Commun. Comput. Phys. Vol. 10, No. 4, pp.

844-866
doi: 10.4208/cicp.310110.161010a October 2011

Picard-Newton Iterative Method with Time Step

Control for Multimaterial Non-Equilibrium
Radiation Diffusion Problem
Jingyan Yue and Guangwei Yuan∗
National Key Laboratory of Science and Technology on Computational Physics,
Institute of Applied Physics and Computational Mathematics, P.O. Box 8009,
Beijing 100088, China.
Received 31 January 2010; Accepted (in revised version) 16 October 2010
Available online 13 June 2011

Abstract. For a new nonlinear iterative method named as Picard-Newton (P-N) itera-
tive method for the solution of the time-dependent reaction-diffusion systems, which
arise in non-equilibrium radiation diffusion applications, two time step control meth-
ods are investigated and a study of temporal accuracy of a first order time integration
is presented. The non-equilibrium radiation diffusion problems with flux limiter are
considered, which appends pesky complexity and nonlinearity to the diffusion coef-
ficient. Numerical results are presented to demonstrate that compared with Picard
method, for a desired accuracy, significant increase in solution efficiency can be ob-
tained by Picard-Newton method with the suitable time step size selection.
AMS subject classifications: 65M06, 65M08, 65M12
Key words: Non-equilibrium radiation diffusion, Picard, Picard-Newton based on linearization-
discretization, nonlinear iterative method, time step control.

1 Introduction
Picard nonlinear iterative method is a globally convergent method and has been used
extensively in many applications. However its iterative convergence rate is only one
order, and its solution efficiency is low in solving some practical problems. So Newton-
like methods with super-linear convergence must be introduced, and these nonlinear
Newton-Krylov solution techniques [2–11] have been developed to ensure nonlinearities
convergence. It is well-known (see, e.g., [5–8]) that there are two main obstacles which
have prevented people from using Newton-like methods for large scale multi-physics
simulations.
∗ Corresponding author. Email addresses: yue [email protected] (J. Yue), yuan [email protected]
(G. Yuan)

http://www.global-sci.com/ 844 c 2011 Global-Science Press

J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 845

The first obstacle in Newton-like methods for solving some practical multi-physical
problems is the need for evaluating the entries of the Jacobian matrix. To overcome
this obstacle we have proposed in [20, 21] a new nonlinear iterative method named as
Picard-Newton (P-N) iterative method for solving numerically nonlinear diffusion equa-
tions and multimaterial non-equilibrium radiation diffusion problems. To construct a
new nonlinear iteration method, we directly design iterative method for the time dis-
cretized nonlinear PDEs instead of nonlinear algebraic system from implicit discretiza-
tion scheme. By linearizing the time discretized nonlinear PDEs first and discretizing
the resulting linear PDEs next, which is named as the linearization-discretization (LD)
approach, we have devised the Picard-Newton iterative scheme which gives a specific
procedure to the formation of (exact or approximated) Jacobian matrix. It follows that
the first obstacle can be overcome through the LD approach.
The second obstacle is that Newton-like methods cannot be expected always to con-
verge since they are locally convergent. In other words, an initial guess inside the ra-
dius of convergence is required for Newton-like methods to converge, while for time-
dependent or transient problems the initial guess is usually the converged solution from
the previous time step. A reliable time step selection is necessary for nonlinear itera-
tive methods to ensure the methods to operate with desired accuracy. Then time step
control is essentially important for fast and accurate numerical solution, especially when
Newton-like method is employed.
In addition to the difficulties of evaluating the Jacobian matrix and the radius of con-
vergence being small, on each iteration step of Newton-like method for nonlinear sys-
tems the cost of solving the linear systems to arrive at the new iteration values might be
very expensive, which depends mainly on the matrix properties of the linear systems. In
other words, since the matrix properties of the linear systems corresponding to Newton
method are usually worse than those corresponding to Picard method, the computational
cost of the original Newton method at each nonlinear iteration step is more expensive
than that of Picard method.
We hope our P-N method can accelerate the existing Picard method. The P-N method
can be formulated by adding certain convective terms of one order to the Picard method.
When the one order terms are discretized by centered scheme, the resulting nonlinear
iteration is named as P-NC iterative scheme, while discretized by upwind scheme, the
resulting nonlinear iteration is named as P-NU iterative scheme. It is worth to point
out that the P-NC may be equivalent to standard Newton method, while the P-NU is
new and is an improvement over Picard and Newton methods. From the construction
of our P-NU method we can see that the diagonal dominance of the matrix from P-NU
is not worse than that from Picard method, and then it is possible to reduce remarkably
computational time for each step of nonlinear iteration.
Since our P-N has overcome the first obstacle through the LD approach, it remains
to demonstrate our P-N need not introduce more strict time steps than that used for
Picard. In this paper, two time step control methods are investigated and a study of
temporal accuracy of a first order time integration is presented for P-N iteration. We are
846 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

primarily concerned with assessing the CPU cost for a desired accuracy as nonlinearity of
the problem concerned increases. Numerical results clearly indicate that compared with
Picard, significant increase in solution efficiency can be obtained by P-N with the suitable
time step size selection. The extend of increase is remarkably larger for more nonlinear
problem. Although we focus on non-equilibrium radiation diffusion, the P-N should also
be applicable to other physical systems described by the equations of parabolic type.
The remainder of the paper is structured as follows. First, the mathematical model for
the coupled system of multimaterial non-equilibrium radiation diffusion with material
conduction equations is introduced in Section 2. Next, Picard-Newton iterative scheme
is presented in Section 3. Then in Section 4, two time step control methods are given.
In Section 5, numerical tests are provided to show the performance comparison between
Picard and P-N methods. Finally, conclusion is given in Section 6.

2 Mathematical model
The mathematical model used in this paper is a system of two-dimensional multimaterial
non-equilibrium radiation diffusion coupled to material conduction equations [1,6,7]. Let
E be radiation energy density and T material temperature. The coupled system is given
by
∂E
−∇·( D ∇ E) = σa ( T 4 − E), (2.1a)
∂t
∂T
−∇·(κ ∇ T ) = σa ( E − T 4 ), (2.1b)
∂t
where  
∂/∂x
∇=
∂/∂y
is the gradient operator and it acts on function u as follows:
 
∂u/∂x
∇u = .
∂u/∂y

The divergence operator ∇· acts on a vector function

 
~ A1
A=
A2

and gives a scalar function

~ = ∂A1 /∂x + ∂A2 /∂y.
∇· A
The energy exchange is controlled by the photon absorption cross section

z3
σa ( T ) = , (2.2)
T3
J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 847

where z is a function of the material and varies as a function of space ( x,y). High values
of z and low values of T lead to higher energy exchange and therefore tighter coupling
between Eqs. (2.1a) and (2.1b). In the more nonlinear problem in Section 5, the two ma-
terials have z = 1 and 10 respectively. In the less nonlinear problems, the two materials
have z = 1 and 5 or 2.5 respectively.
The model for the radiation diffusion coefficient is:
1
D= |∇ E |
. (2.3)
3σa + E

The second part in the denominator, |∇ E|/E, is the flux limiter, which is used to keep the
propagation velocity of a radiation wave front less than the speed of light [1].
The following form of the material (plasma) conduction coefficient from Spitzer and
Harm [16] is used,

κ = c0 T 5/2 , (2.4)

where c0 = 1.0 × 10−2 .

3 Picard-Newton iterative method

Traditionally, the nonlinear iterative methods are designed for the fully implicit dis-
cretization scheme of nonlinear diffusion equations: first,(1st step) fully implicitly dis-
cretize the nonlinear partial differential equations (PDEs); then, (2nd step) linearize the
system of discrete equations (by Picard or Newton-like methods), and get a sequence
of linear system of algebraic equations (maybe the expression of coefficient matrix isn’t
explicit); finally, (3rd step) solve the linear problems. We call this kind of methods as
DL (discretization-linearization) approach. No matter which nonlinear iterative method
is taken, Picard or Newton or some other methods, the discretization has been pre-
determined independently in the DL approach.
To construct a new nonlinear iterative method (P-N), we firstly design an iteration
sequence for the time discretized nonlinear PDEs and then devise the spatial discretiza-
tion scheme for the linearized PDEs, which is named as LD (linearization-discretization)
approach. In the LD approach, the feature of linearized PDEs can be taken into consider-
ation in the construction of discrete schemes. The solution process of our P-N method is
as follows:
Step 1 Design nonlinear iteration sequence for the time-discretized nonlinear PDEs by Newton lin-
earization, and get an iteration sequence of linearized PDEs.

Step 2 Design the spatial discretization scheme for the resulting linearized PDEs, and educe a system
of linear algebraic equations with an explicit coefficient matrix.

Step 3 Solve the linear problems by linear iteration method.

848 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

3.1 Newton linearization of the time discretized PDEs

Consider the first order backward Euler time discretization of Eqs. (2.1a) and (2.1b)
E n +1 − E n
−∇·( D n+1 ∇ En+1 )− f1n+1 = 0, (3.1a)
∆t
T n +1 − T n
−∇·(κ n+1 ∇ T n+1 )− f2n+1 = 0, (3.1b)
∆t
where D n+1 = D ( En+1 ,T n+1 , ∇ En+1 ) and κ n+1 = κ ( T n+1 ) are given by (2.3) and (2.4) re-
spectively, f1n+1 and f2n+1 are given by
f1 = σa ( T 4 − E), f2 = σa ( E − T 4 ).
We directly design an iteration sequence for the time discretized nonlinear PDEs. Intro-
duce the approximations to D n+1 , κ n+1 , f1n+1 , f2n+1 by the first order Taylor expansion as
follows
∂D (s) (s)
D ( En+1 ,T n+1 , ∇ En+1 ) ≈ D ( E(s) ,T (s) , ∇ E(s) )+ ( E ,T , ∇ E(s) )( E(s+1) − E(s) )
∂E
∂D (s) (s)
+ ( E ,T , ∇ E(s) )( T (s+1) − T (s) )
∂T
2
∂D
+∑ ( E(s) ,T (s) , ∇ E(s) )(∂i E(s+1) − ∂i E(s) ), (3.2a)
i =1
∂∂ i E
∂κ (s)
κ ( T n+1 ) ≈ κ ( T (s) )+ ( T )( T (s+1) − T (s) ), (3.2b)
∂T
∂ f 1 (s) (s) ∂f
f1 ( En+1 ,T n+1 ) ≈ f1 ( E(s) ,T (s) )+ ( E ,T )( E(s+1) − E(s) )+ 1 ( E(s) ,T (s) )( T (s+1) − T (s) )
∂E ∂T
3 ( s + 1) ( s ) ( s + 1) ( s ) −4 ( s ) ( s + 1)
=z T − σa E 3
+ 3z ( T ) E ( T − T ( s ) ), (3.2c)
∂ f 2 (s) (s) ∂f
f2 ( En+1 ,T n+1 ) ≈ f2 ( E(s) ,T (s) )+ ( E ,T )( E(s+1) − E(s) )+ 2 ( E(s) ,T (s) )( T (s+1) − T (s) )
∂E ∂T
3 ( s + 1) ( s ) ( s + 1) ( s ) −4 ( s ) ( s + 1)
= −z T + σa E 3
− 3z ( T ) E ( T − T ( s ) ). (3.2d)
Inserting (3.2a)-(3.2d) into (3.1a) and (3.1b), and linearizing the resulting equations, we
can get the following iteration sequence of linearized coupled PDEs
E ( s + 1) − E n  
(s)
−∇· D ( E(s) ,T (s) , ∇ E(s) )∇ E(s+1) −(z3 T (s+1) − σa E(s+1) )
∆t  
−∇· DE′ ( E(s) ,T (s) , ∇ E(s) )( E(s+1) − E(s) )∇ E(s)
 
−∇· DT′ ( E(s) ,T (s) , ∇ E(s) )( T (s+1) − T (s) )∇ E(s)
!
2
−∇· ∑ D∂′ E (E(s) ,T (s), ∇E(s) )(∂i E(s+1) − ∂i E(s) )∇E(s)
i
i =1

− 3z3 ( T (s) )−4 E(s) ( T (s+1) − T (s) ) = 0, s=1,2, ··· , (3.3a)

J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 849

T ( s + 1) − T n  
(s)
−∇· κ ( T (s) )∇ T (s+1) +(z3 T (s+1) − σa E(s+1) )
∆t  
−∇· κ T′ ( T (s) )( T (s+1) − T (s) )∇ T (s)
+ 3z3 ( T (s) )−4 E(s) ( T (s+1) − T (s) ) = 0, s=1,2, ··· , (3.3b)

where E(0) = En , T (0) = T n , and

∂D
DE′ = = (3z3 /T 3 +|∇ E|/E)−2 |∇ E|E−2 ,
∂E
∂D
DT′ = = 9(3z3 /T 3 +|∇ E|/E)−2 z3 T −4 ,
∂T
∂D
D∂′ i E = = −(3z3 /T 3 +|∇ E|/E)−2 |∇ E|−1 E−1 ∂i E,
∂∂i E
∂κ
κ T′ = = 2.5c0 T 3/2 .
∂T
Note that the time level index, n + 1, has been omitted for simplicity. For one dimen-
sional single nonlinear parabolic equation ut −(κ (u)u x ) x = f (u) with Dirichlet boundary
conditions, we can prove that as s →∝, the solution of the iterative sequence u(s) , given
by the above linearization procedure, can converge to the solution of the time discretiza-
tion equation, un+1 , quadratically. Refer to [17, 18] for more details about convergence
analysis.
The next step is to design spatial discretization for the system of convection-diffusion
equations (3.3a) and (3.3b), and it follows that we get a Picard-Newton nonlinear iterative
scheme. Different spatial discretization can lead to different nonlinear iterative scheme.
The properties of coefficient matrix resulted from Newton linearization in DL are often
worse than those from Picard linearization. In our LD approach we can particularly pay
more attention to appropriate spatial discretization schemes for Newton correction terms,
which haven’t been considered in traditional cell-centered discretization schemes in the
DL approach.

3.2 Spatial discretization of the resulting linear PDEs

In the linearized PDEs (3.3a), (3.3b), the spacial derivative terms include not only the
original diffusion terms, but also Newton correction terms. The main diffusion terms
(denoted by FK,e ) are
 
−∇· D ( E(s) ,T (s) , ∇ E(s) )∇ E(s+1) , (3.4a)
 
−∇· κ ( T (s) )∇ E(s+1) , (3.4b)

which appear also in Picard linearization, and can be discretized by the same discretiza-
tion scheme as that used in the spacial discretization of the implicitly semi-discrete PDEs
(3.1a) and (3.1b), so that the resulting scheme is usually Picard nonlinear iterative scheme.
850 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

The Newton correction terms, which arise from Newton linearization, consist of two
parts. One part is the first order differential with respect to E(s+1) or T (s+1) (denoted by
FK,e ) as follows
 
−∇· DE′ ( E(s) ,T (s) , ∇ E(s) )( E(s+1) − E(s) )∇ E(s) , (3.5a)
 
−∇· DT′ ( E(s) ,T (s) , ∇ E(s) )( T (s+1) − T (s) )∇ E(s) , (3.5b)
 
−∇· κ T′ ( T (s) )( T (s+1) − T (s) )∇ T (s) . (3.5c)

The other part is the second order differential with respect to E(s+1) , which comes from
the partial derivative of flux-limited diffusion coefficient with respect to ∇ E (denoted by
̥K,e ) as follows
!
2
−∇· ∑ D∂′ E (E(s) ,T (s), ∇E(s) )(∂i E(s+1) − ∂i E(s) )∇E(s)
i
. (3.6)
i =1

We have found that (3.6) can be written as diffusion term with tensor coefficient:
 
−∇· Λ( E(s) ,T (s) , ∇ E(s) )(∇ E(s+1) −∇ E(s) ) , (3.7)

where
′( s ) ′( s )
!
D ∂1 E ∂1 E ( s ) D ∂2 E ∂1 E ( s )
Λ(s) = ′( s ) ′( s ) . (3.8)
D ∂1 E ∂2 E ( s ) D ∂2 E ∂2 E ( s )

We need to design the appropriate discretization schemes for these terms respectively, in
particular when multimaterial problems are solved.
For expressing discrete flux, we divide spatial domain by rectangular grid and denote
cell by K or L instead of (i, j), and with each cell K we associate one cell center denoted
also by K. The area of the cell K is m(K ). Denote the vertices by A and B, and the cell side
by e (see Fig. 1). If the cell side e is a common edge of cells K and L, and its vertices are A

K E K ,e e L
B

Figure 1: Some geometric notations.

J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 851

and B, then we denote e = K | L = BA. I denotes the midpoint of e = BA. Let ε(K ) be the set
of all cells, ∂K the set of all the cell sides of cell K, ~nK,e outward unit normal to e on K and
~ | AB|. Sometimes we will omit the iteration index (s) or (s + 1) for simplicity
~τAB = AB/
when no confusion occurs.

3.2.1 Flux approximation of the diffusion terms (3.4a)-(3.4b)

We take the term (3.4a) for example to introduce the finite volume discretization. By
integrating (3.4a) over the cell K and using the Green formula, one obtains
Z
→
∑ FK,e = − ∑ D (s) ∇ E(s+1) · n K,e dl,
e ∈∂K e ∈∂K e

where FK,e is the normal face flux on the edge e, and we can get an approximation of FK,e
(s) ( s + 1) ( s + 1)
FK,e ≈ −τK,e ( E I − EK ), (3.9)

where
(s)
(s) DK,e | AB|
τK,e = . (3.10)
| IK |
Similarly, we have
(s) ( s + 1) ( s + 1)
F L,e ≈ −τL,e ( E I − EL ), (3.11)
where
(s)
(s) D L,e | AB|
τL,e = .
| IL|
By the continuity of the normal flux component

FK,e = −F L,e , (3.12)

we can obtain
( s + 1) 1 ( s ) ( s + 1) ( s ) ( s + 1)
EI = (s) (s)
(τK,e EK + τL,e EL ). (3.13)
τK,e + τL,e
Substitute (3.13) into (3.9) to obtain
(s) ( s + 1) ( s + 1)
FK,e ≈ −τe ( EL − EK ), (3.14)

where
(s) (s)
(s) τK,e τL,e
τe = (s) (s)
.
τK,e + τL,e
(s)
DK,e in (3.10) is defined as the limit of the diffusion coefficient to the edge-center in the cell
K. Notice that the energy E and temperature T are continuous and can be defined on the
852 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

edge, while the gradient ∇ E is discontinuous possibly across edge e so that it is defined
on the triangle formed by the cell center and the edge e, and the function of material z is
defined on the cell center. So we define
1
DK,e =   , (3.15)
3

3zK Te3 +|∇ E |K,e Ee

where Ee and Te are evaluated at the cell-edge, and given by Ee = 12 ( E A + EB ), and Te =

1
2 ( TA + TB ). The cell vertex unknowns E A ,EB are approximated by the cell-centered un-
knowns. We use the method in [13] to eliminate the cell vertex unknowns. ∇ EK,e is
defined as a discrete approximation to the gradient of E on the triangle ∆KBA [14]:

EK − E A EK − E B
 
1
∇ EK,e = ~nKB + ~n , (3.16)
sinθK |K A | |KB| AK

where θK is the angle between KB and K A, ~nKB and ~n AK are outward unit normal to edges
~ and AK
KB ~ respectively on ∆KBA.
The derivation of the flux approximation of the term (3.4b) is the same.

3.2.2 Flux approximation of Newton correction terms (3.5a)-(3.5c)

The derivation of flux approximation for Newton correction terms (3.5a)-(3.5c) is similar
to (3.14). Noticing that E is continuous, we can obtain flux approximation of (3.5a) on the
edge e
(s) (s) (s) ( s + 1) (s)
FK,e ≈ −γe [EL − EK ]( Ee − Ee ) , (3.17)

where
(s)
(s) (s) ∂D
(s) γK,e γ L,e (s) | AB| ∂EK,e
γe = (s) (s)
, γK,e = .
γK,e + γ L,e |K I |

In (3.17), E(s+1) − E(s) is approximated on the cell-edge, so the resulting nonlinear iterative
scheme is named as edge-centered Picard-Newton (P-NC). Usually the dominance of the
diagonal elements of the coefficient matrix from the Picard iterative scheme is better than
that from P-NC.
It is well-known that the efficiency for iteratively solving linear system of algebraic
equations depends on the properties of coefficient matrix. So we consider a new nonlin-
ear iterative method to avoid coefficient matrix corrupting. Note that Newton correction
terms (3.5a)-(3.5c) are convective terms. In order to improve the properties of coefficient
matrix and then save CPU time, simultaneously, reflect the character of convective term,
we propose an upwind discretization of (3.5a). Denote

(s) (s) (s)

L1,e = γe ( EL − EK ).
J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 853

Then the upwind discretization of (3.5a) is defined as

(
( s + 1) (s)
L1,e ( EK − EK ), if L1,e ≤ 0 ,
FK,e = − ( s + 1) (s) (3.18)
L1,e ( EL − EL ), if L1,e > 0 .
The treatment of (3.5b) and (3.5c) is similar. The nonlinear iterative scheme using upwind
scheme to the discretization of correction terms (3.5a)-(3.5c) is named as upwind Picard-
Newton (P-NU). Later, we will give numerical comparison among Picard, P-NC and P-
NU iterative schemes.

3.2.3 Flux approximation of Newton correction term (3.6)

We use a nine point scheme [13] to discretize the Newton correction term (3.6). Integrat-
ing (3.6) over the cell K, we obtain
Z
∑ ̥K,e = − ∑ ∇ E(s+1) ·(Λ(s) )t~nK,σ dl. (3.19)
e ∈∂K e ∈∂K e

In the following we shall use the same notation ̥K,e to represent its discrete counterpart.
Notice that there exist scalars α and β depending on Λ, such that

(Λ(s) )t~nK,e = −α(s)~τBA + β(s)~nK,e , (3.20)

where
α(s) =~τAB ·((Λ(s) )t~nK,e ), β(s) =~nK,e ·((Λ(s) )t~nK,e ).
Inserting (3.20) into (3.19), we can get the approximation of flux
h i
(s) ( s + 1) ( s + 1) (s) ( s + 1) ( s + 1)
̥K,e = −τΛ,K,e E I − EK − DΛ,K,e ( E A − EB ) , (3.21)

where
(s) (s)
(s) | AB| β K,e (s) | IK |αK,e
τΛ,K,e = , DΛ,K,e = (s)
.
| IK | | AB| β K,e
Similarly,
h i
(s) ( s + 1) ( s + 1) (s) ( s + 1) ( s + 1)
̥ L,e = −τΛ,L,e E I − EL − DΛ,L,e ( EB − EA ) . (3.22)

( s + 1)
By the continuity of the normal flux component, we can eliminate E I and get the
following approximate flux:
h i
(s) ( s + 1) ( s + 1) (s) ( s + 1) ( s + 1)
̥K,e = −τΛ,e EL − EK − DΛ,e ( E A − EB ) , (3.23)

where
(s) (s)
(s) | AB| (s) | IK |αK,e | IL|α L,e
τΛ,e = | IK| | IL| , DΛ,e = (s)
+ (s)
.
(s) + (s) | AB| β K,e | AB| β L,e
β K,e β L,e
854 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

The cell vertex unknowns E A ,EB in (3.23) are approximated by the cell-centered un-
knowns [13]. From (3.23) we can see that for the diffusion term with tensor coefficient,
there are nine cells in the stencil even on rectangular grid.
For the detailed analysis of the discretization scheme see also [3, 13, 19].
By now we have introduced the discretization of the main diffusion terms and New-
ton correction terms. The finite volume scheme for the discretization of (3.3a) on the cell
K is
n +1,( s +1)
EK − EKn n +1,( s +1) n +1,( s ) n +1,( s +1)
m(K ) −(z3K TK − σa,K EK )m(K )
∆t
n +1,( s ) n +1,( s +1) n +1,( s +1)
− ∑ τe [EL − EK ]
e ∈∂K
n +1,( s ) n +1,( s ) n +1,( s ) n +1,( s +1) n +1,( s )
− ∑ γe [EL − EK ]( EN1 − EN1 )
e ∈∂K
n +1,( s ) n +1,( s ) n +1,( s ) n +1,( s +1) n +1,( s )
− ∑ µe [EL − EK ]( TN1 − TN1 )
e ∈∂K
h i
n +1,( s ) n +1,( s +1) n +1,( s +1) n +1,( s ) n +1,( s +1) n +1,( s +1)
− ∑ τΛ,e EL − EK − DΛ,e (EA − EB )
e ∈∂K
n +1,( s ) −4 n +1,( s ) n +1,( s +1) n +1,( s )
− 3z3K ( TK ) EK ( TK − TK )m(K ) = 0, s = 1,2, ··· , (3.24)

The discrete scheme of (3.3b) on the cell K is

n +1,( s +1)
TK − TKn n +1,( s +1) n +1,( s ) n +1,( s +1)
m(K ) +(z3K TK − σa,K EK )m(K )
∆t
∼n+1,( s ) n+1,( s +1) n +1,( s +1)
− ∑ τe [TL − TK ]
e ∈∂K
∼ n+1,( s ) n+1,( s ) n +1,( s ) n +1,( s +1) n +1,( s )
− ∑ µe [TL − TK ]( TN1 − TN1 )
e ∈∂K
n +1,( s ) −4 n +1,( s ) n +1,( s +1) n +1,( s )
+ 3z3K ( TK ) EK ( TK − TK )m(K ) = 0, s = 1,2, ··· , (3.25)

where
∼ n+1,( s ) ∼ n+1,( s )
∼ n+1,( s ) τ K,e τ ∼n+1,( s ) | AB| n+1,(s)
τe = ∼ n+1,(s) ∼L,en+1,(s) , τ K,e = κ ,
τ K,e + τ L,e |K I | K,e
n +1,( s ) n +1,( s ) n +1,( s )
n +1,( s ) µK,e µ n +1,( s ) | AB| ∂DK,e
µe = n+1,(s) L,en+1,(s) , µK,e = ,
µK,e + µ L,e |K I | ∂T
∼ n+1,( s ) ∼ n+1,( s ) n +1,( s )
∼ n+1,( s ) µ µ L,e ∼ n+1,( s ) | AB| ∂κK,e
µe = ∼ K,e , µ K,e = .
+1,( s ) ∼ n+1,( s ) |K I | ∂T
µ nK,e + µ L,e
J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 855

∂D ∂D ∂κ
The definitions of ∂EK,e , ∂TK,e , ΛK,e , κK,e , ∂TK,e are similar to (3.15). The terms on the 3rd, 4th
lines of (3.24) and the 3rd line of (3.25) are the discretization corresponding to Newton
correction terms (3.5a)-(3.5c). If E(s+1) − E(s) and T (s+1) − T (s) are approximated on the
cell-edge as (3.17) (N1 = e), then the resulting nonlinear iterative scheme is named as
edge-centered Picard-Newton (P-NC). While they are approximated as (3.18) ( N1 = K for
L1,e ≤ 0 or L for L1,e > 0), the resulting nonlinear iterative scheme is named as upwind
Picard-Newton (P-NU). In the following we turn to the emphasis of this work that is to
investigate the effect of time step control on Picard-Newton nonlinear iterative method.

4 Time step control

Because the radiation diffusion problems exhibit multiple time scale, time step control
algorithms are used to increase solution efficiency. Traditionally, the time step control for
methods which do not converge nonlinearities within a time step is based on the relative
change in E or T, i.e. ∆E/E or ∆T/T. If one assumes ∆E/E > ∆T/T, then one can use
the time step control method based on ∆E/E. In this paper the relative change in E is
defined by
!
n |EKn − EKn−1 |
η = max , (4.1)
K ∈ε(K ) EKn−1/2

where EKn−1/2 = ( EKn + EKn−1 )/2. The superscript n is the time index and the subscript K is
the space grid cell index. The time step size is adjusted to achieve a prescribed maximum
relative change in E (ηtarget )

0.5
ηtarget

n +1 n
∆tre = ∆t . (4.2)
ηn

For convenience we call this method ∆E/E method. In this paper, the final time step for
∆E/E method is chosen to limit the rate of growth by 10%

∆tn+1 = min(1.1∆tn ,∆tre

n +1 n +1
,∆tmax ). (4.3)

Therefore, given the initial time step size and “target maximum ∆E/E”, the time step
sizes can be computed from (4.3).
Another kind of time step control method is a thermal front CFL approach which is
originally presented in [12]. The basic idea is to estimate the dominant wave propagation
speed in the problem. In one dimension this is the ratio of temporal to spatial derivative
of the variable. Briefly, one can assume that the following hyperbolic PDE

∂E ∂E
+ vrad = 0 (4.4)
∂t ∂x
856 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

models the problem. One can use the numerical approximation of Eq. (4.4) and get
||△E/△t||1
vrad = − .
||△E/△ x||1
Then the time step size is adjusted to achieve a target CFL number

n +1 CFL k ∆x k
∆tCFL = . (4.5)
vn
rad
In this paper, we let
q
k ∆x k= 2m(K ),

∑ |Ekn − EKn−1 |/∆tn

n K ∈ε(K )
vrad = .
∑ |∇ EKn |
K ∈ε(K )

This is from [6], and the differences are that we include the cells near the boundaries in
the sum, which are excluded in [6], and the approximation to |∇ EKn | is different. We use
(3.16) to discretize the gradient.
Similar to (4.3), the final time step based on CFL method employed in this paper is
defined by
∆tn+1 = min(1.1∆tn ,∆tnCFL
+1 n +1
,∆tmax ). (4.6)
Given the initial time step size and “target CFL”, the time step sizes can be computed
from (4.6).

5 Numerical results
In this section, we present numerical results from Picard and P-N methods. Time step
convergence studies are performed and CPU time will be compared for desired accuracy.
The test problems are solved on the domain 0 ≤ x ≤ 1 and 0 ≤ y ≤ 1. The boundary
condition for the radiation equation at the left boundary (x = 0, 0 ≤ y ≤ 1) is
1 1 ∂E
E− = 1,
4 2D ∂x
and at the right boundary (x = 1, 0 ≤ y ≤ 1) is
1 1 ∂E
E+ = 0.
4 2D ∂x
At the top and bottom boundaries (y = 0, y = 1, 0 ≤ x ≤ 1) we use
∂E
= 0.
∂y
J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 857

For the material conduction equation, all four walls are insulated:
∂T ∂T ∂T ∂T
= = = = 0.
∂x x =0 ∂x x =1 ∂y y =0 ∂y y =1

The value for z is 1 everywhere except in the rectangle defined by 13 ≤ x ≤ 23 , and 13 ≤ y ≤ 32 ,

where z = zhigh . The initial energy is E0 = 1 × 10−5 , and the initial material temperature is
T0 = ( E0 )1/4 . The problem is initialized to equilibrium with radiation temperature being
equal to material temperature. The simulation runs for three units of time (tend = 3).
In this paper, we use Picard and P-N methods to solve the nonlinear system and GM-
RES method with ILUT preconditioner to solve the linear system. The program is writ-
ten in FORTRAN, and run on a windows system. The nonlinear convergence tolerance
within a time step is defined as ||F (us+1 )|| ≤ 1 × 10−6 and ||F(us+1 )|| ≤ 1 × 10−2 ||F(u0 )||,
where F(u) is the residual function. The expression of the residual function for P-N is
the same as that for Picard. We will take the default maximum number of nonlinear it-
erations to be 20, and will also consider other numbers of nonlinear iterations for Picard
method.
Since Picard is a proven solver and we will compare P-N with Picard, a base solu-
tion is computed using Picard for the sake of impartiality. The base solution is com-
puted with a time step of ηtarget = 0.005 and the nonlinear convergence tolerance of
||F(us+1 )|| ≤ 1 × 10−8 and ||F(us+1 )|| ≤ 1 × 10−4 ||F(u0 )||. For the time-step convergence
study, radiation temperature is the variable used to compute the L2 norm between the
base and the computed solution:
s
L2 ( Error) = ∑ (Tr (K )− Trb (K ))2m(K ),
K ∈ε(K )

where Trb is the radiation temperature from the base solution.

The test problems are chosen to represent more nonlinear problem with zhigh = 10 and
less nonlinear problems with zhigh = 5 and zhigh = 2.5.

5.1 More nonlinear problem

In the more nonlinear problem, zhigh = 10. The material is initially cold, and there is the
constant energy flux into the left wall. Therefore, the initial constant energy flux is being
imparted into the initially cold material. At t = 3, it reaches the state shown in Fig. 2. The
energy flux spreads from the left wall until it has engulfed the obstacle of high z material.
Fig. 2 shows that the high z material prevents the radiation from entering the obstacle,
and the center of the obstacle is still cold.

5.1.1 Results for time step of ∆E/E method

The first result to be presented is a time step convergence study with the ∆E/E method
to control the time step. The change in timescales from the ∆E/E method can be seen in
858 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

0.8

0.6

Y
0.4

0.2

0.2 0.4 0.6 0.8 1

X

Figure 2: Contours of material temperature at t = 3.

Figs. 3 and 4. These figures show the time step sizes from P-NU and Picard methods with
ηtarget = 0.2 in (4.3), respectively. The time step sizes from the two methods are almost
the same except around t = 1.6 and t = 2.6, where time step sizes from Picard decrease
significantly.

0.002 0.002

0.0015 0.0015
Time Step

Time Step

0.001 0.001

0.0005 0.0005
0.5 1 1.5 2 2.5 3 0.5 1 1.5 2 2.5 3
t t

Figure 3: Time steps for P-NU with ηtarget = Figure 4: Time steps for Picard with ηtarget =
0.2 for 60 × 60 problem. 0.2 for 60 × 60 problem.

Fig. 5 shows a time step convergence study for Picard and P-N methods on 60 × 60
grid with 0.1 ≤ ηtarget ≤ 1, where ηtarget is depicted on the x axis and L2 norm of error
on the y axis. The dash-dot line represents a zero-order line, which shows zero-order
time accuracy for Picard method. The dotted line represents a first-order line, which
shows first-order time accuracy for P-N. The designed first-order time convergence of
P-N is apparent, however, Picard does not achieve the first-order time convergence and
produces almost 0th order time convergence. This is because Picard method fails to con-
verge within 20 nonlinear iterations when 0.1 ≤ ηtarget ≤ 1. Even we set a maximum of
nonlinear iterations by 100, Picard method fails to converge under this loose tolerance.
J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 859

0 0
10 10

−1
10
−1
10
L2 Error

L Error
−2
10

2
−2
10

−3
10

P−NU
P−NC
Picard P−NU
1st line −3 P−NC
10
0th order Picard
−4
10 2 3 4
−1 0
10 10 10 10 10
Relative Energy Change (target) CPUTime (s)

Figure 5: Time step convergence study with 0.1 ≤ Figure 6: Efficacy with 0.1 ≤ ηtarget ≤ 1 for 60× 60
ηtarget ≤ 1 for 60 × 60 problem. problem.

Therefore the errors build up and the solution drifts away from the correct solution. For
a given ηtarget , the L2 error of Picard is an order of magnitude larger than that of P-N,
and P-NU is slightly more accurate than P-NC.
Fig. 6 shows efficacy for Picard and P-N methods with 0.1 ≤ ηtarget ≤ 1 on 60 × 60 grid.
In this plot, the same error values shown in Fig. 5 are now plotted against the CPU time
of the numerical computation. An efficacy plot can directly show which method gives
the highest level of accuracy for the smallest amount of CPU time. As was expected,
P-NU is the most effective. The properties of coefficient matrix of linear system from P-
NU are better than those from P-NC, so the number of linear iterations of P-NU should
be less than that of P-NC. Table 1 compares average nonlinear iterations per time step
nonlinear linear
( time − step ), average linear iterations per nonlinear iteration step ( nonlinear ) and average
linear
linear iterations per time step ( time − step ). From the table we can see that the numbers of
linear
linear iterations ( nonlinear ) from P-NU with different ηtarget are less than those from P-NC.
Simultaneously, good properties of linear system result in slight decrease on nonlinear
iterations.

Table 1: Algorithm performances as a function of time step (ηtarget ) at t = 3.

nonlinear linear linear
Method and ηtarget time −step nonlinear time −step
P-NU, 0.2 1.852 3.264 6.045
P-NC, 0.2 1.856 3.282 6.092
P-NU, 0.6 2.272 3.516 7.988
P-NC, 0.6 2.273 3.733 8.487
P-NU, 1.0 2.837 3.691 10.472
P-NC, 1.0 2.851 3.863 11.012
860 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

0.005 0.005

0.004 0.004
Time Step

Time Step
0.003 0.003

0.002 0.002

0.001 0.001

0.5 1 1.5 2 2.5 3 0.5 1 1.5 2 2.5 3

t t

Figure 7: Time steps from Picard method Figure 8: Time steps from Picard method
with time step control (5.2) of ηtarget = with time step control (5.2) of ηtarget =
0.2 for 60 × 60 problem. 0.8 for 60 × 60 problem.

When 0.1 ≤ ηtarget ≤ 1 and the maximum number of nonlinear iterations is set by 20,
Picard method can not converge at all time levels under the time step control (4.3). To
ensure the convergence of Picard method, we set a maximum of nonlinear iterations by
50, and once the number of nonlinear iterations is more than 20 we cut down the time
step size by 20%:
n +1
∆tPicard = 0.8∆tn , (5.1)
and the final time step size is computed by

∆tn+1 = min(∆tre
n +1 n +1
,∆tPicard ). (5.2)

In this case, about after t = 1.2, time step size is mainly determined by (5.1) instead of
relative change in E , and time step sizes from different ηtarget are almost the same. The
time step sizes from ηtarget = 0.2 and ηtarget = 0.8 are shown in Figs. 7 and 8 respectively.
In order to compare accuracy and efficiency of P-N with Picard in the case that these
methods are all convergent, firstly we solve the more nonlinear problem by Picard with
time step determined by (5.2), and keep the time step sizes record; then use the time
step sizes recorded to solve the same problem by P-N. Fig. 9 shows an efficacy plot for

−2.5
10

−2.6
10

−2.7
10
L2 Error

−2.8
10

−2.9
10
P−NU
P−NC
Picard
3 4
10 10
CPUTime (s)

Figure 9: Efficacy of Picard and P-N methods with the same time step size.
J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 861

−2 −3
10 10
L Error
2

L2 Error
−3
10

P−NU
−4
10 P−NC P−NU
Picard P−NC
1st line
−4
Picard
−1.6 −1.5 −1.4 −1.3 −1.2 −1.1 10
10 10 10 10 10 10 3.1 3.2 3.3 3.4 3.5 3.6
10 10 10 10 10 10
Relative Energy Change (target) CPUTime (s)

Figure 10: Time step convergence study with Figure 11: Efficacy with 0.03 ≤ ηtarget ≤ 0.07
0.03 ≤ ηtarget ≤ 0.07 for 60 × 60 problem. for 60 × 60 problem.

−2
10
−2
10
L2 Error

L Error
2

−3
10
−3
10

P−NU
P−NC
1st line
P−NU
P−NC
−4
−4 10
10 3 4
−1
10
0
10 10 10
Relative Energy Change (target) CPUTime (s)

Figure 12: Time step convergence study with Figure 13: Efficacy with 0.1 ≤ ηtarget ≤ 1 for
0.1 ≤ ηtarget ≤ 1 for 120 × 120 problem. 120 × 120 problem.

the three schemes with the same time step sizes obtained from Picard method. P-NU
produces the same level of accuracy for the least CPU time.
When time step is determined by ∆E/E method (4.3) with smaller ηtarget , Picard can
converge within 20 nonlinear iterations. Fig. 10 shows a time step convergence study
for Picard and P-N methods on 60 × 60 grid with 0.03 ≤ ηtarget ≤ 0.07. The dotted line
represents a first-order line. Picard is slightly more accurate than P-N when ηtarget is
small. The slopes for P-N are very close to the expected value of 1, while Picard does not
have straight lines on the convergence plot. Fig. 11 shows an efficacy plot corresponding
to Fig. 10. P-NU is still the most effective.
Fig. 12 shows a time step convergence study for P-N on 120 × 120 grid. P-N schemes
achieve the 1st order convergence. Fig. 13 shows the corresponding efficacy plot. The
862 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

two plots show that P-NU is still more accurate and efficacious as the mesh is refined.
Under the same time step, Picard still fails to converge within 20 iterations.

5.1.2 Results for time step of CFL method

In this section, the thermal wave CFL method is used to control the time step. Table 2
compares the CFL number with the corresponding average value of η n for selected runs.

Table 2: A comparison of CFL number and corresponding η.

CFL 0.01 0.02 0.04 0.08 0.1

η 0.1839 0.3603 0.6766 1.1332 1.2770

Fig. 14 shows the time step sizes from P-NU with CFL = 0.01 in (4.6).

0.0025

0.002
Time Step

0.0015

0.001

0.0005

0.5 1 1.5 2 2.5 3

t

Figure 14: Time steps for P-NU with CFL = 0.01 for 60 × 60 problem.

Fig. 15 shows a time step convergence study for P-N on two grids (60 × 60, 120 × 120),
where L2 norm of error (on the y axis) is plotted as a function of CFL number (on the x
axis). The dotted line represents a first-order line. P-NU is slightly more accurate than
P-NC. Fig. 16 shows an efficacy plot, which illustrates that P-NU is more effective.
Similar to the results in the above section, Picard method fails to converge within
20 nonlinear iterations with CFL method (4.6) of the same target CFL numbers used in
Figs. 15 and 16. In order to obtain convergent solution for Picard method, we introduce
time step control (5.1) and the final time step size is rewritten by

∆tn+1 = min(∆tnCFL
+1 n +1
,∆tPicard ). (5.3)

Table 3 presents comparisons for various methods and target CFL numbers. In Table 3,
Picard-(4.6) and P-NU-(4.6) denote Picard and P-NU with time step control (4.6) respec-
tively, and Picard-(5.3) denotes Picard with time step control (5.3). P-NU-(5.3) denotes
P-NU with the same time step sizes as what is obtained by (5.3) from Picard. P-NU is
more efficacious than Picard under the two time step control methods.
J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 863

−1 −1
10 10
P−NU 60
P−NC 60
P−NU 120
P−NC 120

L2 Error
L Error
2

−2
−2 10
10

P−NU 60
P−NC 60
1st line
P−NU 120
P−NC 120
−3
10
2 3 4
−3 10 10 10
10
−2 −1
10 Thermal wave CFL 10 CPUTime (s)

Figure 15: Time step convergence study of P-N Figure 16: Efficacy of P-N with CFL method.
with CFL method.

Table 3: Algorithm performances as a function of CFL at t = 3.

nonlinear linear
Method and CFL Num.of time step time −step time −step L2 error CPU
P-NU-(4.6), 0.02 1196 2.02 6.77 4.96d-3 310
P-NU-(4.6), 0.08 324 3.44 13.31 1.52d-2 126
P-NU-(5.3), 0.02 3368 1.24 3.74 2.03d-3 633
P-NU-(5.3), 0.08 2877 1.21 3.73 3.20d-3 536
Picard-(4.6), 0.02 1125 10.14 25.51 1.2d-1 896
Picard-(4.6), 0.08 300 11.37 33.10 9.8d-2 248
Picard-(5.3), 0.02 3368 15.14 38.77 1.92d-3 3755
Picard-(5.3), 0.08 2877 17.28 44.34 3.16d-3 3524

5.2 Less nonlinear problems

In the less nonlinear problem, zhigh = 5 and zhigh = 2.5. From the foregoing statement,
we can see that P-NU is better than P-NC, so we only compare Picard with P-NU in this
section. Fig. 17 shows a time step convergence study for Picard and P-NU methods with
zhigh = 5 on 60× 60 grid. P-NU achieves the 1st order time convergence. Picard still fails to
converge within 20 iterations for the largest ηtarget (= 0.8) in Fig. 17. When ηtarget ≤ 0.6,
Picard achieves the 1st order time convergence, which cannot be obtained from the more
nonlinear problem. Fig. 18 shows the corresponding efficacy plot. From the plot, it is
clear that for Picard method, the CPU times corresponding to the largest and the second
largest errors are abnormal, and they are larger than the CPU times corresponding to the
smaller errors. This is because the nonlinear iterations increase abruptly when the time
step is enlarged.
Fig. 19 shows a time step convergence study for Picard and P-NU methods with
zhigh = 2.5 on 60 × 60 grid. Both of the two methods achieve the 1st order time conver-
gence. Fig. 20 shows the corresponding efficacy plot. The difference in efficiency between
864 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

−1 −1
10 10

L2 Error
−2
10
L Error

−2
10
2

−3
10

P−NU P−NU
Picard Picard
1st line 10
−3
−4 2 3 4
10 10 10 10
−1 0
10 10
Relative Energy Chang (target) CPUTime (s)

Figure 17: Time step convergence study with Figure 18: Efficacy with ∆E/E method for
∆E/E method for zhigh = 5. zhigh = 5.

−1 −1
10 10
P−NU P−NU
Picard Picard
1st line
L2 Error
L Error
2

−2
10 −2
10

−3
10
−0.8 −0.7 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 −3
10 10 10 10 10 10 10 10 10 10
2 3
Relative Energy Change (target) 10 CPUTime (s) 10

Figure 19: Time step convergence study with Figure 20: Efficacy with ∆E/E method for
∆E/E method for zhigh = 2.5. zhigh = 2.5.

Picard and P-NU goes down, and Picard method behaves well as the problem is easy to
be simulated.

6 Conclusion
A detailed computational cost and efficiency comparison are presented for Picard and
P-N methods with time step control to solve multimaterial non-equilibrium radiation
diffusion problem. By using two time step control methods, our P-N method can yield
high quality numerical results with reduced cost. This is accomplished by running the
simulation at a larger time step, while maintaining accuracy through converging the non-
J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866 865

linearities within less iterations. For our P-N, it does not need to introduce more restricted
time steps than that used for Picard. The behavior of Picard method becomes ruin as the
test problems become more nonlinear, while the good performance of P-NU still remains.
The properties of coefficient matrix of linear system from P-NU are better than those from
P-NC, so P-NU is the most effective among Picard, P-NC and P-NU.

Acknowledgments
This work was partially supported by the Basic Research Project of National Defence
(B1520110011), the Foundation of CAEP (2010A0202010), and the Foundation of National
Key Laboratory of Science and Technology on Computational Physics. The authors thank
the reviewers for their suggestions which improved the presentation of the paper signif-
icantly.

References

[1] R. L. Bowers and J. R. Wilson, Numerical Modeling in Applied Physics and Astrophysics,
Jones & Bartlett, Boston, 1991.
[2] P. N. Brown, D. E. Shumaker and C. S. Woodward, Fully implicit solution of large-scale non-
equilibrium radiation diffusion with high order time integration, J. Comp. Phys., 204:760–
783, 2005.
[3] R. Eymard, T. Gallouët and R. Herbin, Finite Volume Methods, Handbook of Numerical
Analysis, P. G. Ciarlet, J. L. Lions (Eds.), Vol. 7, 713–1020, 1997.
[4] D. A. Knoll, W. J. Rider and G. L. Olson, An efficient nonlinear solution method for non-
equilibrium radiation diffusion, J. Quant. Spectrosc. Radiat. Transfer, 63:15–29, 1999.
[5] D. A. Knoll, W. J. Rider and G. L. Olson, Nonlinear convergence, accuracy, and time step
control in non-equilibrium radiation diffusion, J. Quant. Spectrosc. Radiat. Transfer, 70:25–
36, 2001.
[6] V. A. Mousseau and D. A. Knoll, New physics-based preconditioning of implicit methods
for non-equilibrium radiation diffusion, J. Comput. Phys., 190:42–51, 2003.
[7] V. A. Mousseau and D. A. Knoll, Temporal accuracy of the non-equilibrium radiation diffu-
sion equations applied to two-dimensional multimaterial simulations, Nuclear Science and
Enginerring, 154:174–189, 2006.
[8] V. A. Mousseau, D. A. Knoll and W. J. Rider, Physics-based preconditioning and the Newton-
Krylov method for non-equilibrium radiation diffusion, J. Comp. Phys., 160:743–765, 2000.
[9] C. C. Ober and J. N. Shadid, Studies on the accuracy of time-integration methods for the
radiation-diffusion equations, J. Comp. Phys., 195:743–772, 2004.
[10] G. L. Olson, Efficient solution of multi-dimensional flux-limited non-equilibrium radiation
diffusion coupled to material conduction with second-order time discretization, J. Comp.
Phys., 226:1181–1195, 2007.
[11] W. J. Rider, D. A. Knoll and G. L. Olson, A multigrid Newton-Krylov method for multidi-
mensional equilibrium radiation diffusion, J. Comp. Phys., 152:164–191, 1999.
[12] W. J. Rider and D. A. Knoll, Time step size selection for radiation diffusion calculations, J.
Comp. Phys., 152:790–795, 1999.
866 J. Yue and G. Yuan / Commun. Comput. Phys., 10 (2011), pp. 844-866

[13] Z. Sheng and G. Yuan, A nine point scheme for the approximation of diffusion operators on
distorted quadrilateral meshes, SIAM J. Sci. Comput., 30:1341–1361, 2008.
[14] Z. Sheng, J. Yue and G. Yuan, Monotone finite volume schemes of non-equilibrium radiation
diffusion equations on distorted meshes, SIAM J. Sci. Comput., 31:2915–2934, 2009.
[15] A. Shestakov, J. Greenough and L. Howell, Solving the radiation diffusion and energy bal-
ance equations using pseudo-transient continuation, J. Quantitative Spectroscopy Radiative
Transfer, 90:1–28, 2005.
[16] L. Spitzer and R. Harm, Transport phenomena in a completely ionized gas, Phys. Rev.,
89(5):977, 1953.
[17] G. Yuan, Acceleration techniques of iterative solution methods for nonlinear parabolic equa-
tions, Annual Report of Laboratory of Computaional Physics, 366–413, 2004.
[18] G. Yuan and X. Hang, Acceleration methods of nonlinear iteration for nonlinear parabolic
equations, J. Comput. Math., 24:412–424, 2006.
[19] G. Yuan and Z. Sheng, Analysis of accuracy of a finite volume scheme for diffusion equations
on distorted meshes, J. Comp. Phys., 224:1170–1189, 2007.
[20] G. Yuan, X. Hang, Z. Sheng and J. Yue, Progress in numerical methods for radiation diffusion
equations, Chinese J. Comp. Phys., 26:475–500, 2009.
[21] J. Yue, G. Yuan and Z. Sheng, Picard-Newton iterative method for multimaterial nonequilib-
rium radiation diffusion problem on distorted quadrilateral meshes, Lecture Notes in Engi-
neering and Computer Science, The World Congress on Engineering 2009, Vol. 2, 1180–1185.