Professional Documents
Culture Documents
Heat
Transfer
Volume 1
MaShematScal Modslling
A.A. SAMARSKll
P.N. VABISHCHEVICH
Computational Heat
Transfer
Volume 1 Mathematical Modelling
A. A. Samarskii
P. N. Vabishchevich
Russian Academy of Sciences, Moscow
John Wiley & Sons (SEA) Pte Ltd, 37 Jalan Pemimpin *05-04.
Block B, Union Industrial Building. Singapore 2057
A catalogue record for this book is available fmm the British Library
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Mathematical Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Use of Computers for Mathematical Modelling . . . . . . . . . . . . . . . 3
Computational Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Numerical Modelling of Heat Transfer Processes . . . . . . . . . . . . . . 12
M a t h e m a t i c a l M o d e l s of Physics of H e a t . . . . . . . . . . . . . . . 15
Heat Conduction in Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Closing Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
PhaseTransitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Convective Heat Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Thermal Radiation of Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Thermoelasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Bibliography and Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Analytical M e t h o d s of H e a t Transfer . . . . . . . . . . . . . . . . . . 57
Dimensionless Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Analytical Solution of Linear Problems. . . . . . . . . . . . . . . . . . . . . . 62
Exact Solutions of Nonlinear Problems . . . . . . . . . . . . . . . . . . . . . . 69
Asymptotic Methods of Heat Transfer . . . . . . . . . . . . . . . . . . . . . . 75
Bibliography and Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
S t a t i o n a r y P r o b l e m s of H e a t Transfer . . . . . . . . . . . . . . . . . .
Boundary Value Problems for Second-order Elliptic
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Construction of Difference Schemes . . . . . . . . . . . . . . . . . . . . . . . . .
Uniform Convergence of Difference Schemes . . . . . . . . . . . . . . . . . .
Convergence of Difference Schemes in an Energetic Space . . . . .
Direct Methods for Solving Grid Equations . . . . . . . . . . . . . . . . . .
Iterative Methods of Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . .
Iterative Methods for Solving Grid Equations . . . . . . . . . . . . . . . .
CONTENTS
N o n s t a t i o n a r y P r o b l e m s of H e a t Trasfer . . . . . . . . . . . . . . .
Boundary Value Problems for Second-Order Parabolic
Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Difference Schemes for Nonstationary Problems . . . . . . . . . . . . . .
Uniform Convergence of Difference Schemes for the Heat
Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Theory of Stability of Difference Schemes . . . . . . . . . . . . . . . . . . .
Stability and Convergence of Difference Schemes for the Heat
Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Asymptotical Stability of Difference Schemes for the Heat
Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Hyperbolic Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Regularization of Difference Schemes . . . . . . . . . . . . . . . . . . . . . . .
Nonlinear Nonstationary Problems . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography and Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
H e a t C o n d u c t i o n P r o b l e m s w i t h P h a s e Transi-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Variable Domain Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
Fixed Domain Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
Transformation of Dependent Variables . . . . . . . . . . . . . . . . . . . . . 370
Quasi-Stationary Stefan Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 377
Modelling of Phase Transitions in Binary Alloys . . . . . . . . . . . . . . 388
Bibliography and Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
Dedicated to the memory of
Nikolai Nikolaevich Govorun
Preface
Chapters 4-6 comprise the core of the book. They deal with the solution of
classical heat transfer problems by finite difference methods.
In Chapter 4 we describe the numerical solution of stationary problems.
We first recall some fundamental results of the theory of partial differential
equations. In particular, we formulate the classical maximum principle for
stationary equations, which has a clear thermophysical interpretation. We
construct difference equations so that they preserve the main properties of
a differential problem. We discuss the construction of difference schemes for
stationary problems and describe both direct and iterative methods for finding
an approximate solution.
The same topics are covered in Chapters 5 for nonstationary problems. We
pay particular attention to general stability theory for difference schemes.
In order to examine particular schemes, we verify whether necessary and
sufficient conditions, formulated as simple operator inequalities, are satisfied.
Economical difference schemes (schemes of alternating directions, locally one-
dimensional schemes, etc.) are considered in Chapter 6.
Chapters 7-10 deal with numerical simulation of specific problems of heat
exchange. The presentation is based on the difference methods studied for
solving stationary and nonstationary problems in heat transfer. In Chapter 7
we study heat transfer problems with phase transitions involving a solid
phase. We consider the classical Stefan problem and discuss modelling of
melting/solidification of binary alloys.
In the second volun~eof this book, starting with Chapter 8, solution
techniques for complicated heat transfer problems are presented. In Chapter 8
we investigate special features of heat radiation from the surface of a solid.
For example, when modelling heat exchange in a concave body, it is necessary
to take into account radiation from separate parts of the boundary.
Convective heat transfer problems (Chapter 9) comprise an important class
of heat exchange problems. For these problems we introduce general families
of monotonic difference schemes, for which the maximum principle holds.
The heat/mass transfer is considered in the Boussinesq approximation. We
discuss the solution of these problems both in the natural variables 'velocity-
pressure' and in the variables 'vortex-stream function'. We also suggest the
boundary layer approximation. In Chapter 10 numerical methods are used
to solve problems in thermoelasticity. For example, the thermoelastic state
of a solid with a rectangular cross-section and that of a thin plate are
simulated.
Thermal process control problems are covered in Chapter 11. We briefly
discuss gradient methods, which are the base for numerical solution of the
control problems. We distinguish control problems for steady- and unsteady-
state sources of heat and for various quality criteria (functionals). Problems
of controlling the boundary regime constitute another important class of
applications.
xii PREFACE
The main classes of inverse problems are presented in Chapter 12. The
heat conduction problems are solved approximately for the cases in which
initial values, boundary regimes or coefficients in equations are recovered.
Stability of the solution is studied by perturbation of either original differential
(quasi-inversion methods) or difference (regularization of difference schemes)
equations.
Chapter 13 deals with practice of numerical modelling by the methods
presented. We take actual two-dimensional problems of heat transfer as
examples. First, we pose a problem and separate the dimensionless parameters
of that problem. Then we construct in detail numerical algorithms and give
FORTRAN codes. Finally, we describe computational results and present
program source listings.
Throughout the book, except for the sections Bibliography and Comments,
there are no references to other works. We used more space than usual to make
the presentation complete. As a result, the book became more convenient for
study.
No systematic analysis of existing literature on the subject is attempted.
We only referred to the work that we used in the book and that can be used
by the reader as well to study certain topics more carefully. There are several
reasons for not presenting references to original work. In any case, the choice
of reference literature from the vast variety of publications would reflect the
authors' preferences, demand some research of priority, etc. In particular, the
extended list of references would inevitably include more work by Russian
authors.
We worked on this book in the stimulating and creative atmosphere of
the Chair of Computational Methods at the Department of Computational
Mathematics and Cybernetics at the M. V. Lomonosov Moscow State
University and the Institute of Mathematical Modelling of the Russian
Academy of Sciences. We are grateful to our colleagues and disciples for
their assistance at different stages of the work. The authors are particularly
indebted to M. M. Makarov for his help in preparing the FORTRAN codes
listed in the book.
In view of the complexity of the work undertaken and inevitable omissions,
the authors will appreciate all criticism and suggestions concerning the book.
Aleksander 4 . Samarskii
Petr N. Vabishchevich
Introduction
models are used for qualitative analysis. In this case, we speak of model (or
simplified) problems for the main mathematical problem (models for a model).
When qualitatively investigating a mathematical problem, much attention
is paid to how well it is posed. First of all, the existence of a solution is studied.
The corresponding strict results (existence theorems) ensure that the model
is correct. Besides, constructive proofs of the existence theorem for a problem
can often be used as a basis for approximate methods for solving the problem
itself.
Stability of solution with respect to small perturbations in initial data
is important in applied mathematical modelling. Inverse problems are of-
ten unstable, and this should be taken into account when constructing
approximate solutions.
Nonlinear mathematical models can have multiple solutions. Qualitative
analysis of mathematical models includes the study of branching points,
bifurcations of solutions, problems of how to single out the desired solu-
tion, etc.
Methods of qualitative analysis for different types of mathematical models
have been developed to different extents. For example, qualitative methods led
to the most impressive results for ordinary differential equations. Qualitative
methods are also employed in the theory of partial differential equations,
although to a smaller extent. An important example we should mention is the
maximum principle for second-order parabolic and elliptic equations, which is
related to mathematical models of heat transfer. This principle can be used
to qualitatively investigate mathematical models based on partial differential
equations.
An exact or approximate solution is found by analytical or numerical
methods. Classical examples of analytical methods include the method of
separation of variables and the method of integral relations for linear problems
in mathematical physics.
Methods of linearization and different versions of perturbation methods
are significant for nonlinear mathematical models. This approach is based
on asymptotic expansions with respect to a certain small parameter. Much
consideration is given to these methods when studying singularly perturbed
problems.
Qualitative behaviour of a nonlinear problem can be well represented by
some of its specific solutions. Specific solutions are sought using automodel
variables and the results of group analysis applied to equations that underlie
the mathematical model.
Complex nonlinear multiparameter models can be investigated on a com-
puter by numerical methods. Unlike an analytical solution, which provides
explicit dependence on some initial data of the problem, numerical methods
require repetitive solution of the problem for every value of a parameter.
However, an analytical solution can rarely be found for nonlinear problems,
INTRODUCTION 5
rH-@
TI
computations
experiment
mathematical
model
I
under study
investigation
of the model
Fig. 1.1.
Let us use the simplest model of heat transfer, namely the onedimensional
heat equation, to specify types of computational experiment and describe
main classes of problems of applied modelling.
Let us consider a thin cylindrical rod with thermally insulated lattice
surface. We assume that one end is thermally insulated, while a specific
thermal regime is specified at the other. The temperature T(x, t) is determined
by the heat equation
the inverse problems this implies that we can only restore the boundary
regime with a large error. Thus, in this aspect, optimization and inverse
problems are essentially different. These differences should he emphasized
because identical mathematical formulations of these problems can lead to
wrong methodological views.
1 Mathematical Models of Physics
of Heat
We study temperature field and heat transfer from some parts of a rigid body
to its other parts. The temperature field at a time moment t is specified by
the temperature distribution in the body, i.e. by a function T = T(x, y,z, t),
where (x, y, z ) are Cartesian coordinates. In the simplest case heat flow is
directed from the higher-temperature region to a region in which temperature
is lower.
There are three main types of heat transfer, namely
(a) heat conduction, that is heat transfer due to interaction between micro-
particles of bodies that are in contact;
(b) convection, that is heat transfer, caused by movement of the substance.
This type of heat transfer is observed in moving media (fluids or gases);
(c) mdantaon, that is a process in which energy is transmitted as electromag-
netic waves.
In numerous applied problems heat is transferred by two or three of
these mechanisms concurrently (complex heat transfer). When describing
convection it is necessary to take into account heat conduction between
different parts of a continuous medium (heat and mass transfer). Radiation
can be accompanied by heat conduction, convection, etc. Such complicated
processes occur, for instance, in phase transitions or chemical reactions. In
the case of heat conduction in solids, the thermal effects can be significant.
In this chapter we present reference material and give basic mathematical
models of heat transfer. The main heat equation is written in several
orthogonal coordinate systems. Much attention is paid to boundary conditions
and conditions at the contact between different media (junction conditions).
16 COMPUTATIONAL HEAT TRANSFER
The coefficients and the right-hand side in this equation can vary in space
(inhomogeneous medium). In this case we have c = c ( x ,y, z ) , p = p(x, y, 2 ) .
MATHEMATICAL MODELS OF PHYSICS OF HEAT 17
d
-=
a + v grad,
-
dt at
where v is the local velocity of the medium. Therefore, the heat equation for
a moving medium is written as
g
a a a
-
ax' - -1,
a y ' az
aa,
diva=-+-+-
ax
aa,
ay
aa,
az
Hence, (2) can be rewritten as
18 COMPUTATIONAL HEAT TRANSFER
Other equations are rewritten similarly. For example, to rewrite the Laplace
equation (3) we should take into account that the Laplacim operator A in
the Cartesian coordinates looks like
In the above (see (6)) we wrote the heat equation in the Cartesian
coordinates, for which X I = x, xz = y , x3 = 2, gl = 1, g* = 1, and g3 = 1.
For a cylindrical coordinate system we obtain X I = T , xz = (o, x3 = t , g~ = 1,
gz = T , and g3 = 1. According to (8) and (9), the heat equation (2) becomes
+--6--)+f.alp
T
1
sin o sip
1
sm 0 (13)
Formulae (10) and (11) can also be used to obtain the corresponding
expressions for the curl and Laplacian operators in cylindrical and spherical
coordinates.
+-
a ( k,,
a~ E+
-
aT
k,, - k,,
ay
+ ") +f (15)
instead of (6)
b-,,hkq COMPUTATIONAL HEAT TRANSFER
This equation contains mixed derivatives; we can get rid of them by passing t o
a new Cartesian coordinate system (x',y', z') in which (16) has the simplest
form, namely
The coordinate axes (z',y',zi) are said to be the principal axes of heat
conduction, while the coefficients k,, , k,,, and k,, are said to be the principal
heat conductivities.
where 7, = const is the relaxation time of heat stresses. In order to derive the
heat equation, we substitute (18) into (1) to obtain
2.1.5 PROBLEMS
1. Write the heat equation in an orthogonal coordinate system
for a moving homogeneous medium.
MATHEMATICAL MODELS OF PHYSICS O F HEAT 21
This equation no longer contains terms with first derivatives with respect to
space variables.
22 COMPUTATIONAL HEAT TRANSFER
aT
cp - = div(kgradT)
at
+f, (x, y, z, t ) E Q. (1)
This equation involves partial derivatives with respect to both space and
time variables. Therefore, additional conditions have to be specified on some
sets of the spatial domain fl and the time interval (0, t,,,), i.e. on some sets
of the points from cylinder Q.
Boundary value problems are usually posed for the heat equation. In this
case, additional conditions are specified on the boundary of Q. Conditions
on the lateral surface of the cylinder Q correspond to those with respect to
space variables (on the boundary of the spatial domain), hence they are called
boundary conditions. Conditions on the bottom base of Q are said to be initial
conditions.
More complicated conditions can be specified as well. For instance, instead
of initial conditions for t = 0, we can specify additional conditions on another
section of the cylinder Q, i.e. for some t = t * . In other words, the set of points
on which additional conditions are given can also lie inside Q. Some of these
possibilities are discussed in the following, where we specify the main types
of problems for the heat equation.
It is generally assumed that the temperature field is specified at the initial
moment in time, i.e.
MATHEMATICAL MODELS O F PHYSICS O F HEAT 23
where a is the heat transfer coeficient. The appropriate condition in the case
of anisotropic media [see (a)] is formulated similarly.
The third-kind boundary conditions can be treated as the most general of
the above conditions. These conditions can be written as
where the brackets denote the jump at the interface. In our case [TI =
+
T(x 0, y, z ) - T(x - 0, y, 2). Now we only have to formulate the junction
conditions for the first derivatives of temperature.
In order to derive realistic (natural) junction conditions for the heat
equation (11) with discontinuous coefficients (12) we can separate a bounded
region 6s on the interface and integrate the original heat equation (11) over
the domainof width ZE, namely 0, = {(x, y,z)1 - E < x < E , (y,z) E 65). We
take account of the continuity of temperature (condition (13)) and finiteness
of discontinuities as E -+ 0 and obtain
Since the element 6S is arbitrary, we find the junction condition in the form
The junction conditions (15) and (16) are the conditions of ideal contact. We
emphasize once again that (15) and (16) are natural conditions for (11) with
discontinuous coefficients (12). This implies that they are embedded into the
26 COMPUTATIONAL HEAT TRANSFER
J
n
asP(x, y, z) dxdy dz =
J
s
P(x, y, t)ds
where cs is the lumped heat capacity of the contact. The junction conditions
(15) and (20), in which the surface delta-function is used, are included in the
heat equation similar to (19).
MATHEMATICAL MODELS O F PHYSICS OF HEAT 27
where the coefficient of contact heat transfer a is associated with the contact
conditions.
Optimal control problems for the heat equation are formulated in the same
way. Minimization problem (25), (26) on the set of restrictions (I), (2), (22),
(24) is interpreted in this case as follows. Find a boundary thermal regime
(24) (optimal regime) that would provide necessary quality (functional (26)).
We thus separated three main classes of problems for the heat equation.
The first includes standard boundary value problems, i.e. direct problems of
heat transfer. The second class of applied problems includes inverse problems
of heat transfer. The third class includes optimization and optimal control
problems.
2.2.5 PROBLEMS
1. S u p p o s e we have a n isotropic medium, in which t h e r e is
a homogeneous inclusion with t h e b o u n d a r y S. Let t h e h e a t
conductivity of t h e inclusion b e m u c h greater t h a n t h a t of t h e
medium. T h e inclusion supplies t h e m e d i u m w i t h a n a m o u n t of h e a t
Qs(t). Consider t h e heat transfer in t h e m e d i u m a n d pose necessary
b o u n d a r y conditions o n t h e b o u n d a r y of t h e inclusion.
Solution. Since the heat conductivity of the inclusion is much greater than
that of the medium, we may assume that the temperature on the boundary
S is constant, i.e.
Conditions (27) and (28) are desired nonlocal conditions on the boundary S
of the inclusion.
30 COMPUTATIONAL HEAT TRANSFER
We can take into account the heat capacity of the inclusion to refine
condition (28)
Let us write the corresponding heat equation. In the solid phase we have
where Q- = { ( x ,y, z, t)l ( x ,y, z ) E R-, 0 < t < t,.,). In the liquid phase we
additionally take into account convective heat transfer to get
. .
We are interested in boundary conditions on the interface S of the phase
transition. Firstly, the continuity of temperature is assumed on this boundary
of contact between two media, i.e.
(X,Y,~)ES. (3)
The phase transition is accompanied by a release/absorption of a certain
amount of heat. Therefore the heat flow is discontinuous on the phase
transition boundary and is defined by the equation
Here L is the phase transition enthalpy and V. is the normal velocity of the
phase transition boundary.
As mentioned above, we assume that the phase transition occurs at a
constant temperature T'.Hence the interface is determined a t each moment
in time as follows: S = S ( t ) = { ( x ,y , t) E 0,T ( x ,y, z , t ) = T ' ) or, in other
words, the first-kind conditions
Fig. 2.1.
32 COMPUTATIONAL HEAT TRANSFER
Conditions (1)-(3) are called the Stefan conditions and the corresponding
problem for equations (1) and (2) is referred to as the Stefan problem. In this
problem, processes in both phases are studied; and the problem is also called
the two-phase Stefan problem. The extreme case of this problem is when the
temperature field in one of the phases is known (the temperature equals the
phase transition temperature). We thus should find the temperature field only
for one phase, i.e. we deal with the monophase Stefan problem. In this case,
the unknown phase transition boundary S is external rather than internal.
For example, assume that the domain 0- is filled with the solid phase at
temperature T*. Then in order to find the temperature in the liquid phase,
we use equation (2) in a variable domain R+(t) supplied with the following
conditions on S:
Conditions like (6) and (7) are typical for the monophase Stefan problem.
In some cases the Stefan approximation is inappropriate. Various improve-
ments of mathematical models for phase transitions are actively discus-
sed in the literature. Without going into details, we mention the main
point. The Stefan condition in (5) is based on the assumption that the
temperature instantly levels to the phase transition temperature, which
actually corresponds to the assumption that the velocity of the phase
transition is unbounded. This assumption is not valid in some cases, e.g. in
the problem of describing high-intensity thermal processes (the hyperbolic
heat equation). To avoid this, we can use a more general third-kind condition
instead of the first-kind condition in (5). To derive this condition we should
describe the kinetics of the phase transition. For example, the condition
I
I aTt
k+ -+ a* (T" - T*)= 0, S(t),
(2,y, z ) E (8)
an
which bounds heat flows toward the phase transition boundary, can be used
instead of (6) in the monophase problem. Condition (8) is used along with
( 7 ) , which expresses the law of conservation of energy for any motion of the
interface.
interface in the equation itself (see (19) in Section 2.2). The Stefan problem
is complicated because the unknown phase interface S itself is unknown.
Let us explain how we can pass from equations (1) and (2) equipped with
conditions (3)-(5) to a single heat equation. According to (18) and (19) in
Section 2.2, we can rewrite (1) and (2) as the single equation
cp (g+ v grad T
1 = div(k grad T ) - 6sLV. + f, ( x y z t) E Q. (9)
0 C' I C
Evidently, equation (13) yields equation (11) for a pure substance, where
we have
and T' = Tli, = T.,,. Equation (13) is the basis for the simplest quasi-
equilibrium model the two-phase zone.
In special cases, the width of the two-phase zone may turn out to be much
smaller than a typical linear dimension in the problem. Then we can restrict
our analysis to the simplest assumption of a given temperature of the phase
transition for every concentration (T* = Tli, = T.,I):
C- = k c f , (x, y, z) E S, (17)
The junction conditions (17) and (18) complete diffusion equations (15) and
(16). Of course, we can set different conditions on the boundary of the
computational domain, but these will not he discussed here.
2.3.5 PROBLEMS
1. Formulate conditions for t h e velocity in t h e liquid phase o n t h e
phase transition b o u n d a r y taking into account t h e density j u m p .
Solution. If the density is continuous on the phase transition boundary, we
obtain a natural condition v = 0, (x, y,z) E S , for the velocity in the liquid
phase.
Changes in density (the density typically decreases) when passing from the
solid into the liquid phase lead to a nonzero velocity along the normal to the
phase transition boundary (expansion of substance). At the same time, the
tangent component remains zero, i.e.
This implies
u,, = v n = (I - 5)vn, (x,g,z) E S. (20)
Conditions (19) and (20) completely define the velocity of the liquid phase on
the phase transition boundary.
2. I n t h e quasi-stationary Stefan problem for a homogeneous
medium, transform t h e heat equation s o as t o g e t rid of convective
t e r m s ( t e r m s including directional derivatives along 1).
Solution. The stationary heat equation
where r = xi + yj + zk. Substitution of (22) and (23) into (21) yields the
desired equation
div(RgradT) +R -f = 0, (x, y, z) E RC U 0-.
acp
Thus, we have obtained a self-adjoint equation in each separate subdomain.
2.4 C o n v e c t i v e Heat T r a n s f e r
""
In many cases (if medium motions are slow as compared with the speed of
sound) the density of the fluid can be assumed to be constant for the whole
medium a t any moment in time (the model of an incompressible fluid). In this
case equation (1) becomes
divv = 0. (2)
We use the notation v = (u,v, w) for the velocity components. Equations
of motion for a viscous medium (the Nauier-Stokes equations) are written as
38 COMPUTATIONAL HEAT TRANSFER
In (3), f = (f,, fy, f z ) is the vector of mass forces and a, are components of
the symmetric tensor of viscous tensions.
For an incompressible Newtonian fluid we get
where q is the viscosity. Substitution of (4) into (3) yields the desired equations
of motion for an incompressible viscous fluid.
The Navier-Stokes equations for a homogeneous medium, in which all
characteristics are constant, reduce to
Here c, is the specific heat at constant pressure and q@ is the term that
describes the dissipation of energy to beat due to viscosity or viscous friction.
The quantity @ can be expressed as
The allowance for the compressibility of the medium refines the dissipative
function @ and brings about an additional term to heat equation (7).
The system of equations (2), (5), and (7) is basic for modelling convective
motions of viscous homogeneous media. The unknowns are temperature T,
MATHEMATICAL MODELS O F PHYSICS O F HEAT 39
can be posed on the boundary of the computational domain, i.e. the uwpenet-
ration ( v n = 0) and no-slip ( v x n = 0) conditions of rigid wall are satisfied.
The system of (2), (5), and (7) is peculiar in the sense that there is no equation
for pressure while there are, actually, two equations (vector equation (5) and
scalar equation (2)) for the velocity. The same partially refers to boundary
and initial conditions (see e.g. (9)).
We can eliminate pressure from the equation of motion (5). For this purpose
we use the relation
1
vgradv = gradvZ - v x curlv. (10)
We substitute (10) into (5) assuming that p and 7 are constant and apply the
curl operation to both sides to obtain
a
-- curl(v x curlv) = vA curlv + curlf,
at (11)
Equation (12) implies that the velocity components u and u can be expressed
in terms of the stream function li, (the lines li, = const are said to be
streamlines), namely
We take equations of motion in the form (13) and rewrite the convective term
in (15) in somewhat different form:
We take into account (13) and rewrite the convective term in (15) in a
somewhat different form:
and parabolic equation (15) for the vorticity. However, it is difficult to pose
boundary conditions in these new variables.
For example, the uonleakage and attachment conditions (see (9)) yield two
conditions for the stream function
Ij, = const,
aIj,-- o,
- an.
an
( 5 , ~E ) (17)
We thus have two conditions for equation (16) but no condition for the
vorticity. Besides, the constant in (17) is mast likely to be unknown (e.g. for
flows in multiply connected domains) and to be determined from additional
conditions (see Problem 2).
The system of equations (2), (IS), and (19) is a base for describing free-
convection ruotions of a fluid when equipped with appropriate boundary and
initial conditions.
tube be directed along the z-axis. The velocity of the fluid is directed along
the z-axis and is a function of the coordinates x and y. Similarly, T = T(z, y).
Under these assumptions the continuity equation (2) is an identity, while (5)
yields
Equations (20) and (21) imply that pressure is constant on a section of the
tube, and since u = u(x, y), it follows from (20) that d p l d z = const = 6 p / l ,
where 6 p is the known pressure drop across the tube and 1 is the length of
the tube. The Poisson's equation (20) for longitudinal velocity is considered
on a section of the tube 0 and is augmented by the attachment boundary
condition
= 0, (z, Y) E an. (23)
Similarly, using (7) and (8) we derive the equation for the temperature in
a homogeneous fluid
Equation (24) together with boundary condition (25) specifies the tempera-
ture field, provided that the velocity field is already calculated from equation
(20) equipped with condition (23). In both cases we deal with the classical
Dirichlet problem for the Poisson equation.
Many simplified models are based on the assumption that changes in
velocity and temperature in the longitudinal direction are small, e.g. a thin
layer model and a model of film flows. The most widespread model of this
kind is the boundary layer model.
Consider, a s an example, a flow of a viscous heat-conducting incompressible
fluid over a thin plate (Fig. 2.4). We direct the x-axis along the plate and the
y-axis across the plate. There exists a thin layer near the boundary in which
the flow undergoes substantial changes (the fluid slows down). This thin region
MATHEMATICAL MODELS O F PHYSICS O F HEAT
Fig. 2.4.
Fig. 2.5.
We differentiate (29) with respect to x and (30) with respect to y and take
into account (12) to derive
the boundary y
-=
ap aw
as T%' E 7,
where atas is the longitudinal derivative. Equation (35) follows from equa-
tions of motion (29) and (30) and the attachment and nonleakage conditions
on the boundary y.
MATHEMATICAL MODELS OF PHYSICS O F HEAT 45
E=
7
O
E, dv = uT4.
(3)
We take into account only the radiation of the rigid hody itself and convective
heat exchange with the environment and formulate the third-kind boundary
conditions on the boundary a R
aT
k-
an
+ a ( T - T,) + E U T ~= 0, (x,y,Z) E dR. (6)
MATHEMATICAL MODELS O F PHYSICS OF HEAT 47
+
a , = a + O(T- T,) (T2 T:) . (8)
The boundary value problem for (5) with conditions (7) and (8) is closest
to standard heat transfer problems discussed in Section 2.2.
Fig. 2.6.
Let n(M) denote the normal to the surface aR at a point M and r(M, P )
denote the distance between the points M and P on the surface (Fig. 2.6).
The resulting radiation flux is comprised of the flux of the proper radiation
and the absorbed flux. Therefore, by the law of conservation of energy, we
write
where cos(n(M),r) is the cosine of the angle between the normal and the
line segment that joins the points M and P . The integration in (9) is over
aW = aW(M), i.e. over a part of the boundary 0 that can be seen from the
point M.
We have thus obtained equation (9) to determine the density of flux of
radiation. The temperature field inside the bodies is defined by heat eqna-
tion (5). The boundary condition becomes
The model of ( 5 ) , (9), and (10) is simplified if we consider rigid bodies with
isothermal surfaces. In this case, the radiation flux on the surface of each
distinct body is constant. Consequently, (9) yields algebraic relations that
determine the fluxes.
2.5.4 PROBLEMS
1. Consider two limiting cases for t h e Planck law, namely w h e n
mu << 1 a n d mu >> 1.
Solution. In the first case, in which mu << 1, we use the expansion
In the case of mu >> 1 we neglect the unity in the denominator of the Planck
formula to obtain
The contact heat exchange condition remains the same as usual (see (21) in
Section 2.2):
If the bodies in contact are not absolutely black, the conditions are written
similarly (see the Stefan-Boltzmann law (4)).
2.6 Thermoelasticity
To set the simplest boundary condition for the Lam6 system of equations
would be to specify displacements on the boundary of the computational
domain, e.g.
u(x,y,z,t)=o, (x,y,z)a~. (6)
Thus, given a temperature field in the medium, the stressed state is
determined by the boundary value problem like that in (3)-(6).
A heat equation for a rigid homogeneous body that takes into account
compressibility is somewhat different from usual and looks like (neglecting
internal sources of heat)
MATHEMATICAL MODELS OF PHYSICS O F HEAT 51
The second term in the left-hand side of (7) is small for rigid bodies and is
often neglected. The heat equation (7) with appropriate boundary conditions
can be used to determine the temperature field, given deformations.
Equations (3) and (7) comprise the system of equations of thermoelasticity.
It is logical t o distinguish stationary thermoelasticity problems as a separate
class. For the case of stationary problems the system of (3) and (7) is
significantly simpler:
aZu
p - = (A + 2p) grad div u - p curl curl u - y grad T .
at2
(11)
of the cylinder and consider small deflections of this cylindrical plate under
normal loads and the combined action of heat.
The constant deflections of a fixed-width plate are governed by the Sofi
Jermen equation
DAAw = q - A M T , (15)
where D is the cylindrical rigidity, q = q(x, y) is the load, and MT is the
bending moment due to thermal actions. The value MT can be represented as
2.6.5 PROBLEMS
1. Derive an expression for the deformation of an infinite elastic
medium in which the temperature field is T = T(x, y,z) # To = const,
( x , Y , E~ a,and T ( x , Y , ~=) T O , ( X , Y , ~$)0 .
Solution. A steady thermoelastic state is described by the equation (see
(11))
(A + 2p) grad div u - p curl curl u = y grad T.
The integration of (20) and the substitution of the result into (16) yields
2.7.2 LITERATURE
1. Alifanov 0 .M. (1988) Inverse Problems of Heat Transfer [in Russian]. Mashino-
stroenie, Moscow.
2. Beck J . V., Blackwell B. & St. Clair C., Jr. (1985) Inverse Heat Conduction,
Ill-Posed Problems. John Wiley & Sons, New York.
3. Carslow H. S. & Jaeger J. C. (1959) Conduction of Heot i n Solids, 2nd edn. Oxford
Univ. Press, London.
4. Green A. E. & Zerna W. (1968) Theoretical Elasticity. University Press, Oxford.
5. Isachenko V. P., Osipova V. A. & Sukomel A. S. (1981) Heot Transfer [in Russian].
~ n e r ~ o i z d aMoscow.
t,
6. Jaluria Y. (1980) Natural Convection. Heat and Moss Ransfer. Pergamon Press,
Oxford.
7. Kutateladze S. S. (1979) Basics of Heat 7 h n s f e r [in Russian]. Atomizdat, Moscow.
8. Landau L. D. & Lifshitz E. M. (1987) Fluid Mechanics, 2nd edn. Pergamon Press,
Oxford.
9. Landau L. D. & Lifshitz E. M. (1984) Theory of Elasticity. Pergamon Press,
Oxford.
10. Loitsyanskii L. G. (1973) Mechanics of Fluids and Solids [in Russian]. Nauka,
Moscow.
11. Luikov A. V. (1968) Analytical Heot Difusion Theory. Academic Press, New
York.
12. Luikov A. V. (1972) Heat and Moss linnsfer [in Russian]. Onergiya, Moscow.
13. Schlichting H. (1968) Boundary-layer Theory, 6th edn. Translated by J. Kestin.
McGraw-Hill, New York.
14. Siege1 R. & Howell J. (1981) Thermal Radiative Heat Tmnsfer. McGraw-Hill,
New York.
15. Sparow E. M. & Sess R. D. (1971) Radiative Heat Ronsfer [Russian translation].
~ n e r ~ iLeningrad.
~a,
16. Tikhonov A . N. & Samarskii A. A. (1977) Equations of Mathematical Physics
[in Russian]. Nauka, Moscow.
17. Timoshenko S. P. & Goodier J. P. (1970) Theory of Elasticity. McGraw-Hill,
New York.
Analytical Methods of Heat
Transfer
linear ones. The problem is then reduced to finding general solutions of these
linear problems.
Among approximate methods of applied mathematics we should mention
perturbation methods. An approximate analytical solution can in certain cases
be constructed by separating a small parameter. The most important results
in this way are obtained by homogenization theory. This theory can be nsed
to compute effective characteristics of composite media with small inclusions.
Asymptotic methods are widely nsed in the theory of heat transfer.
As an example, we present the investigation of a regular thermal regime
corresponding to a developed stage of the process.
where we use the standard notation (see Chapter 2). We assume that the base
of the cylinder is subjected to heat exchange with a medium at the tempera-
ture T,
aT
-k
ax + a ( T - T,) = 0,
- x = 0. (2)
Let the heat flow be specified on the other end of the rod, i.e
In addition, we assume that the initial temperature of the rod is other than
that of the environment, e.g.
These are all the necessary boundary and initial conditions for (1)
aT' - a2T'
at1
Os,
(a~')~
+ 0 < sf < 1, t' > 0.
3.1.4 PROBLEMS
1. S e p a r a t e a dimensionless p a r a m e t e r responsible for convective
h e a t transfer.
Solution. The model equation is
2. Find the conditions under which the heat of phase change can
be neglected while modelling heat transfer.
Solution. Let us reduce the appropriate one-dimensional monophase Stefan
problem to a dimensionless form. The latent heat L is taken into account by
the condition (see Section 2.3)
The problem of (5) and (7) has nontrivial solutions only for some X
and is referred to as a spectral problem (the Sturm-Lioville problem). The
corresponding values of X are said t o be eigenvalues and the corresponding
solutions u(x) are called eigenfunctions.
Let us number the eigenvalues of the problem of (5) and (7) in ascending
order so that
All the eigenvalues are positive, i.e. X1 > 0. We denote the corresponding
eigenfunctions by u,(x), n = 1 , 2 , .. . .
Notice the main properties of the eigenfunctions vn(x). Let us consider a
Hilbert space 71 = Lz(12) equipped with the scalar product
and the norm llyll = (y,y)'/2. Similarly, for positive c(x) we define a weight
Hilbert space 71, such that
64 COMPUTATIONAL HEAT TRANSFER
is Cronecker's delta.
Given a solution of the spectral problem of (5) and (7), we can determine
the general solution of (6) as
u(x,t) = ~ ( ~ ~ , ~ ~ ) , e x p ( - ~ , t ) ~ , ( x ) . (9)
n=1
We thus obtained the general solution (10) of the heat transfer problem
(1)-(3). The cases of the first- or second-kind boundary conditions, mixed
boundary conditions, etc. are proceeded with similarly. Since the solution is
represented as an infinite series, it is often necessary to simplify the original
problem to get a simpler solution.
The general solution is constructed given the solution of the spectral
problem in (5) and (7). Note that the solution of this problem is known only
ANALYTICAL METHODS OF HEAT TRANSFER 65
in a few cases, and most textbooks on heat transfer do not present these
solutions. The same is true for stationary problems of heat transfer.
For example, let us find the solution of the simplest one-dimensional
problem in which R = (0,l). We consider the heat equation
The solution of (11) is thus presented according to (lo), (14) and the
conditions in (2) and (3).
The Green function (a source function) for this problem is defined as the
solution of the equation
aG
C(Z)- = div (k(x) grad G)
at
+ 6(z - z', t - t'), (16)
with homogeneous initial and boundary conditions (3). Here 6(z, t) is the
delta function (see Section 2.3). Thus, the Green function G(z,zf;t , t') is the
temperature field due to the instantaneous (at time t') source of heat located
at the point 2'.
The solution of the general problem in (I), (2), and (5) is uniquely
represented in terms of the Green function. The problem is thus reduced
to searching for the Green function. This function can be constructed in some
specific problems of heat transfer (in particular, when stationary problems are
considered in this way).
66 COMPUTATIONAL HEAT TRANSFER
The first, second and third terms in the right-hand side of (17) are res-
ponsible for nonhomogeneous initial temperature, nonhomogeneous boundary
condition and nonhomogeneous equation ( I ) , respectively.
We separate an important case in which heat transfer is considered in
an unlimited isotropic medium. In this case, Cl is the whole space Rm and
the bounded solution is defined by equation (1) and initial condition (2) (an
initial value problem). Under these conditions G ( x ,x'; t , t') is the fundamental
solution and the corresponding integral representation of the solution is
u ( x ,t ) =
I
Rm
u o ( x l ) G ( x ,x'; t , 0 ) dx'.
(
G ( x ,x'; t , 0 ) = ( 2 ( a a t )1,2 ) -3 exp --T ~ ? ; ~ X ' ) ) ,
where T ~ ( x , x '=) C z 1 ( x ,- x : ) ~ .
Note that the Green function can be constructed by the method of
separation of variables. For this purpose we set the right-hand side f ( x , t )
of ( 1 0 ) to be the delta function, which yields
m
G ( x ,x'; t , t ' ) = c(x')u,(x')u,(x) exp ( - X,(t - t'))
n=1
In particular, we can conclude from this equation that the Green function is
symmetrical with respect to the spatial variables for c ( x ) = const.
ANALYTICAL METHODS OF HEAT TRANSFER 67
>
exists for ( &, given a continuous function f (t) of a real argument t such
that /f(t)l 5 c o n s t e ~ p ( < ~ tGiven
). the image F(p), we can define the inverse
Laplace transform
There is a lot of reference material about the Laplace transform and some
other integral transforms.
Without going into detail we mention that the Laplace transform can be
used to find exact solutions of heat transfer problems. Let us consider the
problem (1)-(3). We multiply (1) by the kernel exp(-pt) of the Laplace
transform and integrate the resultant with respect to t from 0 to m. Let
We thus come to solving the Dirichlet problem in (18) and (19) for a se-
cond-order elliptic equation that includes a parameter p. If this problem can
be solved, the next step is to determine the inverse Laplace transform.
In this case we use the Laplace transform with respect to time. Of course,
in certain problems it is more convenient to apply the Laplace transform with
respect to some of the spatial variables. For example, this is the case for
stationary and nonstationary heat transfer in space. Obviously, there are a
lot of other examples.
3.2.4 PROBLEMS
1. A n isotropic b o d y with a fixed initial t e m p e r a t u r e exchanges
h e a t w i t h t h e environment whose t e m p e r a t u r e varies i n time.
R e d u c e t h e solution of t h i s problem t o t h e solution of a problem
with fixed boundary conditions.
Solution. We need to find the solution to the equation
au
c(x) - = div (k(x) gradu), x E 0, t > 0, (20)
at
that satisfies the conditions
au
c(x) - = div (k(x) gradu), x E R, t > 0,
at
U(X,O) = 0, E n,
I~(x)
au + a(u
an - 1) = 0, E an.
c ) ( + ) = div ( ( x ) g r a d ) r s R, t >o
instead of (6). Solving this equation with the initial conditions in (2) and (24)
taken into account results in the sought-after solution of the hyperbolic heat
equation written as
where a ( u ) = IZ(u(u))/c(u(u)),
i.e. we come to a problem with constant specific
heat capacity.
Quasilinear parabolic equations (4) and ( 6 ) can be considered as model
equations of nonlinear heat transfer.
When describing convective heat transfer, dependent variables in a model
equation of continuum mechanics are changed in a more complicated way. An
important example of such an equation is the Burgers equation
The initial condition for ( l o ) ,with ( 8 ) and ( 9 ) taken into account, is written
u ( x ,t ) = vo
i G ( x - c , t )exp
(-- dc.
72 COMPUTATIONAL HEAT TRANSFER
Thus, in order to construct the exact solution of the nonlinear heat equation
( 1 3 ) , we need to solve the ordinary differential equation ( 1 5 ) . In many
important cases the latter can be solved analytically.
Let us construct a travelling-wave self-similar solution of equation ( 1 3 ) :
This i m ~ l i e s
The equation
74 COMPUTATIONAL HEAT TRANSFER
immediately follows from (13) and (19). Hence, under transform (19) every
solution of the linear equation
3.3.4 PROBLEMS
1. F i n d t h e solution of t h e two-phase Stefan p r o b l e m i n a
homogeneous m e d i u m filling a half-space, given first-kind b o u n d a r y
conditions.
Solution. We search for the solution of the equation
u ( 0 ,t ) = U = const. (20)
The initial condition of constant temperature V is written as
u ( x , 0 ) = v. (21)
The phase change occurs at the point x = q ( t ) > 0 , say, at the zero
temperature. Then the junction conditions
are satisfied.
We seek the solution that depends on the self-similar variable < (see (14)).
We determine u(E) from equation (15) with k(u) = 1, namely
The constants
..
az = - V erf ( a )
bz =
v
1 - erf ( a )' 1 - erf ( a )
We set the constant of integration to be zero and find v > 0 for < < 0.
Equation ( 2 6 ) yields
in more detail the asymptotic behaviour of the solution to the heat equation
over time.
Let the temperature be defined by the equation
au
c(z) - = div ( k ( ~ ) ~ r a d u ) , z E $2, t > 0,
at (1)
equipped with the initial condition
k(x) -
av + au(x) = 0, x E 80.
an (11)
ANALYTICAL METHODS O F HEAT TRANSFER 77
To study the behaviour of the temperature over short times we should take
into account all harmonics in expansion (9),and the solution substantially
depends on the initial distribution. The regular regime of heat transfer occurs
at the developed stage of the process for sufficiently large times. In this
case, since the eigenvalues A, grow fast with the number of harmonic i, the
contribution of higher harmonics is negligible; therefore it is typical of this
stage that
w(x, t ) = ~ ( xt), = ( u ~ , u ~ ) , e x p ( - A ~ t ) ~ ~ ( x ) , (12)
i.e. the behaviour of the solution is defined only by the first harmonic. At
this stage the influence of the initial conditions and the distribution of the
initial temperature are of secondary importance. Finally, the third stage of
the process corresponds to the stationary regime with w(x, t) = 0.
For the regular regime
Condition (13) is not satisfied for singularly perturbed problems, and typically
The boundary value problem (15), (16) describes the temperature field
taking into account convective heat transfer. Our aim is to understand special
features of problems with dominating convection.
Evidently, the exact solution of (15), (16) has the form
to find u t
ANALYTICAL METHODS OF HEAT TRANSFER 79
The problem
for the solution to problem (19), (20). Obviously, this representation also
results from (21).
Even for small values of the perturbation parameter the solution of the
perturbed problem is not identical to that of the degenerate problem in a
certain region (namely, in a vicinity of the point x = 0). This is because of
the additional boundary conditions posed for the perturbed problem. This
phenomenon is typical of problems with small parameters at the leading
derivatives.
If a boundary layer is present, asymptotic approximations should be
constructed in a different way. These approximations, along with regular
terms, contain singular terms (boundary functions) that take into account
the behaviour of the solution inside the boundary layers.
This situation is also typical of more general heat transfer problems, namely
for nonstationary, multidimensional and nonlinear problems. The methods of
a small parameter are used to construct approximate solutions of nonlinear
heat transfer problems. In this case, each term of the asymptotic series is a
solution of a linear problem, which is simpler than the original problem.
80 COMPUTATIONAL HEAT TRANSFER
We assume that the cylinder exchanges heat with the environment according
to the law
au
k - + a u ( x ) = 0, (XI,x2) E aD. (24)
an
In addition, let some boundary conditions be given on the faces, e.g
We obtain the following problem for the first term in the asymptotic expan-
sion (29):
is satisfied. This identity with (32) and (33) taken into account yields the
desired equation for the temperature distribution in the thin rod
properties of the material are nonhomogeneous over a cube, they are identical
for each cell of the composite medium. In the simplest case, the composite
medium is a periodic structure with inclusions of another material.
We consider the heat equation
The heat capacity and thermal conductivity of the material are periodic
fnnctions cC(x/&)and kE(x/&)with the period E in each direction.
The asymptotic solution of problem (34)-(36) is constructed, provided that
the size of inclusions is small as compared with that of the domain R. Let us
denote y = Z/E. We construct a two-scale expansion for the solution ue(x)
where
is the heat capacity averaged over an elementary cell. The heat conductivities
are determined by averaging the solutions of special problems in the cells. Let
wyY) be the periodic solution of the problem
akE
div (kE(y)gradwi) = --a y i
-
satisfying w1(y) = 0. Then we get
ANALYTICAL METHODS OF HEAT TRANSFER 83
Thus, we should first solve problems for a single cell in order to find effective
characteristics of the composite material.
We now consider heat exchange through a fine-grained boundary. Let QE
have a fine-grained (wavy) boundary aRL with a typical wave size &. Let the
body exchange heat with the environment through this boundary according
to the law
auE
an
+
k - u ~ u ~ (=x 0, ) Eav. (39)
When averaging the boundary, we consider an effective third-kind condition
with a different heat exchange coefficient
Asymptotic analysis yields u = aE,y for the effective heat exchange coefficient,
where ,y stands for the local ratio of the length of the boundary of 80' to the
length of the averaged boundary 8 0 .
3.4.5 PROBLEMS
1. Find an expression for the rate of cooling XI as a measure of
nonuniformity of the temperature over the surface of a body and
over its volume.
Solution. According to (12) we should find an expression for the first
eigenvalue of problem (lo), (11). Equation (10) yields
We take into account the boundary condition (43)and rewrite the last term
Since the body is thin, we can assume that the temperature is constant over
a cross-section. Consequently, substitution of (45)into (44)yields the desired
heat equation
This equation for the stationary case was discussed from the viewpoint of the
asymptotic analysis.
3.5.2. LITERATURE
1. Bakhvalov N. S. & Panasenko G. P. (1984) Averaging of Processes i n Periodic
Media [in Russian]. Nauka, Moscow.
2. Belyaev N. M. & Ryadno A. A . (1982) Methods of the Heat Transfer Theory [in
Russian], in two vols. Vysshaya Shkola, Moscow.
3. Cherdakov P. V. (1975) The Theory of Regular Regime [in Russian]. Energiya,
Moscow.
.. Mathematics. Blaisdell Publishinr-
4. Colle J. 119681 Perturbation Methods i n Applied
cornpan;, w i t h a m (USA).
5. Gukhman A . A . (1973) Introduction to the Similarity Theory [in Russian].
Vysshaya Shkola, Moscow.
6. Ibragimov N. Kh. (1983) Groups of Transforms i n Mathematical Physics [in
Russian]. Nauka, Moscow.
7. Kartashov E. M. (1979) Annlyticol Methods i n the Heat Transfer i n Solids [in
Russian]. Vysshaya Shkola, Moscow.
8. Ovsyannikov L. V. (1978) Group Analysis of Ordinary Differential Equations [in
Russian]. Nauka, Moscow.
9. Sanchez-Palencia E. (1980) Non-homogeneous Media and Vibmtion Theory.
Springer, Heidelberg.
10. Tikhonov A. N. & Samarskii A . A . (1972) Equations of Mathematical Physics
[in Russian]. Nauka, Moscow.
11. Vasil'eva A. B. & Butuzov V. F. (1990) Asymptotic Methods i n the Theory of
Singular Perturbations [in Russian]. Vysshaya Shkola, Moscow.
Stationary Problems of Heat
Transfer
-
" a ( k
au ) f x E ~ .
a=1
is the derivative along the conormal and cos(n,x,) are the direction cosines
of the external normal n. In the case of equations (4) and (5), condition (7)
is simplified because
The Neumann problem (I), (7) is soluble neglecting a constant, provided that
Theorem 4. Given a(x) 2 rro > 0, the solution of problem (3), (9) can be
estimated as
Simple a priori bounds in (12), (13) and (15) can be extended in different
ways (see literature on partial differential equations).
Since k,p = kp,, we find condition ( l a ) ,which says that the operator A is
self-adjoint.
We also obtain (22) for the operator A defined on the set of functions
satisfying the condition
result in
Thus, the operator of the third-kind boundary value problem is also self-
adjoint.
Now let us check that the operator A is positive definite on the set of
functions in (17), i.e. let us show that ( 1 9 ) is satisfied. According to ( 2 2 ) and
the ellipticity condition ( 2 ) , we find the inequality
holds, where the positive constants cl and c2 depend only on the domain. For
the functions u ( x ) that vanish on the boundary a R inequality ( 2 7 ) becomes
We take into account the Friedrichs inequality (27) and get the conditions
of positive definiteness of the operator A for the third-kind boundary value
problem
m 00
( A u , ~2) min (-, -)11~11~. (30)
c1 c2
Consideration of the operator of the Neumann problem (16), (23) is
essentially based on the Poincari inequality
where the constants c3 and c4 do not depend on u(x). The solution of the
second-kind boundary value problem is defined neglecting a constant. Let us
separate the unique solution u(z) such that
The corresponding operator A' is self-adjoint. Given c(x) 2 co, this operator
is of a fixed sign if the inequality
is satisfied. In the simplest case co 2 0 both operators A and A' are positive
definite.
STATIONARY PROBLEMS OF HEAT TRANSFER 95
The bound in (34) also holds for the solution of the Neumann problem ( I ) ,
( 7 ) ,given the normalization condition (32).
We can derive a similar bound in the case in which the original equation
has a divergent term in the right-hand side. Instead of ( 1 ) we consider the
equation
Bounds in other norms can be derived using embedding theorems from the
bounds of the solution in W t ( n ) . For instance, bounds (36) and (35) yield
bounds in the homogeneous norm in the one-dimensional case (m = 1).
4.1.5 PROBLEMS
LW 2 I, x E n,
W(X) 2 0, E an.
96 COMPUTATIONAL HEAT TRANSFER
+
4 5 ) = u ( ~ ) l l f ( ~ ) l l c ( Inl)d ~ ) l l c ( a n ) 4%). (39)
We take (38) into account to derive w ( x ) > 0 on the boundary a R . Inside the
domain, (39) immediately yields
au x
''a
( x )( )x = -
i
n
u 2 ( r )div b d x = 0
Therefore ( A 1 u , u )= ( A u , u ) 2 ( n l / c l ) l l u ~ l 2where
, nl and cl are constants
in (2) and (28).
STATIONARYPROBLEMS OFHEATTRANSFER
Lu = f (x), x E R,
LU = g(x), E an
We relate to the continuous problem ( I ) , (2) the difference problem
where Lh and 11, are some difference operators that approximate the operators
L and I of the differential problem (I), (3).
In projective methods for solving problems of mathematical physics func-
tions of a. continuous argument are also approximated by functions of a
continuous argument, which is the principal difference of these methods as
compared with the grid methods. We pass to a finite-dimensional problem in
the following way.
Let the solution of the problem u(x) belong to a space X.In the latter
space we separate a subspace X M which is a linear span of the elements
98 COMPUTATIONAL HEAT TRANSFER
that appears when we substitute approximate solution (5) into equation (8).
Coefficients of expansion are specified in different ways to provide minimal
discrepancy (6).
Let us consider the problem in a common Hilbert space 71 = Lz(R). We
specify some weight functions yk, k = 1,2,. . . , M by the M conditions
The Ritz method is close to the latter (for self-adjoint methods they are simply
identical). Let the solution of problem ( I ) ,(2) with the homogeneous boundary
conditions be equivalent to the solution of the minimization problem for the
STATIONARYPROBLEMS O F H E A T T R A N S F E R 99
Here Qh = p h - Lhuh is the residual in the equation for the discrepancy, which
is called the error of approximation of equation (1) by difference equation (3).
Respectively, uh = x h - lhuh is the residual in the boundary condition for
the discrepancy, which is called the e n o r of approximation of the boundary
condition (2) by the difference boundary conditions (4). We stress that these
errors of approximation are considered on the solutions of the original problem
( I ) >(2).
The solution of difference problem (3), (4) converges to the solution of
problem (I), (2) if llzhll~h= llYh - uhlllh -+ 0 as h -+ 0 in some grid norm
I/ . llIh. The difference scheme converges at the rate O(hk)),k > 0, (i.e. it has
the kth order of accuracy) if for any h 5 ho, however small, l l ~ h l l l h5 M h k ,
where the positive constant M is independent of h.
Difference scheme (3), (4) approximates problem ( I ) , (2) if ll$h1(2h -+ 0 as
h -+ 0 and IIVh113h -+ 0 as h -* 0, where 11. l l ~ hand II.113h are some grid norms.
Stability of the solution with respect to small perturbations of the right-
hand side and boundary conditions is an important characteristic of a
difference problem. The stability of the difference scheme for linear problem
(3), (4) follows from the relation
Thus the operator Lh approximates the operator L with the first order
($ = O(h) at each internal node) for u(x) E C(')(0).
Similarly, the right-hand side difference derivative looks like
which approximates the derivative with the second order if u(x) E C(3)(0).
Similar algebra for the second derivative Lu = d2u/dx2 yields
102 COMPUTATIONAL HEAT TRANSFER
The latter operator approximates the second derivative with the second order
if u(x) E C(4)(R).
The approximation error is separately estimated at each node. In order to
evaluate the approximation error on the grid w, we should use grid norms.
For example, we can use the norms in C(w) and L z ( w ) ,namely
In the above examples (14)-(17) the approximation error had the same order
in both norms. Of course, this is not always the case. The approximation
order can be different in different norms. As an example, let us consider an
approximation of the operator of the second derivative on a nonuniform grid
yj = {XIx = xi, i = 1 , 2 , . . . , N , xo = 0, X N = 1) with the steps hi = xi-xi-1.
We introduce the difference operator
Therefore IIvllc = O(ho) and II$II = O(ho), where ho = maxh,, i.e. we only
rEu
get the first-order approximation.
Investigation of the structure of the approximation error shows that
operator (18) still has second-order approximation in an appropriate norm.
For instance, let
and find
STATlONARYPROBLEMS O F H E A T T R A N S F E R 103
for the principal term of the error. According to (19) we represent the
+
approximation error in the form $, = $: $f, where
We thus obtain
We thus can define the kth derivative as a solution of integral equation (21),
given u(x).
For example, let us construct difference operators that approximate the first
derivative (k = 1 in (21)). On a uniform grid w we have
/
T.+1
u -1 = j ( t ) dt. (22)
Ti-,
If we use quadrature formulae with several internal nodes, we obtain the so-
called compact difference opemtors. For example, the Simpson formula applied
to (22) leads to
104 COMPUTATIONAL HEAT TRANSFER
We take into account equation (23) and boundary condition (24) and derive
the following second-order approximation:
Fig. 4.1.
Fig. 4.2.
It is easy to check that the operators A;p and A& have the first-order and
Amp has the second-order approximation.
For cu = we have
1\12 Y = -$ ( k 1 2 ~ , ) , , .
Let us show that scheme (45), (46) diverges in the class of piecewise constant
thermal conductivities
110 COMPUTATIONAL HEAT TRANSFER
= + 0 ) - u(- 0 ) = 0 ,
[u] u(E
kg] =O
are satisfied at the discontinuity point x = E. Under these conditions problem
(41), (42) has a piecewise linear solution
+ +
where a, = (5k1 - k2)/4, a,+' = (kl 3k2)/4, b, = ( 3 k ~ k2)/4 and b,+l =
+
(5k2 - k 1 ) / 4 . We take into account that x , = - Oh and x,+l = E ( 1 - 8 ) h
and derive from (48)
We compare (49) with (47) and find that the limiting function %(x) is
identical to the exact solution u(x) of the problem only for Eo = a0 and
Po = Do, which is only possible if x = 1, i.e. kl = kz. Therefore, the difference
scheme (45), (46) diverges if kl # kz.
It can easily be seen that the function E(x) is the solution of problem
(42), (43) with the piecewice constant thermal conductivity that satisfies the
nonhomogeneous junction conditions
The balance relation (50) reflects the law of conservation of the amount of
< <
heat for the interval xi-1/2 x x , + ~ / The
~ . quantity qi-l/2 is the amount
of heat incoming through the cross-section and qitl/z is the amount
of heat outgoing through the section X,+~,Z.The unbalance of these fluxes
is due to distributed sources (the right-hand side of (50)). Although in the
integreinterpolation method we started from equation (23), in the method of
control volume the same relation (50) is derived immediately.
112 COMPUTATIONAL HEAT TRANSFER
The quantity a, in (52) has the sense of a mean thermal resistance of the
interval x,-lj2 < <
x xi+lj2. According to (50) and (53) we write difference
scheme (27) with the right-hand side
corresponds to the use of the rectangle formula for computing the integral ( 5 2 )
and the third formula to the trapezoid formula. Comparing with the scheme
( 2 7 ) , (30), (32), we can conclude that the conservative difference scheme has
the second-order approximation for smooth coefficients and solutions. This
scheme belongs to the class of homogeneous differenceschemes (the coefficients
of the difference equation and boundary conditions are computed by the same
formulae for every node of the grid).
The difference scheme for the model heat transfer is similarly constructed
by the balance method in the case of a nonuniform grid. Let = xi-hi12
and x , + ~ = / ~x, + hi12 for the nonuniform grid. Then we obtain the difference
equation
for the internal nodes of the grid. The coefficients in (55) are computed by
the same formulae ( 5 2 ) ,(54) as in the case of the uniform grid. The boundary
conditions are of the form ( 3 0 ) , (32).
Let us introduce the Steklou averaging operators by the relations
In the notation of (56) the coefficients of the conservative scheme are defined
~ ~ ( x=) S f ( % ) .The balance equation (50) is
as a ( x ) = ( S - ( ~ - ' ( X ) ) ) -and
obtained if we act on the original equation ( 2 2 ) by the averaging operator S ,
namely
for equation ( 2 3 ) .
Difference schemes for multidimensional problems are also simply construc-
ted by the integrointerpolation method. Certain more complicated cases can
be considered separately. We now dwell on the stationary problem (33)-(35)
of heat transfer in a homogeneous medium. Similarly to ( 5 6 ) , we specify
averaging operators along certain directions, e.g.
n,, = { X I X = ( X I , X Z ) x, ~ . 5~x1 -5 ~ ~ ~
X2.j-112 < 22 5~2,~+1/2).
SLu = S f ( x ) , x E w.
STATIONARY PROBLEMS OF HEAT TRANSFER
The version in which the averaging operator T ((27) and (58) in the
onedimensional case) acts on the original equation is considered similarly.
The same technique can he used for considering difference schemes for the
general stationary heat equation (33), (39), including problems with third-
kind boundary conditions.
We use the uniform grid Gi with the step h. We employ the Ritz method to
construct a projective-difference scheme.
The problem (23), (60) is equivalent to minimization of the functional
Fig. 4.3.
Fig. 4.4.
In this version of the finite element method the difference scheme has the
form (36),(38) with a1 = a2 = 1 and
We can employ the Taylor expansion to show that the operator A approxi-
mates the operator L with the fourth order, provided u ( x ) E C(". We thus
relate to (65) the difference equation
Consequently,
h2
$=f+r-Lu+-Lf +O(h4)
12
because the approximation on the solutions is L2u = Lf. In order to get the
fourth-order approximation, it is sufficient, for example, to put
I
and get
h: h$
AIY + h y - 7
+A I ~=
Yd x ) , (72)
where, e.g.
h: hi
) 12f z W 1 +
d x )= f ( ~+ fz211- (73)
Difference equation (72), (73) approximates the Poisson equation on the
solutions with the fourth order, provided u(x) E C(6)(R)and f(x) E d 4 ) ( R ) .
4.2.9 PROBLEMS
1. S h o w t h a t t h e relations
du du d2u
S+ -(x) = u,(x), S- -(x) = us(x), T ( x ) = ( x ) (74)
dx dx dx
hold for averaging o p e r a t o r s (48) a n d (49).
Solution. According to the definition of the operator S+ (see (56)) we have
STATIONARY PROBLEMS O F HEAT TRANSFER 121
The last two equations yield difference equation (27), provided that the
coefficient a is defined according to (52) and the right-hand side is
=.+I
1
P' -
-h ( I + d d ) . (76)
z. =. =.-I
If we use quadrature formulae for computing the integrals in (52) and (76),
we obtain a difference equation with the desired approximation error.
122 COMPUTATIONAL HEAT TRANSFER
For the boundary nodes we can consider m f ( x ) to be the empty set. Then
boundary condition (5) is rewritten in the canonical form ( 1 ) as
A ( X ) = 1, F ( X )= g ( x ) , E aw. (7)
Thus, difference problem ( 4 ) , (5) is written in the canonical form (1) with
A(%)> 0, x E (z),
B(x,O > 0, A(x) =
CE~C.)
x
B(x,F), x E w. (8)
are satisfied for the coefficients of the grid operator A . Then difference equation
( 1 ) on the subset of nodes W can be rewritten as
Ay(x) = F(x), x E W. (12)
The maximum principle holds for grid equation (12) (cf. Theorem 1 in
Section 4 . 1 ) .
124 COMPUTATIONAL HEAT TRANSFER
be satisfied. Assume that the maximal positive value of the grid functions
y ( x ) is attained at a point x' E W , i.e.
Since y ( x ) f const, there exists a point x" E W such that y(xl') < y ( x l ) .
Since the grid W is connected, there exists a sequence of points x l , x z , . . . ,xk
such that each subsequent term belongs to a neighhourhood of the preceding
one. We take into account (16)to conclude that y(x1) = y(xl).We repeat the
above reasoning first for the node X I , then for the node xz and so on. As a
result, we get y(xl) = y(x1) = y(x2) = . . . = y ( x k ) Therefore
Let us now consider the case in which the grid contains no boundary nodes.
Corollary 2. Let conditions (11) be satisfied for dzflerence operator (91,
(10) on a connected grid 3 and let there exist at least one node I' of the grid
w where
D(xl) > 0, x E w. (18)
Then difference problem (1) is uniquely solvable.
We can apply the maximum principle to homogeneous equation (17) ( 3 =
w). Therefore y(x) 5 0 (y(x) >
0), i.e. the only possibility is y(x) = const,
x E w. For such y(x) we have
hold for grid functions y(x) and z(x). Then y(x) < z(x) at the internal nodes
of the grid.
To prove this, let us consider the function v(x) = y(x) - z ( x ) . According
< <
to (12) Av 0 and y(x) 0 , x E aw. By the maximum principle u(x) 0, <
x E w , which completes the proof.
The following corollary of the comparison theorem is useful for estimating
the difference solution.
Corollary 3. Let y(x) and Y(x) be the solutions of the problems
In this case, if
>
However, Y ( x l ) Y(E),therefore D ( x ' ) Y ( x l )5 IF(xf)l.Consequently
and y ( 2 ) ( x is
) the solution of the nonhomogeneous equation with homogeneous
boundary conditions
130 COMPUTATIONAL HEAT TRANSFER
We thus derive a bound for the solution of problem (32) using Corollary 5,
namely
I l ~ ( " ( ~ ) l lS
c (I~l d~ ~ ) l l c ( a ~ ) . (34)
In order to derive bounds for the solution of problem (33) for the
nonhomogeneous equation we need to construct the corresponding majorant
function Z(x). Let us have (cf. the Dirichlet problem ior an elliptic equation,
Problem 1 in Section 4.1) a grid function w(x) such that
where M = I l w ( ~ ) l I c ( ~ ) .
In view of (31), (34) and (36) we find the a priori bound
for the solution of difference Dirichlet problem (22). The constant M in this
bound is defined as the maximal solution of (35). In order to estimate a
difference solution like (37) it is important to show that the constant M can
be specified independently of the grid.
Let us find a typical bound for M. Let R be a rectangle whose sides 11 and
l2 satisfy the inequality 11 5 12. We assume that the function w(x) satisfying
conditions (35) only depends on the variable xl. We can take into account the
original difference scheme (4), (5) and determine w(x1) as the solution of the
following one-dimensional problem:
in the a priori bound in (37). Hence, the constant M in (37) does not depend
on the grid steps. The bound in (37) reflects stability of the difference solution
with respect to the right-hand side and boundary conditions.
STATIONARY PROBLEMS OF HEAT TRANSFER 131
Because of the stability of (37), (40) we can also derive appropriate esti-
mates for the discrepancy of the difference solution. Indeed, let
where ,$(x) is the error of approximation on the solutions of problem (2), (3).
We have
5 MI (hf + h;)
II@(x)IIc(~~ (42)
for smooth coefficients and solutions (see Section 4.2). Because of (42) and
the bound in (37) for the solution of problem (41) we derive the bound for
the error
Thus, difference scheme (4), (5) uniformly converges with the second order.
The grid operator A in (44) is defined by (4) for x E w(') and corresponds to
the approximation (20) for x E ~ ( ' 1 .
Again, we represent the solution of the difference problem in form (31). The
difference problem for y(')(x) becomes
A =F ( ) x E~('1, Ay(') = 0, x E J2). (45)
The difference solution y(')(x) is defined by the conditions
hy(') = 0 , x w = F(')(%), x E w('). (46)
Let us first derive a bound for the solution y(2)(z).We define a majorant
function Y(2)(x)as the solution of the problem
132 COMPUTATIONAL HEAT TRANSFER
i.e. the constant M does not depend on the grid steps. This leads to estimate
(50).
Combining (48) and (50) we find an a priori bound for the solution of
problem (44), namely
This bound is in agreement with the corresponding bound for the solution of
the differential problem (see Section 4.1).
We can use (53) to determine the convergence rate of the corresponding
difference scheme in the usual way. The problem for the error is posed as
follows:
Az = $(z), x E w ( ' ) , h z = u(x), x E w('),
where $(x) is the error of approximation of the equation and u(x) is the error
of approximation of the third-kind boundary conditions. An estimate like (43)
holds for difference scheme (4), (20), i.e. this scheme converges with the second
order.
4.3.7 PROBLEMS
1. S h o w t h a t t h e solution of t h e difference p r o b l e m
where hi = x,. We sum this relation with respect to i from 1 to k, take into
134 COMPUTATIONAL HEAT TRANSFER
We now take into account the second condition in (55). For i = N we get the
expression for the constant c:
where A-u = u ~ ~ a. =~1 , 2 . Let us rewrite (58) in the canonical form (1)
and verify conditions ( l l ) , which provide that the maximum principle holds.
After some algebra we find
STATIONARYPROBLEMS OFHEATTRANSFER 135
We thus get D ( x ) = 0, x E w , and A(%) > 0 for any grid step. However, the
conditions B ( x , E ) are only satisfied if
Condition ( 5 9 ) shows that we should use a rectangular grid whose steps are not
very different so that the maximum principle holds for the difference Dirichlet
problem with increased order of approximation.
in H A becomes
2
( A Y I , Y Z5) ( A Y I , Y I ) ( A Y ~ , Y ~ ) .
If A is a positive definite operator, i.e. A 2 6E, 6 > 0, then there exists
the inverse operator A-I and IIA-lll 5 6-'. In a finitedimensional space
any positive operator is positive definite; therefore its inverse does exist. The
11 = IIAlIn for a positive definite operator A holds for any integer n.
identity / [ A n
We can introduce the negative n o m
We use condition (2) to estimate the left-hand side of (4) and apply the
Cauchy-Schwarz inequality (or Bunyakowskii inequality) I(ip, y)l 5 llipll . llyll
to estimate the right-hand side of (4). We derive the desired bound (3) for
the norm of the solution. Estimate (3) provides stability of the solution of (1)
with respect to small perturbations of the right-hand side.
If A is a self-adjoint positive operator ( A = A* > 0), we can derive an a
priori bound of the solution of equation (1) in the energy space, namely we
find the exact bound
IIYIIA = Iliplla-1. (5)
We take into account y = A-lip and obtain (5) directly from (4).
If A is not a self-adjoint operator, we can derive bounds like (5) in the
energy space generated by the self-adjoint positive part of the operator A . Let
This is because the inequalities A 2 yA and A-' 2 A-' are equivalent for
self-adjoint positive operators.
w={xlz=ai=ih, i = l , 2 ,...,N - I } ,
wt = { x / x= xi = ih, i = 1,2,, , , ,N } ,
w-={xlz=x,=ih, i = O , l , ...,N - I }
STATIONARYPROBLEMS OFHEATTRANSFER 139
for separate parts of the grid. We denote by &(xi) = &i the average grid step,
namely
-
We define a Hilbert space H(G) on the grid G equipped with the scalar
product
(Y,~)YE Y ( x ) ~ x ) ~ ( x ) .
zEu
For grid functions y(x) and u(x) that vanish for x = 0 and x = 1, formulae
(13) and (14) are simplified and become
in the rectangle fl
STATIONARY PROBLEMS O F HEAT TRANSFER 141
where
2
AY = x A , y , LY= (22)
,=l
We take into account the homogeneous condition (21) and introduce the
space H of grid functions that are defined on G and vanish on am. We define
the scalar product in H by the relation
We denote
2
A ~ = A ~ = E A~ , Ev ,H (23)
a=l
where
A,y = A m y , y E H, a = 1,2
Let us show that each of the operators A,, a = 1,2, possesses the desired
properties, i.e.
A, = A: > 0, a = l,2. (25)
For example, for the operator A2 we have
Note that the bound in (31) is established for any rectangular grid in 0 with
an improved constant M = ~(1;~+1;~).The inequalities with this constant M
for a uniform grid can be based on the spectral problem for the grid Laplace
problem (see Section 4.6).
Let us consider a one-dimensional function v(xl) = y(xl, xz). We take into
account the homogeneous boundary conditions and obtain expressions for
u(x1):
To estimate the right-hand sides of (32) and (33) we apply the Cauchy-
Schwarz inequality in the form
We put ak = %,,kh:'2 and bk = h;l2 and derive from (32) and (33) that
We now multiply (36) by hi, sum the resultant with respect to i (trapezoid
formula for a segment of a parabola) and derive the following bound from (36):
A similar relation holds for the bound of the solution of the norm of the
difference derivative with respect to 2 2 . Thus, we take into account (28) and
derive the Friedrichs inequality (31).
By the Friedrichs inequality (31) we find from (30) that
We take into account (28) and (29) to derive the bound for the operator A
Let us relate the norm (30) with the corresponding negative norm 11 . 11-1
defined by the formula
I(ip>Y)l
IIvII-1 = SUP -.
y+o IIyII1
Similarly to (6) and (9), let us derive an a priori bound for difference
problem (20)-(22) written as the operator equation (I), (23).
Theorem 1. The a priori bound
We take into account bounds (37) and (40) for the solution of (39) and find
lialll< M31hI2, M3 = n;lMlMz. Thus, the difference scheme (20)-(22)
converges with the second order in the energy space.
STATIONARY PROBLEMS OF HEAT TRANSFER 145
We estimate the error from equation ( 3 9 ) when (42) and (43) define the
structure of the error. We thus have
The second term is estimated by ( 4 1 ) . We take into account ( 4 4 ) and find for
the first term:
(2,So)= x2
u=1
( 2 , %=.). (45)
The error is also represented in the desired form (42)-(44) when solving
problems of stationary heat transfer (17)-(19) with a discontinuous heat
conductivity. This is obtained using the integreinterpolation method. Accor-
ding to Section 4.2, we find for the difference scheme (20)-(22) for problem
(17)-(19) that
= 4 S z L u - Au = ~ ( s ~ s- A,u).
~ L , (47)
o=1
Consider, for instance, the first term in the right-hand side of ( 4 7 ) . We apply
the properties of the averaging operators to find
SlSzL1u - A I U = -
- (s2~ix)
8x1
*) ;-1,2,j
)+ (aluz, L,= n ~, ,
where
(48)
are satisfied on this interface. In each separate subdomain, X I > xlk and
x1 < xlk, qa(x) = O(lhlZ) because of (48) and (49), provided that the
coefficients and the solution are smooth. Therefore the homogeneous difference
scheme (20)-(22), (46) converges with the second order. A more complex
situation arises when the discontinuity line xl = const does not pass through
the grid nodes and, moreover, when the heat conductivity breaks down on an
arbitrary curve.
Similar reasoning is valid in the situation with a surface heat source. In this
case, we consider nonhomogeneous junction conditions
To rewrite the difference problem in the operator form, we use the notation
of (23) and (24). We take into account the second-order approximation
of boundary conditions (53) on the solutions of the problem (see (20) in
Section 4.3) and put
Under such a choice of the operators A,, a = 1,2, the operator A possesses
the desired properties, namely the following statement is true.
Lemma 4. The operator A defined b y (241,(54) and (55) is self-adjoint and
positive in H , given u ( x ) > 0 .
Let us prove that (25) holds. In our case,
( A z Y , ~=) x x
zlEu,
fLl
22EY2
AZY(X)U(X)~LZ. (56)
STATIONARY PROBLEMS OF HEAT TRANSFER
IIYIIL = x
s,EY,
( Y ~ ( X ~ +y2(x1,12))
,O) h.1 + x
22EY2
(y2(0,x2) + y2(11,x2)) h.2.
The following Friedrichs inequality holds for the grid functions that do not
vanish on aw.
150 COMPUTATIONAL HEAT TRANSFER
where the positive constants m and M are independent of the grid, holds for
any grid function y ( x ) defined on the grid Z.
We do not prove this and only note that bounds like (58) can be derived in
a way similar to the proof of Lemma 3. Instead of (32) and (33) it is sufficient
to use the relations
where M I = Ml(tcl, uo, m, M ) , hold for the problem ( I ) , (241, (54) and (55),
provided that u ( x ) 2 uo > 0 .
We take into consideration the Fkiedrichs inequality (58) and find
We combine (60) and (61)and derive the bound from below, namely ( A y ,y) >
M;'llyllt, which, as usual, yields the desired a priori bound (59).
Convergence of a difference problem with third-kind boundary conditions
is studied like that of the Dirichlet problem, hence we do not dwell on it.
4.4.6 PROBLEMS
1. Show energetic equivalence of the difference operator of the
problem of stationary heat transfer in anisotropic media with first-
kind boundary conditions.
STATIONARY PROBLEMS O F HEAT TRANSFER 151
~ ( A Y , Y5 )( A V , Y )5 ~ A Y , Y ) .
The operator A+ has the same properties. Thus, the difference operator
of the Dirichlet problem for an elliptic equation with mixed derivatives is
energetically equivalent to the difference Laplace operator with constants of
equivalence nl and n2.
152 COMPUTATIONAL HEAT TRANSFER
2. Derive t h e b o u n d
maximally take into account special characteristics of the problem are involved
to solve it. It is typical of grid problems that the corresponding matrix of the
system of algebraic equations is sparse, i.e. contains many zero elements and
has a band structure. When solving multidimensional problems, the matrix
is of a large order, which is equal to the total number of the grid nodes. For
example, we consider a two-dimensional problem on a grid with Nl nodes
in one direction and Nz nodes in the other. Then the matrix has the order
Nl N 2 . Therefore, grid problems require a lot of memory and an improved fast
response of computers if standard mathematical software is used.
We rewrite a system of linear algebraic equations in the form
A = LU, (2)
where
The solution of problem (1) using the expansion in (2)-(4) is reduced to the
solution of two simple problems
154 COMPUTATIONAL HEAT TRANSFER
4 . 5 . 2 THOMAS ALGORITHM
When approximately solving one-dimensional problems of heat transfer, we
encounter grid problems with tridiagonal matrices. For these problems, special
STATIONARY PROBLEMS O F HEAT TRANSFER 155
versions of the Gauss method are used that take into account special features
of the problem. We rewrite the difference problem in the form
C I Y I- Biyz = Fi,
+ -
- A .& .- I C,yi Biyi+i = F,, i = 2 , 3 , .. . , m - 1 ,
-A,U,-I + C,Y, = F,.
The solution of difference problem ( 9 )is sought in the form
Y , =a;+~yi+l
+iRi+l, =2 . - 1 , y, =iR,+l.
a2 = C c 1 B 1 , i = I , & . . ., m - 1,
aitl = (C, - ~ , a , ) - ' B;,
PI = C;'F1, @,+I = (C, - Aiai)-' (F, +A$,), i = 1,2,. , , , m .
Lemma. Let all real coeficients of the system of equations (9) satisfy the
conditions
and at least one strict inequality hold in (13). Then in (10) and (11)
not taking into account the sparse structure of the matrix, the computational
cost would be Q = O(N:NZ) arithmetic operations. The m a t k Thomas
algorithm, which is a generalization of the above scalar Thomas algorithm,
allows us to substantially decrease the amount of comput&tional work.
Let us introduce the vectors
i.e. the vector y; corresponds to the unknowns at the nodes of the ith column of
the grid G. Then difference scheme (14), (15) can be rewritten as a three-point
vector equation (9), where A;, Bi and C, are square matrices of the order Nz-l
and m = Nl - 1. In our case, A, = (a?) and B, are diagonal matrices, while
C; is a tridiagonal matrix. Computational formulae of the matrix Thomas
algorithm have the form (lo), ( l l ) , where a; are square matrices and 0, are
(N2 - 1)-dimensional vectors.
The matrix Thomas algorithm requires a large amount of computer memory.
In the process of computations the computer must store all the matrices at,
hence the desired amount of memory is P = 0(N2Nl). Computational work
for inverting of the matrices C, - A,=, (see (11)) is Q = O(NzNl). The
method can be used for problems with Nz << Nl (the number of nodes in
the xl-direction is substantially larger than that in the xz-direction). Besides,
the matrix Thomas algorithm is successfully applied for solving of systems of
differential equations of a low dimension.
The reduction method is one of the most widespread methods for solving
elliptic problems. It is used for solving boundary value problems with
separating variables. In the case considered of model problem (14), (15) we
assume k = k(xz). We rewrite the grid problem as the following three-point
vector equation:
where rn = Nl.
The general (see e.g. (11)) three-point vector equation is essentially
simplified because A, and Bi are the identity matrices (A, = B; = E) and
the matrices C; = C are constant. For simplicity, we only consider first-kind
equations (see (18)). Currently, the reduction method is improved t o include
the cases of grid problems with second- and third-kind conditions and periodic
conditions, assuming that the variables are separated.
The reduction method is a special version of the Gauss method for system of
equations (30), (31) with m = 2'. It is based on subsequent elimination of the
unknowns yi from equations (30). Firstly, the unknowns with odd numbers
are eliminated, then the unknowns with even numbers are eliminated from
158 COMPUTATIONAL HEAT TRANSFER
the rest of the equations, then the unknowns whose numbers are multiples of
four, etc. Under such elimination, the number of unknowns is reduced after
each step. As a result, a unique equation remains for determining y,,,. The
inverse procedure is involved to determine all the unknowns.
The reduction method requires Q = 0(N2Nl log, NI) arithmetic opera-
tions. This estimate is better than that for the matrix Thomas algorithm.
The computational cost is O(log, NI) arithmetic operations per node which
is close to the optimal asymptotic estimate O(1). Therefore the reduction
method is said to be a fast method for solving grid equations.
A marching algorithm is optimal ( Q = 0 ( N l N 2 ) ) for grid elliptic problems
with constant coefficients.
Nowadays, linear algebra software packages are widespread. They are
oriented on grid problems that arise when solving problems in mathematical
physics by difference and projective-grid methods. This applied software
includes well-developed and tested routines for solving grid problems by the
above methods.
where
4.5.5 PROBLEMS
1. Derive c o m p u t a t i o n a l formulae of t h e cyclic T h o m a s a l g o r i t h m
for solving t h e s y s t e m of equations
Since the first equation in (31) includes a term with A. and the last equation
in (31) includes a term with Bm-l, we cannot use the usual formulae of the
T h o m a algorithm. In order to solve (31) we use the cyclic Thomas algorithm,
which is based on the well-known method of linear algebra called bordering.
We seek the solution of problem (31) in the form
Solution. We directly verify that the eigenvalues of problem (35), (36) are
4 .,rk
XE = - sin - k = l , 2 , ..., m - l
h2 2m'
162 COMPUTATIONAL HEAT TRANSFER
x = ( )' I 2 rkzj
sin -
1
AY = f (1)
This is the canonical form of a two-level iterative method. Given yo, all
subsequent approximations are found by (3).
To estimate the accuracy of the approximate solution it is logical to
introduce the discrepancy t k = yk - y. An iterative method converges in
the energy space generated by a self-adjoint positive definite operator D in H
if /(zk(lo-+ 0 as k -+ m. The relative error E is usually taken as a measure of
convergence of iterations, i.e.
The convergence criterion (5) corresponds to the choice D = A*A in (4). Let
no() denote the minimal number of iterations that provide the accuracy E
((4) or (5) is satisfied).
When constructing an iterative method, we should try to minimize compu-
tational work needed to determine the approximate solution of problem (1)
with a given accuracy. Let Qk be the number of arithmetic operations required
to find the approximation yk and let n > no(^) iterations be performed. Then
the total cost is evaluated by the quantity Q(E) = Qk. AS applied to
two-level iterative method (3), Q(E) is minimized by choosing the operators
B k and iterative parameters T ~ + I .The operators B k are usually defined by
some reasoning, while the iterative parameters are chosen to optimize iterative
method (3).
Two approaches are accepted in the theory of iterative methods. The first
involves a priori information about operators of an iterative scheme (namely,
Bk and A in (3)). In the second approach (variational iterative methods),
iterative parameters are determined on each iteration by minimizing several
functionals, while the a priori information about the operators is not used.
We first describe in general iterative methods not specifying the structure of
the grid operators Bk. Specific results for grid elliptic operators are presented
in the next section.
We consider the generic problem (I), in which the operator A is self-adjoint
and positive definite (A = A* > 0) in a finite-dimensional Hilbert space H.
We consider the iterative process
assuming that
A=A'>O, B=B*>O. (8)
The iterative method in (7) is said to be steady-state.
Let us have the a priori information about the operators A and B in the
form of the operator inequality
<
7 l B A 5 %B, 71 > 0, (9)
i.e. the operators A and B are energetically equivalent
Theorem 1. Iterative method (7)-(9) converges in Ho, D = A, B , for
0 < T < 2/71. The optimal value of the iterative parameter is T = 70. The
number of iterations n necessary to obtain the accuracy E can be estimated as
where
Note that no(&) in (16) is, generally speaking, noninteger, and n is the
>
minimal integer such that n no(&)is satisfied. Theorem 1 suggests how to
optimize convergence of iterative process (7), (8) by choosing the operator B
according to (9),i.e. the operator B should be close to A with respect to the
energy. In this case, the most favourable situation corresponds to the choice
B = A and TO = 1, n = 1 (a direct method).
The formula
where
Note that in the Chebyshev method (see (11) and (12)) the iterative para-
meters are computed using the given total number of iterations. Obviously,
n = 1 corresponds to the above method of a simple iteration.
The practical realization of the Chebyshev method is associated with
the problem of computational stability. The thing is that the norm of the
transition operator at several iterations is greater than unity, and consequently
the local discrepancy can grow up and lead to an abend. The problem of
computational stability is solved by a special ordering of iterative parameters
(by choosing pk from the set Mk). Several algorithms were suggested to
determine optimal sequences of the iterative parameters ~ k given , the total
number of iterations rk.
We pass from the kth iteration to the next, ( k + l ) t h iteration in two stages. At
the first stage we determine the intermediate value yktl12 from the equation
and T = 2w,
T h e o r e m 3. The optimal value of the iterative parameter in the method
of alternating directions (14)-(16) with = 6~ = 6 and A1 = Az = A is
w = (&A)-'.The estimate
where p is defined by
We compare (20) and (21) and associate the parameters X k and pk with the
parameters a k and r k :
This implies, for example, that the method of conjugate gradients converges
in a finite number of iterations, which is not larger than the dimension of the
finite-dimensional space H.
In the case D = A , according to (22), the iterative parameters are computed
by the formulae
4.6.7 PROBLEMS
1. Show t h a t t h e n o r m of t h e o p e r a t o r
+
) (E WA)-'(E
~ ( w= - WA) (24)
with
A = A*, 6E 5 A 5 AE, 6>0 (25)
is m i n i m a l for w = wo = (6A)-' a n d is equal t o
We take into account (64) and find (A-' +w)A 5 B 5 (6-' +w)A. Therefore,
the constants of energetic equivalence are written as
We take into account the definition of C, denote the residual T & = Ayk - f
and the correction wk = B - ~ T ~ derive from (28) a computational formula
and
for iterative parameters in the desired form (17).
where we denoted
7
In this equation y(x) = 0 for x E aw, i.e. at the boundary nodes this function is
not identical to the difference solution. We take a nonhomogeneous boundary
condition by an additional term in difference equation (1). The right-hand
side of (4)is different from that of difference equation (1) only at the nodes
+ +
neighbouring the boundary. It is easy to check that f = i p i p ~ h ; ~i p ~ h ; ~ ,
where
holds
for approximate solution of grid problem (1)-(3) written as the equation (4),
assuming that an initial approximation yo is given. The choice of iterative
parameters was discussed in the above. We now consider the choice of operator
B in iterative method (7).
Given A = A* > 0 and B = B ' > 0, the convergence rate is defined by the
constants of energetic equivalence yl and 72, namely
or, more exactly, by the ratio = yl/yz. Therefore, the operator B chosen,
on one hand, should be easily invertible and, on the other, should maximize
the value E .
Different classes of easily invertible operators are discussed in the following.
Among these, we can separate the simplest class of operators of multiplication
by a given grid function, which have diagonal matrices. Triangular grid
operators that correspond to lower or upper triangular matrices are easily
invertible as well. The same is evidently true for products of these operators
(factorized operators B). The method of variable directions is also based on
using a factorized operator.
A regularization principle for iterative rnethodsis important for construction
of iterative methods. When choosing B, we start with an operator R = R* > 0
(a regularizer) that is energetically equivalent to the operator A, i.e.
Let us now construct the operator B using the operator R, i.e. according to
the inequality
0
7 l B 5 R YzB,< YI > 0. (10)
Conditions (9) and (10) yield the desired inequality (8) with yl = cl+l and
0
y1 = c ~ Ywhere
~ ,
0 0 0
E = (cl/cz)E, E =7 1 1 ~ ~ . (11)
The regularizer R is chosen so that the ratio cl/cz is independent of the grid
steps and has a simple structure. In the case of problem (1)-(3), it is logical
to take the Laplace operator Y as the regularizer. Because of (6) we have
c, = K,, a = 1,2, i.e. the constants are chosen according to the given thermal
conductivity. Consequently, the ratio c ~ / c zdoes not depend on parameters of
the grid.
Two possibilities arise in this choice of the regularizer R. The first possibility
is when the ratio nl/nz in a problem of stationary heat transfer is not very
small. The second is associated with too small a ratio nl/nz when the number
of iterations depends on this ratio inappropriately. In the latter case, the
STATIONARY PROBLEMS O F HEAT TRANSFER 173
ht '4
and put in (12)
b(x) = 2 (h;' + h;') . (13)
In order to derive the bound in (10) (for R = A) under the choice (12), (13),
we can use the least (6) and greatest (A) eigenvalues of the Laplace operator,
namely
0
6 E 5 A 5 AE. (14)
The eigenvalues of the problem Ay = Xy are composed of the eigenvalues of
the corresponding operators with respect to the variables xl and x2 and are
represented in the form (see Problem 2 in Section 4.5)
2
X I , = A(,:) + -Cg
Ira -
4 . k ah,
s1n2 *'
,=I
From (14) and (15), given (12) and (13), we derive the bound (10) with
where lhI2 = h: + hi, According to (16) and (11) we derive the corresponding
bounds for the number of iterations of the method (7), (12), (13).
According to Section 4.6, the number of iterations of the method of a simple
iteration with the optimal value T = TO = 2(y1 +yz) and that of the method of
steepest descent with the operator B chosen by (12) and (13) can be evaluated
The method with the Chebyshev set of iterative parameters converges much
faster than that of a simple iteration.
In the case of a large ratio x = nzlnz (in problems of stationary heat
transfer in compound bodies with different-scale thermal parameters) the
number of iterations (18) can inappropriately depend on X. Therefore, it is
reasonable to consider the iterative method with the diagonal operator B
constructed directly using A (without a regulaxizer).
Let us consider iterative method (7) for problem (4) with the operator B
specified by (12), where according to (1)-(3)
( A v , v ) = 2 ( B u , u ) - ( A u ,u )
taking into account (19). The analysis of boundary value problems (27) shows,
for example, that for media with piecewise-constant thermal conductivity
Let the operator D' correspond to the diagonal part of A and let C correspond
+ +
to the lower triangular matrix. Then since A = C 'D C*, we find
which is called the treangular iterative method. The method in (31) represented
in the canonical form corresponds to the choice
Note that in the above reasoning, D is not necessarily the diagonal part of
the operator A.
The over-relaxation method thus corresponds to the iterative parameter T
close to the optimal value. The bound in (34) for the number of iterations
also holds for the over-relaxation method.
Let us illustrate the advantages of triangular iterative methods (31) by the
0
grid Dirichlet problem for the Poisson equation (equation (4) with A = A).
We thus consider problem (1)-(3) with a,(x) = 1, a = 1,2. We take D to be
0
This bound is close to that for the Chebyshev iterative method (see (18)) with
a diagonal D .
The Seidel method is c h a r ~ t e r i z e dby T = 2. The number of iterations is
estimated as
no(.) = o ( + - l n t ) .
Hence, the Seidel method converges much more slowly than the over-
relaxation method. The number of iterations in the twedimensional problems
considered, as in the method of a simple iteration, is proportional to the total
number of nodes.
Since then operator V is positive, we derive (By, y) 2 2w(Ay, y), i.e. A 5 yzB,
where
1
yz = -. (43)
2w
Now we can choose the parameter w in (40) from the condition of maximal
( = ((w) = yl/yz. According to (42) and (43) we obtain
71
((w) = - = 2w6
h 1 + w6 + w26A/4'
The maximum of ( ( w ) is attained for
and is equal to
The constant 6 in the first inequality in (33) is the least eigenvalue of the
operator A, namely
4 . ahl, 4 . ah2
6 = - sin - + - sm -
ht 211 hi 212
The operators A,, a = 1,2, are defined by (36). According to the above
estimate (37) we conclude
We take into account (48) and (49) to find the optimal value of the iterative
parameter w for which the number of iterations in the Chebyshev iterative
method is estimated as
Thus, the number of iterations is proportional to the square root of the number
of nodes in one direction (to the fourth root of the total number of nodes in
our two-dimensional problem). The bound in (50) is preferable as compared
with the above bounds for other iterative methods.
When solving problem (1)-(3) with variable coefficients, the alternating
triangle method can be constructed using a regularizer R, e.g. the grid Laplace
operator. In this case, instead of the bound in (50) we find
+
We use the notation a:'(x) = al(x1 h1,xz) and a Z 1 ( x ) = a z ( x l , x 2 + hz).
The operator 'D is given as 'Dy = d ( x ) E , where
The grid functions b, and c,, a = 1,2, are determined by solving three-point
boundary value problems.
In order to find b, = max v m ( x )we solve the problem
z,E%
is satisfied for problem ( 5 5 ) .The strict inequality holds for nodes ueighbonring
theboundary ( i = l , N l - 1 , j = l , N ~ - 1 ) .
Let us derive a formula for computation of the elements gij of the grid
operator 2)taking into account that the operators A and B are 'close' because
of (54). We obtain
therefore
Since the row sums are equal to each other according to ( 5 4 ) ,we derive
STATIONARYPROBLEMSOFHEATTRANSFER 185
.,
We can then rewrite (57) a s gij = d - a ,,p,-1,
. . - b,jpij-l. This formula is
transformed as follows:
Relation (59) together with (56) can be used to prove by induction that the
conditions gij > 0, pij 5 1 are satisfied, i.e. B = B* > 0.
To investigate the convergence rate of the alternating triangle method in (7),
(53), (54), let us estimate the constants of energetic equivalence in inequality
(8). According to (57), we have the representation
+ +
We apply the -inequality ( a b)' 5 ( 1 ) a Z+ (1 + E-')b2 with E =
ai+l,,/bi,j+l to the right-hand side of (63) and get
4.7.6 PROBLEMS
1. Let p(x) > 0 , p(x) $ 0 and a ( x ) > n > 0 on a uniform grid w . Show
that the estimate
(a& 2 M(m,vLI (64)
where M-' = max,~, u ( x ) and u ( x ) is the solution of the problem
holds for any grid function y ( x ) such that y(0) = 0 and y(1) = 0.
Solution. Let us define a grid operator A = A* > 0 on the functions that
vanish at the nodes of the closed interval [O,I] by the relation Ay = -(a%).,
x E w. Let us consider the eigenvalue problem
STATIONARY PROBLEMS OF HEAT TRANSFER 187
for all y(x), where ();. = (., .),. Equality holds on an eigenfunction
corresponding t o the minimal eigenvalue Amin. We denote this eigenfunction
by w(x). Since (Ay, y) = (ay$, I ) + , because of (67) the desired hound in (64)
will be established if we show that M 5 X,i,.
First of all, let us explain that w(x) > 0, x E w (more exactly, the function
w(x) has a constant sign). Let us assume the contrary, i.e. w(x) changes its
sign on the grid w. We consider the function lw(x)l and obtain
This inequality contradicts the assumption that w(x) minimizes the ratio
(my$, l ) , + l b y 2 , 1).
Let us now consider the relation Aw = X,i,p(x)w. This, because of
W(X)> 0. x E w, yields
+ +
B = E - w ( A i + Az) W ~ A I A Z Zw(A1 + A2)
= ( E - wAt)(E - wA2) + 2wA.
Consequently,
x
Equation (1) in this case corresponds to the use of polar coordinate system
(T,91, namely
Equations (2) and (3) can be rewritten in the following divergent form:
Except for the notation, equations (4) and (5) belong to the above
considered class of heat equations written as second-order self-adjoint elliptic
equations with variable coefficients, namely
and replaced xlf (2)by f (x).Equation (5)is written in the form (6)with the
anisotropic 'heat conductivity'
natural to require the solution of equation (6), (7) (or (6), (8)) to be bounded
for zl = 0, which is equivalent to the requirement
au
lim zlk(z) - + 0.
=I-0 axl
We have to solve the boundary value problem for equation (6),(7) (or (6),
(8)) in the rectangle R = {XI x = (XI,ZZ),1; < z, < 1;) for 1; = 0. The
equation is equipped with boundary condition (9). It is convenient to use the
following quasi-uniform grid with respect to the variable z l grid
i.e. the uniform grid shifted by half a step (a flux grid). Further, difference
approximation of (6), (7) (or (6), (8)) is constructed by the integro-
interpolation method (Section 4.2).
We denote ql(x) = -xlk(x)au/ax and integrate equation (6) with respect
t o xl from 0 to hl (over a neighhourhood of the node XI = hl). We thus get
Fig. 4.5.
Fig. 4.6.
192 COMPUTATIONAL HEAT TRANSFER
Fig. 4.7.
define H as the set of grid functions vanishing on aw equipped with the scalar
product
(Y,w) = x y ( x ) w ( x ) h ~ h z , (13)
rEu
It can be shown that this modification does not make the monotonicity
properties of the difference scheme worse, its accuracy is preserved, while
the operator A is self-adjoint in a simple Hilbert space H with scalar product
(13). We can thus choose (12), (14) or (13), (15).
We suggested several possibilities for solving problems of stationary heat
transfer on irregular orthogonal (rectangular) grids. It is natural that
nonorthogonal grids provide a wider range of possibilities. For instance,
one can use nonorthogonal quadrilateral grids. This technology has its
own peculiarities both in construction of difference schemes and in their
investigation. These problems are beyond the subject of this monograph.
194 COMPUTATIONAL HEAT TRANSFER
Fig. 4.8.
STATIONARYPROBLEMSOFHEATTRANSFER 195
u E ( x )+ 0 as E-
temperature in R I level and the conditions on the boundary (17) yield
0. The corresponding bound for the discrepancy between
the approximate and exact solutions is 11u,(x) - U ( X ) I [ ~ ; ( ~ ) 5 ME'.
Another common version of the method of fictitious domains for appro-
ximate solution of the Dirichlet problem for elliptic equations is the version
with extension with respect to lowest terms. In this case, the second
derivatives are continuously extended into the fictitious domain. In order
to approximately solve problem ( l o ) , (11) with homogeneous boundary
conditions, we involve the boundary value problem (16), (17) with the
coefficients
The corresponding bound for accuracy of the version (16), (17), (19) has the
form
IIuc(x) - 4 x ) l l w ; ( n ) 5 M E . (20)
This version of the method of fictitious domains allows us to consider problems
like ( l o ) , (11) with nonhomogeneous Dirichlet conditions. For this purpose,
we extend the function g ( x ) into the fictitious domain Ol and define the right-
hand side of equation (16) to be f,(x) = ~ - ' g ( x ) for x E 0 1 .
Boundary conditions on 80, which lies inside the extended domain 00,are
most naturally specified in the version of the method of fictitious domains
with a surface delta function. We extend f ( x ) into the fictitious domain 01
and instead of ( l o ) , (11) consider the problem for the equation
Here (.I stands for the jump of a function when passing through the boundary
8 0 . Similarly, for problem (21), (17) we have
We subtract (25) from (26), take into account the boundary conditions ( l l ) ,
(23) and come to the relation
We discard the first term in (27) and derive from (27) and (28) that
Here n,, a = 1,2, is the external normal relative t o the domain S1,, u = 1,2.
Let us construct the simplest iterative process to explicitly refine the first-kind
boundary condition on y taking into account (30) by the formula
where now
where
Iterative methods like (33), (34) are known as iterative methods of capacity
matrix.
Let us mention the basic topics of investigation of convergence of iterative
decomposition methods. It is convenient to consider the iterative process for
the discrepancy wk(x) = uk(x) - u(z). In this case (33) yields
where the operator Aw defined by (34) is given for x E 7, where (see (lo),
(32))
Let us show that the operator A defined on the set of functions satisfying
(36) is self-adjoint in 71 = L2(y). In this case, because of Green's formula
where h is the grid step with respect to the variable xz (Fig. 4.9a) or XI
(Fig. 4.9b). The bounds in (39) are used to estimate the convergence rate of
specific iterative methods.
When realizing the iterative decomposition methods, it should be taken into
account that on each iteration the conditions are refined only on a part of the
boundary, and it is necessary to find the solution of the grid problem in a
subdomain in a portion of nodes. This problem in several cases can be solved
STATIONARY PROBLEMS O F HEAT TRANSFER 201
by direct methods faster than the problem of determining the solution at all
nodes of the grid.
Other versions of the decomposition method became widespread in com-
putational practice. Among these, we can mention approaches with different-
type conditions on common boundaries of the subdomains (the Dirichlet-
Neumann condition). When decomposition methods are solved by parallel
computers, increased attention is paid to the dependence of convergence rate
of the corresponding iterative processes on the number of subdomains.
where
I
The convergence rate of iterative process (40)-(45) in C ( R ) is established
by the maximum principle for second-order elliptic operators (see Section 4 . 1 ) .
We introduce the notation
'
Lv, = 0 , x E R,,
v,(x)=O, xEr,,
v , ( x ) = 1, x E y,.
STATIONARY PROBLEMS OF HEAT TRANSFER 203
Let G,(x, a ' ) denote Green's function of the Dirichlet problem for equation
(46) in the domain a,, a = 1,2. According to (43)-(45) we have the
representation
w ~ + ' ( x=
) w:(z), x E 7.2
Awl = w i ( x ) - . ? ~ S ~ w l ( x ) , x E 71
where vl(x) is the solution of problem (46)-(48). We thus derive from (55)
the bound
l l ~ i l l c (5~~i11~i11c(~,)~
~) <
IISill 41. (56)
Similarly, from (52) we get
method by parallel computers. The asynchronous method does not have this
disadvantage and is based on the iterative procedure
rather than on (43). That is, the boundary condition on yl is refined by the
solution in the domain Rz on the kth iteration.
As compared with the classical Schwarz method (Problem 2), the con-
vergence rate of the asynchronous version is much slower (twice as slow for
Q1 = qz).
The maximum principle for grid problems (see Section 4.3) can be applied
to investigation of grid analogues of the alternating Schwarz method, as was
shown in the above.
4.8.6 PROBLEMS
1. Consider t h e a p p r o x i m a t i o n of t h e third-kind b o u n d a r y con-
dition
au
- +
a n U(X)U= g(x),
E an (63)
for e q u a t i o n (10) o n a consistent grid.
Solution. Let us consider a node 0 on the boundary (Fig. 4.5). We rewrite
(63) in the form
where cos(n,x,), a = 1,2, are the direction cosines of the external normal.
The simplest approximation of (64) yields
Thus, (65) approximates boundary condition (64) with the first order.
In several cases, the order of approximation on the solutions of equation
a = 1,2, of the
(10) can be increased by specially choosing the grid steps h*,
consistent grid. We rewrite (66) in the form
206 COMPUTATIONAL HEAT TRANSFER
and write the error in the form $(a) = 0((hi)2 + (h;)2). Of course, the
consistent grid and, what is more, the grid satisfying (67) can only be
constructed in several special cases.
2. Study the convergence rate of the asynchronous version of the
alternating Schwarz method for the case in which the boundary
condition is refined according to (62).
Solution. Let us study convergence of the alternating Schwarz method (40)-
(42), (44), (45), (62) on the direct sum of the spaces 'H = 'Hi 'Hz, where
H
' , = C(y,), a = 1,2. We define the norm of an element U = (u1,uz) E 'H by
+
the expression [IUll = IIulllcct,) l l ~ l l ~ ( + ~ The
! . iterative process (40)-(42),
(44), (45), (62) for the error W = (wl,w2) is written as
We take into account the above bounds for S,, a = 1,2 to derive from (69),
(70) that
Hence, iterative process (40)-(42), (44), (45), (62) converges for 0 < T <
+
2/(1 q), where q = maxq,,m
at the rate of geometric progression to the
k
solution of problem (lo), (11). The bound llu:llc(,,)+ lluzllC(72)5 Mqk holds
for the optimal value of the iterative parameter T = TO = 1. We compare this
bound with that in (49) and conclude that the asynchronous version of the
Schwarz method converges much slower than the synchronous version. The
number of iterations for the former is at least twice that for the latter.
are satisfied.
Let us first establish sufficient conditions of uniqueness of the solution of
( I ) , (2). Uniqueness of solutions to nonlinear problems is studied using the
corresponding linear problems. Let us illustrate this general statement by the
example of problem ( I ) , (2).
We assume that two solutions exist, namely u4(x), 0 = 1,2, i.e.
the classical maximum principle (Section 4.1) holds for equation (5). We
thus conclude that boundary value problem ( 6 ) , ( 7 ) has only trivial solution
w ( x ) = 0, i.e. the solution of problem ( I ) , (2) is unique, provided that the
thermal conductivity is linear and ( 8 ) is satisfied.
The situation with k = k ( x , u ) is more interesting. It turns out that the
maximum principle also holds for linear equations like ( 6 ) (see Problem 1).
This is the case if the derivative of the thermal conductivity with respect to
temperature is bounded, i.e.
The right-hand side of (10) for smooth functions f (z, u) can be rewritten as
where
Because of the maximum principle (Section 4.3) for grid problem (la), (19)
we obtain v(x) = 0, x E a,provided that (8) is satisfied. That is, if (8) is
satisfied, we have a solution of nonlinear difference problem (10)-(12), (17).
Let us now formulate the problem for the error z(x) = y(x) - u(x), x E ii.
We have
STATIONARY PROBLEMS O F HEAT TRANSFER 211
Ild~)llcc~)
C Mll$(~)llc(w) (23)
holds for the solution of difference problem (21), (22), provided that ( 8 ) is
satisfied. Here M = I l w ( ~ ) l l ~and
( ~ w) ( x ) is defined by the conditions
AW 2 1, E w, W(X) 2 0, E aw.
Thus, if (8) is satisfied, nonlinear difference scheme (10)-(12), (17) converges
with the second order.
Note that the maximum principle yields that the solution of the difference
problem is bounded. We thus have
where
We rewrite difference equation (10)-(12) taking into account (24) and get
results in the iterative process that is based on the solution of the difference
equation
where the notation of Section 4.7 is used. The iterative relaxation method with
iterative parameter w = 1 corresponds to specifying the new approximation
by solving the problem
On the set of grid functions y(x) vanishing on the boundary we define the
grid operators in the usual way:
Bu = Av, A ~= af
v AV- -(x, -
yk)u, x E W. (35)
ay
The following bound holds for the error: IIuk+111 5 pollvkII, where PO =
(1 -()/(I + E), E = yl/yz. Thus, convergence of the simplest iterative process
(27), (33) is in general like that for convergence of the method of a simple
iteration (Section 4.6) for linear problems.
Let us now consider the Newton method for approximate solution of
difference problem (10)-(12), (17). Because of (28), the new approximation is
determined from the difference equation
completed by boundary conditions (27). We have the problem for the error
On the basis of the investigation of linear problem (40), (41) we can draw
a conclusion about the convergence rate of the Newton method. According to
the maximum principle (see (23)) we derive the bound
This implies that the Newton method converges if the initial approximation
<
is close to the solution. Namely, if qllyo - yllc(,) 1, the iterations converge
at a quadratic convergence rate.
When considering iterative processes (including those for nonlinear prob-
lems), considerable attention is paid t o monotonicity of the approximate
solution. For example, a n~onotonicprocess can give an approximation from
<
below, i.e. yo 5 yl 5 . . . 5 yk yt+t 5 . . . 5 y. An example of a monotonic
process for nonlinear problem (10)-(12), (17) is considered in Problem 2. The
Newton method gives an approximation from above or below depending on
whether the function f(x, y) is convex or concave with respect to the second
argument. For instance, let f ( x , y) be a concave function, i.e.
Then (41) (the maximum principle) implies that uk+l = yk+l - y 5 0, i.e. the
Newton method gives an approximation from below (yk+l 5 y).
216 COMPUTATIONAL HEAT TRANSFER
4.9.6 PROBLEMS
1. Show t h a t t h e Dirichlet problem for t h e elliptic equation
>
h a s a unique solution if c(x) 0 a n d lb.(x)I M. <
Solution. We should show that the homogeneous problem
<
has only a trivial solution. For this purpose, let us show that u(x) 0, x E 0,
<
for f ( x ) 5 0, x E R, and u(x) 0, x E a0 (the maximum principle). It can
similarly be shown that u(x) 2 0, x E R, for f (x) 2 0, x E 0, and u(x) 0, >
x E a n . Under these conditions problem (44), (45) has only a trivial solution.
Let us prove this by contradiction. We assume that for f ( x ) <
0, x E a,
<
and U(X) 0, x E 8 0 ,
u(x) > 0, x E R+ (46)
in a subdomain n+( 1 + c 0).We integrate equation (43) over the domain
n,. Since u(x) = 0, x E an+,we have
- / I(,),
au dx + / C(Z)Udx = / f (x) dx.
an+ an+ an+
Under assumptions (46), the left-hand side of this inequality is positive while
the right-hand side is nonnegative. We thus come to a contradiction, which
shows that u(x) 5 0 in the entire domain R.
2. Use t h e i t e r a t i v e m e t h o d of a s i m p l e iteration (33) for
a p p r o x i m a t e solution of p r o b l e m (10)-(12), (17) t o c o n s t r u c t a
m o n o t o n i c iterative process, assuming (8) t o b e satisfied.
Solution. Instead of (33) we consider the following iterative process:
and let the boundary conditions be satisfied exactly. For the error we have
+
( A c(z)E) + Ayo - f ( z ,yo) = 0 :
therefore wl >_ 0 because of assumption (48). From (47) we derive
where
+
Fk = ( 1 - T ) ( A c(z)E)wr: r + (50)
AYO- YO) 2 0 ,
the iterative process gives a monotonic approximation from above.
218 COMPUTATIONAL HEAT TRANSFER
diminished [7]. The most widespread method for band matrices (in
particular, tridiagonal) is the Thomas algorithm and its versions. Detailed
presentation of these algorithms and the grounds for them are given
in (281. Special direct methods are used for grid elliptic problems
with separating variables, e.g. the reduction method and fast Fourier
transform. These methods are discussed in detail from various viewpoints
in 110, 281.
4.6 We present the general theory of difference schemes following [28]
but more concisely. In particular, we do not cover topics related t o
computational stability of Chebyshev iterative methods. Earlier, the
method of variable directions was very popular. However, an optimal
set of iterative parameters can only be constructed for grid elliptic
problems with separating variables, for which the direct methods are
faster. Therefore we only consider the general case of the method of
variable directions with noncommuting operators. Other advantages of
the methods of variable directions are thoroughly studied in [24, 281. In
order to make the presentation of variational (both two- and three-layer)
systematic, we choose the iterative parameters that minimize the residual
on the next iteration, which seems to be natural. In other books on
iterative methods (see e.g. [9]), more attention is paid t o a more complex
and less clear interpretation of the method of conjugate gradients. In this
chapter, we do not discuss the solution of grid problems with operators
which are not self-adjoint.
4.7 We estimate the effectiveness of iterative methods and choose iterative
parameters of iterative methods following 128). The dependence of
the convergence rate of classical iterative methods on discontinuous
coefficients is presented in [31]. The alternating triangle method was
suggested by A. A. Samarskii in 1964. Its more detailed presentation
is given in [28]. The method of approximate factorization, which in the
symmetrical version was treated as an alternating triangle method, is
described in [4]. The general theory of iterative methods in subdomains
is studied in [12].
4.8 Curvilinear orthogonal coordinates and locally irregular grids are tra-
ditionally widely used in computational practice [24, 27, 281. We do
not touch upon problems of generation of nonorthogonal grids and
solution of boundary value problems on such grids. The method of
fictitious domains has been developed since the 1960s. This approach
is thoroughly considered in [31]. Decomposition methods are almost
absent in monographs and textbooks. The only exception is (151, where
one chapter is devoted to this method. Decomposition methods without
overlapping are discussed in the review [I].
4.9 Methods for approximate solution of nonlinear boundary value problems
for elliptic equations are considered from different points of view in
220 COMPUTATIONAL HEAT TRANSFER
4.10.2 LITERATURE
Constraints are imposed on the time step for other schemes (conditionally
stable difference schemes).
In studying the stability in Hilbert spaces, we use the general theory
of stability for difference schemes. The latter is based on writing rnany-
level difference schemes in the canonical form and formulating the stability
conditions in one or other norms as operator inequalities. The theory is
accurate in the sense that the necessary and sufficient conditions coincide.
The general theory of stability for difference schemes is applied in studying
two- and three-level schemes for the heat equation. Thus, in particular,
conventional schemes with weights are studied. The corresponding results
on accuracy of difference schemes are formulated on the basis of stability
investigation.
The regular mode of heat conduction is investigated separately. The
notion of asymptotical stability is introduced to describe the developed stage
correctly. The asymptotical stability of difference schemes is investigated for
the heat equation. In particular, we show that the conventional symmetrical
difference scheme is unconditionally stable in the common sense and is
asymptotically stable.
Difference schemes are also considered for the hyperbolical heat equation.
Using the simplest boundary value problems of heat conduction as examples,
we illustrate the specific features in using difference schemes for approximate
solution of nonlinear problems.
where
NONSTATIONARYPROBLEMS OFHEATTRANSFER 225
Equation (I), (2) is the classical linear parabolic second-order equation, which
is classified as a basic equation of t,he mathematical physics.
In the general case of a movable medium equation (1) is written as
as basic problems. Let us write the heat equation separately for a homogeneous
medium where both the heat conductivity k(x) and the specific heat capacity
c(x) are constant. Reducing the problem (see Section 3.1) to a dimensionless
form, we arrive at the simplest second-order parabolic equation with constant
Equation (1) (or (3) and (6)) is complemented by the necessary boundary
conditions. The most attention is paid to the Dirichlet problem, where
where u(x, t ) > 0. In (8) we use the notation from Chapter 4. For
nonstationary equations we can point out the problem with u(x, t ) = 0 as
a separate problem with boundary conditions of the second kind.
The well-posed problem for equation (1) is that with the known initial
temperature
u(x,O) = uo(x), x c R. (9)
226 COMPUTATIONAL HEAT TRANSFER
hold for the functions u(x, t) and ~ ( t ) ;then u(x, t) v(x, t) over the whole
domain Q.
Let us present a priori estimates in the uniform norm for parabolic boundary
value problems. The estimates are based on the maximum principle.
NONSTATIONARY PROBLEMS OF HEAT TRANSFER 227
We connect the Hilhert space 'Ka with the operator B so that llulli = ( B u , ~ ) .
Taking into account that the operator d is positive, and using the inequality
( f , ~ 5) IlfIla-1 I I ~ B ,
from (19) we obtain
d
- Il~ll5
dt
< Il4la-1.
By virtue of the lemma, from this we obtain the estimate of stability with
respect to the initial data and right-hand side for problem (13)-(16)
follows from this. Unlike (20), here we estimate the squared norm of the
solution, and use the norm of 'R in estimating the right-hand side.
A similar estimate for problem (13)-(16) can also be obtained (see
Problem 2) in 'Ha. Let us also show how to obtain estimates of stability
in some general spaces ED generated by a constant operator D ' = 'D* > 0.
Assume that the operators B and A are permutahle. By virtue of conditions
(16), there exist the inverse operator B-', and we can switch from equation
(13) t o the equation
where A' = B-'A. By virtue of (15) and (16), for the permutable operators
A and B we obtain A' = (A')' > 0. The estimate (20) for equation (22) is
now like
Il.(t)ll <
Il.(O)ll + t m p IID-lf II. (23)
The operators A and B are permutable for the considered problems of heat
conduction if c ( x ) ? coust. If, additionally, the operator D
' is permutable with
A', then from (22) we obtain the estimate
5.1.4 PROBLEMS
1. O b t a i n t h e a priori estimate in t h e uniform n o r m for t h e first
b o u n d a r y value problem for parabolic equation
230 COMPUTATIONAL HEAT TRANSFER
>
If the condition d(x, t ) - pc(x) is satisfied, then the a priori estimate like
(12) holds for this equation.
For the solution to the first boundary value problem (25), (7), and (9) at
d(x, t) - pc(x) 2 0, p < 0, we have the estimate
holds for the problem in hand. Multiplying equation (15) in a scalar way by
duldt, we obtain
for the right-hand side. Substituting these in (28), we obtain the inequality
Yn+l - Yn
Y=Yn, (P='Pn, Yt=
T
The scheme (3) can be written in the canonical fonn of the three-level
difference scheme as
assuming that the time grid is uniform. In order to switch from (3) t o ( 6 ) , we
Put
Y; = + Y:)
$ ( ~ t =
Yntl - Yn-1
> %t =
Y n t l - 2yn + Yn-I
2T T2
By.
t
+ .r2RZt+ Ay = ip. (7)
In order to determine the solution a t the new time level in case (4),we
should solve the equation
NONSTATIONARY PROBLEMS O F HEAT TRANSFER 233
Difference scheme (10) is stable with respect to the initial data if the inequality
is satisfied.
It is logical to introduce the notion of stability with respect to the right-hand
side for a nonuniform difference scheme with zero initial data. The difference
scheme
Yn+l - Yn + A y , = i p n , n = 0 , 1 , ..., y o = O ,
r
(12)
234 COMPUTATIONAL HEAT TRANSFER
is satisfied.
The introduced notions of stability with respect to the initial data and
right-hand side reflect more precisely the sense of the two terms in the general
condition of stability (9). In some cases this allows us to restrict ourselves by
investigating only the stability with respect to the initial data.
The stability of many-level schemes can be investigated by reducing to
an equivalent two-level scheme. Further we discuss the ways to do this in
considering three-level difference schemes. This allows us to restrict ourselves
to considering only the stability of twdevel difference schemes.
>
where c 0 does not depend on h, T , and n. Using (15), for the constant MI
we obtain ( M I 2 p") M I = exp(cT).
The uniform difference scheme (10) can be written a s
where
s = E - TB-'A (17)
is the operator of transition from one time level into another. The operator S ,
generally speaking, can depend on n. By virtue of (16) and because y, E Hh
NONSTATIONARY PROBLEMS OF HEAT TRANSFER 235
are arbitrary, the uniform stability of difference scheme (10) with respect t o
the initial data is equivalent to that the norm of operator S is bounded by
the constant p, namely
IISll 5 P . (18)
The estimates of stability like (9) for two-level difference scheme (4) can be
obtained on the basis of the Gronwall diffewnce lemma (cf. Section 5.1).
L e m m a . Let E, and 3" be nonnegative functions defined on the grid w,,
and let p > 0. Then the inequality
we arrive at estimate (9) of stability with respect to both the initial data and
right-hand side at l)(ok)/2,, = JJB;lp*lJlh and M2 = M I T . Thus the theorem
is proved.
Definitions of stability can use the corresponding estimates for the squared
norm of the difference solution (see estimates (21) and (27) for the differential
problem in Section 5.1). Instead of (9),we can require that the inequality
is satisfied. In this case, however, we cannot obtain the stability with respect
to the right-hand side from that with respect to the initial data. One can
pass to estimates baaed on the squared norms by transforming the difference
Gronwall lemma (Problem 1).
The vector Y ncan also he chosen in other ways (a variant different from (25)
is considered, for example, in Problem 2).
Calculating directly, for the operators of difference scheme (24) we obtain
the formulae
Difference scheme (4) has the approximation O(JhJm T'), m + > 0, 1 > 0,
on the solutions u(x, t), if
max ll$k112h
OSkSN
< M3 ((him + 7') , (30)
238 COMPUTATIONAL HEAT TRANSFER
Assume that the initial condition is approximated within the same accuracy,
i.e. the estimate
<
IIyo - 4 J ) l l l h M4 (Ihlm T')+ (31)
is satisfied. If difference scheme (4) is stable, i.e. estimate (9) is satisfied, then
for the error we have
5.2.7 PROBLEMS
1. Show t h a t if t h e e s t i m a t e
holds w i t h a r b i t r a r y E.
Solution. It follows from (33) that
B2Yn+1+B1Yn+BoY,-~=9,, n = l , 2 ,....
and for the right-hand side we have an = {-O,B;'~,}. Comparing (36) with
(6),we have
Formulae (37) and (38) determine the operator of transition in the twcblevel
difference scheme (35).
Equations (1) and (2) are complemented by the boundary and initial
conditions
t"
We assume (see Section 4.2) that the grid elliptic operator A approximates
the differential operator L within the second-order accuracy, i.e.
242 COMPUTATIONAL HEAT TRANSFER
-
J - (A?i - Ln)
=b(~)ri+ + (v- f ) - (u- 4) rAZ + 0 ( r 2 )
If, in addition to (16)-(19), we assume that 9 = f + 0 ( r 2+ lhI2), then for
the error we have the formula
We use the same notation for the pattern with respect to spatial variables
Then the difference scheme (10) is written as
For the latter we can formulate conditions sufficient for the maximum principle
to be satisfied. The condition that all the coefficients in the right-hand side
of (22) are nonnegative results in
Condition (23) is quite logical and does not restrict the time step. Condition
(24) is satisfied for all T only if a = 1, i.e. the maximum principle is satisfied
unconditionally only for the purely implicit scheme.
From the other a from (24) we have
Taking into account the formulae for the coefficients of difference scheme ( l o ) ,
from (25) we obtain the sufficient conditions for the maximum principle to be
satisfied for the scheme with weights, namely
1 minc(x)
75- ( h c 2+ h ~ 2-I)
2(1 - a ) max k(x)
Estimate (26) shows that the maximum principle is satisfied for schemes with
weights o # 1 under severe conditions for the time step T = 0(lh12).
the stability of the scheme with respect to the initial data and right-hand
side. The estimate can be obtained in the norm llyll = max Ily,llccw). In
O<k<N
considering nonstationary problems, we try to obtain the estimates in the
norm of difference solution at one time level. This, in fact, corresponds to
using the maximum principle at a separate time level (for the grid elliptic
operator).
+
In order to estimate the solution at the (n 1)th time level (n 2 0), we
rewrite (22) as
where
Under the formulated conditions (23) and (24) for the parameters of the
difference scheme, we consider the grid elliptic problem (27) and (28). It is
most simple to obtain the estimate of solution to the difference equation (27)
with uniform boundary conditions (g(z, t , + ~ )in ( l l ) ) , which is sufficient for
investigation of convergence of the difference scheme.
For equation (27) with uniform boundary conditions we have (see Corol-
lary 6 in Section 4.3) the estimate
In our case
If (23) and (24) are satisfied, then for such D(x) from (28) we have
Using the difference Gronwall lemma, from (29) we obtain the required
estimate for the difference scheme (27) and (28) with the uniform boundary
conditions, namely
The latter reflects the stability of difference scheme (10)-(12) with weights
with respect to the initial data and right-hand side in the uniform norm.
Recall that the estimate is obtained assuming that the maximum principle is
satisfied (i.e. the conditions (23) and (24) are satisfied for the parameters of
the difference scheme).
In studying the accuracy of difference scheme (10)-(12),we formulate the
corresponding problem for the error 2, = y,-u(x, t,), x E w. From (10)-(12)
we obtain
q X )&+I - Z" A(u2,+1 ( 1 - a ) z n ) = ,
7
+ + *, (31)
X E W , n = 0 , 1 , ...,
with the uniform conditions
Taking into account estimate (21) for the approximation error of difference
scheme (10)-(12), we obtain
schemes, one should store the solution at the (n - 1)th level. Instead of (10)
we use the difference scheme
where iL = u n _ ~under
, the additional conditions (11) and (12). In order to
begin the calculations by the three-level difference scheme, we should know
yl along with yo. In order to find y l , one can use some two-level schemes.
Difference scheme (36) is characterized by three weight parameters 8, a1,
and 0 2 . Two classes of three-level schemes are usually used, each including
only one parameter. First we mention the symmetrical difference schemes
which are written as
The scheme (37) corresponds to the weights 6' = 0.5 and a1 = 02 = a.Taking
into account (19), it is easy to check that the symmetrical three-level scheme
provides the second order of spatial and time approximations.
The other class of three-level difference schemes for the heat equation
corresponds to the purely implicit approximations of the elliptic operator.
In this case the difference equation is
i.e. a1 = 1 and u 2 = 0 in (36). For the approximation error of the scheme (38)
we have
0(-r2
+ + lhI2), a = 1.5,
$ n = { o(7 (hI2), 0 # 1.5.
Thus the class of three-level schemes (38) also includes that providing the
second order of approximation.
It can be shown (Problem 2) that one cannot formulate simple sufficient
conditions of the maximum principle like (23) and (24) for symmetrical three-
level schemes (37) at a # 0.5. The asymmetrical threolevel difference scheme
is more fruitful here.
To investigate the difference scheme (38), let us write it in a form similar
to (22), namely
Hence it can be seen that the maximum principle is satisfied if the conditions
>
26' - 1 0 and 1 - 9 > 0 are satisfied, i.e. at 0.5 5 6' 5 1. However, the
NONSTATIONARY PROBLEMS O F HEAT TRANSFER 247
5.3.6 PROBLEMS
1. C o n s t r u c t t h e two-level difference s c h e m e of h i g h e r approxi-
+
m a t i o n o r d e r ($ = 0 ( r 2 h4))for t h e one-dimension u o n s t a t i o n a r y
p r o b l e m of h e a t conduction
The simplest estimates of stability for the two-level difference scheme follow
from estimates of stability with respect to the initial data. We discuss this
and other issues later. The stability of difference scheme with respect to the
initial data is considered as the uniform stability in the sense of the definition
in Section 5.2.
We study the stability by using the method of energetical inequalities.
Therefore first we obtain the simplest energetical identity for the difference
scheme (1).Let us multiply the latter in a scalar way by 2ryt = 2(yn+1 - y,).
This results in the inequality
Taking into account the formula (3), we transform the second term in the
right-hand side of (3):
T
Yn
1
+
= ? ( ~ " + l yn) - 5Y t .
Substituting (4) into (3) we obtain
is satisfied. The main result of the stability theory for difference schemes is
formulated by the following theorem.
250 COMPUTATIONAL HEAT TRANSFER
If ( 8 ) is satisfied, we obtain
Taking into account that the scheme is stable with respect to the initial data,
we have
are necessary and suficient for the difference scheme (1) and (2) to be p-stable
in HA for A > 0 (in Hg for B > 0 ) .
We prove the theorem by obtaining the corresponding operator inequalities.
In doing so the implicit scheme (1) is reduced to an explicit one, and we
estimate the corresponding transition operator.
We restrict our proof to the case of B > 0 in Hg. Since B > 0, for the
uniform difference scheme (1)we have
where 2, = B 1 f Z y ,and
S=E-rC.
is a self-adjoint operator.
The inequality for the transition operator S, namely
-E<c<- ~ + P E .
7 T
Substituting (15) into (20) and multiplying this from both sides by BIIZ (in
doing so the inequality still holds), we just obtain the two-side inequality (11).
The stability in HA is similarly proved by passing to the explicit scheme
for s, = A1JZy,.
Estimates (21) and (22) follow immediately from the proved general
estimate (23) from Section 5.2 (with MI = 1, (1 . l l l h = 11 . ID, D = A,B).
The estimates of stability of the difference solution with respect to the right-
hand side in other norms are useful. Some of these have been mentioned for
solution of the differential problem in Section 5.1. Similar estimates can be
obtained by making condition (8) more approximate.
NONSTATIONARY PROBLEMS O F HEAT TRANSFER
B > E E + ~ A ,
2
Substituting this into (6) and using condition (23), we arrive at the inequality
Taking into account the difference Gronwall lemma, from this inequality we
obtain the required estimate (24).
Theorem 5. If A > 0 and the inequality
I+
B>- TA,
2
+
Taking into account (25), we choose 0 so that 1 E = 1/(1 - 0 ) . Then (28)
gives the following estimate of the solution on a level:
mn = {%>01
NONSTATIONARY PROBLEMS O F HEAT TRANSFER 255
then the following estimate of stability with respect to the initial data holds
for the difference scheme i n (29) and (30):
We should check the inequality (36) and condition A > 0, in order to obtain
the estimate (40). If inequality (39) is satisfied, then, by virtue of (37), we
have A > 0. From the definitions of operators A and B (see (33) and (34)),
we have
B--A=
T
2 I B
-T ( R - + A )
((B-$A)V,V) = ( B V I , U I ) + T ( ( R - + A ) ~ ~ , W I )
- T ( ( R - : A ) U I , ~ ,=
) (BV,,VI)
(38) is not only sufficient, but also necessary for the difference scheme (29)
and (30) to be stable.
The presented result can be obtained on the basis of corresponding
energetical identity for three-level difference schemes, and by estimating the
norm of transition operator in writing the difference scheme a s in Problem 2,
Section 5.2. Anyway this entails cumbersome calculations, and therefore we
restrict ourselves to the considerations presented.
is applied to difference scheme (29). The simplest estimate for the norm of
a new grid function vn+l like [lun+l[[5 [I D , [ [ corresponds to the estimate
Ily,+l(l 5 plly,ll, which is related to the pstability. A similar situation occurs
in using norms in HZ.
By taking into account (41), the three-level difference scheme (29) is
transformed to
g ""+I - "n-1 +72E U"+1 - 221, + -
ZT r2
Un-1
+ Av, = &,
(42)
n = 1 , 2 , ... .
The operators in (42) are determined in terms of the operators of initial
difference scheme (29) by means of the relations
Let us see how the conditions of stability for difference scheme (42) are
transformed into operator inequalities for the initial difference scheme (29).
A
By taking into account (43), the condition > 0 is reduced to
NONSTATIONARY PROBLEMS O F HEAT TRANSFER
where
,vn - n"-1 (49)
IlY"+'II,- 5 PllY"II,-.
Thus we can formulate the following statement.
258 COMPUTATIONAL HEAT TRANSFER
T h e o r e m 8. I f A > 0 and the inequalities (38) and (39) are satisfied, then
the estimate "
IIynt'll~5 llylll~ C
+ IlB~l~nll~ (54)
k=1
and it only remains for us to transform the second term in (55). We have
NONSTATIONARY PROBLEMS O F HEAT TRANSFER 259
Substituting (56) and (57) in (55), we obtain the proved estimate (54).
As for two-level difference schemes, we can extend the set of estimates of
stability with respect to the right-hand side by using various norms for the
right-hand side of (29) and (30).
5.4.7 PROBLEMS
1. O b t a i n t h e estimate of stability with respect t o t h e initial d a t a
i n H s by t h e m e t h o d of energetical inequalities for t h e difference
scheme (1) a n d (2) w i t h positive operators A a n d B assuming t h a t
t h e condition (8) is satisfied.
Solution. In order to obtain a new energetical identity, let us multiply the
uniform difference scheme (1) in a scalar way by 27G, namely
and therefore estimate II$IIB 5 llylle follows from (59), i.e. the difference
scheme (1) and (2) is stable in H s .
2. O b t a i n t h e conditions of pstability for t h e two-level difference
scheme (1) a n d (2) o n t h e basis of transformation (41).
Solution. The difference scheme (1) is written as
260 COMPUTATIONAL HEAT TRANSFER
Here the operator A is defined on the set of grid functions being zero on aw
and
is now
D + ( o - i)T A 2 0. (9)
>
Let D 0; then condition (9) is satisfied for all u 2 0.5 independent of the
time step (the time stability). If we know additionally a positive constant 6
such that
A AD, < (10)
then condition (9) is satisfied at
This inequality can be interpreted as a condition for the time step for u < 0.5,
namely
1
T<To=
A(0.5 - a)' (1'4
Thus, for example, for the explicit scheme the time step is like
We specify the obtained condition for the difference scheme (1) and (2)
written in the form (5) and (6). In this case the mentioned properties of
operators A and D clearly hold. Therefore the difference scheme (1) and
(2) with weights is unconditionally stable with respect to the initial data at
>
u 0.5 in H A . This, in particular, holds for the symmetrical scheme. Earlier
(see Section 5.3) we proved the unconditional stability (in the uniform norm)
for only the purely implicit scheme.
If o < 0.5 the scheme with weights is conditionally stable. Let A be the
constant in the inequality
<
A Ab(x)E, (14)
262 COMPUTATIONAL HEAT TRANSFER
then the stability condition is as in (11) and (12). Making (14) more
approximate we can obtain explicit constraints for the time step depending
on the spatial parameters of the grid. Let co = min c(x) and ~2 = max k(x).
Taking into account the inequalities
b(z)E 2 COE, A 5 KZA5 KzA,&, (15)
where A is the Laplace grid operator, and A,, is its maximum eigenvalue.
For A,. (see Section 4.7) we can use the estimate
A,, < 4(hy2 + h i 2 ) . (16)
Taking into account estimates (15) and (16), we see that the inequality (14)
is satisfied with
K2
A = 4-(hc2 + hy2). (17)
CO
From (12) and (17) we obtain the following constraints for the time step:
The presented estimate (for u < 0.5) can be compared with the estimate of
stability of difference schemes with weights (see (26) in Section 5.3) in uniform
norm
If u < 0.5, the limit time step in estimate (18) exceeds the time step in (19)
(the latter can be increased by (1 - u ) / ( l - 20) times). Estimate (18) reflects
(like (19)) the essential dependence of the maximum step on h (T = O(lhI2)).
X E ~ , n = 0 , 1 , ...,
zo(x) = 0, x E W . (21)
For the approximation error we have (see Section 5.3)
$L(x) = O ( r Y+ llhl12). (22)
where Y = 2 for u = 0.5 and Y = 1 if 0 # 0.5.
The corresponding estimates for the error can be obtained by using the
results on stability of the difference scheme (20) and (21) with respect to the
right-hand side (see Theorems 3-5 in Section 5.4). Here we use an analogue
of Theorem 4 for the difference scheme (5) and (6).
NONSTATIONARYPROBLEMSOFHEATTRANSFER 263
From this we obtain the proved estimate (23) of stability for the difference
scheme (5) and (6).
Now let us apply the theorem to the problem for error (2) and (21). Taking
into account (6), we obtain
Taking into account (22), the estimates (24) and (25) ensure that the difference
scheme (1) and (2) converges with second order with respect to spatial
variables and with order v with respect to time.
By checking directly we see that in this case we should take the coefficients
The operators B, R, and A are self-adjoint in (27) and (28), with A > 0. The
stability of the scheme with respect to the initial data is ensured (Theorem 6
from Section 5.4) by the inequalities
We consider the difference scheme (27) and (28) under the additional
>
constraint 8 0.5. The other case (8 < 0.5) should be considered separately
(see Problem 2) and is not of interest for the problems on heat conduction in
hand. Taking into account the operator inequality (14), from (28) and (29)
we obtain
For the latter the conditions of stability (30) and (31) are reduced to the
simplest inequality
1
u>z. (33)
The class of purely implicit threelevel schemes corresponds to the parameters
NONSTATIONARY PROBLEMS OF HEAT TRANSFER 265
where
1
& = R + - B27
,
It follows from estimate ( 3 6 ) that the accuracy of solutions at the zero and first
levels should be consistent with the approximation of the threelevel scheme.
Let us obtain a simpler estimate for the norm of approximation error.
Taking into account ( 3 7 ) and the positive definiteness of operators B and
R , we have
5.5.4 PROBLEMS
1. S t u d y t h e convergence of difference scheme (5) for A = A', if
By virtue of this, the difference scheme (5) and (38) is unconditionally stable
>
at a 16.
The latter differs from (39)by replacing y, in the diagonal part of the operator
A by a half-sum of solutions from the (n+l)th and ( n - 1 ) t h levels. It is written
in the canonical form (27) with
The scheme (27) and (41) is stable with respect to initial data, provided that
the inequality
R-+A=~(zD-A)>o (42)
NONSTATIONARY PROBLEMS O F HEAT TRANSFER 267
is satisfied. For the grid elliptic operators A determined by (3) and (4) we
have the estimate A < 2 0 , which has been proved in Section 4.7. This implies
that the condition of stability (42) is satisfied. The disadvantages of difference
scheme (40) are caused by the fact that it has the conditional approximation,
+ +
namely $, = 0 (.r2 lhI2 ~ ~ I h l - i.e. ~ ) , = O(lhI2) at T = O(lhI2), as for
the conventional explicit scheme.
au
-=C-axu ~ 0 ,t > O ,
at ax; '
o=l
For the asymptotically developed stage of regular heat transfer the solution
is specified by the formula
where XI is the first eigenvalue, and the eigenfunction wl(x) corresponds to it.
268 COMPUTATIONAL HEAT TRANSFER
holds. The latter holds if the corresponding difference scheme is pstable with
are satisfied.
Let us check the conditions (13) of pstability for the scheme with weights
( 1 1 ) . The latter can he written in the canonical form (12) with the operators
First we consider the right-hand side of the two-side inequality ( 1 3 ) .For (14)
we have
A< l + p ~ + ~ ( l + ~ ) ~ . (15)
T
since sinh(P) > p for fi > 0. Thus we cannot point out the weight u
under which both the conditions (16) and (18) are satisfied, i.e. no two-level
difference scheme with weights (11) is unconditionally asymptotically stable.
We can, however, mention some classes of conditionally pstable schemes
with weights. Along with (17), for the operator A we have the upper estimate
A 5 AE, (19)
In this case
where 17 = 6/A.
The admissible time steps are determined by the condition 01 - uo 2 0, i.e
by the inequality
being nonnegative.
In investigating the nonnegativeness of functions, we use the statement
proved in Problem 1 (see later). In our case we have
condition for p, we can obtain stable schemes under weak constraints for the
time step.
Let us consider the purely implicit scheme as an example. In this case
inequality (16) (and even more, inequality (20)) is satisfied. Let us put
p = exp(-y&), where 0 < y < 1 (less restrictive constraints for the solution
decrease). Then let us choose the maximum step TO so that the inequality (18)
is also satisfied. This is the case for u = 1 if
Table 1.
7 0.5 0.75 0.9 0.95 0.975 0.99
r 1 2.513 0.734 0.230 0.107 0.052 0.020
5.6.4 PROBLEMS
1. S h o w t h a t f(x) 2 0 for all x 2 0 if
is satisfied.
Solution. In order to show this, it is sufficient to consider the expansion of
function f ( x ) in the Taylor series, namely
Solution. The right-hand side of (13) is satisfied for all p > 0, and, taking
into account (31), the left-hand side can be reduced to
Taking into account estimate (7) and p = exp(-6r), we obtain the inequality
$ p2 + p 2 exp(p) - 1, p = 67.
The latter is not satisfied for any p. As in the purely implicit scheme we can
< >
take p = exp(-y6~) at ip 1. The function ~ ( y ) 6.r is given in Table 2.
Table 2.
Y 1 0.5 0.75 0.9 0.95 0.975 0.99 1
In (1) V is the relaxation parameter of the heat flow. The closing relations are
the initial and boundary conditions
where, taking into account the uniform boundary conditions (3), the value
determines the norm of the solution to (1)-(5). Applying the Gronwall lemma
(see Section 5.1) to (8), we obtain the estimate
The difference scheme in hand differs from that with weights for the heat
equation only by the additional term V T - ~ ~ ( X )inE the operator R, which
improves the stability conditions.
The operators B, R, and A in (12) are self-adjoint, with A > 0.Let us check
the conditions of stability with respect to the initial data for the scheme ( l l ) ,
namely
B ~ O , R > ~ A . (13)
It follows from (12) and (13) that the scheme (11) is stable if
NONSTATIONARY PROBLEMS O F HEAT TRANSFER 277
It follows, in particular, from (14) that the scheme is absolutely stable under
the conventional (see Section 5.5) constraints for the weights, namely
The norms of the difference solution are specified according to the general
theory of stability for difference schemes (Section 5.4).
5.7.3 S Y M M E T R I C A L SCHEMES
Among schemes with weights we select a oneparametric family of symmet-
rical difference schemes for which
In this case difference scheme (11) has the approximation error $, = O(T' +
lhI2).
, , .
If (16), then the stability condition (14) is like
we write the symmetrical difference scheme (11) and (16) for a = 0.25 as
6 + u = -2( w
T " T .
-w), i, - u = qw
2
+w),
where
a = v(o(~g,,)+ ( A F, q) . (24)
Inequality (23)is the differenceanalogue to (9), and the norm (24)is consistent
with the norm (10) of the differential problem.
The estimate obtained ensures the stability with respect to the initial data
and right-hand side. Therefore we have shown that the symmetrical scheme
converges to the precise solution in quite simple norms. The same holds for
the corresponding difference scheme for the conventional heat equation (V = 0
in (11)).
5.7.4 PROBLEMS
1. For t h e hyperbolic h e a t equation (1) a n d (2) c o n s t r u c t a
difference scheme providing t h e f o u r t h o r d e r of t i m e approximation.
Solution. Expanding the equation at the point ( x ,t,), x E w , we have
NONSTATIONARY PROBLEMS O F HEAT TRANSFER
Therefore
au
c(x) at+Vc(z)-
aZu+ Lu - f (x, t)
at2
Due to the constructions, the difference scheme (26) has the approximation
+
error 0 ( r 4 lhl2).
2. Investigate the stability of the difference scheme (26).
Solution. Scheme (26) is written in the canonical form with the operators
By taking into account the inequality A < Ab(x)E, the condition of stability
can be transformed to
280 COMPUTATIONAL HEAT TRANSFER
holds. Hence it follows that the difference scheme in hand is stable for
V 5 ll(2A).
where
as the initial one. The operator A in (5) is determined on the set of grid
functions being zero on aw, and, for example,
A < Ab(x)E.
We construct absolutely stable difference schemes based on the conditional
stable scheme (5)-(7). According to the formulated principle, we write (5) in
the canonical form of two-level difference schemes, namely
282 COMPUTATIONAL HEAT TRANSFER
The proof is based on checking the necessary and sufficient condition (12).
For the scheme (13) we have
Taking into account (9), and substituting (15) into (12) with R = A, we
obtain
Hence it follows that for R = A' and a > r2/(16%), ~g = minc(x), the
regularized scheme (13) is absolutely stable. An example of such a scheme
with b(x) = 1 has been considered earlier (see Problem 1, Section 5.5).
We have improved the properties of explicit scheme (5) by perturbing the
grid operator B in the two-level difference scheme (10). The operator A could
also be perturbed here (see Problem 1).
Hence it follows that the regularized difference scheme (13) with the regula-
rizator R = A is stable if a 5 n2(7/2 - A-I). This condition is more rigid
than (14).
284 COMPUTATIONAL HEAT TRANSFER
and therefore the scheme is stable for a 2 nzr2/(16co). Thus in this case the
regularization parameter a is proportional to the squared maximum of heat
conductivity coefficient.
b(x)y.
t
+ Ay = ip.
(17)
This scheme provides the second order of time and space approximation and
is written in the canonical form
By.
t
+ T'RR, + Ay = ip (18)
with the operators
The scheme (18) and (19) is absolutely unstable because the second inequality
in (20) is not satisfied. Therefore we can regularize it by perturbing the grid
operator R.
Starting from (17) we construct the regularized scheme
Let us mention some variants in choosing the regularizators. The simplest way
is to put R equal to D, where D is the diagonal part of the grid operator A.
Taking into account (Section 4.7) the inequality A < 2 0 , for R = D we obtain
R - ' 4A = ~ D - $ A > ( ~ - $ ) D .
NONSTATIONARYPROBLEMSOFHEATTRANSFER 285
5.8.5 PROBLEMS
1. C o n s t r u c t a regularized difference scheme o n t h e basis of
explicit scheme (13) by regularizing t h e o p e r a t o r A.
Solution. We consider the scheme
1
a <
A min c ( x )
The explicit scheme (23) is conditionally stable. It follows from (24) and (25)
that the maximum time step (for optimal a) is constrained by the condition
8 maxc(x)
r<-
A min C(X).
Comparing (26) with the stability condition (T 5 2/A) for the explicit scheme
(13), we see that the regularized scheme (23) for R = A2 admits a larger time
step.
2. Consider t h e regularized scheme (21) w i t h t h e regularizator
R = E ( t h e Dufort-Frankel scheme).
Solution. In this case the regularized scheme has the canonical form with
The condition (20) results in the following constraints for the regularization
parameter: a > max k(z) (h;' +
h;').
5.9 N o n l i n e a r N o n s t a t i o n a r y Problems
where
Let us mention conditions under which the problem bas a unique solution.
The investigation is based on the corresponding statement on the uniqueness
of the solution to the corresponding linear problem. Assume that u(x,t) in Q
satisfies the parabolic equation
under the corresponding initial and boundary conditions. For the difference
of solutions w(x) = u2(x)= -ul(x) we obtain (see Section 4.9)the boundary
value problem
Here we denote
are satisfied. Thus the solution of the nonlinear problem of heat conduction
(1)-(4) is unique in the class of bounded nonlinear coefficients.
the solution a t the new time level being found by solving a linear difference
problem. In the simplest scheme the coefficients are taken from the previous
time level. We can cite the difference scheme
+
as an example. This scheme, clearly, has the approximation error O(T lhI2).
The further development of scheme (15)-(17) is that with quasi-linearized
right-hand side, namely
The latter is linear, but has a larger reserve of stability with respect to the
nonlinear right-hand side.
Linearized schemes can be constructed on the basis of predictor-corrector
difference schemes. In this case difference scheme (15) is linearized in the
following way. The explicit scheme
is used at the predictor stage for calculating the coefficients and the right-hand
side. Therefore the correction stage may correspond to the scheme
The correction stage may also use the linearization scheme (18). Therefore, it
is worthwhile to use the scheme
290 COMPUTATIONAL HEAT TRANSFER
instead of (20).Instead of (21) we can also use another variant of the quasi-
linearization, namely
If 0 = 1.5 this implicit scheme provides the second order of time and space
approximation, and the corresponding linear scheme is unconditionally stable.
The difference schemes presented are direct analogues to the implicit schemes
for the linear boundary value problem of nonstationary heat conduction.
In this case the new iterative approximation wk+l to yntl is found by solving
the boundary value problem for the difference equation
The linearized difference scheme (15) coincides with the iterative scheme (24)
if we restrict ourselves to one iteration.
Equation (24) can be modified by linearizing its right-hand side and using
the equation
In doing so the grid elliptic operator for determining the new approximation
remains self-adjoint.
The Newton method is widely applied in realizing nonlinear difference
schemes for nonstationary problems. In this case the new approximation is
found from the difference equation
Here (see Section 4.9) the grid elliptic operator A'(wk) is like
is satisfied for the time step; then the maximum principle is satisfied for the
problem (28)-(30) (see Section 5.3), and the latter has only the trivial solution.
If constant v is negative in (27) then the solution is unique for any time step.
In order to investigate the accuracy of the difference scheme (26), (16)
and (17), we formulate the corresponding problem for the error zn(x) =
y,(x) - u(x, tn), x E a. Let &(x) be the approximation error, with &(x) =
O(T+ lhj2), lhI2 = hf + h;. The solution error satisfies the equation
and uniform boundary and initial conditions (29) and (30), respectively.
Assume that v 5 0 in (27); then (see Section 5.3), from the maximum principle
we obtain the estimate
Thus we have proved that the purely implicit difference scheme converges if
(27) is satisfied with a nonpositive constant v.
5.9.6 PROBLEMS
1. Show that the a priori estimate
I l ~ ( ~ , t ) l l c (5q max
) { l l ~ ( ~ ~ t ) l l cIluo(~)llccn)}
(r),
+ \v\-l\\f(">t , O ) l \ ~ ( ~ )
holds for the problem (1)-(4) with c = c(x) and k = k(x) if the
inequality
-a(fx , t , u ) < v < O
au
is satisfied.
Solution. Using the expansion
Further we consider separately the problem for the uniform equation (the
stability with respect to the boundary and initial conditions) and that with
uniform boundary conditions (the stability with respect to the right-hand
side).
2. C o n s t r u c t t h e linear s c h e m e of predictor-corrector providing
t h e second o r d e r of t i m e a n d s p a c e a p p r o x i m a t i o n for t h e p r o b l e m
of h e a t conduction (1)-(4) w i t h c = c(x) a n d k = k(x).
Solution. The predictor stage is used for calculating the nonlinear right-
hand side at the time t , + ~ ~ ,The . corrector corresponds to using the linearized
symmetrical scheme
5.10.2 LITERATURE
1. Bakhvalov N. S., Zhidkov N. P. & Kobel'kov G. M. (1987) Numerical Methods [in
Russian]. Nauka, Moscow.
2. D'yakonov E. G . (1972) Difference Methods for Solving Boundary Value Problems.
Part 2. Nonstotionory Problems [in Russian]. Izd-vo Mosk. un-ta. Moscow.
3. Fedotov E. M. (1987) Difference Schemes for Nonlinear Nonstationary Problems
[in Russian]. Izd-va Kazan. un-ta, Kazan'.
4. Friedman A. (1964)
~,
Cliffs, New Jersey.
-
Partial Differential Eouotions o f Parabolic Tuue. -
". Endewood-
5. Karchevskii M. M. & Lvashko A. D. (1976) , . Differrnce
- Schemes for Nonlinear
Problems of Mathematical Physics [in Russian]. Izd-vo Kaz. un-ta. Kazan'.
6. Krein S. G. (1967) Linear Differential Equations i n Bonoch Spnces [in Russian].
Nauka, Moscow.
7. Ladyzhenskaya 0 . A. (1973) Boundary Value Problems of Mathematical Physics
[in Russian]. Nauka, Moscow.
8. Ladyzhenskaya 0. A , , Solonnikov V. A. & Ural'tseva N. N. (1973) Linear and
Quasilineor Equotions of Parabolic Type [in Russian]. Nauka, Moscow.
9. Marchuk G. I. (1975) Methods of Numerical Mathematics. Springer-Verlag, New
York.
10. Richtmyer R. (1957) Difference Methods for Initial-value Problems. Interscience,
New York.
11. Richtmyer R . & Morton K. (1967) Difference Methods for Initial Value Problems.
Springer, New York.
12. Samarskii A. A. (1971) Introduction to Theory of Difference Schemes [in
Russian]. Nauka, Moscow.
13. Samarskii A. A. (1983) Theory of Difference Schemes [in Russian]. Nauka,
Moscow.
14. Samarskii A. A. (1987) Introduction to Difference Methods [in Russian]. Nauka,
Moscow.
15. Samarskii A. A,, Galaktianov V. A., Kurdyumov S. P. & Mikhailov A. P. (1987)
Models with Peaking i n Problems of Quasy-linear Porobolie Equations [in Russian],
Nauka, Moscow.
16. Samarskii A. A. & Gulin A. V. (1973) Stability of Difference Schemes [in
Russian]. Nauka, Moscow.
17. Samarskii A. A. & Gulin A. V. (1989) Numerical Methods [in Russian]. Nauka,
Moscow.
18. Saul'ev V. K. (1960) Integmtion of Equations of Pombolic Type by the Grid
Method [in Russian]. Fizmatgiz, Moscow.
19. Tikhonov A. N. & Samarskii A. A. (1972) Equations of Mathematical Physics
[in Russian]. Nauka, Moscow.
20. Vladimirov V. S. (1976) Equations of Mathematical Physics [in Russian]. Nauka,
Moscow.
21. Wasow W . & Forsythe G. (1960) Finite-difference Methods for Partial Differrn-
tial Equations. John Wiley & Sons, New York.
Economical Difference Schemes for
Nonst ationary Heat Conduction
Problems
Two- and three-level difference schemes for the two-dimensional heat equation
have been classified in Chapter 5. When implicit difference schemes are used
to find the solution on a new time level it is necessary to invert a grid elliptic
operator, and various direct and iterative methods of solving grid elliptic
problems are employed to this end. Computational costs can significantly
exceed in this case the costs involved when the solution is determined by an
explicit scheme.
For explicit schemes, the number of arithmetic operations per grid node is
independent of the total number of the nodes, i.e. the computational costs
are asymptotically optimal. Difference schemes that possess this property are
called economical difference schemes. However, the magnitude of the time
step for explicit difference schemes is heavily restricted by considerations
of stability. The computational costs for the implicit difference schemes
considered above are in general higher in the case of multidimensional
nonstationary problems, however, some of these schemes are unconditionally
stable. Therefore, we are faced with a problem of constructing difference
schemes that would he as economical as explicit schemes and would retain
the unconditional stability inherent in some of the implicit schemes. This
chapter deals with this problem of constructing stable economical difference
schemes for the heat equation.
We consider here how to implement implicit difference schemes for multi-
dimensional nonstationary problems on a computer. When solving grid elliptic
300 COMPUTATIONAL HEAT TRANSFER
problems iteratively on the top time level, the convergence rate depends not
only on the number of space steps but also on the time step.
The well-known schemes of alternating directions are simple examples of
economical difference schemes for the heat equation. It should be noted that it
is difficult to extend the method of alternating directions to three-dimensional
problems with nonseparable variables.
Factorized schemes, in which the operator is represented on the top level as a
product of two economical operators, are widely used to construct economical
difference schemes. Such schemes can be constructed by using the principle of
regularization of difference schemes.
The economical schemes we consider here belong to the class of additive
difference schemes. A feature of these schemes is that the main grid operator
(the heat conduction operator A in our case) can be represented as a sum of
several simpler operators, e.g. one-dimensional grid operators. The notion of
a sum approximation is used to construct and analyse the general classes of
additive difference schemes.
Two- and three-level economical difference schemes for the heat equation
are analysed on the basis of general results of the theory of additive schemes.
Splitting into the sum of one-dimensional grid operators leads us t o locally
one-dimensional difference schemes.
+
The accuracy of this scheme is O ( r Y lhI2),where v = 2 if o = 0.5 and v = 1
otherwise.
Let A be a constant in the inequality
ECONOMICAL DIFFERENCE SCHEMES.. . 301
then the scheme with weights (1)-(3) is stable (see Section 5.5) provided that
To find the solution on a new time level from (1)-(3),we have to solve the
grid elliptic problem ( a > 0) for the equation
equipped with the boundary condition (2).The right-hand side in (6)has the
form
The grid elliptic problem (2),(6) explicitly involves the time step T and the
weight parameter a. Therefore, it is worthwhile analysing the convergence
rate of iterative methods for the problem (2),(6) versus these parameters.
Hence, we have
A=A'>O (9)
for T > 0 and a > 0.
In order to solve the problem in (7) approximately, we shall consider, for
definiteness, a two-level iterative process with the Chebyshev set of iterative
parameters ~ k namely
,
Provided that
B=B*>O (11)
302 COMPUTATIONAL HEAT TRANSFER
we have (see Section 4.6) the following estimate for the number of Chebyshev
iterations (9)-(11) necessary to achieve the prescribed relative accuracy E :
n 2 no = ln(2~-')/ln(~;'), (13)
where
We have (13), (14) and, therefore, when solving the problem (7), (8) iteratively
we are most interested in how E depends on T and u.
Let us consider the simple example of an iterative process in which the
operator B (see Section 4.7) is the diagonal part of the operator A. For the
grid operator (8) the diagonal part D = d(x)E has the form
Let us give some estimates for the constants "i,, a. = 1,2, in (12), given the
selection of the operator B = D.
As previously (see Section 4.7), we have the equality
where
u(a,v~JZm +-
' 4 ~ )v = -d(x)
T
> XEw,
va(x) = 0, 2, = 0,1,.
ECONOMICAL DIFFERENCE SCHEMES.. . 303
where A is the grid Laplace operator on a square grid. Taking into account
the inequality
where 6 = O(1) is the least eigenvalue, we derive from (17) the following
expression for yl:
i.e. the number of iterations does not depend explicitly on the number of
nodes. Of course, the relative error E of the iterative solution should be
adjusted to the approximation error, and, thereby, to the grid parameters.
However, in any case this dependence is weak (for E = TP, /3 > 0, nO(c) =
O(ln(r-I)) and for this reason the iterative method can be ascribed to the
class of economical methods for solving nonstationary problems, given this
relation of space and time steps.
304 COMPUTATIONAL HEAT TRANSFER
To this end we make use of the decomposition of the operator A into one-
dimensional operators, namely
Then, taking into account (8), we represent the operators A,, a = 1 , 2 , in the
The two components A1, Az of the operator A in (22) obey (see Section 4.7)
inequalities similar to (25), namely
ECONOMICAL DIFFERENCE SCHEMES.. . 305
where -
6, = 0 ( 1 ) , i,= O(hZ,), a = 1,2. (27)
It immediately follows from (24)-(26) that
min b ( x ) min b ( x )
61 =p-
T
+a&, 62=(1-P)7 + 2 2 , (28)
max b(x) max b ( x )
Al=B a , 1 - +ax2. (29)
where
A = A1 + A2, A1 = A;
Taking into account (30), we derive
where 6 and A are positive constants that appear in the operator inequalities
To derive these estimates for the operator in (30), we take advantage of the
corresponding estimates for the Laplace operator (Section 4.7):
where
s" = 0 ( 1 ) , = 0(lhl-2). (3'3)
Taking into account (30), (32), and (35), we derive the following for the
constant 6 in the inequality (34):
and, hence,
2%
~ = -T + a k ~ A
It follows from (37), (38),and (36) that 17 = & / A= 0 ( l h l - 2 ~ )and
, for the
number of iterations in the alternating triangular method we therefore derive
On the basis of (44) we can infer that the rate of convergence of iterative
process (12), (42) is independent of the space and time steps and is solely
determined by the difference in heat conductivity and heat capacity.
6.1.6 PROBLEMS
1. Formulate t h e grid elliptic p r o b l e m o n t h e t o p level for t h e case
i n which three-level schemes w i t h weights for t h e h e a t e q u a t i o n are
used.
Solution. For Y,+~ we obtain equation (7) with the operator (see Section 5.3)
Hence, the above arguments about the dependence of the rate of convergence
of the iterative process on the time step remain valid.
ECONOMICAL DIFFERENCE SCHEMES.. . 309
The following representation is, for example, used in equation (5) for the
operator A:
2
Av = C A,u, a = 1,2, (8)
a=1
A,u = - (
with b(x) = c(x) for x E w.
Difference schemes in the method of alternating directions for heat equation
(I), (2) are based on the representation of the operator A with respect to
space variables in the form of the sum of two operators A1 and A2, each of
which is one-dimensional. Difference schemes based on the inversion of one-
dimensional operators belong to the class of economical schemes, because the
computational costs per node (three-point Thomas algorithm, band matrices)
are independent of the total number of nodes.
The classical difference scheme of alternating directions (the Peaceman-
Rackford scheme) for the problem in (5)-(7) consists of two steps. First, given
some y,, we determine an auxiliary grid function (which we will denote by
~ , + ~ from
j ~ )the equation
+
If we interpret yn+l12 as a solution at the time t,+l/2 = t,, r / 2 , it can be
noticed that (9) amounts to finding the solution by a purely implicit scheme
with respect to the variable xl (the operator Al) and by an explicit scheme
with respect to the variable x2 (the operator Az).
It is clear that in this transcription the second step in the method of
alternating directions corresponds to the use of the equation
Thus, the second step corresponds to the use of an explicit scheme with respect
to the first variable and a purely implicit scheme with respect to the second
variable.
The implementation of the scheme of alternating directions (9), (10) cor-
responds to determining yn+,lz and y,+~ from the equations
ECONOMICAL DIFFERENCE SCHEMES.. . 311
The notation (11) indicates that the solution is found by inverting the
appropriate one-dimensional grid operator (using the Thomas algorithm) first
with respect to one variable (one direction) and then with respect to the
other variable (the other direction). For this reason, the scheme of alternating
directions (9), (10) is sometimes called the longitudinal-transverse difference
scheme for the heat equation.
if D = B-'
We have
Adding (13) to (14) and taking into account (12) for o = 0.5, we obtain
On the basis of a grid analogue of the Gronwall lemma, we derive the following
estimate of the stability of the scheme of alternating directions (9), (10)with
respect to initial data and right-hand side in HD for D = B-':
Put in (19),(20)
$;= -141
%,+I + u, - AZU" + ip,
- b(z)U"+1 -fin 7
-- hz
U"+l -
+ 0(T2).
T 2 7
ECONOMICAL DIFFERENCE SCHEMES
we obtain
$2" +
= $; = 0(r2 lh12). (22)
Thus, given a special definition of the intermediate solution (see (21)), the
difference scheme of alternating directions (9), (10) is second-order accurate
in space and time.
To analyse the errors, we consider the grid problem (I?), (18). Taking into
account (22) and using the estimate (16) for the exact specification of initial
conditions, we have
where D = B-'. On the basis of the estimate in (23) we can draw a conclusion
that the scheme of alternating directions (9), (10) converges at a rate of
0 ( r 2 + [hi2)in the appropriate norm.
In the first step (24), the heat equation is approximated over the entire time
interval, and the second step (25) is only introduced for stability reasons.
Therefore, the schemes like that in (24), (25) are being called now stabilization-
correction schemes. It is easy t o verify that the scheme (24), (25) has an error
of O(T + Ih12) and is absolutely stable (because it is an instance of factorized
schemes considered in Section 6.3).
Difference schemes of alternating directions for problems with alternating
coefficients, like the problem in (1)-(4), belong to the class of unconditionally
stable. However, construction of such schemes for three-dimensional problems
encounters considerable difficulties. In order to explain this, it is sufficient to
314 COMPUTATIONAL HEAT TRANSFER
dv
-
dt
+ Av = 4 ( x ,t ) , x E w, 0<t 5T (26)
When using the method of alternating directions for equations (26), (27), it
is crucial that the operators A,, a = 1,2,3, are self-adjoint, nonnegative, and
pairwise-commutative, i.e.
6.2.5 PROBLEMS
1. Write down the scheme of alternating directions for heat
equation (I), (2) with the nonhomogeneous boundary condition
u ( x ,t ) = g ( x ,t ) , x E r. (29)
In these conditions, the solution error satisfies the above estimates, and the
scheme (9), (10) with boundary conditions (30) will converge and will be
second-order accurate in time and space.
2. P r o v e t h e stability of t h e scheme of a l t e r n a t i n g directions for
t h e h e a t conduction p r o b l e m i n a homogeneous m e d i u m , using t h e
g e n e r a l t h e o r y of stability of difference schemes.
Solution. In a homogeneous medium, the scheme of alternating directions
has the form (see (26))
where
2
Now we only have to verify the stability condition. Taking into account the
commutativity and positiveness of the operators A,, a = 1,2, we have
The factorized scheme is obtained when we select the operator B in the form
+
B1 = b(x)E U T A I ,
(4)
+
Bz = b-I (x)(b(x)E a ~ A 2 ) .
Now we add and subtract from the right-hand side of (6) the term (20)-'B2y,
to get
2rJ - 1 1
Bzyn+t = -Bzy,
2u
+
- (Bz - 2u~B;~A)y,.
20 (7)
Taking into account (2), (4), we derive
where, taking into account (a), we have the following representation for the
operator Q:
Hence we represent Q as
Q = sls;,
318 COMPUTATIONAL HEAT TRANSFER
where
We can give a more convenient representation for the operators S , from (12),
namely
+
S , = ( E CT~[,)-'(E + CTL),
(13)
i[ = b - 1 2 ( ) ~ b - 1 2 ( ) a, = l , 2 .
-
Because of ( 2 ) ,the operators A, possess analogous properties, i.e.
Now we shall formulate the following assertion, known as the Kellog lemma.
Lemma. Let an operator C 2 0 in H . Then the following inequality holds
whatever 2 0 :
I~(E +
- P C ) ( E ac)-'115 1. (15)
+
Let S = (E - p C ) ( E +PC)-' and G = ( E PC). Then (15) is equivalent
t o the operator inequality J = E - S'S 2 0. We have
and the inequality in (15) therefore takes place for nonnegative operators C .
Due to the lemma, from (13), (14) we derive IISJ 5 1, a = 1,2. Because
of ( l l ) ,the norm of the operator Q is also less than unity, and it follows from
>
(10) that I I v , , + ~ ~ ~ 5 IIwnll, which holds for all c 0.5. Thus we arrive at the
desired estimate (9) for the difference solution and we have established the
unconditional stability of the factorized scheme in (1)-(4).
class of problems in which the grid function b(x) is constant is important and
essential.
In a more general case we employ a simple transformation of the equation
dw
+
b(z) .;ii. Av = $(x, t), x E w, 0 <t 5T
to the form in which the factor that multiplies the time derivative equals
unity, while preserving the operator's properties of being self-adjoint and
nonnegative with respect to space variables. To this end it is sufficient t o
( ~ ,x E w . After this the
introduce a new grid function w(x, t) = b 1 / 2 ( ~ ) w t),
equation takes on the form
where
A' = b-'/2~b-'iz, $'(x, t) = b - ' / 2 ( ~ ) $ ( ~t).
,
Because of (2), the operator A' possesses the following properties:
Let the scheme in ( I ) , (19) be stable, i.e. let the condition (5) be satisfied.
On the basis of this scheme we construct a economical scheme (which is also
stable) with the factorized operator
Note that we restrict consideration to two terms only for simplicity, and a more
general case in which the operator R is decomposed into a greater number of
operators can be considered in exactly the same way.
Let R be the sum of economical operators
320 COMPUTATIONAL HEAT TRANSFER
Taking into account (19), (21) we construct the factorized operator (20) with
Let us show that if the original scheme ( I ) , (19) is stable and the operators
R p , 0 = 1,2, are self-adjoint, nonnegative (Rp = R; > 0, 0 = 1,2) and
commutative (R1R2= R z R l ) , then the factorized scheme (2), (22) is also
stable. Then, the principle of regularization would allow us to construct
economical stable difference schemes given a stable inefficient difference
scheme.
In order to prove the above it is sufficient to verify the inequality (5).
The operators R U , 0 = 1,2, are commutative and, hence, we have RlRz =
RzRl 2 0 and from (20)-(22) we obtain
The analysis of the regularized scheme ( I ) , (19) shows that for stability it
is sufficient, for example, that the regularization parameter o satisfies the
condition 7
a > nz-.
-2 (24)
In the following we shall restrict ourselves to the simplest case in which
"
b(x) = 1. We use the following representation for the operator A:
Note that there exists another possibility of decomposing (21) in the case
considered. Taking into account (23), the stability conditions will be satisfied
for a self-adjoint nonnegative opemtor R if in (21) we have
The implementation of the factorized scheme (24), (25) corresponds t o the use
of the so-called 'running-calculation' formulx. It is interesting t o note that
such a factorized scheme can also be constructed for a general problem with
b ( x ) # const (see problem 1).
Hence, the conditions sufficient for the stability of the scheme in (33) are met
subject to the assumption in (32), because
6.3.5 PROBLEMS
1.Analyse t h e stability of t h e factorized s c h e m e ( I ) , (3), (4) u n d e r
t h e condition
A=A, A = + A;=Az. (36)
Solution. Taking into account. (3), (4), (36), we have the following for the
operator B:
Because the operators AD, P = 1 , 2 (see (36)) are adjoint, it follows from (37)
that
B 2 ~uTA, (38)
and therefore the factorized scheme (1) will be stable provided that u 0.25. >
Note that we derived the estimate (38) earlier, in Section 4.7, when considering
the alternating-triangular iterative method.
2. C o n s t r u c t a factorized scheme o n t h e basis of t h e following
implicit three-level scheme, which i s second-order a c c u r a t e i n t i m e
a n d s p a c e ( B a k e r a n d Oliphant):
Solution. The scheme (39) has the form (28) with the operator R = (1/3)A.
The factorized scheme is constructed in accordance with (30) and (31), and
we have
0
Rp=Ap, A ~ Y = - Y E ~ Z , ,p = 1,2.
The factorized scheme proper takes on the form
Taking into account that the operators Ap, 0 = 1 , 2 , are positive, self-
adjoint and commutative, we can easily verify the condition of stability for
the scheme (4).
324 COMPUTATIONAL HEAT TRANSFER
where
Let A,, a = 1 , 2 , . . . ,m, denote the difference heat conduction operator in the
direction x, and let us define this operator as follows:
Difference schemes in which the transition from one time level to another
is associated with solving a sequence of problems for the operators A,, a =
1 , 2 , . . . ,m, will be called additive difference schemes. The above (Section 6.3)
examples of economical difference schemes belong to the class of additive
difference schemes.
In ( 9 ) we split the operator A into one-dimensional grid operators A,,
a = 1 , 2 , . . . ,m . Other operators can also be taken as primitive operators. For
example, we have considered schemes with splitting into triangular operators.
When considering more general nonstabionary problems, the operators A,,
a = 1,2,. . . ,m, can have other meanings. In the general consideration of
additive difference schemes, we can forget about what the separate operators
A,, a = 1 , 2 , . . . , m, in the expansion ( 9 ) signify specifically.
b ( x ) du'
--
m dt
+ Alu' - d l ( x , t ) = 0, t, < t 5 t,+ll,,
% ,
When implementing (14), (15), we first solve the equation for ul(x, t) ( a = 1
in (14)), given the initial condition ul(x,tn) = vm(x,tn), and determine
ul(x, t,+l), which is later used as the initial condition when determining
v2(x, t), and so on. We take vm(x, tn+l) as an approximate solution of the
problem (5)-(7) at the time t,+i.
It should be noted that the two approaches to constructing additive
difference schemes (namely, (12), (13) and (14), (15)) are closely connected.
If the coefficients of the grid equation (5) do not depend on time and the
right-hand sides $"(z,t) = 0,then the problems (12), (13) and (14), (15)
are actually identical. It is this situation that occurs in our case of a heat
conductio~~ problem in nonhomogeneous medium (equations ( I ) , (2)), when
we are modelling heat conduction without inner sources of heat (f (x, t ) = 0
in (1)). In more general cases it is preferable to employ the second approach,
when auxiliary problems are solved over integer time steps.
Now we should answer the question of what is the error within which the
system of equations (14), (15) allows us to find the approximate solution to
the problem in (5)-(7). In order to answer this question we should extend the
notion of approximation.
dv
$'(x, t ) = -b(x) -
dt
- Alv + & ( x , t ) , (18)
we obtain
where 6,' is Kronecker's delta. For the solutions of the problem (5)-(7) we
have
x"%
,=I
o
$"(x, t ) = x
"'
o=l
P ( x ,t ) - x
"'
o=l
dv
A,v - b ( x )- = 0 ,
dt
and therefore for the summarized approximation error we get
In this case we have the estimate llum(x, t ) - x(x, t,)ll = 0 ( r 2 ) under some
additional restrictions on smoothness.
The above additive scheme (21) has the canonical form (22) for
We shall take that the additive scheme (23) possesses the property of
summarized approximation, i.e.
Taking into account (30), we rewrite this estimate in the final form as
+ T max
O<k<"
CC
1
1, I
a=1 &
Amp
m
C
1=p+l
B-'$;:
' lI1h
332 COMPUTATIONAL HEAT TRANSFER
This estimate for the difference problem (23) indicates that under the
conditions of summarized approximation (24)-(26) the additive difference
scheme (22) converges, provided that some additional conditions are satisfied,
Let us recall that the estimate (31) of the error of the additive difference
scheme has been derived under assumptions about the stability of this scheme
in the ordinary sense (the inequality in (27)). In such a context, we can say
that the stability and the summarized approximation of an additive difference
scheme imply its convergence.
Other a priori estimates for two-level additive difference schemes (22) are
currently available. These estimates make it possible, for example, to weaken
additional conditions like that in (32).
6.4.6 PROBLEMS
1. P r o v e t h a t t h e intermediate problems (14), (15) yield t h e e x a c t
solution to t h e p r o b l e m
The solution is
$n=$~+$~+~~+$~=~(~2+lh12),
334 COMPUTATIONAL HEAT TRANSFER
Here we have adopted that the operator A, defines the heat conduction in
the direction x,, i.e.
with
A=A1+A2+ ...+A,. (3)
Each of the operators A,, a = 1,2,. . . ,m in the additive representation (3) is
one-dimensional and, therefore, additive difference schemes (1)-(3) are called
locally one-dimensional difference schemes.
It has been shown in Section 6.4 that we can analyse the convergence of
additive difference schemes by deriving appropriate estimates of stability and
by studying the summarized approximation. Let us make a couple of remarks
about derivation of a priori estimates of stability of additive differenceschemes
using the example of locally one-dimensional scheme (1)-(3) and restricting
consideration, for the sake of simplicity, to the stability with respect to initial
data in some Hilbert space H D , generated by an operator D = D* > 0.
We need to derive the a priori estimate of pstability
with some constant p > 0 for the difference problem (1) with vp, = 0,
a = 1 , 2 , . . . , m. There are two options we can use to achieve this. The first
ECONOMICAL DIFFERENCE SCHEMES.. . 335
with some p,, a = 1 , 2 , .. . , m . When deriving the estimates (5), the additive
scheme (1) is considered as an ordinary two-level difference scheme for every
fixed a = 1 , 2 , .. . ,m. Then we can derive from (5) the desired estimate (4)
with
P= npm.
m
0=l
(6)
For instance, using the estimates (4), (6), we can analyse the asymptotic
stability of locally one-dimensional schemes for the heat equation.
Let us stress one important point. For every a = 1 , 2 , .. . ,m the stability
estimates (5) have to be derived in the same norm if we do not consider a
more general situation with the consecutive embedding of separate spaces. We
can achieve this for the scheme (1) if we consider stability in norms that are
not associated with the operators A,, a = 1 , 2 , .. . , m .
Let us find conditions under which the norm of the operator S does not exceed
unity (the stability of scheme with p = 1). To this end it is necessary t o verify
the inequality J = S'S - E < 0 which takes the following form if we allow
for (10):
It follows from (11) that llSll < 1 subject to the usual condition (see
Section 5.5)
a > -1 - - 1 (12)
2 AT'
where A is the constant in the estimate
A 5 AE. (13)
we have a problem like that in ( 7 ) ,(15) for every fixed a = 1,2,. . . , m and on
the basis of the estimate (18) we conclude
The estimate (22) is the sought-for estimate of stability of the locally one-
dimensional scheme (1).
In order to study the convergence of the scheme (1) we write the following
problem for the error z+,/, = yn,+, - u(x, t n + ~ ) ,x E w :
Here u(x,
t) is the exact solution of the problem
where
for m
C fm(x,t)= f(">t).
o=1
where
For problems with fairly smooth solutions, the estimates ( 3 0 ) , ( 3 1 ) , and the
summarized approximation conditions (27)-(29) entail
The estimate of stability for the scheme (33) is derived on the basis of
the maximum principle in the usual way (see Section 5.3). For every fixed
a = 1,2, .. . , m we obtain
These functions are then used to find the solution on a new time level:
On every time level we solve m one-dimensional problems and for this reason
these schemes are ascribed t o the class of locally one-dimensional additive
difference schemes. Note that the solutions c,,+,lm,
a = 1 , 2 , . . . , m , can be
determined independently (asynchronous calculations). Then we average these
solutions (see (36)). Because of this procedure, the schemes like (35), (36) are
called additiwely averaged difference schemes.
It is easy t o derive (see problem 2) the estimate (22) of stability of the
additively averaged difference scheme (35), (36) with respect to the right-
hand side and initial data.
Now let us consider the appropriate problem for the error:
-zn+a/m = Yn+o/m
-
- e(x. tn+,),
2, = y, - u(x, t,), x Ew
-
For $: we have
*
$: = $; - ~ m u ~ ~ ~ b - l ( z ) & (43)
Using the estimate ( 2 2 ) for the problem in (41)-(43) and taking into account
that z, = B,, we obtain
ECONOMICAL DIFFERENCE SCHEMES.. . 343
The estimate (44) ensures the convergence of the additively averaged scheme
(35), (36) with the first order in time and the second order in space for a fairly
smooth solution of the problem (1)-(4) from Section 6.4.
6.5.5 PROBLEMS
1. For t h e m o d e l p r o b l e m w e have considered w r i t e d o w n a locally
one-dimensional difference s c h e m e which is second-order a c c u r a t e
i n time.
Solution. According to Section 6.4, we use the bicyclic organization of
computations, which corresponds to a symmetric splitting of the following
form:
2m
Taking into account (46), this inequality yields the stability estimate (45),
which coincides with the estimate in (22) for the usual locally one-dimensional
scheme (1).
6.6.2 LITERATURE
1.D'yakonov E. G. (1972) Difference Methods for Solving Boundary Value Pmblems.
Issue 2: Nonstotionory Problems [in Russian]. Moscow State University, Moscow.
2. Marchuk G. I. (1975) Methods of Numerical Mathematics Springer-Verlag, New
York.
3. Marchuk G. I. (1988) Splitting Methods [in Russian]. Nauka, Moscow.
4. Richtmyer R. (1957) Diffwenee Methods for Initial Value Problems. Interscience,
New York.
5. Richtmver R. & Morton K. 119721
~ , Difference
- Methods for Initial Value Problems.
Interscience, New York.
6. Samarskii A. A. (1971) Introduction to the Theory of Difference Schemes [in
Russian]. Nauka, Moscow.
-
7. Samarskii A. A. (1983) Theow. o.f Difference Schemes lin Russianl. Nauka.
Moscow.
8. Samarskii A. A. (1987) Introduction to Numerical Methods [in Russian]. Nauka,
Moscow.
9. Samarskii A. A. & Gulin A. V. (1973) Stability of Difference Schemes [in Russian].
Nauka, Moscow.
10. Samarskii A. A. & Gulin A. V. (1989) Numerical Methods [in Russian]. Nauka,
Moscow.
11. Samarskii A. A. & Nikolaev E. S. (1989) Numerical Methods for Grid Equations.
Birkhauser-Verlag, Basel.
12. Saul'ev V. K. (1960) Integration of Equations of Parabolic Type by the Grid
Method [in Russian]. Fizrnatgiz, Moscow.
13. Vabishchevich P. N. (19871
~, Numerical Methods for Solvino Free-boundaru Prob-
lems [in Russian]. Moscow State University, Moscow.
14. Wazow W. & Forsythe G. (1960) Finite-difference Methods for Partial Differen-
tial Equations. John Wiley & Sons, New York.
15. Yanenko N. N. (1967) The Method of Fmetionol Steps. Springer-Verlag, New
Yark.
Heat Conduction Problems with
Phase Transit ions
are bound to the phase interface is nsed. When discussing these approaches,
the guideline is to formulate the problem in terms of new independent variables
in which the computational domain is regular. For simplicity we consider a
single-phase Stefan problem.
For multidimensional problems with phase transformations, variable do-
main methods often encounter difficulties with algorithms of implementation
and large computational costs. For this reason fixed domain methods have
become popular for approximate solution of these problems. A generalized
formulation of the classical Stefan problem is nsed to this end. Appropriate
numerical methods for solving the Stefan problem are constructed on the basis
of methods for solving quasi-linear heat conduction problems. The enthalpy
formulation of the Stefan problem, in which enthalpy rather than tempera-
ture is an independent variable, is also used in these problems. Economical
difference schemes are used to solve multidimensional problems.
Progress in the theoretical research of problems with a free (unknown)
boundary has recently been achieved by considering these problems as
variational. For the single-phase Stefan problem, a new independent variable
(Baiocchi transform) is introduced, which makes it possible to simplify
the problem. A fixed domain method for the Stefan problem can also be
constructed on the basis of the penalty method applied to an appropriate
variational inequality.
The quasi-stationary Stefan problem is described by the stationary heat
equation. We shall mention below the peculiarities of numerical methods used
to solve such problems. For instance, fixed domain methods can be constructed
by separating out a singularity of the solution on the phase interface.
We shall specially stress the features of mathematical models of phase
transformations in alloys. Models with an equilibrium two-phase zone will
be formulated. In some cases it is necessary to take into account diffusion of
impurities in a liquid or even in a solid phase.
The nonlinearity of problems with phase transitions is, first of all, accounted
for by the presence of an unknown (free) phase interface. Moreover, similar
to ordinary heat conduction problems, the nonlinearity can be due to the
dependence of thermal parameters on temperature. In order t o highlight the
features of problems with phase transformations, we shall take liquid and solid
phases to be homogeneous media with constant thermal parameters.
We shall assume the phase transition temperature t o be zero (u* = 0). Then
in the solid phase, which occupies the domain W ,we have u(x, t) = 0, and
in the liquid phase, which occupies the domain R+, we have u(x, t ) > 0. In
order to determine the temperature in the liquid phase we consider the heat
equation (a homogeneous medium)
Let us recall (Section 2.3) that the constant X is connected with the enthalpy
of phase transition.
By virtue of the above assumptions about boundary and initial conditions
in the single-phase Stefan problem (1)-(5), the speed of the phase front
v, = dqldt is positive, i.e. the domain occupied by the liquid phase is
expanding. A monotonic rise of the function q(t) follows from the maximum
principle (see Section 5.1) for parabolic problems.
Let w be the set of the inner nodes of the grid. We shall use the nonuniform
time grid
Now we only have to approximate the condition in (5). Taking into account
the assumption that the front is shifted by one space step (by one node), we
In order to retain the second-order accuracy with respect to h (see Section 4.2),
we approximate the condition in (5) as follows:
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS 351
The new independent variable ,$ varies within a finite interval from 0 (at
the left-hand endpoint x = 0) to 1 (at the interface we have x = q(t)). When
352 COMPUTATIONAL HEAT TRANSFER
~ ( c0 ,) = uo( E ( I ) ( ~ ) ) ) , (15)
while the boundary conditions ( 3 ) ,( 4 ) are transformed to become
It follows from (14) and (18) that for < = 1 we have the equation
Then, in order to refine Qntl, we use the relation in (23). For instance, the
simplest linearization yields
354 COMPUTATIONAL HEAT TRANSFER
Naturally, we can use other formulae: for example, we can determine the new
approximation;::8 from the appropriate quadratic equation.
be partitioned by the free boundary S = S ( t ) into two parts: R+(t) (the liquid
phase) and W ( t ) (the solid phase). We shall assume that the phase interface
S is described by the equation xl = q(x2,t), 0 < x2 5 l z , and during the
whole time period considered it preserves its structure (the function q(x2, t)
is single-valued for every fixed t-the free boundary without self-intersections,
0 < q(x2,t) < 11).
The single-phase Stefan problem for a homogeneous medium is described
by the equation
Fig. 7.1.
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS 355
Here n is the outward normal and V,, is the speed at which the free boundary
moves along the normal to S.
Under these conditions we can straighten the front with respect to the
variable X I by using the transformation $1 = E1q(xZ,t). Then the problem
(24)-(28) is written in the new independent variables (G,C2) with
In order to write the Laplace operator in the new variables, we shall use
general ideas of tensor analysis instead of direct transformations. Let goo,
a,/3 = 1,2, be the components of the metric tensor, J be the Jacobian of
the transformation (in our case we have J = gllgzz - glzgzl). We denote by
a,/3 = 1,2, the components of the associate tensor (the inverse of the
metric tensor). Then, in an arbitrary reference frame, we have for the Laplace
operator the following:
Hence we have
356 COMPUTATIONAL HEAT TRANSFER
and we arrive at the following expression for the Jacobian of the transfor-
mation:
J = q-1. (33)
The substitution of (30), (31) into (24) yields the desired equation
in which the coefficients are defined by (32), (33). Equation (34) is considered
in the fixed regular domain
where vl is the component of the front velocity along the axis XI. It follows
from the equation
au 1 av
from the equation of the free boundary XI = q(x2, t), and from (35) we derive
Note that the condition (35) on the free boundary exactly coincides with a
similar condition for the one-dimensional problem (see (18)).
Now the problem is considered in the new variables exactly in the same
way as in the one-dimensional case. The iterative refinement of the unknown
boundary on every time level is now associated with refining a function
rather than a single parameter. For this reason, we may find it useful to take
advantage of simplified noniterative schemes, which correspond, for example,
to using a single iteration in the above iterative procedures for the one-
dimensional problem.
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS 357
New variables are chosen so that the free boundary is fixed. For instance, the
original irregular domain W ( t )is mapped at every moment in time onto the
regular one
Fig. 7.2.
358 COMPUTATIONAL HEAT TRANSFER
7.1.6 PROBLEMS
1. For t h e m o d e l single-phase Stefan p r o b l e m (1)-(5), consider
t h e case i n which t h e front i s caught into a n o d e of a n irregular
s p a c e grid.
Solution. We shall use an irregular space grid. A new node in this grid will
be defined from the condition that the phase interface falls to this new node.
Let the grid consist of the nodes x, = + h,, i = 1,2,. . . ,in+lat a time
t = t,+, and let the position of the boundary grid node x,, m = be
unknown (we do not know the step h,). Let us relate to equation (1) the
difference equation
In order t o solve the nonlinear problem (38)-(40) for yn+l and h,, we
shall use the Thomas algorithm (see Section 4.5). We shall use the notation
u, = yn+l(x,), i = 0,1,. . . , m . The difference problem (38)-(40) is written in
the form of the tridiagonal system of linear equations
For the coefficients of the Thoma* algorithm we use the following formulae:
Using the expression for P,, we arrive at the following cubic equation for h,:
360 COMPUTATIONAL HEAT TRANSFER
An additional analysis shows that this equation has one real and two complex-
conjugate roots. After we have found the grid step, we restore the solution of
the difference problem by the reverse Thomas algorithm.
It should be stressed that the front-catching algorithm described is nonite-
rative, although the problem (38)-(40)is nonlinear.
2. Write heat equation (24) in terms of new independent variables
(&,t2) for a given dynamic transformation of variables (37).
Solution. We put u ( x ,t ) = v(C,t ) and get
au au
- = - + - - + -at,
- av at2 av
at at at at,at a t ~
and, therefore,
at1 = -1
- ax2 axl axl axz
- - - --
at 1( a t at2 at ac2
Here
Here the coefficients of the metric tensor of the transformation (37)are defined
by the formulae (43).
The phase transition temperature is taken to be zero and the free boundary
S = S(t) is therefore defined by
Two conditions must be satisfied on the free boundary (see Section 2.3): these
are the temperature continuity and the law of heat conservation:
[u] = 0, x E S(t),
= v x.,),
where X is the enthalpy of the phase transition and V, is the speed at which
the free boundary propagates along the normal to S(t).
This two-phase Stefan problem can be written as a single common heat
equation in the whole domain R. Let 6(u) be the Dirac delta-function. Then,
instead of equation (1) and the junction conditions (4)-(6), we can consider
the single heat equation
The heat conductivity and specific heat capacity are discontinuous in this
equation and have the form
According to (2) and (3), equation (7) is supplemented with the conditions
The peculiarity of the Stefan problem reveals itself in the fact that there
is a term with a 6-function in the left-hand side of equation (7). The release
and sink of heat, which accompanies a phase change, corresponds to a lumped
heat capacity on the interface. The boundary value problem (7)-(9) itself is
not very much different from the above quasi-linear heat conduction problems
(see Section 5.9). This fact allows us to proceed to the construction of required
difference schemes.
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS 363
Various approximate formulae for 6(u, A) which are constructed from the
assumption of the heat balance on the interval [-A, A] have been widely used
in computational practice. The simplest of the kind is to put
The condition
5
-n
is evidently fulfilled for this approximation. Numerical experiments provide
evidence that the error of the approximate solution only weakly depends on
which specific function is selected as an approximation to the 6-function, for
instance, on whether we select (12) or (13). The choice of the smoothing
parameter A is more significant. Naturally, the choice depends on the grid
used and is most often determined empirically from auxiliary computational
experiments.
364 COMPUTATIONAL HEAT TRANSFER
Here
b(y) = E(Y),
and
2
Nv)y = Cu
o=l
v ) ~
Y =- ( x E w.
The boundary and initial conditions ( X ) , (9) become
" , + +-
I
b(wk) Yn + A(wk)wk+l= 0, XEw, (20)
w k + l ( s )= g ( s , L + ~ ) ,z E aw, n = 0,1,.. . . (21)
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS 365
As a rule, a few iterations (20), (21) suffice to ensure the proper smoothness.
Instead of the iterative process (20), (21), we can use the Newton method,
which leads us to a nonself-adjoint grid elliptic problem for the new
approximation.
Among practitioners, a linearized purely implicit difference scheme has
become popular in which the coefficients are taken from the previous time
level. For the problem (a), (9), ( l l ) , at the inner grid nodes we use the equation
for yntl12 and y,+,. To this end, we can use an iterative process like (20),
(21) or the Newton method. The corresponding systems of linear equations
are solved by the Thomas algorithm.
The linearized difference scheme (18), (22) will be related to a difference
scheme with the equations
and the boundary conditions (24), (26). Instead of (28), the following equation
is frequently used
Introducing enthalpy corresponds (see Section 3.3) to the use of the Goodman
transform. In our case of homogeneous liquid and solid phases, in accordance
It follows from (30) that the new unknown function w has a discontinuity on
the interface (for u = 0).
The heat equation (7) with allowance for (30) takes the form
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS
Fig. 7.3.
where w(A) = c+A + X and w(-A) = -c-A. Instead of (32), we can use
more intricate approximation formulae.
The exact equation (31) is replaced by the approximate equation
The resulting problem is then solved numerically. If the function WA(U)is used
as an unknown, one speaks about approximate solution of the Stefan problem
in enthalpy formulation. Eliminating wA(u) from (32), (33), we arrive at a
smoothing scheme in the above temperature formulation (problem 2).
It immediately follows from (32) that for determining u(x, t) by a given
wA(u) we have the formula
368 COMPUTATIONAL HEAT TRANSFER
where n ( w ~ =) k(u(wa)).
If the functional transform (34) is given, we can easily formulate initial
and boundary conditions for equation (35). It should he noted that simple
functional relations (32), (34) between u(x, t) and WA(X,t ) take place for
prescribed constant heat capacities. In a more general case of c = ~ ( u )the
,
transition to enthalpy formulation is not that easy.
In order to solve (35) numerically, we can use the difference schemes
described in the above when we were considering fixed domain methods in
temperature formulation. To identify the free boundary, we use the condition
w(0) =
w(-A) + w(A) = const,
2
which corresponds to (4).
7.2.6 PROBLEMS
1. F o r t h e p r o b l e m (a), (9), ( l l ) , c o n s t r u c t a linearized analogue
of t h e Douglas-Rachford difference scheme.
Solution. According to Section 6.2, an analogue of this scheme would be
The implementation of the correction (the second stage in the scheme (36),
(37)) can be done on the basis of the difference equation
This equation is distinct from (37) only in that the coefficients are calculated
by the solution y,+l/2 rather than the time level t = t,.
2. C o n s t r u c t a linearized difference s c h e m e i n t e m p e r a t u r e for-
mulation o n t h e basis of t h e s m o o t h i n g (32).
Solution. For the problem (32), (33) with smoothed coefficients we use the
difference scheme
We supplement this equation with the simplest initial and boundary con-
ditions
i.e. the temperature u ( x , t ) has the meaning of the rate of change of the
function w(x, t) in time.
For the derivatives with respect to space variables we have
Taking into account condition (6) on the free boundary, defined by (8), we
derive
372 COMPUTATIONAL HEAT T R A N S F E R
Similarly, we have
W ( X ,t ) = 0 , x E R-(t). (13)
On the free boundary the relations ( 6 ) , (10) yield
w ( x ,t ) = 0 , x E S(t), (14)
while the second condition in (7) is reduced to the form
w(x,O) = 0 , x E a. (16)
Now we only have to formulate boundary conditions on aS2. Taking into
account ( 3 ) , ( 5 ) ,we write them in the form
W ( X ,t )=
i
O
g(x, 6 ) dB, x E 80.
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS 373
Thus, we arrive at the problem with free boundary (12)-(17) for the new
unknown function w(x, t ) .
The problem (12)-(17) differs from the original problem (I)-(?) for the
temperature because this time the solution itself as well as its first derivatives
are continuous over the whole domain 0.
Combining (12) and (18) and taking into account the homogeneous conditions
(14), (15) on the free boundary, we can write a single general equation for the
whole computational domain, namely
Equations (19), (20) are supplemented by the boundary conditions (17) and
the initial condition (16).
We can attempt the construction of uniform fixed domain computational
algorithms on the basis of the problem (16), (17), (19), (20). However, the
discontinuous right-hand side in (20) significantly complicates the search for
an approximate solution for w ( x ,t).
In theoretical research (unique solvability, smoothness of the free boundary),
formulation of the problem (16), (17), (19), (20) as a variational inequality has
been successfully used. A penalty method is used for approximate solution of
the variational problem. Without giving the variational formulation of the
problem concerned, we shall only provide a problem for the approximate
solution derived by the penalty method.
Let w,(x, t ) denote an approximate solution found by the penalty method
for the problem (16), (17), (19), (20), where E > 0 is a penalty parameter.
The following problem is formulated for w,(z, t). In the domain 0 we solve
374 COMPUTATIONAL HEAT TRANSFER
The initial and boundary conditions for (21), (22) have the usual form (see
(16))( 1 7 ) )
w,(x,O) = 0 , x E R, (23)
Then we have
n:(t) = { ( ~ , t ) l ~ , (>~ O, ,t ) x E 01,
In order to verify that this definition of the subdomains R;(t) and @(t)
is feasible, we shall considet the boundary value problems in separate
subdomains. In the subdomain R f ( t ) we have an equation (see (21), (22))
that coincides with the original equation (12) and the corresponding boundary
conditions. Therefore, in @ ( t ) we indeed have w,(x,t) > 0. For the
subdomain R; ( t ) we derive
2
a2w, 1
at - ~ , : i y W xr ~=
~ - t- ,o~
c t .i r
a=1
y d x ) = 0, E W, (27)
) g ( x , t n + ~ ) , x E aw,
Y , + ~ ( X= n = 0 , ~. .. . (28)
Here, as usual,
2
AY = Amy, Y =- xE W.
m=1
The implementation of the nonlinear difference scheme can be based on the
iterative process
7.3.5 PROBLEMS
1. Transform the condition (7) on an interface S ( t ) defined by the
relation t = q ( x ) to the condition (9).
Solution. Taking into account (6), we derive from ( 7 )
where V, is the component of the velocity of the free boundary along the
direction z, a = 1,2. For the point on the interface we have
We multiply (31) by a q / a x , and sum over a = 1,2. Then, taking into account
(32),(33), we derive
Differentiating this equality with allowance for (11) and (34),we derive
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS
Thus, we arrive a t the third-kind boundary condition for the new function
W(X>t).
for x = (XI,XZ),0 < X I < 11, -m < x2 < ca.Here the heat conductivity and
the specific heat capacity are assumed constant within each phase, i.e.
Let us formulate the boundary conditions on the edges of the plate. We assume
that the lower part of the plate is heat-insulated, that is
while the upper part experiences a heat flow (modelling of the heat source):
378 COMPUTATIONAL HEAT TRANSFER
Fig. 7.4.
For definiteness, we shall assume that the power of the heat source is sufficient
for the total melting of the plate (a tentative pattern of isotherms is presented
in Fig. 7.4).
In this quasi-stationary Stefan problem (1)-(4), the computational domain
is infinite. In order to solve this problem approximately, we single out a large
enough rectangle
and consider eqnation (1) in 0 after supplementing this equation with the
following conditions on the side boundaries:
where E(u) = ~ ( u+) 6(u, A) if we use the ordinary approximation formula for
For the numerical solution of the problem (2), (3), ( 5 ) , (6) we can use
the simplest iterative process of successive refinement of coefficients. In each
iteration we solve the linear boundary value problem
-
k(vk) =o, x1=11,
- auk+'
k(uk) 7- g(x) 2 0, 21 = 0,
uk+'(x) = -1, 1x21 = 12.
Thus, we are solving a linear boundary value pn3blem for an elliptic equation
with convective terms.
S = { X ~ X E Q , u(x)=o) (12)
boundary during every next iteration, we have to consider the problem with
nonhomogeneous junction conditions (14). Similar approaches can also be
used when constructing fixed domain difference schemes for the approximate
solution of nonstationary problems. Because the problems are nonstationary,
we can employ noniterative versions of this approach as well (which essentially
means that we go with just one iteration).
Let us describe the iterative process of successive refinement of the free
boundary on a differential level. It is logical to use the approach in which a
new approximation is determined from the solution of the linear equation
when
Sk = { X ~ XE 0 , vk(x) = 0). (I6)
The junction conditions (13), (14) yield
We select such a function z(x), smooth in R: and R;, that the junction
conditions for w(x) on S k are homogeneous. Hence, we should construct such
382 COMPUTATIONAL HEAT TRANSFER
which does not have any singularities all over the domain n.
Now the whole problem reduces to constructing such a smooth function
z(x) that would satisfy the conditions in (Zl), (22). It is also necessary
that the computational costs involved in computing this function be at least
comparable to those involved in solving the boundary value problem for
w(x). Without attempting any detailed description of possible approaches
to constructing the function ~ ( x in
) this problem, we shall only mention the
additive separation by a method of harmonic potentials.
Taking into account the properties of simplelayer potential, we put
mixing. If the condition of complete melting of the plate is satisfied (Fig. 7.4),
we can consider separately the singlophase problem for x2 > 0 and the
problem for x2 < 0 (only the sign of velocity is changed). For definiteness, we
shall consider the problem for x2 > 0, i.e. in the domain
where q = Kc-/k-.
On the free boundary S defined by (12) we have
Taking into account that n is the outward normal, the second condition on
the free boundary has the form
Let the equation for the free boundary have the form xz = q(xl), i.e.
Similarly,
Taking into account equation (24) and the boundary condition in (25), we
transform this relation to the form
The boundary condition (3) depends on the point at which the free boundary
S intersects the top boundary of the plate (on the point 2 2 = ~ ( 0 ) A) similar
situation also takes place when we are considering a quasi-stationary single-
phase Stefan problem with nonhomogeneous conditions of the first kind.
The problem (33)-(38) is further considered by the scheme from Section 7.3.
In particular, it is convenient to use the penalty method (problem 1) when
constructing computational fixed domain algorithms for the approximate
solution of the problem for the new function w(z). The solution of the
appropriate nonlinear elliptic problem can be found by an iterative method,
similar to those that have been considered previously for the nonstationary
Stefan problem.
on the passage to new coordinates, where the solution to the problem itself
serves as an independent variable. In the problem (24)-(29) we pass from the
variables X I ,x 2 , in terms of which the computational domain is known, to the
new independent variables
The coefficients of the metric tensor (see problem 2 in Section 7.1) will have
the form
g11= (2)'.(2)' + (2) =1
2
,
axl axl axz axz
g12=g21=--+--=-- axz ax2
(41)
ah aEz ah at2 a x l av '
and the corresponding expression for the Laplace operator in the new
coordinates, we switch from equation ( 2 4 ) to the following equation for the
function xz(x1, v ) :
which follows from (25), (26). This fact makes it possible to employ the
following second-kind boundary condition in terms of the new variables:
7.4.7 PROBLEMS
1. Formulate t h e b o u n d a r y value problem for t h e a p p r o x i m a t e
solution of t h e problem w i t h free b o u n d a r y (33)-(38) o n t h e basis
of t h e penalty m e t h o d (see Section 7.3).
Solution. We seek an approximate solution w,(x) in the rectangle 0 < X I <
11, 0 < x2 < l z as a solution of the equation
Fig. 7.5.
boundary (the boundary of the solid phase) there is a domain in which the
phase change proper takes place at the constant phase transition temperature.
This domain, in which the two phases coexist, is called the two-phase zone. In
order t o describe this zone it is logical to introduce a new unknown, namely,
the fraction of the solid phase.
Thus, given a large velocity of motion of the melting/crystallization
boundary, the phase transformation occurs in some space domain (the t w e
phase zone). Slow motions of the phase transition boundary make the t w e
phase zone become thinner and, as a limit case, we have an expressed phase
transition boundary, i.e. the classical Stefan problem. The questions of how to
allow for a finite rate of the phase change and the very kinetics of the growth
of the new phase are beyond the scope of present consideration.
A similar situation takes place when considering melting/crystallization of
simple binary alloys (two components, one admixture). In this case new time
scales appear, namely, there is a time scale of diffusion of the admixture
(equalization of the concentration of the admixture) in the solid and liquid
phases. This appreciably widens the spectrum of possible directions in which
phase changes may occur. Here we shall discuss some typical modes of
melting/crystallization.
At a constant temperature, a binary alloy can be in three different states.
We shallassume that the temperature of a phase transition in a pure substance
(C = 0) equals zero and for small concentrations of the admixture the solidus
and liquidus temperatures are straight lines (Fig. 7.5), with the liquidus
temperature falling as the concentration of the admixture increases. At a
temperature u = ua = const the alloy can be in a solid state (for C < C.,I)
or in a liquid state (for C > Cliq), depending on the concentration. For
C.,I < C < Cli,) there exists an intermediate equilibrium state when the solid
and liquid phases coexist. This twephase zone appears because the alloy is
multicomponent.
390 COMPUTATIONAL HEAT TRANSFER
where $(u) is the volume fraction of the solid phase, defined from the
equilibrium phase diagram of a particular alloy. It is natural to assume the
function $(u) to he continuous and to satisfy the conditions $(u) = 1 for
<
u uar $(u) = 0 for u 2 U I , and 0 < $(u) < 1 for U. < u < UI.
In the approximation we shall use, the temperature field arising during
phase transitions of a binary alloy is described by the usual heat equation
considered in Section 7.2. Some other difference schemes are also given in this
section. Therefore, we need not dwell on these here.
HEAT CONDUCTION PROBLEMS WITH PHASE TRANSITIONS 391
u*(x,o) = U O ( X ) , x E n*(o),
u f ( 2 ,t ) = g(x, t ) , x E y*, 0 < t 5 T,
[u]= 0, x E S(t),
[,g] = -A, x E S,)
392 COMPUTATIONAL HEAT TRANSFER
The only difference is that when specifying condition (4) on the free boundary
we have instead of the ordinary Stefan condition
Now let us formulate the problem for the distribution of the admixture. We
shall assume that the coefficients of diffusion of the admixture are constant
for every subdomain, therefore, the diffusion equation has the form
We assume that the admixture cannot leave the material and, therefore, use
the boundary condition.
ac*
D*-(x,t)=0, XET*, 0<t<T.
an (11)
In addition to the relation in (3), the following condition (see Section 2.3) is
satisfied on the interface:
7.5.5 PROBLEMS
1. Formulate t h e problem of diffusion of a n a d m i x t u r e i n t h e liquid
phase assuming t h a t t h e a d m i x t u r e is insoluble in t h e solid phase.
Solution. The conditions of the problem actually say that D- = 0. The
diffusion equation in the liquid phase has the form (see (10))
The boundary condition for the solidification follows from the junction
condition (12) with allowance for D- = 0. This leads us to the third-kind
boundary condition
on the interface.
2. Formulate conditions for defining t h e melting t e m p e r a t u r e for
t h e case in which t h e r e is n o diffusion of t h e a d m i x t u r e in t h e solid
phase a n d t h e a d m i x t u r e is completely mixed in t h e liquid phase.
Solution. In our case the phase transition temperature is constant over the
interface a t every moment in time, i.e.
Balance relations are used in order to determine the function C + = C+(t). Let
C-(x) denote the concentration of the admixture in the solid phase. When a
solid region is melted down, the admixture that has been frozen in this region
changes the concentration in the liquid phase. Hence, for C+(t) we derive
7.6.2 LITERATURE
1. Avdonin N. A. 11980)
, -
, Mathematical Description o.f Crvstollizotion Processes lin
Russian]. Zinatne, Riga.
2. Danilvk I. I. (1985)
. . On the Stefan Problem,. Uspekhi
. Mot. Nnuk 40, Issue 5 (245),
. .
133-185.
3. Duvaut G. & Lions J. L. (1972) Les inequations en mechoniques et physique.
Dunod, Paris.
4. Fridman A. (1982) Variational Principles and Free-boundary. John Wiley & Sons,
New York.
5.Glowinski R., Lions 1. J. & Tr&molikres(1976)Anolyse nume'rique des ine'quntions
variotionnelles. Dunod, Paris.
6.Meirmanov A. M. (1986) The Stefan Problem [in Russian]. Nauka, Novosibirsk.
7. Rubinsbtein L. I. (1967) The Stefan Pmblem [in Russian]. Zvaigzne, Riga.
8.Vabishchevich P. N.(1987)Numekcol Methodsfor Problems with o Free Boundary
[in Russian]. Moscow State University, Moscow.
Index
algorithm
forward-backward algorithm 155
marching algorithm 158
matrix Thomas algorithm 157
Thomas algorithm 155
approximation
approximation of the boundary condition on the solutions 105
approximation on the solutions 119
Boussinesq approximation 41
b o u n d a r y conditions 22
approximation of the boundary condition on the solutions 105
Dirichlet boundary conditions 23
error of approximation of the boundary condition 100
first-kind boundary conditions 23
mixed boundary conditions 27
Nenmann boundary conditions 23
second-kind boundary conditions 23
third-kind boundary condition 24
canonical f o r m
canonicd form of a difference equation 122
canonical form of a three-level iterative method 167
canonical form of a two-level iterative method 163
canonical form of the three-level difference scheme 232
canonical form of the two-level difference scheme 232
Cholesky decomposition 154
coefficient
coefficient of linear expansion 50
coefficient of thermal expansion 41
heat transfer coefficient 24
Lamk coefficients 50
Poisson coefficient 52
complex s c h e m e 274
c o m p u t a t i o n a l experiment 6
COMPUTATIONAL HEAT TRANSFER
condition
boundary conditions 22
connected grid 123
conservative difference scheme 109
consistent grid 191
continuity equation 37
convection 15
decomposition m e t h o d 197
density 16
diagnostic computational experiment 11
difference equation
canonical form of a difference equation 122
difference problem 97
difference scheme 100
additive difference schemes 325
additive difference schemes with fractional steps 326
additive difference schemes with integer steps 326
additively averaged difference schemes 341
asymptotically stable difference scheme 267
canonical form of the three-level difference scheme 232
canonical form of the two-level difference scheme 232
conservative difference scheme 109
Crank-Nicolson difference scheme 241
difference scheme of alternating directions 310
difference scheme with advancing 241
difference scheme with weights 241
Dauglas-Rachford difference scheme 313
Dufort-F'rankel difference scheme 266
economical difference scheme 299
exact difference scheme 119
factorized difference scheme 316
homogeneous difference scheme 113
locally one-dimensional difference scheme 334
Peaceman-Rackford difference scheme 310
Peaceman-Rackford scheme) 310
predictor-corrector difference scheme 289
purely implicit difference scheme 241
regularization principle for difference schemes 280
Richardson difference scheme 266
stability of the difference scheme 100
stability of the difference scheme with respect to the initial data 233
stability of the difference scheme with respect to the right-hand side 234
INDEX
m a j o r a n t function 128
marching algorithm 158
mathematical model 2
m a t r i x T h o m a s algorithm 157
m a x i m u m principle
maximum principle for parabolic equations 226
maximum principle for second-order elliptic equations 90
method
alternating Schwarz method 201
alternating triangle iterative methods 180
approximate factorization method 183
balance method 111
canonical form of a three-level iterative method 167
canonical form of a twelevel iterative method 163
Chehyshev iterative method 164
decomposition method 197
finite element method 99
fixed domain method 347
front straightening method 352
Gauss method 153
integro-interpolation method 111
iterative method 162
iterative method of capacity matrix 199
iterative method of conjugate directions 168
Jacobi iteration method 173
method of a control volume 111
method of a simple iteration 164
method of a square root 154
method of alternating directions 165
method of catching of the front into a space grid node 350
method of conjugate gradients 168
method of fictitious domains 194
method of lines 240
method of minimal corrections 167
method of separation of variables 63
method of steepest descent 167
Newton method 212
one-step (two-level) iterative method 162
penalty method 373
projective methods 97
projective-grid methods 99
projective-grid methods 99
quasi-linearization method 212
INDEX
searching experiment 11
second difference Green formula 140
second-kind boundary conditions 23
Seidel method 177
self-adjoint operator 135
self-similar solutions 72
singularly perturbed problem 78
skew-symmetric operator 135
solidus temperature 34
solutions
approximation of the boundary condition on the solutions 105
approximation on the solutions 119
self-similar solutions 72
sparse matrix 153
specific heat capacity 16
spectral problem 63
stability of t h e difference scheme 100
stability of t h e difference scheme with respect t o t h e initial data 233
stability of t h e difference scheme with respect t o t h e right-hand side 234
stabilization-correction difference scheme 313
stable difference scheme 233
steady-state iterative method 164
Stefan number 62
Stefan problem 32
Stefan-Boltzmann law 46
Steklov averaging operators 113
stencil 101
stream function 39
stability
stability of the difference scheme 100
stability of the differencescheme with respect to the initial data 233
stability of the difference scheme with respect to the right-hand side 234
uniform stability with respect to the initial data 234
summarized approximation 327
symmetrical difference scheme (the Crank-Nicolson scheme) 241
symmetrical over-relaxation method 180
temperature
liquidus temperature 34
solidus temperature 34
thermal conductivity 16
thermal diffusivity 17
tbermodiffusion Stefan problem 392
406 COMPUTATIONAL HEAT TRANSFER