,ezz,exz,eyz,exy)
sum=0.0
do 8200 m3=1,3
do 8200 m=1,ns
sum=sum+u(m,m3)*b(m,m3)
8200 continue
if(ii.eq.1) gb(ns+1,1)=gb(ns+1,1)+sum
if(ii.eq.2) gb(ns+1,2)=gb(ns+1,2)+sum
if(ii.eq.3) gb(ns+1,3)=gb(ns+1,3)+sum
if(ii.eq.4) gb(ns+2,1)=gb(ns+2,1)+sum
if(ii.eq.5) gb(ns+2,2)=gb(ns+2,2)+sum
if(ii.eq.6) gb(ns+2,3)=gb(ns+2,3)+sum
8100
continue
return
end
7250

c

Subroutine that carries out the conjugate gradient relaxation process

157

subroutine dembx(nx,ny,nz,ns,Lstep,gg,gtest,ldemb,kkk)
real u(8002,3),gb(8002,3),b(8002,3)
real h(8002,3),Ah(8002,3)
real dk(100,8,3,8,3),zcon(2,3,2,3)
real lambda,gamma
integer*4 ib(8002,27)
integer*2 pix(8002)
common/list2/h,Ah
common/list3/ib
common/list4/pix
common/list5/dk,b,C,zcon,Y
common/list6/u
common/list7/gb
nss=ns+2
c Initialize the conjugate direction vector on first call to dembx only.
c For calls to dembx after the first, we want to continue using the value
c of h determined in the previous call. Of course, if npoints
c is greater than 1, then this initialization step will be run each time
c a new microstructure is used, as kkk will be reset to 1 every time
c the counter micro is increased.
if(kkk.eq.1) then
do 500 m3=1,3
do 500 m=1,nss
h(m,m3)=gb(m,m3)
500
continue
end if
c
c

Lstep counts the number of conjugate gradient steps taken
in each call to dembx
Lstep=0
do 800 ijk=1,ldemb
Lstep=Lstep+1

do 290 m3=1,3
do 290 m=1,nss
Ah(m,m3)=0.0
290
continue
c Do global matrix multiply via small stiffness matrices, Ah = A * h
c The long statement below correctly brings in all the terms from
c the global matrix A using only the small stiffness matrices.
do 400 j=1,3

158

do 400 n=1,3
do 400 m=1,ns
Ah(m,j)=Ah(m,j)+h(ib(m,1),n)*( dk(pix(ib(m,27)),1,j,4,n)
&+dk(pix(ib(m,7)),2,j,3,n)
&+dk(pix(ib(m,25)),5,j,8,n)+dk(pix(ib(m,15)),6,j,7,n) )+
&h(ib(m,2),n)*( dk(pix(ib(m,27)),1,j,3,n)
&+dk(pix(ib(m,25)),5,j,7,n) )+
&h(ib(m,3),n)*( dk(pix(ib(m,27)),1,j,2,n)+dk(pix(ib(m,5)),4,j,3,n)+
&dk(pix(ib(m,13)),8,j,7,n)+dk(pix(ib(m,25)),5,j,6,n) )+
&h(ib(m,4),n)*( dk(pix(ib(m,5)),4,j,2,n)
&+dk(pix(ib(m,13)),8,j,6,n) )+
&h(ib(m,5),n)*( dk(pix(ib(m,6)),3,j,2,n)+dk(pix(ib(m,5)),4,j,1,n)+
&dk(pix(ib(m,14)),7,j,6,n)+dk(pix(ib(m,13)),8,j,5,n) )+
&h(ib(m,6),n)*( dk(pix(ib(m,6)),3,j,1,n)
&+dk(pix(ib(m,14)),7,j,5,n) )+
&h(ib(m,7),n)*( dk(pix(ib(m,6)),3,j,4,n)+dk(pix(ib(m,7)),2,j,1,n)+
&dk(pix(ib(m,14)),7,j,8,n)+dk(pix(ib(m,15)),6,j,5,n) )+
&h(ib(m,8),n)*( dk(pix(ib(m,7)),2,j,4,n)
&+dk(pix(ib(m,15)),6,j,8,n) )+
&h(ib(m,9),n)*( dk(pix(ib(m,25)),5,j,4,n)
&+dk(pix(ib(m,15)),6,j,3,n) )+
&h(ib(m,10),n)*( dk(pix(ib(m,25)),5,j,3,n) )+
&h(ib(m,11),n)*( dk(pix(ib(m,13)),8,j,3,n)
&+dk(pix(ib(m,25)),5,j,2,n) )+
&h(ib(m,12),n)*( dk(pix(ib(m,13)),8,j,2,n) )+
&h(ib(m,13),n)*( dk(pix(ib(m,13)),8,j,1,n)
&+dk(pix(ib(m,14)),7,j,2,n) )+
&h(ib(m,14),n)*( dk(pix(ib(m,14)),7,j,1,n) )+
&h(ib(m,15),n)*( dk(pix(ib(m,14)),7,j,4,n)
&+dk(pix(ib(m,15)),6,j,1,n) )+
&h(ib(m,16),n)*( dk(pix(ib(m,15)),6,j,4,n) )+
&h(ib(m,17),n)*( dk(pix(ib(m,27)),1,j,8,n)
&+dk(pix(ib(m,7)),2,j,7,n) )+
&h(ib(m,18),n)*( dk(pix(ib(m,27)),1,j,7,n) )+
&h(ib(m,19),n)*( dk(pix(ib(m,27)),1,j,6,n)
&+dk(pix(ib(m,5)),4,j,7,n) )+
&h(ib(m,20),n)*( dk(pix(ib(m,5)),4,j,6,n) )+
&h(ib(m,21),n)*( dk(pix(ib(m,5)),4,j,5,n)
&+dk(pix(ib(m,6)),3,j,6,n) )+
&h(ib(m,22),n)*( dk(pix(ib(m,6)),3,j,5,n) )+
&h(ib(m,23),n)*( dk(pix(ib(m,6)),3,j,8,n)
&+dk(pix(ib(m,7)),2,j,5,n) )+
&h(ib(m,24),n)*( dk(pix(ib(m,7)),2,j,8,n) )+
&h(ib(m,25),n)*( dk(pix(ib(m,14)),7,j,3,n)
&+dk(pix(ib(m,13)),8,j,4,n)+

159

400

&dk(pix(ib(m,15)),6,j,2,n)+dk(pix(ib(m,25)),5,j,1,n) )+
&h(ib(m,26),n)*( dk(pix(ib(m,6)),3,j,7,n)
&+dk(pix(ib(m,5)),4,j,8,n)+
&dk(pix(ib(m,27)),1,j,5,n)+dk(pix(ib(m,7)),2,j,6,n) )+
&h(ib(m,27),n)*( dk(pix(ib(m,27)),1,j,1,n)
&+dk(pix(ib(m,7)),2,j,2,n)+
&dk(pix(ib(m,6)),3,j,3,n)+dk(pix(ib(m,5)),4,j,4,n)
&+dk(pix(ib(m,25)),5,j,5,n)+
&dk(pix(ib(m,15)),6,j,6,n)+dk(pix(ib(m,14)),7,j,7,n)+
&dk(pix(ib(m,13)),8,j,8,n) )
continue

c
c
c
c
c
c

The above accurately gives the second derivative matrix with respect
to nodal displacements, but fails to give the 2nd derivative terms that
include the macrostrains [ du d(strain) and d(strain)d(strain) ].
Use repeated calls to bgrad to generate mixed 2nd derivatives terms,
plus use zcon in order to correct the matrix multiply and correctly bring
in macrostrain terms (see manual, Sec. 2.4).
do 8100 ii=1,6
e11=0.0
e22=0.0
e33=0.0
e13=0.0
e23=0.0
e12=0.0
if(ii.eq.1) e11=1.0
if(ii.eq.2) e22=1.0
if(ii.eq.3) e33=1.0
if(ii.eq.4) e13=1.0
if(ii.eq.5) e23=1.0
if(ii.eq.6) e12=1.0
call bgrad(nx,ny,nz,ns,e11,e22,e33,e13,e23,e12)
c now fill in terms from matrix multiply
c right hand sides, 1 to ns
do 3333 m=1,ns
do 3333 m1=1,3
if(ii.eq.1) Ah(m,m1)=Ah(m,m1)+b(m,m1)*h(ns+1,1)
if(ii.eq.2) Ah(m,m1)=Ah(m,m1)+b(m,m1)*h(ns+1,2)
if(ii.eq.3) Ah(m,m1)=Ah(m,m1)+b(m,m1)*h(ns+1,3)
if(ii.eq.4) Ah(m,m1)=Ah(m,m1)+b(m,m1)*h(nss,1)
if(ii.eq.5) Ah(m,m1)=Ah(m,m1)+b(m,m1)*h(nss,2)
if(ii.eq.6) Ah(m,m1)=Ah(m,m1)+b(m,m1)*h(nss,3)
3333 continue
c now do across bottom, 1 to ns
do 3334 m=1,ns

160

if(ii.eq.1) Ah(ns+1,1)=Ah(ns+1,1)+b(m,1)*h(m,1)+
+b(m,2)*h(m,2)+b(m,3)*h(m,3)
if(ii.eq.2) Ah(ns+1,2)=Ah(ns+1,2)+b(m,1)*h(m,1)+
+b(m,2)*h(m,2)+b(m,3)*h(m,3)
if(ii.eq.3) Ah(ns+1,3)=Ah(ns+1,3)+b(m,1)*h(m,1)+
+b(m,2)*h(m,2)+b(m,3)*h(m,3)
if(ii.eq.4) Ah(nss,1)=Ah(nss,1)+b(m,1)*h(m,1)+
+b(m,2)*h(m,2)+b(m,3)*h(m,3)
if(ii.eq.5) Ah(nss,2)=Ah(nss,2)+b(m,1)*h(m,1)+
+b(m,2)*h(m,2)+b(m,3)*h(m,3)
if(ii.eq.6) Ah(nss,3)=Ah(nss,3)+b(m,1)*h(m,1)+
+b(m,2)*h(m,2)+b(m,3)*h(m,3)
3334 continue
c now do righthand corner terms, ns+1 to nss
do 3335 m=1,2
do 3335 m1=1,3
if(ii.eq.1) Ah(ns+1,1)=Ah(ns+1,1)+zcon(1,1,m,m1)*h(ns+m,m1)
if(ii.eq.2) Ah(ns+1,2)=Ah(ns+1,2)+zcon(1,2,m,m1)*h(ns+m,m1)
if(ii.eq.3) Ah(ns+1,3)=Ah(ns+1,3)+zcon(1,3,m,m1)*h(ns+m,m1)
if(ii.eq.4) Ah(nss,1)=Ah(nss,1)+zcon(2,1,m,m1)*h(ns+m,m1)
if(ii.eq.5) Ah(nss,2)=Ah(nss,2)+zcon(2,2,m,m1)*h(ns+m,m1)
if(ii.eq.6) Ah(nss,3)=Ah(nss,3)+zcon(2,3,m,m1)*h(ns+m,m1)
3335 continue
8100

continue

530

hAh=0.0
do 530 m3=1,3
do 530 m=1,nss
hAh=hAh+h(m,m3)*Ah(m,m3)
continue

540

lambda=gg/hAh
do 540 m3=1,3
do 540 m=1,nss
u(m,m3)=u(m,m3)-lambda*h(m,m3)
gb(m,m3)=gb(m,m3)-lambda*Ah(m,m3)
continue

550

gglast=gg
gg=0.0
do 550 m3=1,3
do 550 m=1,nss
gg=gg+gb(m,m3)*gb(m,m3)
continue

161

if(gg.lt.gtest) goto 1000

570

gamma=gg/gglast
do 570 m3=1,3
do 570 m=1,nss
h(m,m3)=gb(m,m3)+gamma*h(m,m3)
continue

800

continue

1000

continue
return
end

c

Subroutine that computes the six average stresses and six average strains
subroutine stress(nx,ny,nz,ns)
real u(8002,3),uu(8,3)
real T(8002,3),eigen(100,6)
real dndx(8),dndy(8),dndz(8),es(6,8,3),cmod(100,6,6)
integer*4 ib(8002,27)
integer*2 pix(8002)
common/list1/strxx,stryy,strzz,strxz,stryz,strxy
common/list3/ib
common/list4/pix
common/list6/u
common/list8/cmod,T,eigen
common/list11/sxx,syy,szz,sxz,syz,sxy
nxy=nx*ny
nss=ns+2
exx=u(ns+1,1)
eyy=u(ns+1,2)
ezz=u(ns+1,3)
exz=u(nss,1)
eyz=u(nss,2)
exy=u(nss,3)

c set up single pixel strain matrix
dndx(1)=-0.25
dndx(2)=0.25
dndx(3)=0.25

162

dndx(4)=-0.25
dndx(5)=-0.25
dndx(6)=0.25
dndx(7)=0.25
dndx(8)=-0.25
dndy(1)=-0.25
dndy(2)=-0.25
dndy(3)=0.25
dndy(4)=0.25
dndy(5)=-0.25
dndy(6)=-0.25
dndy(7)=0.25
dndy(8)=0.25
dndz(1)=-0.25
dndz(2)=-0.25
dndz(3)=-0.25
dndz(4)=-0.25
dndz(5)=0.25
dndz(6)=0.25
dndz(7)=0.25
dndz(8)=0.25
c
c

Build average strain matrix, follows code in femat, but for average
strain over a pixel, not the strain at a point
do 2799 n1=1,6
do 2799 n2=1,8
do 2799 n3=1,3
es(n1,n2,n3)=0.0
2799 continue
do 2797 n=1,8
es(1,n,1)=dndx(n)
es(2,n,2)=dndy(n)
es(3,n,3)=dndz(n)
es(4,n,1)=dndz(n)
es(4,n,3)=dndx(n)
es(5,n,2)=dndz(n)
es(5,n,3)=dndy(n)
es(6,n,1)=dndy(n)
es(6,n,2)=dndx(n)
2797 continue
c now compute average stresses and strains in each pixel
sxx=0.0
syy=0.0
szz=0.0
sxz=0.0

163

syz=0.0
sxy=0.0
strxx=0.0
stryy=0.0
strzz=0.0
strxz=0.0
stryz=0.0
strxy=0.0
do 470 k=1,nz
do 470 j=1,ny
do 470 i=1,nx
m=(k-1)*nxy+(j-1)*nx+i
c load in elements of 8-vector using pd. bd. conds.
do 9898 mm=1,3
uu(1,mm)=u(m,mm)
uu(2,mm)=u(ib(m,3),mm)
uu(3,mm)=u(ib(m,2),mm)
uu(4,mm)=u(ib(m,1),mm)
uu(5,mm)=u(ib(m,26),mm)
uu(6,mm)=u(ib(m,19),mm)
uu(7,mm)=u(ib(m,18),mm)
uu(8,mm)=u(ib(m,17),mm)
9898 continue
c Correct for periodic boundary conditions, some displacements are wrong
c for a pixel on a periodic boundary. Since they come from an opposite
c face, need to put in applied strain to correct them.
if(i.eq.nx) then
uu(2,1)=uu(2,1)+exx*nx
uu(2,2)=uu(2,2)+exy*nx
uu(2,3)=uu(2,3)+exz*nx
uu(3,1)=uu(3,1)+exx*nx
uu(3,2)=uu(3,2)+exy*nx
uu(3,3)=uu(3,3)+exz*nx
uu(6,1)=uu(6,1)+exx*nx
uu(6,2)=uu(6,2)+exy*nx
uu(6,3)=uu(6,3)+exz*nx
uu(7,1)=uu(7,1)+exx*nx
uu(7,2)=uu(7,2)+exy*nx
uu(7,3)=uu(7,3)+exz*nx
end if
if(j.eq.ny) then
uu(3,1)=uu(3,1)+exy*ny
uu(3,2)=uu(3,2)+eyy*ny
uu(3,3)=uu(3,3)+eyz*ny
uu(4,1)=uu(4,1)+exy*ny

164

uu(4,2)=uu(4,2)+eyy*ny
uu(4,3)=uu(4,3)+eyz*ny
uu(7,1)=uu(7,1)+exy*ny
uu(7,2)=uu(7,2)+eyy*ny
uu(7,3)=uu(7,3)+eyz*ny
uu(8,1)=uu(8,1)+exy*ny
uu(8,2)=uu(8,2)+eyy*ny
uu(8,3)=uu(8,3)+eyz*ny
end if
if(k.eq.nz) then
uu(5,1)=uu(5,1)+exz*nz
uu(5,2)=uu(5,2)+eyz*nz
uu(5,3)=uu(5,3)+ezz*nz
uu(6,1)=uu(6,1)+exz*nz
uu(6,2)=uu(6,2)+eyz*nz
uu(6,3)=uu(6,3)+ezz*nz
uu(7,1)=uu(7,1)+exz*nz
uu(7,2)=uu(7,2)+eyz*nz
uu(7,3)=uu(7,3)+ezz*nz
uu(8,1)=uu(8,1)+exz*nz
uu(8,2)=uu(8,2)+eyz*nz
uu(8,3)=uu(8,3)+ezz*nz
end if
c

stresses and strains in a pixel
str11=0.0
str22=0.0
str33=0.0
str13=0.0
str23=0.0
str12=0.0
s11=0.0
s22=0.0
s33=0.0
s13=0.0
s23=0.0
s12=0.0
c********compute average stress and strain tensor in each pixel*************
c First put thermal strain-induced stresses into stress tensor
do 465 n=1,6
str11=str11-cmod(pix(m),1,n)*eigen(pix(m),n)
str22=str22-cmod(pix(m),2,n)*eigen(pix(m),n)
str33=str33-cmod(pix(m),3,n)*eigen(pix(m),n)
str13=str13-cmod(pix(m),4,n)*eigen(pix(m),n)
str23=str23-cmod(pix(m),5,n)*eigen(pix(m),n)

165

str12=str12-cmod(pix(m),6,n)*eigen(pix(m),n)
continue
do 466 n3=1,3
do 466 n8=1,8
c compute non-thermal strains in each pixel
s11=s11+es(1,n8,n3)*uu(n8,n3)
s22=s22+es(2,n8,n3)*uu(n8,n3)
s33=s33+es(3,n8,n3)*uu(n8,n3)
s13=s13+es(4,n8,n3)*uu(n8,n3)
s23=s23+es(5,n8,n3)*uu(n8,n3)
s12=s12+es(6,n8,n3)*uu(n8,n3)
do 466 n=1,6
c compute stresses in each pixel that include both non-thermal
c and thermal strains
str11=str11+cmod(pix(m),1,n)*es(n,n8,n3)*uu(n8,n3)
str22=str22+cmod(pix(m),2,n)*es(n,n8,n3)*uu(n8,n3)
str33=str33+cmod(pix(m),3,n)*es(n,n8,n3)*uu(n8,n3)
str13=str13+cmod(pix(m),4,n)*es(n,n8,n3)*uu(n8,n3)
str23=str23+cmod(pix(m),5,n)*es(n,n8,n3)*uu(n8,n3)
str12=str12+cmod(pix(m),6,n)*es(n,n8,n3)*uu(n8,n3)
466
continue
c Sum local stresses and strains into global stresses and strains
strxx=strxx+str11
stryy=stryy+str22
strzz=strzz+str33
strxz=strxz+str13
stryz=stryz+str23
strxy=strxy+str12
sxx=sxx+s11
syy=syy+s22
szz=szz+s33
sxz=sxz+s13
syz=syz+s23
sxy=sxy+s12
470
continue
465

c

Volume average global stresses and strains

strxx=strxx/float(ns)
stryy=stryy/float(ns)
strzz=strzz/float(ns)
strxz=strxz/float(ns)
stryz=stryz/float(ns)
strxy=strxy/float(ns)
sxx=sxx/float(ns)

166

syy=syy/float(ns)
szz=szz/float(ns)
sxz=sxz/float(ns)
syz=syz/float(ns)
sxy=sxy/float(ns)
return
end
c

Subroutine to count volume fractions of various phases
subroutine assig(ns,nphase,prob)
integer*2 pix(8002)
real prob(100)
common/list4/pix

do 999 i=1,nphase
prob(i)=0.0
999 continue
do 1000 m=1,ns
do 1000 i=1,nphase
if(pix(m).eq.i) then
prob(i)=prob(i)+1
end if
1000
continue
do 998 i=1,nphase
prob(i)=prob(i)/float(ns)
998 continue
return
end
c

Subroutine to set up image of microstructure
subroutine ppixel(nx,ny,nz,ns,nphase)
integer*2 pix(8002)
integer*4 ib(8002,27)
common/list3/ib
common/list4/pix

c
c

(USER) If you want to set up a test image inside the program, instead
of reading it in from a file, this should be done inside this subroutine.

167

200

nxy=nx*ny
do 200 k=1,nz
do 200 j=1,ny
do 200 i=1,nx
m=nxy*(k-1)+nx*(j-1)+i
read(9,*) pix(m)
continue

c

Check for wrong phase labels--less than 1 or greater than nphase
do 500 m=1,ns
if(pix(m).lt.1) then
write(7,*) 'Phase label in pix < 1--error at ',m
end if
if(pix(m).gt.nphase) then
write(7,*) 'Phase label in pix > nphase--error at ',m
end if
500
continue
return
end

168

9.3.4 DC3D.F
c
c

*************************
BACKGROUND

dc3d.f

*********************************

c
c
c
c
c
c
c
c

This program accepts as input a 3-d digital image, converting it
into a real conductor network. The conjugate gradient method
is used to solve this finite difference representation of Laplace's
equation for real conductivity problems.
Periodic boundary conditions are maintained.
In the comments below, (USER) means that this is a section of code
that the user might have to change for his particular problem.
Therefore the user is encouraged to search for this string.

c

PROBLEM AND VARIABLE DEFINITION

c
c
c
c
c
c
c
c
c
c
c
c

The mathematical problem that the conjugate gradient algorithm solves
is the minimization of the quadratic form 1/2 uAu, where
u is the vector of voltages, and A is generated from the bond
conductances between pixels. Nodes are thought of as being in the
center of pixels. The minimization is constrained by maintaining an
general direction applied electric field across the sample.
The vectors gx,gy,gz are bond conductances, u is the voltage array,
and gb,h, and Ah are auxiliary variables, used in subroutine dembx.
The vector pix contains the phase labels for each pixel.
The small vector a(i) is the volume fraction
of the i'th phase, and currx, curry, currz are the total volume-averaged
currents in the x,y, and z directions.

c

DIMENSIONS

c
c
c
c
c
c
c
c
c
c
c
c
c
c

The vectors gx,gy,gz,u,gb,h,Ah,list,pix are all dimensioned
ns2 = (nx+2)*(ny+2)*(nz+2). This number is used, rather than the
system size nx x ny x nz, because an extra layer of pixels is
put around the system to be able to maintain periodic boundary
conditions (see manual, Sec. 3.3). The arrays pix and list are also
dimensioned this way.
At present the program is set up for up to 100
phases, but that can easily be changed by the user, by changing the
dimension of sigma, a, and be. Note that be has both dimensions
equal to each other. The parameter nphase gives the number of
phases being considered. The parameter ntot is the total number
of phases possible in the program, and should be equal to the
dimension of sigma, a, and be.
All arrays are passed to subroutines in the call statements.

169

c

STRONGLY RECOMMENDED:

READ MANUAL BEFORE USING THE PROGRAM!!

c
c
c

(USER) Change these dimensions for different system sizes. All
dimensions in the subroutines are passed, so do not need to be changed.
The dimensions of sigma, a, and be should be equal to the value of ntot.
real gx(10648),gy(10648),u(10648),gz(10648)
real gb(10648),h(10648),ah(10648)
real currx,curry,currz,sigma(100,3)
real a(100),be(100,100,3)
integer*2 pix(10648)
integer*4 list(10648)

c
c

(USER) Unit 9 is the microstructure input file, unit 7 is the
results output file.
open(unit=9,file='microstructure.dat')
open(unit=7,file='outputfile.out')

c

(USER) real image size is nx x ny x nz
nx=20
ny=20
nz=20
write(7,1111) nx,ny,nz,nx*ny*nz
1111 format(' Image is ',3i6,' No. of real sites = ',i8)
c auxiliary variables involving the system size
nx1=nx+1
ny1=ny+1
nz1=nz+1
nx2=nx+2
ny2=ny+2
nz2=nz+2
L22=nx2*ny2
c computational image size ns2 is nx2 x ny2 x nz2
ns2=nx2*ny2*nz2
c
c
c
c

(USER) set cutoff for norm squared of gradient, gtest. gtest is
the stopping criterion, compared to gb*gb. When gb*gb is less
than gtest=abc*ns2, then the rms value of the gradient at a pixel
is less than sqrt(abc).
gtest=1.0e-16*ns2

c
c
c
c

(USER) nphase is the number of phases being considered in the problem.
The values of pix(m) will run from 1 to nphase. ntot is the total
number of phases possible in the program, and is the dimension of
sigma, a, and be.

170

nphase=2
ntot=100
c
c
c

Make list of real (interior) sites, used in subroutine dembx. The 1-d
labelling scheme goes over all ns2 sites, so a list of the real sites
is needed.
nlist=0
do 103 i=2,nx1
do 102 j=2,ny1
do 101 k=2,nz1
m=i+(j-1)*nx2+(k-1)*L22
nlist=nlist+1
list(nlist)=m
101
continue
102
continue
103
continue
c Compute average current in each pixel.
c (USER) npoints is the number of microstructures to use.

npoints=1
do 8000 micro=1,npoints
c Read in a microstructure in subroutine ppixel, and set up pix(m)
c with the appropriate phase assignments.
call ppixel(pix,nx2,ny2,nz2,a,ns2,nphase,ntot)
c output phase volume fractions
do 99 i=1,nphase
write(7,299) i,a(i)
299
format(' Phase fraction of ',i3,' = ',f12.6)
99
continue
c

(USER) Set components of applied field, E = (ex,ey,ez)
ex=1.0
ey=1.0
ez=1.0
write(7,*) 'Applied field components:'
write(7,*) 'ex = ',ex,' ey = ',ey,' ez = ',ez
c Initialize the voltage distribution by putting on uniform field.
do 30 k=1,nz2
do 30 j=1,ny2
do 30 i=1,nx2
m=(k-1)*nx2*ny2+nx2*(j-1)+i
u(m)=-ex*i-ey*j-ez*k
30
continue
c (USER) input value of real conductivity tensor for each phase

171

c

(diagonal only). 1,2,3 = x,y,z, respectively.
sigma(1,1)=1.0
sigma(1,2)=1.0
sigma(1,3)=1.0
sigma(2,1)=0.5
sigma(2,2)=0.5
sigma(2,3)=0.5

c

Subroutine bond sets up conductor network in gx,gy,gz 1-d arrays
call bond(pix,gx,gy,gz,nx2,ny2,nz2,ns2,sigma,be,nphase,ntot)

c
c

Subroutine dembx accepts gx,gy,gz and solves for the voltage field
that minimizes the dissipated energy.
call dembx(nx2,ny2,nz2,ns2,gx,gy,gz,u,ic,gb,h,ah,list,nlist,gtest)

c

find final current after voltage solution is done
call current(nx2,ny2,nz2,ns2,currx,curry,currz,u,gx,gy,gz)
write(7,*) 'Average current in x direction= ',currx
write(7,*) 'Average current in y direction= ',curry
write(7,*) 'Average current in z direction= ',currz
write(7,*) ic,' number of conjugate gradient cycles needed'
call flush(7)
8000 continue
end
c
c

Subroutine that performs the conjugate gradient solution routine to
find the correct set of nodal voltages

&

c
c
c
c
c

subroutine dembx(nx2,ny2,nz2,ns2,gx,gy,gz,u,ic,gb,h,Ah,
list,nlist,gtest)
real gx(ns2),gy(ns2),u(ns2),gb(ns2)
real Ah(ns2),h(ns2),gz(ns2)
real gg,hAh,lambda,gglast,gamma,ravg,currx,curry,currz
integer*4 list(ns2),ncheck

Note: voltage gradients are maintained because in the conjugate gradient
relaxation algorithm, the voltage vector is only modified by adding a
periodic vector to it.
L22=nx2*ny2
First stage, compute initial value of gradient (gb), initialize h, the
conjugate gradient direction, and compute norm squared of gradient vector.
call prod(nx2,ny2,nz2,ns2,gx,gy,gz,u,gb)
do 20 i=1,ns2
h(i)=gb(i)

172

20
continue
c Variable gg is the norm squared of the gradient vector
gg=0.0
do 105 k=1,nlist
m=list(k)
gg=gb(m)*gb(m)+gg
105
continue
c
c

205

50
c
c
c
c
c
c

Second stage, initialize Ah variable, compute parameter lamdba,
make first change in voltage array, update gradient (gb) vector
if(gg.lt.gtest) goto 44
call prod(nx2,ny2,nz2,ns2,gx,gy,gz,h,Ah)
hAh=0.0
do 205 k=1,nlist
m=list(k)
hAh=hAh+h(m)*Ah(m)
continue
lambda=gg/hAh
do 50 i=1,ns2
u(i)=u(i)-lambda*h(i)
gb(i)=gb(i)-lambda*Ah(i)
continue
third stage: iterate conjugate gradient solution process until
gg < gtest criterion is satisfied.
(USER) The parameter ncgsteps is the total number of conjugate gradient steps
to go through. Only in very unusual problems, like when the conductivity
of one phase is much higher than all the rest, will this many steps be
used.
ncgsteps=30000

do 33 icc=1,ncgsteps
gglast=gg
gg=0.0
do 305 k=1,nlist
m=list(k)
gg=gb(m)*gb(m)+gg
305
continue
call flush(7)
if(gg.lt.gtest) goto 44
gamma=gg/gglast
c update conjugate gradient direction
do 70 i=1,ns2
h(i)=gb(i)+gamma*h(i)

173

70

continue
call prod(nx2,ny2,nz2,ns2,gx,gy,gz,h,Ah)
hAh=0.0
do 401 k=1,nlist
m=list(k)
hAh=hAh+h(m)*Ah(m)
401
continue
lambda=gg/hAh
c update voltage, gradient vectors
do 90 i=1,ns2
u(i)=u(i)-lambda*h(i)
gb(i)=gb(i)-lambda*Ah(i)
90
continue
c
c
c
c
c

(USER) This piece of code forces dembx to write out the total current and
the norm of the gradient squared, every ncheck conjugate gradient steps,
in order to see how the relaxation is proceeding. If the currents become
unchanging before the relaxation is done, then gtest was picked to be
smaller than was necessary.
ncheck=30

if(ncheck*(icc/ncheck).eq.icc) then
write(7,*) icc
write(7,*) ' gg = ',gg
c call current subroutine
call current(nx2,ny2,nz2,ns2,currx,curry,currz,u,gx,gy,gz)
write(7,*) ' currx = ',currx
write(7,*) ' curry = ',curry
write(7,*) ' currz = ',currz
end if
call flush(7)
33
continue
write(7,*) ' Iteration failed to converge after',ncgsteps,' steps'
44
continue
ic=icc
return
end
c The matrix product subroutine
subroutine prod(nx2,ny2,nz2,ns2,gx,gy,gz,xw,yw)
real gx(ns2),gy(ns2),gz(ns2),xw(ns2),yw(ns2)
c

xw is the input vector, yw = (A)(xw) is the output vector

174

c

auxiliary variables involving the system size
nx1=nx2-1
ny1=ny2-1
nz1=nz2-1
nx=nx2-2
ny=ny2-2
nz=nz2-2
L22=nx2*ny2

c
c

Perform basic matrix multiplication, results in incorrect information at
periodic boundaries.
do 10 i=1,ns2
yw(i)=0.0
10
continue
do 20 i=L22+1,ns2-L22
yw(i)=-xw(i)*(gx(i-1)+gx(i)+gz(i-L22)+gz(i)+gy(i)+gy(i-nx2))
yw(i)=yw(i)+gx(i-1)*xw(i-1)+gx(i)*xw(i+1)
+ +gz(i-L22)*xw(i-L22)+gz(i)*xw(i+L22)+gy(i)*xw(i+nx2)
+ +gy(i-nx2)*xw(i-nx2)
20
continue
c

Correct terms at periodic boundaries (Section 3.3 in manual)

c

x faces
do 30 k=1,nz2
do 30 j=1,ny2
yw((k-1)*L22+nx2*(j-1)+nx2)=yw((k-1)*L22+nx2*(j-1)+2)
yw((k-1)*L22+nx2*(j-1)+1)=yw((k-1)*L22+nx2*(j-1)+nx1)
30
continue
c

40

y faces
do 40 k=1,nz2
do 40 i=1,nx2
yw((k-1)*L22+i)=yw((k-1)*L22+ny*nx2+i)
yw((k-1)*L22+ny1*nx2+i)=yw((k-1)*L22+nx2+i)
continue

c

z faces
do 50 m=1,L22
yw(m)=yw(m+nz*L22)
yw(m+nz1*L22)=yw(m+L22)
50
continue
return

175

end
c
c

Subroutine that determines the correct bond conductances that are used
to compute multiplication by the matrix A
subroutine bond(pix,gx,gy,gz,nx2,ny2,nz2,ns2,sigma,be,nphase,ntot)
real gx(ns2),gy(ns2),gz(ns2),sigma(ntot,3),be(ntot,ntot,3)
integer*2 pix(ns2)

c

auxiliary variables involving the system size
nx=nx2-2
ny=ny2-2
nz=nz2-2
nx1=nx2-1
ny1=ny2-1
nz1=nz2-1
L22=nx2*ny2

c
c
c

Set values of conductor for phase(i,m)--phase(j,m) interface,
store in array be(i,j,m), m=1,2,3. If either phase i or j
has zero conductivity in the m'th direction, then be(i,j,m)= 0.0.

10

do 10 m=1,3
do 10 i=1,nphase
do 10 j=1,nphase
if(sigma(i,m).eq.0.0) then
be(i,j,m)=0.0
goto 10
end if
if(sigma(j,m).eq.0.0) then
be(i,j,m)=0.0
goto 10
end if
be(i,j,m)=1./(0.5/sigma(i,m)+0.5/sigma(j,m))
continue

c
c
c
c

Trim off x and y faces so that no current can flow past periodic
boundaries. This step is not really necessary, as the voltages on the
periodic boundaries will be matched to the corresponding real voltages
in each conjugate gradient step.
do 20 k=1,nz2
do 15 j=1,ny2
gx((k-1)*L22+nx2*(j-1)+nx2)=0.0
15
continue
do 16 i=1,nx2

176

16
20

gy((k-1)*L22+ny1*nx2+i)=0.0
continue
continue

c
c

Set up conductor network
bulk--gz
do 30 i=1,nx2
do 30 j=1,ny2
do 30 k=1,nz1
m=(k-1)*L22+(j-1)*nx2+i
i1=i
j1=j
k1=k+1
m1=(k1-1)*L22+(j1-1)*nx2+i1
gz(m)=be(pix(m),pix(m1),3)
30
continue
c

bulk---gy
do 40 i=1,nx2
do 40 j=1,ny1
do 40 k=2,nz1
m=(k-1)*L22+(j-1)*nx2+i
j1=j+1
i1=i
k1=k
m1=(k1-1)*L22+(j1-1)*nx2+i1
gy(m)=be(pix(m),pix(m1),2)
40
continue
c

bulk--gx
do 50 i=1,nx1
do 50 j=1,ny2
do 50 k=2,nz1
m=(k-1)*L22+(j-1)*nx2+i
i1=i+1
j1=j
k1=k
m1=(k1-1)*L22+(j1-1)*nx2+i1
gx(m)=be(pix(m),pix(m1),1)
50
continue
return
end
c

Subroutine that sets up the image, either by reading it from file,

177

c

or generating it internally
subroutine ppixel(pix,nx2,ny2,nz2,a,ns2,nphase,ntot)
real a(ntot)
integer*2 pix(ns2)

c
c

(USER) If you want to set up a test image inside the program, instead
of reading it in from a file, this should be done inside this subroutine.

c

auxiliary variables involving the system size
nx=nx2-2
ny=ny2-2
nz=nz2-2
L22=nx2*ny2

c

Initialize phase fraction array.
do 120 i=1,nphase
a(i)=0.0
120
continue
c Use 1-d labelling scheme as shown in manual
do 100 k=2,nz2-1
do 100 j=2,ny2-1
do 100 i=2,nx2-1
m=(k-1)*L22+(j-1)*nx2+i
read(9,*) pix(m)
a(pix(m))=a(pix(m))+1.0
100
continue
do 220 i=1,nphase
a(i)=a(i)/float(nx*ny*nz)
220
continue
c

now map periodic boundaries of pix (see Section 3.3, Figure 3 in manual)
do 110 k=1,nz2
do 110 j=1,ny2
do 110 i=1,nx2
if(i.gt.1.and.i.lt.nx2) then
if(j.gt.1.and.j.lt.ny2) then
if(k.gt.1.and.k.lt.nz2) then
goto 110
end if
end if
end if
k1=k

178

if(k.eq.1) k1=k+nz
if(k.eq.nz2) k1=k-nz
j1=j
if(j.eq.1) j1=j+ny
if(j.eq.ny2) j1=j-ny
i1=i
if(i.eq.1) i1=i+nx
if(i.eq.nx2) i1=i-nx
m=(k-1)*L22+(j-1)*nx2+i
m1=(k1-1)*L22+(j1-1)*nx2+i1

110

pix(m)=pix(m1)
continue

c

Check for wrong phase labels--less than 1 or greater than nphase
do 500 m=1,ns2
if(pix(m).lt.1) then
write(7,*) 'Phase label in pix < 1--error at ',m
end if
if(pix(m).gt.nphase) then
write(7,*) 'Phase label in pix > nphase--error at ',m
end if
500
continue
return
end
c

Subroutine to compute the total current in the x, y, and z directions
subroutine current(nx2,ny2,nz2,ns2,currx,curry,currz,u,gx,gy,gz)
real u(ns2),gx(ns2),gy(ns2),gz(ns2),currx,curry,currz

c

c

c

auxiliary variables involving the system size
nx=nx2-2
ny=ny2-2
nz=nz2-2
L22=nx2*ny2
initialize the volume averaged currents
currx=0.0
curry=0.0
currz=0.0
Only loop over real sites and bonds in order to get true total current
do 10 k=2,nz2-1

179

do 10 j=2,ny2-1
do 10 i=2,nx2-1
m=L22*(k-1)+nx2*(j-1)+i
c cur1, cur2, cur3 are the currents in one pixel
cur1=0.5*( ( u(m-1) - u(m) )*gx(m-1)+( u(m)-u(m+1) )*gx(m) )
cur2=0.5*( ( u(m-nx2) - u(m) )*gy(m-nx2)+( u(m)-u(m+nx2) )*gy(m) )
cur3=0.5*( ( u(m-L22) - u(m) )*gz(m-L22)+( u(m)-u(m+L22) )*gz(m) )
c sum pixel currents into volume averages
currx=currx+cur1
curry=curry+cur2
currz=currz+cur3
10
continue
currx=currx/float(nx*ny*nz)
curry=curry/float(nx*ny*nz)
currz=currz/float(nx*ny*nz)
return
end

180

9.3.5 AC3D.F
c
c

************************
BACKGROUND

ac3d.f ********************************

c
c
c
c
c
c
c
c

This program accepts as input a 3-d digital image, converting it
into a complex conductor network. The conjugate gradient method
is used to solve this finite difference representation of Laplace's
equation for complex conductivity problems.
Periodic boundary conditions are maintained.
In the comments below, (USER) means that this is a section of code
that the user might have to change for his particular problem.
Therefore the user is encouraged to search for this string.

c

PROBLEM AND VARIABLE DEFINITION

c
c
c
c
c
c
c
c
c
c
c
c

The mathematical problem that the conjugate gradient algorithm solves
is the minimization of the quadratic form 1/2 uAu, where
u is the vector of voltages, and A is generated from the bond
conductances between pixels. Nodes are thought of as being in the
center of pixels. The minimization is constrained by maintaining an
general direction applied electric field across the sample.
The vectors gx,gy,gz are bond conductances, u is the voltage array,
and gb,h, and Ah are auxiliary variables, used in subroutine dembx.
The vector pix contains the phase labels for each pixel.
The small vector a(i) is the volume fraction
of the ith phase, and currx, curry, currz are the total volume-averaged
complex currents in the x,y, and z directions.

c

DIMENSIONS

c
c
c
c
c
c
c
c
c
c
c
c
c

The vectors gx,gy,gz,u,gb,h,Ah,list,pix are all dimensioned
ns2 = (nx+2)*(ny+2)*(nz+2). This number is used, rather than the
system size nx x ny x nz, because an extra layer of pixels is
put around the system to be able to maintain periodic boundary
conditions (see manual, Sec. 3.3). The arrays pix and list are also
dimensioned this way. At present the program is set up for 100 phases,
but that can easily be changed by the user, by changing the dimensions
of sigma, a, and be. Note that be has both dimensions equal to
each other. The parameter nphase gives the number of phases
being considered. The parameter ntot is the total number of phases
possible in the program, and should be equal to the dimension
of sigma, a, and be. All arrays are passed to subroutines in
the call statements.

181

c

STRONGLY RECOMMENDED:

c
c
c
c

(USER) Change these dimensions for different system sizes. All
dimensions in the subroutines are passed, so do not need to be
changed. The dimensions of sigma, a, and be should be equal to
the value of ntot.
complex gx(10648),gy(10648),u(10648),gz(10648)
complex gb(10648),h(10648),ah(10648)
complex currx,curry,currz,sigma(100,3),be(100,100,3)
real a(100)
integer*2 pix(10648)
integer*4 list(10648)
(USER) Unit 9 is the microstructure input file, unit 7 is the
results output file.
open(unit=9,file='microstructure.dat')
open(unit=7,file='outputfile.out')

c
c

READ MANUAL BEFORE USING THE PROGRAM!!

c

(USER) real image size is nx x ny x nz
nx=20
ny=20
nz=20
c auxiliary variables involving the system size
nx1=nx+1
ny1=ny+1
nz1=nz+1
nx2=nx+2
ny2=ny+2
nz2=nz+2
L22=nx2*ny2
write(7,1111) nx,ny,nz,nx*ny*nz
1111 format(' Image is ',3i6,' No. of real sites = ',i8)
c computational image size ns2 is nx2 x ny2 x nz2
ns2=nx2*ny2*nz2
c

defines the value of pi for later use
pi=4.0*atan(1.0)

c
c
c
c

(USER) set cutoff for norm squared of gradient, gtest. gtest is
the stopping criterion, compared to gb*gb. When gb*gb is less
than gtest=abc*ns2, then the rms value of the gradient at a pixel
is less than sqrt(abc).
gtest=1.0e-16*ns2

c
c

(USER) nphase is the number of phases being considered in the problem.
The values of pix(m) will run from 1 to nphase. ntot is the total

182

c
c

number of phases possible in the program, and is the dimension of
sigma, a, and be.
nphase=2
ntot=100

c
c
c

Make list of real (interior) sites, used in subroutine dembx. The 1-d
labelling scheme goes over all ns2 sites, so a list of the real sites
is needed.
nlist=0
do 103 i=2,nx1
do 102 j=2,ny1
do 101 k=2,nz1
m=i+(j-1)*nx2+(k-1)*L22
nlist=nlist+1
list(nlist)=m
101
continue
102
continue
103
continue
c Compute average current in each pixel.
c (USER) npoints is the number of microstructures to use.
c nfreq is the number of frequencies to be computed.
c The program is set up assuming that the effective
c conductivity is going to be solved for at several different
c frequencies on the same microstructure.
npoints=1
do 400 micro=1,npoints
c Read in a microstructure in subroutine ppixel, and set up
c pix(m) with the appropriate phase assignments.
call ppixel(pix,nx2,ny2,nz2,a,ns2,nphase,ntot)
c output phase volume fractions
do 99 i=1,nphase
write(7,299) i,a(i)
299
format(' Phase fraction of ',i3,' = ',f12.6)
99
continue
c

c
c

(USER) Set components of applied field, E = (ex,ey,ez)
ex=1.0
ey=1.0
ez=1.0
write(7,*) 'Applied field components:'
write(7,*) 'ex = ',ex,' ey = ',ey,' ez = ',ez
Initialize the voltage distribution by putting on uniform field.
Only do this for the first frequency considered, thereafter use the

183

c

previous frequency's voltages as a starting point.
do 30 k=1,nz2
do 30 j=1,ny2
do 30 i=1,nx2
m=(k-1)*nx2*ny2+nx2*(j-1)+i
u(m)=-ex*i-ey*j-ez*k
30
continue
c

c
c
c
c
c

c
c

(USER) Set how many frequencies need to be computed.
nfreq=50
Loop over desired frequencies.
do 300 nf=1,nfreq
(USER) Define frequency to use each time. Alter this statement to get
different frequencies. The frequencies are in Hz, according to
the units used for sigma.
w=10.**((nf-1)*11.4/49.-3.)
convert to angular frequency
w=w*2.*pi
write(7,*) 'No.',nf, ' angular frequency = ',w,' radians'
(USER) input value of complex conductivity tensor for each phase
(diagonal only). 1,2,3 = x,y,z, respectively.
sigma(1,1)=cmplx(1.0,10.*w)
sigma(1,2)=cmplx(1.0,10.*w)
sigma(1,3)=cmplx(1.0,10.*w)
sigma(2,1)=cmplx(0.5,1.*w)
sigma(2,2)=cmplx(0.5,1.*w)
sigma(2,3)=cmplx(0.5,1.*w)

c

bond() sets up conductor network in gx,gy,gz 1-d arrays
call bond(pix,gx,gy,gz,nx2,ny2,nz2,ns2,sigma,be,nphase,ntot)

c
c
c
c
c
c

Subroutine dembx accepts gx,gy,gz and solves for the voltage field
that minimizes the dissipated energy. As a starting point for u,
the voltage vector, each frequency uses the voltages left over from the
previous minimization. This can often reduce total run time dramatically,
compared to starting with a new voltage vector each time, as in do loop
30 above.
call dembx(nx2,ny2,nz2,ns2,gx,gy,gz,u,ic,gb,h,ah,list,nlist,gtest)

c

find final current after voltage solution is done
call current(nx2,ny2,nz2,ns2,currx,curry,currz,u,gx,gy,gz)
write(7,*) 'Average current in x direction= ',currx
write(7,*) 'Average current in y direction= ',curry
write(7,*) 'Average current in z direction= ',currz
write(7,*) ic,' number of conjugate gradient cycles needed'

184

300
400

c
c

call flush(7)
continue
continue
end
Subroutine that performs the conjugate gradient solution routine to
find the correct set of nodal voltages
subroutine dembx(nx2,ny2,nz2,ns2,gx,gy,gz,u,ic,gb,h,Ah,
& list,nlist,gtest)
complex gx(ns2),gy(ns2),u(ns2),gb(ns2)
complex Ah(ns2),h(ns2),gz(ns2)
complex gg,hAh,lambda,gglast,gamma,ravg,currx,curry,currz
integer*4 list(ns2),ncheck

c
c
c
c
c

Note: voltage gradients are maintained because in the conjugate gradient
relaxation algorithm, the voltage vector is only modified by adding a
periodic vector to it.
L22=nx2*ny2
First stage, compute initial value of gradient (gb), initialize h, the
conjugate gradient direction, and compute norm squared of gradient vector.

call prod(nx2,ny2,nz2,ns2,gx,gy,gz,u,gb)
do 20 i=1,ns2
h(i)=gb(i)
20
continue
c Complex variable gg is the norm squared of the gradient vector
gg=cmplx(0.0,0.0)
do 105 k=1,nlist
m=list(k)
gg=gb(m)*gb(m)+gg
105
continue
c
c

205

Second stage, initialize Ah variable, compute parameter lamdba,
make first change in voltage array, update gradient (gb) vector
if(abs(real(gg)).lt.gtest) goto 44
call prod(nx2,ny2,nz2,ns2,gx,gy,gz,h,Ah)
hAh=cmplx(0.0,0.0)
do 205 k=1,nlist
m=list(k)
hAh=hAh+h(m)*Ah(m)
continue
lambda=gg/hAh
do 50 i=1,ns2

185

50
c
c
c
c
c
c

u(i)=u(i)-lambda*h(i)
gb(i)=gb(i)-lambda*Ah(i)
continue
third stage: iterate conjugate gradient solution process until
real(gg) < gtest criterion is satisfied.
(USER) The parameter ncgsteps is the total number of conjugate gradient steps
to go through. Only in very unusual problems, like when the conductivity
of one phase is much higher than all the rest, will this many steps be
used.
ncgsteps=30000

do 33 icc=1,ncgsteps
gglast=gg
gg=cmplx(0.0,0.0)
do 305 k=1,nlist
m=list(k)
gg=gb(m)*gb(m)+gg
305
continue
call flush(7)
if(abs(real(gg)).lt.gtest) goto 44
gamma=gg/gglast
c update conjugate gradient direction
do 70 i=1,ns2
h(i)=gb(i)+gamma*h(i)
70
continue
call prod(nx2,ny2,nz2,ns2,gx,gy,gz,h,Ah)
hAh=cmplx(0.0,0.0)
do 401 k=1,nlist
m=list(k)
hAh=hAh+h(m)*Ah(m)
401
continue
lambda=gg/hAh
c update voltage, gradient vectors
do 90 i=1,ns2
u(i)=u(i)-lambda*h(i)
gb(i)=gb(i)-lambda*Ah(i)
90
continue
c (USER) This piece of code forces dembx to write out the total current and
c the norm of the gradient squared, every ncheck conjugate gradient steps,
c in order to see how the relaxation is proceeding. If the currents become
c unchanging before the relaxation is done, then gtest was picked to be
c smaller than was necessary.
ncheck=30

186

if(ncheck*(icc/ncheck).eq.icc) then
write(7,*) icc
write(7,*) ' gg = ',gg
c call current subroutine
call current(nx2,ny2,nz2,ns2,currx,curry,currz,u,gx,gy,gz)
write(7,*) ' currx = ',currx
write(7,*) ' curry = ',curry
write(7,*) ' currz = ',currz
end if
call flush(7)
33
continue
write(7,*) ' Iteration failed to converge after',ncgsteps,' steps'
44
continue
ic=icc
return
end
c The matrix product subroutine
subroutine prod(nx2,ny2,nz2,ns2,gx,gy,gz,xw,yw)
complex gx(ns2),gy(ns2),xw(ns2),gz(ns2),yw(ns2)
c

xw is the input vector, yw = (A)(xw) is the output vector

c

auxiliary variables involving the system size
nx1=nx2-1
ny1=ny2-1
nz1=nz2-1
nx=nx2-2
ny=ny2-2
nz=nz2-2
L22=nx2*ny2

c
c

Perform basic matrix multiplication, results in incorrect information at
periodic boundaries.
do 10 i=1,ns2
yw(i)=cmplx(0.0,0.0)
10
continue
do 20 i=L22+1,ns2-L22
yw(i)=-xw(i)*(gx(i-1)+gx(i)+gz(i-L22)+gz(i)+gy(i)+gy(i-nx2))
yw(i)=yw(i)+gx(i-1)*xw(i-1)+gx(i)*xw(i+1)
+ +gz(i-L22)*xw(i-L22)+gz(i)*xw(i+L22)+gy(i)*xw(i+nx2)
+ +gy(i-nx2)*xw(i-nx2)
20
continue

187

c

Correct terms at periodic boundaries (Section 3.3 in manual)

c

x faces
do 30 k=1,nz2
do 30 j=1,ny2
yw((k-1)*L22+nx2*(j-1)+nx2)=yw((k-1)*L22+nx2*(j-1)+2)
yw((k-1)*L22+nx2*(j-1)+1)=yw((k-1)*L22+nx2*(j-1)+nx1)
30
continue
c

40

y faces
do 40 k=1,nz2
do 40 i=1,nx2
yw((k-1)*L22+i)=yw((k-1)*L22+ny*nx2+i)
yw((k-1)*L22+ny1*nx2+i)=yw((k-1)*L22+nx2+i)
continue

c

z faces
do 50 m=1,L22
yw(m)=yw(m+nz*L22)
yw(m+nz1*L22)=yw(m+L22)
50
continue
return
end
c
c

Subroutine that determines the correct bond conductances that are used
to compute multiplication by the matrix A
subroutine bond(pix,gx,gy,gz,nx2,ny2,nz2,ns2,sigma,be,nphase,ntot)
complex gx(ns2),gy(ns2),gz(ns2),sigma(ntot,3),be(ntot,ntot,3)
integer*2 pix(ns2)

c

auxiliary variables involving the system size
L22=nx2*ny2
nx=nx2-2
ny=ny2-2
nz=nz2-2
nx1=nx2-1
ny1=ny2-1
nz1=nz2-1

c
c
c

Set values of conductor for phase(i)--phase(j) interface,
store in array be(i,j,3). If either phase i or j has zero conductivity,
then be(i,j,3)=0.

188

10

do 10 m=1,3
do 10 i=1,nphase
do 10 j=1,nphase
if(real(sigma(i,m)).eq.0.0.and.aimag(sigma(i,m)).eq.0.0) then
be(i,j,m)=cmplx(0.0,0.0)
goto 10
end if
if(real(sigma(j,m)).eq.0.0.and.aimag(sigma(j,m)).eq.0.0) then
be(i,j,m)=cmplx(0.0,0.0)
goto 10
end if
be(i,j,m)=1./(0.5/sigma(i,m)+0.5/sigma(j,m))
continue

c
c
c
c

Trim off x and y faces so that no current can flow past periodic
boundaries. This step is not really necessary, as the voltages on the
periodic boundaries will be matched to the corresponding real voltages
in each conjugate gradient step.
do 20 k=1,nz2
do 15 j=1,ny2
gx((k-1)*L22+nx2*(j-1)+nx2)=cmplx(0.0,0.0)
15
continue
do 16 i=1,nx2
gy((k-1)*L22+ny1*nx2+i)=cmplx(0.0,0.0)
16
continue
20
continue
c
c

Set up conductor network
bulk--gz
do 30 i=1,nx2
do 30 j=1,ny2
do 30 k=1,nz1
m=(k-1)*L22+(j-1)*nx2+i
i1=i
j1=j
k1=k+1
m1=(k1-1)*L22+(j1-1)*nx2+i1
gz(m)=be(pix(m),pix(m1),3)
30
continue
c

bulk---gy
do 40 i=1,nx2
do 40 j=1,ny1
do 40 k=2,nz1
m=(k-1)*L22+(j-1)*nx2+i

189

40

j1=j+1
i1=i
k1=k
m1=(k1-1)*L22+(j1-1)*nx2+i1
gy(m)=be(pix(m),pix(m1),2)
continue

c

bulk--gx
do 50 i=1,nx1
do 50 j=1,ny2
do 50 k=2,nz1
m=(k-1)*L22+(j-1)*nx2+i
i1=i+1
j1=j
k1=k
m1=(k1-1)*L22+(j1-1)*nx2+i1
gx(m)=be(pix(m),pix(m1),1)
50
continue
return
end
c
c

Subroutine that sets up the image, either by reading it from file,
or generating it internally
subroutine ppixel(pix,nx2,ny2,nz2,a,ns2,nphase,ntot)
real a(ntot)
integer*2 pix(ns2)

c
c

(USER) If you want to set up a test image inside the program, instead
of reading it in from a file, this should be done inside this subroutine.

c

auxiliary variables involving the system size
nx=nx2-2
ny=ny2-2
nz=nz2-2
L22=nx2*ny2
c Initialize phase fraction array.
do 120 i=1,nphase
a(i)=0.0
120
continue
c Use 1-d labelling scheme as shown in manual
do 100 k=2,nz2-1
do 100 j=2,ny2-1
do 100 i=2,nx2-1

190

100

220
c

m=(k-1)*L22+(j-1)*nx2+i
read(9,*) pix(m)
a(pix(m))=a(pix(m))+1.0
continue
do 220 i=1,nphase
a(i)=a(i)/float(nx*ny*nz)
continue
now map periodic boundaries of pix (see Section 3.3, Figure 3 in manual)
do 110 k=1,nz2
do 110 j=1,ny2
do 110 i=1,nx2
if(i.gt.1.and.i.lt.nx2) then
if(j.gt.1.and.j.lt.ny2) then
if(k.gt.1.and.k.lt.nz2) then
goto 110
end if
end if
end if
k1=k
if(k.eq.1) k1=k+nz
if(k.eq.nz2) k1=k-nz
j1=j
if(j.eq.1) j1=j+ny
if(j.eq.ny2) j1=j-ny
i1=i
if(i.eq.1) i1=i+nx
if(i.eq.nx2) i1=i-nx
m=(k-1)*L22+(j-1)*nx2+i
m1=(k1-1)*L22+(j1-1)*nx2+i1

110

pix(m)=pix(m1)
continue

c

Check for wrong phase labels--less than 1 or greater than nphase
do 500 m=1,ns2
if(pix(m).lt.1) then
write(7,*) 'Phase label in pix < 1--error at ',m
end if
if(pix(m).gt.nphase) then
write(7,*) 'Phase label in pix > nphase--error at ',m
end if
500
continue

191

return
end
c

Subroutine to compute the total current in the x, y, and z directions
subroutine current(nx2,ny2,nz2,ns2,currx,curry,currz,u,gx,gy,gz)
complex u(ns2),gx(ns2),gy(ns2),gz(ns2),currx,curry,currz
complex cur1,cur2,cur3

c

c

auxiliary variables involving the system size
nx=nx2-2
ny=ny2-2
nz=nz2-2
L22=nx2*ny2
initialize the volume averaged currents
currx=cmplx(0.0,0.0)
curry=cmplx(0.0,0.0)
currz=cmplx(0.0,0.0)

c

Only loop over real sites and bonds in order to get true total current
do 10 k=2,nz2-1
do 10 j=2,ny2-1
do 10 i=2,nx2-1
m=L22*(k-1)+nx2*(j-1)+i
c cur1, cur2, and cur3 are the currents in one pixel
cur1=0.5*( ( u(m-1)-u(m) )*gx(m-1)+( u(m)-u(m+1) )*gx(m) )
cur2=0.5*( ( u(m-nx2)-u(m) )*gy(m-nx2)+( u(m)-u(m+nx2) )*gy(m) )
cur3=0.5*( ( u(m-L22)-u(m) )*gz(m-L22)+( u(m)-u(m+L22) )*gz(m) )
c sum pixel currents into volume averages
currx=currx+cur1
curry=curry+cur2
currz=currz+cur3
10
continue
currx=currx/float(nx*ny*nz)
curry=curry/float(nx*ny*nz)
currz=currz/float(nx*ny*nz)
return
end

192

9.3.6 GAUSS.F
c***********************

gauss.f

***********************************

real x(4000),w(4000),xi(4000),wi(4000)
open(unit=62,file='gauss20.dat')
data pi/3.14159 26535 89793 23846 26434/
c***********************************************************************
c
This program (gauss.f) evaluates zeros and weights of Legendre
*
c
polynomials to be used in Gaussian integrals.
*
c
Pick N to be the order of the Gaussian quadrature that you
*
c
wish to use. File output will contain Gaussian points in the
*
c
first column, weights in the second column.
*
c
The limit on N is currently 4000. If a larger value of N is
*
c
desired (not recommended), then change the dimensions of
*
c
x,w,xi, and wi.
*
c
This program was supplied by Professor W.W. Repko, of the
*
c
Michigan State University Physics and Astronomy Department (1983). *
c***********************************************************************
c

Number N of Gaussian points desired.
N=20
M=(N+1)/2

c

Subroutine grule calculates weights and points
call grule(N,x,w)

10
c

3
c

do 10 i=1,M
xi(i)=-x(i)
xi(i+M)=x(M+1-i)
wi(i)=w(i)
wi(i+M)=w(M+1-i)
continue
output weights and points
xc=0.0
do 20 i=1,n
write(62,3) xi(i),wi(i)
format(' ',2f20.12)
test with integral from -1 to 1 of exp(x), should be e - 1/e = 2.3504024

193

20

20

30

10

10

xc=xc+wi(i)*exp(xi(i))
continue
print *,' Numerical value of integral of exp(x) from -1 to 1 = '
print *,xc
print *,' Actual value is 2.3504024'
end
subroutine grule(N,x,w)
real x(N),w(N)
data pi/3.14159 26535 89793 23846 26434/
data eps/1.e-14/
dn=float(N)
M=(N+1)/2
e1=dn*(dn+1.0)
do 10 i=1,M
t=(4.0*float(i)-1.0)*pi/(4.0*dn+2.0)
x0=(1.0-(1.0-1.0/dn)/(8.0*dn*dn))*cos(t)
call legendr(N,x0,pn,pnm1,pnp1)
den=1.0-x0*x0
d1=dn*(pnm1-x0*pn)
dpn=d1/den
d2pn=(2.0*x0*dpn-e1*pn)/den
u=pn/dpn
v=d2pn/dpn
x1=x0-u*(1.0+0.50*u*v)
if((x1-x0).lt.eps) go to 30
x0=x1
go to 20
x0=x1
call legendr(N,x0,pn,pnm1,pnp1)
x(i)=x0
w(i)=2.0*(1.0-x0**2)/(dn*pnm1)**2
if(M+M.gt.N) x(M)=0.0
return
end
subroutine legendr(N,x,pn,pnm1,pnp1)
pkm1=1.0
pk=x
do 10 k=2,N
t1=x*pk
pkp1=t1-pkm1-(t1-pkm1)/float(k)+t1
pkm1=pk
pk=pkp1
pn=pk
pnm1=pkm1

194

t1=x*pn
pnp1=t1-pnm1-(t1-pnm1)/float(k+1)+t1
return
end

195

9.3.7 BURN3D.F
c**************************

burn3d.f

*******************************

c

PROBLEM DEFINITION

c
c
c
c
c
c
c
c

In a random multi-phase structure, a question that is important
is whether a particular phase percolates through the microstructure
or not. This program is designed to answer that question, for
a general 3-D multi-phase random microstructure. The burning
algorithm checks whether a percolation threshold exists in a periodic
image. The program searches all three directions, using periodic
boundary conditions in the two perpendicular directions for
each burn.

c

Variables

c
c
c
c
c
c
c
c
c

There are a maximum of "ntot" phases possible, numbered 1,2,3,...
The label of the phase being burned is "phase", and is an input variable.
The value assigned to burned pixels is "burned" and is equal to ntot+1.
To burn on more than one phase at a time, just use
"phase2", "phase3", etc., and check for these values too, whenever
the value "phase" is checked for. (See manual)
A value of 0 is assigned to the variables percx, percy, and
percz for non-continuity, and 1 for percolation (continuity) in the
given direction.

c

Dimensions

c
c
c
c
c
c

(USER) The variable pix is dimensioned the size of the system.
The variables old and new are dimensioned 1/10 the size of the system,
but with three components. (A minimum of 1000 should be used
to dimension these variables, so that small systems will have enough
computation room.) These array dimensions should be changed
simultaneously in all subroutines using a global replacement statement.

c

(USER) Dimensions of main arrays
integer*2 pix(1000000),old(100000,3),new(100000,3)
integer*2 burned,phase

c

(USER) Unit = 9 is the input file, unit 7 is the output file
open(unit=9,file='microstructure.dat')
open(unit=7,file='output.out')

c

(USER) system size ns = nx x ny x nz

196

nx=100
ny=100
nz=100
ns=nx*ny*nz
c
c
c

(USER) Identify the label of the phase to be burned
phase=1
(USER) Total number of phases allowed in problem
ntot=100
Label of burned pixel
burned=ntot+1

c

Read in microstructure file
do 330 k=1,nz
do 330 j=1,ny
do 330 i=1,nx
m=nx*ny*(k-1)+nx*(j-1)+i
read(9,*) pix(m)
330
continue
c
c
c

c
c
c
c

Call the subroutine that actually does the burning
call fire(pix,nx,ny,nz,percz,percy,percx,new,old,phase,burned)
Output the values of perc*, showing the continuity of the three
principal directions
write(7,*) ' percx = ',percx
write(7,*) ' percy = ',percy
write(7,*) ' percz = ',percz
end
This subroutine does the actual burning. The burning starts with the
k=1 or j=1 or i=1 plane, and then iteratively burnes through the system,
until there are no more acessible pixels to be burned. To be burned,
a pixel must be next to a previously burned pixel.
subroutine fire(pix,nx,ny,nz,percz,percy,percx,new,old,
& phase,burned)
integer*2 pix(1000000)
integer*2 old(100000,3),new(100000,3)
integer*2 in(6),jn(6),kn(6),phase,burned

c

System size
ns=nx*ny*nz

c

Direction labels to check for burning path (nearest neighbor information

197

c

in 3-D digital system).
in(1)=-1
in(2)=1
in(3)=0
in(4)=0
in(5)=0
in(6)=0
jn(1)=0
jn(2)=0
jn(3)=-1
jn(4)=1
jn(5)=0
jn(6)=0
kn(1)=0
kn(2)=0
kn(3)=0
kn(4)=0
kn(5)=-1
kn(6)=1

c

Initialize percolation flags
percx=0.0
percy=0.0
percz=0.0

c
c
c

do 3000 ijk=1,3
Build up first burned pixels from i or j or k=1 layer, according to
choice of ijk (ijk = 1, k = 1; ijk = 2, j = 1; ijk = 3, i = 1).
Store(i,j,k) labels in array old().
iold=0
if(ijk.eq.1) then
n2=ny
n1=nx
end if
if(ijk.eq.2) then
n2=nz
n1=nx
end if
if(ijk.eq.3) then
n2=ny
n1=nz
end if

198

do 1000 jj=1,n2
do 1000 ii=1,n1
if(ijk.eq.1) then
i=ii
j=jj
k=1
end if
if(ijk.eq.2) then
i=ii
j=1
k=jj
end if
if(ijk.eq.3) then
i=1
j=jj
k=ii
end if
m=nx*ny*(k-1)+nx*(j-1)+i

1000
c
c

if(pix(m).eq.phase) then
pix(m)=burned
iold=iold+1
old(iold,1)=i
old(iold,2)=j
old(iold,3)=k
end if
continue

If no pixels burned in first layer, then phase can't possibly percolate,
so move to next direction
if(iold.eq.0) goto 3000

c Now start building up new burned pixels from old set of burned pixels,
c thus propagating the fire.
60
inew=0
do 100 i=1,iold
ii=old(i,1)
jj=old(i,2)
kk=old(i,3)
c check all six nearest neighbors of previously burned pixel
do 90 n=1,6
i1=ii+in(n)
j1=jj+jn(n)
k1=kk+kn(n)

199

c
c
c
c
c
c

Periodic boundary conditions
(USER) Can replace with hard boundary conditions
if desired to remove periodicity. Keep the ijk if statements and the
goto 90 statements, and change the other if statements to have
goto 90 as well. That way the program does not allow "wrappping"
around to find a neighbor. (See manual)
if(ijk.eq.1) then
if(k1.lt.1.or.k1.gt.nz) goto 90
if(i1.lt.1) i1=i1+nx
if(i1.gt.nx) i1=i1-nx
if(j1.lt.1) j1=j1+ny
if(j1.gt.ny) j1=j1-ny
end if
if(ijk.eq.2) then
if(j1.lt.1.or.j1.gt.ny) goto 90
if(i1.lt.1) i1=i1+nx
if(i1.gt.nx) i1=i1-nx
if(k1.lt.1) k1=k1+nz
if(k1.gt.nz) k1=k1-nz
end if
if(ijk.eq.3) then
if(i1.lt.1.or.i1.gt.nx) goto 90
if(k1.lt.1) k1=k1+nz
if(k1.gt.nz) k1=k1-nz
if(j1.lt.1) j1=j1+ny
if(j1.gt.ny) j1=j1-ny
end if

c

Store (i,j,k) labels of newly burned pixels in array new().
m1=nx*ny*(k1-1)+nx*(j1-1)+i1
if(pix(m1).eq.phase) then
pix(m1)=burned
inew=inew+1
new(inew,1)=i1
new(inew,2)=j1
new(inew,3)=k1
end if
90
continue
100
continue
c
c

If new pixels were burned, then transfer labels to old() array, start
burning process over again.
if(inew.gt.0) then
iold=inew
do 150 i=1,inew

200

150

c
c
c

old(i,1)=new(i,1)
old(i,2)=new(i,2)
old(i,3)=new(i,3)
continue
goto 60
end if
If no new burned pixels, then check to see if the last layer of the image
has any burned pixels in it. If so, then there is continuity. If not,
then there is no continuity.
if(ijk.eq.1) then
n2=ny
n1=nx
end if
if(ijk.eq.2) then
n2=nz
n1=nx
end if
if(ijk.eq.3) then
n2=ny
n1=nz
end if
do 30 jj=1,n2
do 30 ii=1,n1
if(ijk.eq.1) then
i=ii
j=jj
k=nz
end if
if(ijk.eq.2) then
i=ii
j=ny
k=jj
end if
if(ijk.eq.3) then
i=nx
j=jj
k=ii
end if
m=nx*ny*(k-1)+nx*(j-1)+i
if(pix(m).eq.burned) then
if(ijk.eq.1) percz=1.0
if(ijk.eq.2) percy=1.0

201

if(ijk.eq.3) percx=1.0
end if
continue

30
c

Restore burned pixels back to their original label
call restore(pix,ns,phase,burned)

3000

continue
return
end

c This subroutine restores the burned pixels back to their original, unburned
c value (phase).
subroutine restore(pix,ns,phase,burned)
integer*2 pix(1000000),phase,burned

10

do 10 m=1,ns
if(pix(m).eq.burned) pix(m)=phase
continue
return
end

202



---

Source Exif Data:

File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.2
Linearized                      : No
Page Count                      : 210
Create Date                     : 1998:12:16 11:07:46
Producer                        : Acrobat Distiller Command 3.01 for Solaris 2.3 and later (SPARC)
Creator                         : dvipsk 5.58f Copyright 1986, 1994 Radical Eye Software
Title                           : ir_cover.dvi

EXIF Metadata provided by EXIF.tools