1D Finite Difference Method

Excerpt from GEOL557 Numerical Modeling of Earth Systems by Becker and Kaus (2016)
1
1.1
Finite difference example: 1D implicit heat equation

Boundary conditions Neumann and Dirichlet
We solve the transient heat equation
T
c p
=
t
x
T
k
x

(1)
on the domain L/2 x L/2 subject to the following boundary conditions for fixed
temperature
T ( x = L/2, t) = Tle f t
T ( x = L/2, t) = Tright
(2)
T ( x < W/2, x > W/2, t = 0) = 300

T (W/2 x W/2, t = 0) = 1200,
(3)
(4)
with the initial condition
where we have again assumed a hot dike intrusion for W/2 x W/2.
Boundary conditions (BCs, see also sec. ??) for PDEs that specify values of the solution
function (here T) to be constant, such as eq. (2), are called Dirichlet boundary conditions.
We can also choose to specify the gradient of the solution function, e.g. T/x (Neumann
boundary condition). This gradient boundary condition corresponds to heat flux for the
heat equation and we might choose, e.g., zero flux in and out of the domain (isolated
BCs):
T
( x = L/2, t) = 0
x
T
( x = L/2, t) = 0.
x
1.2
(5)
Solving an implicit finite difference scheme
As before, the first step is to discretize the spatial domain with n x finite difference points.
The implicit finite difference discretization of the temperature equation within the medium
where we wish to obtain the solution is eq. (??). Starting with fixed temperature BCs
(eq. 2), the boundary condition on the left boundary gives
T1 = Tle f t
(6)
Tnx = Tright .
(7)
and the one on the right
See geodynamics.usc.edu/~becker/Geodynamics557.pdf for complete document.
Eqs. (??), (6), and (7) can be written in matrix form as

Ax = b.
(8)
For a six-node grid, for example, the coefficient matrix A is
1
0
0
0
0
0
s (1 + 2s)
s
0
0
0
0
s
(1 + 2s)
s
0
0
A=
0
0
s
(1 + 2s)
s
0
0
0
0
s
(1 + 2s) s
0
0
0
0
0
1
(9)
the unknown temperature vector x is
x=
and the known right-hand-side vector b is
b=
T1n+1
T2n+1
T3n+1
T4n+1
T5n+1
T6n+1
Tle f t
T2n
T3n
T4n
T5n
Tright
(10)
(11)
Note that matrix A will have a unity entry on the diagonal and zero else for each node
where Dirichlet (fixed temperature) boundary conditions apply; see derivation below and
eqs. (??) and (??) for how to implement Neumann boundary conditions.
Matrix A also has an overall peculiar form because most entries off the diagonal are
zero. This sparseness can be exploited by specialized linear algebra routines, both in
terms of storage and speed. By avoiding computations involving zero entries of the matrix, much larger problems can be handled than would be possible if we were to store
the full matrix. In particular, the fully implicit FD scheme leads to a tridiagonal system
of linear equations that can be solved efficiently by LU decomposition using the Thomas
algorithm (e.g. Press et al., 1993, sec. 2.4).
1.3
MATLAB implementation
Within MATLAB , we declare matrix A to be sparse by initializing it with the sparse

function. This will ensure a computationally efficient internal treatment within MATLAB.Once the coefficient matrix A and the right-hand-side vector b have been constructed,
MATLAB functions can be used to obtain the solution x and you will not have to worry
about choosing a proper matrix solver for now.
First, however, we have to construct the matrices and vectors. The coefficient matrix
A can be constructed with a simple loop:
A = sparse(nx,nx);
for i=2:nx-1
A(i,i-1) = -s;
A(i,i ) = (1+2*s);
A(i,i+1) = -s;
end
and the boundary conditions are set by:
A(1 ,1 ) = 1;
A(nx,nx) = 1;
(Exercise: Improve on the loop formulation for A assembly by using MATLAB vector
functionality.)
Once the coefficient matrix has been constructed, its structure can be visualized with
the command
>>spy(A)
(Try it, for example by putting a break-point into the MATLAB code below after assembly.)
The right-hand-side vector b can be constructed with
b = zeros(nx,1);
b(2:nx-1) = Told(2:nx-1);
b(1) = Tleft; b(nx) = Tright;
The only thing that remains to be done is to solve the system of equations and find x.
MATLAB does this with
x = A\b;
The vector x is now filled with new temperatures T n+1 , and we can go to the next time
step. Note that, for constant t, , and x, the matrix A does not change with time. Therefore we have to form it only once in the program, which speeds up the code significantly.
Only the vectors b and x need to be recomputed. (Note: Having a constant matrix helps a
lot for large systems because operations such as x = A\b can then be optimized further
by storing A in a special form.)
1.4
Exercises
a) Save the script heat1Dexplicit.m from last section as heat1Dimplicit.m. Program

the implicit finite difference scheme explained above. Compare the results with
results from last sections explicit code.
b) Time-dependent, analytical solutions for the heat equation exists. For example, if
the initial temperature distribution (initial condition, IC) is

x 2
(12)
T ( x, t = 0) = Tmax exp
where Tmax is the maximum amplitude of the temperature perturbation at x = 0 and

its half-width of the perturbance (use < L, for example = W). The solution is
then

Tmax
x2
T ( x, t) =
(13)
exp
2 + 4t
1 + 4t/2
(for T = 0 BCs at infinity). (See Carslaw and Jaeger, 1959, for useful analytical solutions to heat conduction problems).
Program the analytical solution and compare the analytical solution with the numerical solution with the same initial condition. Compare results of the implicit
and FTCS scheme used last section to the analytical solution near the instability
region of FTCS,
1
t
< .
(14)
s=
2
2
(x )
Note: Eq. (13) can be derived using a similarity variable, x = x/xc with xc t.
Looks familiar?
c) A steady-state temperature profile is obtained if the time derivative T/t in the
heat equation (eq. ??) is zero. There are two ways to do this.
(i) Wait until the temperature does not change anymore.
(ii) Write down a finite difference discretization of 2 T/x2 = 0 and solve it. (See
the limit case consideration above.)
Employ both methods to compute steady-state temperatures for Tle f t = 100 and
Tright = 1000 . Derive the analytical solution and compare your numerical solutions accuracies. Use the implicit method for part (a), and think about different
boundary conditions, and the case with heat production.
d) Apply no flux boundary conditions at | x | = L/2 and solve the dike intrusion problem in a fully implicit scheme. Eqs. (??) and (??) need to replace the first and last
columns of your A matrix.
e) Derive and program the Crank-Nicolson method (cf. Figure ??C). This best of both
worlds method is obtained by computing the average of the fully implicit and fully
explicit schemes:

n +1
n +1
n +1
n 2T n + T n
n +1
n
+
T
T
2T
+
T
Ti
Ti
i +1
i
i 1
i +1
i
i 1
.
=
(15)
2
t
2
(x )
This scheme should generally yield the best performance for any diffusion problem,
it is second order time and space accurate, because the averaging of fully explicit
and fully implicit methods to obtain the time derivative corresponds to evaluating
the derivative centered on n + 1/2. Such centered evaluation also lead to second
order accuracy for the spatial derivatives.
Compare the accuracy of the Crank-Nicolson scheme with that of the FTCS and fully
implicit schemes for the cases explored in the two previous problems, and for ideal
values of t and x, and for large values of t that are near the instability region of
FTCS.
Hint: Proceed by writing out eq. (15) and sorting terms into those that depend on
the solution at time step n + 1 and those at time step n, as for eq. (??).
f) Bonus question: Write a codefor the thermal equation with variable thermal con
T
ductivity k: c p T
t = x k x . Assume that the grid spacing x is constant. For
extra bonus, allow for variable grid spacing and variable conductivity.
Bibliography
Carslaw, H. S., and J. C. Jaeger (1959), Conduction of Heat in Solids, 2nd ed., Oxford University Press, London, p. 243.
Press, W. H., S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery (1993), Numerical Recipes
in C: The Art of Scientific Computing, 2 ed., Cambridge University Press, Cambridge.

1D Finite Difference Method

Uploaded by

Copyright:

Available Formats

1D Finite Difference Method

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1D Finite Difference Method

Uploaded by

Copyright:

Available Formats

Excerpt from GEOL557 Numerical Modeling of Earth Systems by Becker and Kaus (2016)

Finite difference example: 1D implicit heat equation

We solve the transient heat equation

T ( x < W/2, x > W/2, t = 0) = 300

with the initial condition

Solving an implicit finite difference scheme

and the one on the right

See geodynamics.usc.edu/~becker/Geodynamics557.pdf for complete document.

Eqs. (??), (6), and (7) can be written in matrix form as

For a six-node grid, for example, the coefficient matrix A is

the unknown temperature vector x is

and the known right-hand-side vector b is

Within MATLAB , we declare matrix A to be sparse by initializing it with the sparse

See geodynamics.usc.edu/~becker/Geodynamics557.pdf for complete document.

a) Save the script heat1Dexplicit.m from last section as heat1Dimplicit.m. Program

where Tmax is the maximum amplitude of the temperature perturbation at x = 0 and

See geodynamics.usc.edu/~becker/Geodynamics557.pdf for complete document.

See geodynamics.usc.edu/~becker/Geodynamics557.pdf for complete document.

You might also like