Annals of Nuclear Energy 63 (2014) 83–99

Contents lists available at SciVerse ScienceDirect

Annals of Nuclear Energy

journal homepage: www.elsevier.com/locate/anucene

An advancement in iterative solution schemes for three-dimensional,

two-fluid modeling of two-phase flow in PWR fuel bundles
Maryam Mohitpour a,⇑, Gholamreza Jahanfarnia a, Mehrzad Shams b
a
Department of Nuclear Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran
b
Department of Mechanical Engineering, K.N. Toosi University of Technology, Tehran, Iran

a r t i c l e i n f o a b s t r a c t

Article history: This paper outlines a fully three-dimensional two-fluid one-pressure model with a semi-implicit finite
Received 12 April 2013 difference scheme coupled with heat conduction which can be applicable to thermal non-equilibrium
Received in revised form 4 July 2013 two-phase flow field in subchannel geometry of Pressurized Water Reactors (PWR). The system of equa-
Accepted 9 July 2013
tions was linearized using the Newton–Raphson method and was collapsed into the pressure equations
Available online 15 August 2013
forming a system of the Poisson type. Then, two-phase flow modeling was combined with Krylov meth-
ods as advanced computing techniques to investigate the feasibility of implementing preconditioned Kry-
Keywords:
lov subspace solvers as the numerical scheme to solve pressure equations. Six popular Krylov subspace
Two-phase flow
Two-fluid model
solvers were considered: GMRES, FGMRES, DQGMRES, CGNR, BCG, and TFQMR combined with the block
Semi-implicit scheme incomplete LU factorization with a dual truncation strategy (BILUT) preconditioner. These proposed iter-
PWR subchannel analysis ative solvers were applied to the constructed linear pressure equations in the inner iteration in combina-
Krylov subspace methods tion with the outer-Raphson iteration loop. Evaluation was performed in two stages. First, two-fluid
numerical scheme capability was evaluated against OECD/NRC NUPEC PWR Bundle tests (PSBT Bench-
mark). The results for steady-state (PSBT) bundle show that an overall agreement can be found. At the
second stage, convergency, stability, and accuracy of the proposed schemes were studied based on PSBT
steady-state data through a comparison of utilized Krylov solvers and the direct inversion method as the
pressure solution modeling options. It was found that the Krylov solvers do not introduce deviations to
the two-phase flow results. Moreover, the convergence history of candidate Krylov solvers demonstrated
that GMRES, FGMRES, and DQGMRES have a more efficient and stable convergence performance.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction will not be practical in the near future. Large-scale, global, three-
dimensional, two-phase two-fluid modeling capability already ex-
In the nuclear reactor safety area, during the last 30–40 years, ists in certain system codes such as TRACE (U.S. NRC), RELAP-3D
thermal hydraulics has been one of the key disciplines for simula- (U.S. NRC), and CATHARE 3 (Emonot et al., 2011), and it is used
tion and analysis of transient and accident scenarios. The topic of to describe two-phase flows in certain parts of the primary system,
two-phase flow in fuel rod bundles, due to its safety implications, downcomer, core, etc. However, the meshing used is still very
has received a continuous focus from the research and has sus- coarse, and the effects of turbulence are normally not included
tained its high relevance in nuclear technology. To filter the dimen- (Yadigaroglu et al., 2003). COBRA-TF (Pacific Northwest Labora-
sions smaller than the subchannel size, the physical modeling tory) is an advanced subchannel code applicable to both PWR
requires a 3D local two-fluid model. This approach may be proved and BWR analyses. The code represents a two-fluid, three-field
useful when the knowledge of the details of the flow inside the (continuous vapor, continuous liquid, and entrained liquid drops)
subchannel has a direct impact on the phenomena under model of the two-phase flow. The conservation equations for mul-
consideration. tidimensional flow are discretized using a semi-implicit finite-dif-
The capabilities of the CFD codes are currently restricted to ference technique with donor cell differencing for the convective
certain two-phase flow regimes. Furthermore, they incur a huge quantities. On the other hand, momentum equations for the phases
computational cost. Consequently, it seems that the direct use of are written for vertical and transverse directions. FIDAS code mod-
two-phase CFD codes in the transient analysis of nuclear systems els the two-phase flow as three fluids (three fields) and uses Eule-
rian solution scheme (Sugawara and Miyamoto, 1990). The code
solves three continuity, three energy, and six momentum equa-
⇑ Corresponding author. Tel.: +98 917 3087852.
tions, the latter for axial and lateral velocities. As an outcome of
E-mail addresses: [email protected] (M. Mohitpour), [email protected]
(G. Jahanfarnia), [email protected] (M. Shams).
the development of sophisticated system analysis codes, it became

0306-4549/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.anucene.2013.07.007
84 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99

Nomenclature

A area (m2) T temperature (K)

c specific heat (J/kg K) V velocity (m/s)
D diameter (m) Wt mass exchange due to turbulent (W/m3 s)
e internal energy (J/kg)
Fi vapor–liquid interfacial momentum exchange rate (Pa/ Greek symbols
m) a void fraction
Ft turbulent momentum exchange rate (Pa/m) c phase exchange rate (W/m3 s)
Fw wall frictional force (Pa/m) q density (kg/m3)
g gravitational constant (m/s2) e/l turbulent velocity (m/s)
G mass flux (kg/m2 s)
h heat transfer coefficient (W/m2 K)
Subscripts
i enthalpy (J/kg) g vapor
k thermal conductivity (W/m K) i interfacial
P pressure (Pa) l liquid
q00 heat flux (W/m2)
s saturation
q000 heat source density (W/m3) t turbulent
Qi interfacial heat transfer rate (W/m3) w wall
Qt turbulent heat transfer rate (W/m3)
Qw wall heat transfer rate (W/m3)
Re Reynolds number

quite obvious that the 6-equation, 2-fluid model forms a suitable equations coupled with nonlinear heat conduction using the Jaco-
basis together with the semi-implicit numerical solution scheme bian-free Newton–Krylov (JFNK) method, which employs the oper-
to predict complicated, highly-transient flows considering non- ator-split semi-implicit method (OSSI) as a physics-based
equilibrium conditions. preconditioner. Cuervo et al. (2006) implemented the BiCGSTAB
In addition to work on the two-phase flow physics, new ap- method combined with ILUT preconditioner in COBRA-TF code
proaches have focused on improving performance and efficiency for solving the linear system of pressure equations.
of numerical methods in order to obtain reasonable running times The objective of this paper is to employ field conversation equa-
for large problems with good accuracy and stability of the numer- tions to model thermal non-equilibrium two-phase flow in rod
ical scheme which is often based on knowledge of the characteris- bundle on a subchannel basis of Pressurized Water Reactor
tics of the partial differential equations. The memory and (PWR). According to the brief review above, we take THERMIT
computational requirements for solving three-dimensional Partial (Kelly et al., 1981) as the basic tool for the development of numer-
Differential Equations (PDEs) may seriously challenge the most ical simulation of two-phase flow in rod bundles because of its
efficient direct solvers which are available today. It was found that main features specified below: two-fluid 3-D model, heat conduc-
the combination of preconditioning and Krylov subspace iterations, tion in fuel rods, realistic modeling of thermal and mechanical
categorized as nonstationary iterative methods, could provide effi- non-equilibrium between the phases, representation of momen-
cient and simple ‘‘general-purpose’’ procedures that could compete tum equations for each phase in x, y, and z directions (six momen-
with direct solvers (Saad, 2000). tum equations), and implementation of turbulent mixing terms
The superior performance of Krylov solvers, as compared to the (transfer of mass, energy, and momentum across the interchannel
stationary iterative methods, has been well documented for ther- interfaces). The mixing terms are one of the most important phe-
mal–hydraulic problems in nuclear reactors. The application of nomena that must be taken into account in subchannel analysis.
preconditioned conjugate gradient methods to the linearized pres- Further, the implementation of the preconditioned Krylov method
sure equation is presented in (Turner and Doster, 1991). Allaire is introduced as an accurate and efficient technique in the solution
(1995) implemented a preconditioned Krylov solver for linearized of the linear algebraic system of pressure equations.
three-dimensional two-phase flow equations in the vessel ther- The present research differs from previous studies in two ways:
mal–hydraulics code FLICA-4 developed at CEA, France. Joo and
Downar (1996) implemented the preconditioned BiCGSTAB solver, 1. We solve a 10-field equation model for two-phase flow coupled
with both block incomplete LU factorization and domain decompo- with heat conduction in the form of first-order finite difference
sition preconditioners, in the NESTLE code (Turinsky et al., 1994). with a semi-implicit solution technique. The emphasis is on
Hotta et al. (1998) utilized four solvers (ILUBCG, MILUBCG, ILUCR, important phenomena associated with subchannel analysis.
and MILUCR) as the pressure iteration scheme for solving the com- 2. The idea of using different Krylov subspace solvers, all of which
pressible two-fluid one-pressure (six equations) model. Downar are introduced for non-symmetric linear systems, is examined
and Joo (2001) applied the BiCGSTAB method as the continuity as the pressure iterative scheme.
equation solution scheme in VIPRE-02 (1994), which is a two-fluid
two-field code for subchannel analysis. Kastanya (2002) imple- The remainder of the paper is organized as follows. In Section 2,
mented three Krylov solvers (CGS, BiCGSTAB and GMRES) within we present a physical model for three-dimensional two-fluid for-
a BWR core simulator code FORMOSA-B. Gan et al. (2003) em- malism coupled with the heat conduction equation. In Section 3,
ployed a Jacobian-free Newton method to solve a transient which the discretized version of the physical model is presented which
couples a two-phase flow code (TRAC-M, 2000) with a 3-D neutron includes both spatial and temporal differencing methods. Section 4
kinetics code (PARCS, 1998). Frepoli et al. (2003) used Newton is dedicated to the numerical scheme in the solution of constructed
method to solve an eight-equation model for transient two-phase nonlinear equations. Section 5 briefly discusses the Krylov sub-
flow. Mousseau (2004) presented a solution for the two-phase flow space method. Section 6 gives a description of the implementation
 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99 85

of Krylov iterative solvers for solving the pressure matrix. Feasibil- Conservation of vapor momentum:
ity of solving the two-fluid equations using the proposed solvers is !
analyzed via the NUPEC PSBT test case in Section 7. Lastly, Section 8 @Vv ! !
aqv þ aqv V v  r V v þ arP
presents the conclusion and offers ideas for future work. @t
! ! !
¼  F wv  F iv þ aqv ~
g  F tv ð5Þ
2. Physical model
Conservation of liquid momentum:
In this section, we present the field and constitutive equations of !
the two-fluid model. We are aware that the potential inherent in @Vl ! !
ð1  aÞql þ ð1  aÞql V l  r V l þ ð1  aÞrP
the two-fluid approach can only be utilized if validated constitutive @t
! ! !
relations are available and is only justified if the presented physics ¼  F wl  F il þ ð1  aÞql~
g  F tl ð6Þ
of the two-phase interaction involve phasic mechanical and ther-
mal non-equilibrium. The complexity introduced by using a flow The above-mentioned transport processes will be described la-
regime map is evident from the need for a large number of consti- ter. The field equations are usually derived by assuming that inter-
tutive relations. A simpler approach to the flow regime dependence face separating the phases has zero thickness and zero mass and
of the interfacial exchange models reduces the number of required hence cannot store momentum or kinetic thermal energy. There-
constitutive equations (and associated logic) and at the same time fore, conservation of mass, momentum, and energy imply that
maintains a good representation of the available data. the jump conditions at the interface are:
The interfacial exchange models described here have been
incorporated in the THERMIT-2 (Kelly and Kazimi, 1981). It is 1. Mass
worth noting that some two-fluid models of two-phase flow are Cv þ Cl ¼ 0 : C ¼ Cv ¼ Cl ð7Þ
ill-posed in the mathematical sense. The equations displayed be-
low are one such ill-posed system, and no attempt was made in 2. Energy
this study to address this problem. A stability analysis of the Q iv þ Q il ¼ 0 : Q i ¼ Q iv ¼ Q il ð8Þ
semi-implicit method for the two-fluid model was carried out by
3. Momentum
Stewart (1979).
F iv þ F il ¼ 0 : F i ¼ F iv ¼ F il ð9Þ
2.1. Two-fluid conservation equations

In total, 12 unknowns appear in the equations: void fraction a,

Field equations, mass, energy, and momentum conservation
pressure P, densities qv and ql, specific internal energies ev and el,
equations for each phase can be obtained by applying the appropri-
three components of Vv and Vl. The exchange terms are assumed
ate time and space averaging operators to the local, instantaneous
to depend upon these variables and phase temperatures Tv and Tl.
balance equations (Delhaye and Achard, 1976). Other procedures
Thus, there are 14 unknowns. We assume that any other fluid prop-
(for example, Boltzmann statistical averaging) may be followed
erty appearing in the exchange terms has a known, functional
to obtain usable field equations, but the given cases are the most
dependence on the above-mentioned unknowns. Since a total num-
common. Fortunately, all the averaging techniques produce effec-
ber of ten equations are represented, we must add four additional
tively identical sets of equations.
relations for closure. They are the equations of state, in the form:
The major simplifying assumptions in the averaging procedure
are as follows: qv ¼ qv ðP; T v Þ; ql ¼ ql ðP; T l Þ; ev ¼ ev ðP; T v Þ; el ¼ el ðP; T l Þ ð10Þ
(1) Viscous stress and energy dissipation can be neglected due 2.2. Constitutive relations
to the relatively small value of these terms for reactor condi-
tions, and Completion of the analysis requires constitutive relations that
(2) The liquid and vapor pressures are assumed to be equal describe the thermodynamic properties of the materials involved
within any control volume because the size of the control and interactions between the phases and also between each phase
volume is not too large. and the medium in which the flow occurs.
This study uses the simplified approach to interfacial exchange
The derivation procedure results in the following conservation expressions (Kelly and Kazimi, 1981). It is not obvious if a more
equations: sophisticated approach produces more accurate results. Also, since
Conservation of vapor mass: the models are physically based, non-equilibrium phenomena can
@ ! be predicted in a realistic manner. The interphase mass transfer is
ðaqv Þ þ r  ðaqv V v Þ ¼ Cv  W tv ð1Þ
@t modeled according to Relap computer code (RELAP5/MOD3.3 Code
Manual, 2001), where it is formulated such that the net interfacial
Conservation of liquid mass:
mass transfer rate is composed of two components: the mass
@ !
½ð1  aÞql  þ r  ½ð1  aÞql V l  ¼ Cl  W tl ð2Þ transfer rate in the bulk fluid and the mass transfer rate near the
@t wall. The Constitutive models are summarized in Table 1, where
Conservation of vapor energy: the subscript FD denotes the fully-developed distribution term,
@ ! ! @a Sij is the gap between rods, and Aij is the area of node i.
ðaqv ev Þ þ r  ðaqv ev V v Þ þ Pr  ða V v Þ þ P Detailed representation of the constitutive relations is available
@t @t
¼ Q w v þ Q iv  Q t v ð3Þ in Kelly and Kazimi (1981).

Conservation of liquid energy: 2.3. Heat conduction model

@ ! ! @a
½ð1  aÞql el Þ þ r  ½ð1  aÞql el V l Þ þ Pr  ½ð1  aÞ V l   P In general, a fuel rod model consists of an approach to solving
@t @t
the general three-dimensional, time-dependent Poisson’s equation
¼ Q wl þ Q il  Q tl ð4Þ
(heat conduction equation):
86 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99

Table 1
A brief description of closure relations.

Interfacial heat exchange rate Q i ¼ Hiv ðT S  T V Þ þ Cig

 h i
Interfacial momentum exchange rate F i ¼ 1:01 a 1:01a l þ qv jV v V l j ðV  V Þ
aDh aDh l 2 v l

Turbulent vapor mass exchange rate P Sij e 

W tv ¼ Aj l ½ðaqv Þi  ðaqv Þj  ½ðaqv Þi  ðaqv Þj FD 

P Sij e 
Turbulent liquid mass exchange rate W tl ¼ Ai l ½ðð1  aÞql Þi  ðð1  aÞql Þj  ½ðð1  aÞql Þi  ðð1  aÞql Þj FD 
Turbulent vapor energy exchange rate P Sij e 
Q tv ¼ Ai l ½ðaqv ev Þi  ðaqv ev Þj  ½ðaqv ev Þi  ðaqv ev Þj FD 

P Sij e 
Turbulent liquid energy exchange rate Q tl ¼ Ai l ½ðð1  aÞql el Þi  ðð1  aÞql el Þj  ½ðð1  aÞql el Þi  ðð1  aÞql el Þj FD 
Turbulent vapor momentum exchange rate P Sij e 
F tv ¼ Ai l ½Gv i  V v j  ½Gv i  Gv j FD 
Turbulent liquid momentum exchange rate P Sij e 
F tl ¼ Ai l ½Gli  V lj  ½Gli  Glj FD 
Wall frictional force on the liquid phase 0:2 ð1aÞql V l De
F wl ¼ 0:18
2De ðRel Þ ql V l jV l j Rel ¼ ll
Wall frictional force on the vapor phase aqv V v De
0:2
F wv ¼ 2D e
ðRev Þ0:2 qv V l jV v j Rev ¼ lv

dT ! ! equations so that ai+1/2 refers to ai+1/2,j,k. A superscript n or n + 1 re-

qc ¼ r  ðK rT Þ þ q000 ð11Þ
dt fers to the time level at which the dependent variables are evalu-
ated, and n + 1/2 indicates semi-implicit formulation. This index
where K represents the conductivity tensor, K = Kij, i = 1, 2, 3, j = 1, 2,
is used only for the exchange terms which are functions of both
3, and q000 is the heat source density which represents the amount of
new and old time values of variables. In the following equations,
heat generated in the material per unit volume per unit time. Gen-
the area A and the volume V represent the net flow area and net
erally, the cylindrical shape of the fuel rod dictates that a cylindrical
fluid volumes of the cell. With these background, the difference
coordinate system must be chosen to expand the first term in the
equations (Kelly, 1980) are as follows:
RHS of Eq. (11). Further simplification of this equation is possible
Vapor mass equation
by assuming that axial and circumferential heat conduction effects
are neglected and all physical properties are isotropic. However, the ½ðaqv Þnþ1  ðaqv Þn =Dt
temperature dependence of thermal conductivity and heat capacity 
nþ1 nþ1
þ ð1=WÞ ½Aðaqv Þn ðV xv Þ iþ1=2  ½Aðaqv Þn ðV xv Þ i1=2
is considered (TRAC-M, 2000). Therefore, one-dimensional heat
conduction equation can be written in cylindrical coordinates for nþ1 nþ1
þ½Aðaqv Þn ðV yv Þ jþ1=2  ½Aðaqv Þn ðV yv Þ j1=2
the fuel region as 
n z nþ1 n z nþ1
  þ½Aðaqv Þ ðV v Þ kþ1=2  ½Aðaqv Þ ðV v Þ k1=2
@T 1 @ @T
qc  rK ¼ q000 ð12Þ ¼ Cnþ1=2  W tnþ1=2 ð13Þ
@t r @r @t v

The thermal source density q000 depends on both position and Liquid mass equation
time. The clad region is assumed to be free of heat source.
½ðð1  aÞql Þnþ1  ðð1  aÞql Þn =Dt

nþ1 nþ1
þ ð1=WÞ ½Aðð1  aÞql Þn ðV xl Þ iþ1=2  ½Aðð1  aÞql Þn ðV xl Þ i1=2
3. Discrete equations
nþ1 nþ1
þ½Aðð1  aÞql Þn ðV yl Þ jþ1=2  ½Aðð1  aÞql Þn ðV yl Þ j1=2
This section will present the discrete equations that are solved 
n nþ1 n nþ1
computationally. The spatial discretization is based on a first-order þ½Aðð1  aÞql Þ ðV zl Þ kþ1=2  ½Aðð1  aÞql Þ ðV zl Þ k1=2
spatial scheme, donor cell method, and staggered-mesh approach.
¼ Cnþ1=2  W nþ1=2
tl ð14Þ
In donor cell differencing, all centered quantities will be trans-
ferred to face-centered positions. On a staggered mesh, fluid prop-
erties such as a, P, Tv, Tl, densities, and energies are defined at the
center of each rectangular mesh cell while each velocity compo-
nent is defined at a different location on the cell faces normal to
that component. On cell faces, subscripts for the cell-centered
quantities are i, j and k, which are used to indicate positions along
the x, y, z axes, respectively. While on cell faces, half increments are
used (e.g. i + 1/2, j, k). Particular mesh cell is illustrated in Fig. 1.
The temporal differencing that is based on the semi-implicit dif-
ferencing scheme is in order one in time. If we categorize the phys-
ical phenomena in Eqs. (1)–(6) as interphase exchanges, sonic
propagation, and fluid convection, then the semi-implicit method
(Liles and Reed, 1978) treats the first two (which typically have
short time scales) implicitly, while convection (having a somewhat
longer time scale) is treated explicitly. Each of these terms may be
a complicated function of the fluid state. Once this function is spec-
ified, it is usually possible to evaluate certain parts at the new time.
It turns out that if difference equations are written in a form which
follows this philosophy, the coupling of new-time unknowns in the
difference equations has a relatively simple structure. To simplify
the nomenclature, we supress indices in writing difference Fig. 1. Fluid mesh cell (cell-centered variables P, Tv, Tl, ev, el, qv, ql, a).
 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99 87

Vapor energy equation akþ1=2 ¼ ½akþ1 Dzkþ1 þ ak Dzk =ðDzkþ1 þ Dzk Þ; ðqv Þkþ1=2
½ðaqv ev Þ nþ1 n
 ðaqv ev Þ =Dt ¼ ½ðqv Þkþ1 Dzkþ1 þ ðqv Þk Dzk =ðDzkþ1 þ Dzk Þ ð19Þ
n n n x nþ1
þ ð1=WÞð½P þ ðq a
v ev Þiþ1=2 ½A ðV v Þ iþ1=2 For (aqv)k+1/2 and other similar terms in the mass and energy
n n n x nþ1 equations, the donor cell logic is used. If C stands for any cell-
 ½P þ ðq a
v ev Þi1=2 ½A ðV v Þ i1=2 centered quantity, the value of Ck+1/2 is determined as
n
þ ½P þ ðq n n
a y nþ1 8
v ev Þjþ1=2 ½A ðV v Þ jþ1=2
< C kþ1
> if V kþ1=2 < 0
nþ1 :
 ½Pn þ ðqv ev Þnj1=2 ½Aan ðV yv Þ j1=2 C kþ1=2 ¼ Ck if V kþ1=2 P 0 ð20Þ
>
:
n n n z nþ1
þ ½P þ ðq a
v ev Þkþ1=2 ½A ðV v Þ kþ1=2
n
 ½P þ ðq n n
a z nþ1 It should be noted that ðv xv Þkþ1=2 ; ðv yv Þkþ1=2 are not at the proper
v ev Þk1=2 ½A ðV v Þ k1=2 Þ
nþ1=2
defined locations, so an averaging must be performed. These
nþ1=2
þ Pn ðanþ1  an Þ=Dt ¼ Q w v þ Qi  Q tnþ1=2
v ð15Þ velocities are defined by
h i
Liquid energy equation ðV yv Þkþ1=2 ¼ ð1=4Þ ðV yv Þj1=2 þ ðV yv Þjþ1=2 þ ðV yv Þj1=2;kþ1 þ ðV yv Þjþ1=2;kþ1 ;
h i
½ðð1  aÞql el Þnþ1  ðð1  aÞql el Þn =Dt ðV xv Þkþ1=2 ¼ ð1=4Þ ðV xv Þi1=2 þ ðV xv Þiþ1=2 þ ðV xv Þi1=2;kþ1 þ ðV xv Þiþ1=2;kþ1
nþ1
þ ð1=WÞð½Pn þ ðql el Þniþ1=2 ½Að1  aÞn ðV xl Þ iþ1=2 ð21Þ
n n n nþ1
 ½P þ ðq l el Þi1=2 ½Að1  aÞ ðV xl Þ i1=2 The convective operators of the forms ðDx v v =DxÞkþ1=2 ; z

n n n nþ1 ðDy v zv =DyÞkþ1=2 ; ðDz v zv =DzÞkþ1=2 which designate the difference

þ ½P þ ðq l el Þjþ1=2 ½Að1  aÞ ðV yl Þ jþ1=2
approximation of @ v zv =@x; @ v zv =@y, @ v zv =@z associated with the point
n nþ1
n
 ½P þ ðq l el Þj1=2 ½Að1  aÞn ðV yl Þ j1=2 (i, j, k + 1/2) are defined according to the donor cell scheme. The
n n n nþ1 convective derivatives are given by:
þ ½P þ ðq l el Þkþ1=2 ½Að1  aÞ ðV zl Þ kþ1=2 8 .
>
< ½ðV zv Þiþ1;kþ1=2  ðV zv Þkþ1=2  Dxiþ1=2 if ðV xv Þkþ1=2 < 0
n nþ1
n
 ½P þ ðq l el Þk1=2 ½Að1  aÞn ðV zl Þ k1=2 Þ z
ðDx V v =DxÞkþ1=2 ¼ .
>
: ½ðV zv Þkþ1=2  ðV zv Þi1;kþ1=2  Dxi1=2 if ðV xv Þkþ1=2 P 0
nþ1=2
n
 P ða nþ1 n
 a Þ=Dt ¼ Q wl  Q inþ1=2  Q tlnþ1=2 ð16Þ
ð22Þ
Now let us consider only Z – direction of the vapor momentum
equation because the additional equations are found by permuta- 8 .
>
< ½ðV zv Þjþ1;kþ1=2  ðV zv Þkþ1=2  Dyjþ1=2 if ðV yv Þkþ1=2 < 0
tion. According to the staggered mesh scheme, each momentum
ðDy V zv =DyÞkþ1=2 ¼ .
equation is differenced between the centers of the two appropriate >
: ½ðV zv Þkþ1=2  ðV zv Þj1;kþ1=2  Dyj1=2 if ðV yv Þkþ1=2 P 0
cells. Therefore, the mesh used in the momentum equations (Fig. 2)
is different than the one used in the mass and energy equations. ð23Þ
Vapor Z – direction momentum equation (
h z
½ðV zv Þkþ3=2  ðV zv Þkþ1=2 =Dzkþ1 if ðV zv Þkþ1=2 < 0
nþ1 n
ðaqv Þnkþ1=2 ½ðV zv Þ  ðV zv Þ kþ1=2 =Dt þ ðaqv Þnkþ1=2 ðV xv Þkþ1=2 ðDx V zv =DxÞkþ1=2 ðDz V v =DzÞkþ1=2 ¼
½ðV zv Þkþ1=2  ðV zv Þk1=2 =Dzk if ðV zv Þkþ1=2 P 0
in
þðV yv Þkþ1=2 ðDy V zv =DyÞkþ1=2 þ ðV zv Þkþ1=2 ðDz V zv =DzÞkþ1=2 ð24Þ
nþ1=2 nþ1=2
þankþ1=2 ½ðPkþ1  Pk Þ=Dzkþ1=2 nþ1 ¼ ðF zwv Þkþ1=2  ðF ziv Þkþ1=2 where the mesh spacing is evaluated as:

ðaq n n
ðF ztv Þkþ1=2 ðDXÞiþ1=2 ¼ ðDX iþ1 þ DX i Þ=2; ðDYÞjþ1=2 ¼ ðDY jþ1 þ DY j Þ=2 ð25Þ
v gÞkþ1=2  ð17Þ
Similar structure is obtained for the liquid momentum equation
In the momentum equations, (aqv)k+1/2 and similar expressions
and for the momentum equations in the x, y directions. Now, rep-
are given by:
resentation of the fluid dynamics difference equations is complete
ðaqv Þkþ1=2 ¼ akþ1=2 ðqv Þkþ1=2 ð18Þ with the exception of exchange terms. These terms are briefly dis-
cussed in Section 2.2. To decide on the proper finite difference
Where
approximation for the rod conduction equation, the fluid dynamics
is treated semi-implicitly; therefore, a steady-state calculation
must be performed as an unperturbed transient. Semi-implicit for-
mulation dictates the limitation on time step size of the fluid flow
solution which then seems to need a large number of time steps.
On the other hand, the generally-large thermal inertia of the fuel
rods requires a rather long time to elapse before reaching a steady
state. If the temporal derivative in calculations of rod heat conduc-
tion is removed, a treatment which is equivalent to setting the
thermal capacity of the rod material at zero (Schor et al., 1984),
the amount of heat generated in the rod will be entirely and
instantaneously transferred to the fluid. Consequently, the time re-
quired to reach steady state will be significantly shortened. To ob-
tain the difference equation of heat conduction, the fuel and the
clad are now divided into an arbitrary number of mesh cells, but
mesh spacing in the same region is of the same size. One cell is as-
sumed for the gap. The temperatures are located at the edges of
Fig. 2. Typical control volumes Momentum equation control volume Mass or mesh cells represented by the subscript i. Material properties are
Energy equation control volume. evaluated in the centers of the mesh cells, i + 1/2 (Fig. 3).
88 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99

Fig. 3. Fuel rod discretization modeling.

We obtain finite difference approximation of the heat conduc-

tion by integrating Eq. (12) between the centers of two adjacent
mesh cells, i.e. from i  1/2 to i + 1/2. Since an implicit scheme
(Granziera and Kazimi, 1980) has been chosen for temporal dis-
cretization of Eq. (12), the difference equation for the interior
points in the fuel rod becomes
 
h    i T nþ1
i  T ni
n n
ð1=2Þ r 2i  r 2i1=2 ðqcÞi1=2 þ r 2iþ1=2  r 2i ðqcÞiþ1=2
Dt
 n   rkn   
rk nþ1 nþ1 nþ1 nþ1
 T  Ti þ T  T i1
Dr iþ1=2 iþ1 Dr i1=2 i
h    i
nþ1=2 nþ1=2
Fig. 4. Pressure in each cell is coupled to the pressures in six mesh cells that ¼ ð1=2Þ r 2i  r 2i1=2 ðq000 Þi1=2 þ r 2iþ1=2  r 2i ðq000 Þiþ1=2
surround it.
ð26Þ

Fig. 5. NUPEC PWR bundle test facility, bundle heater rod cross section, and B5 rod bundle.
 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99 89

nþ1=2 n nþ1=2
ðq00wv Þ ¼ hv c ðT nþ1 nþ1 00
w  T v Þ þ ðqv Þ ;
nþ1=2 n n 00 nþ1=2
ðq00wl Þ ¼ hfc ðT nþ1
w  T nþ1
l Þ þ hnb ðT nþ1 nþ1
w  T sat Þ þ ðql Þ ð29Þ

q00wv ; q00wl ; q00v ; q00l ;

hvc, hfc, hnb are wall heat flux to vapor, wall heat flux
to liquid, heat flux to vapor for transition boiling, heat flux to liquid
for transition boiling, heat transfer coefficient to vapor, heat trans-
fer coefficient to liquid (forced convection), and heat transfer coef-
ficient to liquid (nucleate boiling), respectively. It should be noted
that hvc is equal to zero because the film boiling and single-phase
vapor regimes do not occur in our test sample (PWR condition). Un-
der some circumstances, including a high heat transfer coefficient
Fig. 6. Lateral power distribution. and/or a low heat capacity of the clad or the fluid, semi-implicit
evaluation of the wall heat flux may give rise to instabilities. This
problem can be solved by a reduction in time step size. In Eq.
Table 2 (26), the temperature for all cells must be solved simultaneously
Data range of PSBT void tests. because the temperature at a cell i with its neighbors i + 1 and
Quality Range i1 are coupled. In the matrix form, Equation set 26 becomes
Pressure 4.9–16.6 MPa 
Mass velocity 550–4150 Kg/m2 s AT nþ1 ¼ b ð30Þ
Inlet coolant temperature 140–345 °C T
where A is the tridiagonal matrix, T ¼ and ½T nþ1 nþ1
1 ; . . . ; Ti ; . . . ; T nþ1
W 
 include the old time temperatures, the sur-
the elements of matrix b
Where
face heat flux, and the volumetric heat source. Matrix Eq. (30) is
ðDrÞiþ1=2 ¼ riþ1  r i and ðDrÞi1=2 ¼ r i  ri1 ð27Þ easily solved by a direct forward-elimination backward-substitu-
tion technique. Thus, specifications of the finite difference equa-
The difference equation at the center of the fuel pin is obtained tions are finished. The procedure for solving the nonlinear
by setting r1 = r1/2 = 0 in Eq. (26). By introducing the clad surface difference equations will be discussed in Section 4.
heat flux q00 and integrating Eq. (12) from r = rI1/2 to r = rI = rod
outer radius, the difference equation becomes
4. Solution scheme of nonlinear equations
ðT nþ1 T nI Þ
ð1=2Þðr 2I  r2I1=2 Þð
q cÞnI1=2 I The above-mentioned finite difference equations combined
Dt
 n with constitutive relations and the equations of state form a large
rk 000 nþ1=2
þ ðT nþ1  T nþ1 00 2 2
I1 Þ þ q r I ¼ ð1=2Þðr I  r I1=2 Þðq ÞI1=2 : ð28Þ system of nonlinear, complicate equations with the following new
Dr I1=2 I
time variable as the unknown: a, p, qv, ql, ev, el, Tv, Tl and six com-
A semi-implicit formulation of the surface heat flux q00 was also ponents of Vv, Vl in x, y and z directions.
developed (Reed et al., 1979) where the heat transfer coefficient is A numerical iterative technique is practically the only option for
evaluated explicitly, and wall temperature Tw, vapor temperature solving nonlinear equations (Reed et al., 1979). At the first stage,
Tv and fluid temperature Tl are treated implicitly. Eq. (29) represent the density and internal energy in the conversation equations are
these relations. eliminated in favor of the pressure and temperatures by using

Fig. 7. Test Series 5 bundle-averaged void fraction results – lower elevation.

90 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99

the equations of state. Extracting an analytic relationship for the The Jacobian matrix J(xm) is defined as Jðxm Þ  ð@F=@xÞm
x ; we ob-
density and internal energy in terms of the fundamental unknowns tain a set of linear equations for the corrections Dxmþ1 that move
necessitates linearization the equations of state. The point about each function closer to zero simultaneously, namely J(xm)Dxm+1 =
which these equations are linearized is indicated by the symbol F(xm). The corrections are then added to the solution vector
() placed over the variables. The linearized equations of state xm+1 = xm + Dxm+1 where Dxm+1 is the iteration error, and F(xm)
are given by the following equations: is called a residual error. Applying the block Newton–Raphson
 n method, we can obtain the matrix equation below:
@q ~l
qnþ1
l
~ nl þ
¼q ðpnþ1  p ~n Þ
@p
 n  
@q
~l
þ T nþ1
l  Te nl ;
@T l
   
@q~ v n nþ1 ~ v n nþ1 e n
@q
qnþ1
v ¼q ~ nv þ ðp ~n Þ þ
p ðT v  T v Þ ð31Þ
@p @T v
 n  n
@ ~el @ ~el
enþ1
l ¼ ~enl þ ðpnþ1  p ~n Þ þ ðT nþ1
l  Te nl Þ;
@p @T l
 n  n
@ ~ev @ ~ev en
enþ1
v ¼ ~env þ ðpnþ1  p~n Þ þ ðT nþ1
v  TvÞ ð32Þ
@p @T v
The thermodynamic equations of state and their derivatives are
taken from TRAC-M (2000). We denote the system of difference
equations for a given time step by F(x) = 0, where x is a vector con-
taining the ten principal unknown variables for the new time level
at all mesh points. Then, ð35Þ

2 3
@Rm 1 @Rm
1 @Rm 1 @Rm 1 @Rm
1 @Rm
1 @Rm
1 @Rm
1 @Rm
1 @Rm
1 @Rm 1 @Rm 1 @Rm1 @Rm1 @Rm
1 @Rm1 @Rm
1 @Rm1 @Rm1 @Rm
1 @Rm
1 @Rm
1
6 @ðV xv Þ
i 1 @ðV Xl Þ 1 @ðV xv Þ
iþ1
@ðV xl Þ
iþ1
Y
@ðV v Þ
j1
@ðV Yl Þ 1 Y
@ðV v Þ
jþ1
@ðV Yl Þ 1 Z
@ðV v Þ
k1
@ðV Zl Þ 1 @ðV zv Þ
kþ1
@ðV zl Þ
kþ1
@P i;j;k @P iþ1 @Pi1 @P jþ1 @Pj1 @P kþ1 @P k1 @a @T v @T l 7
6 2 i
2 2 2 2
j
2 2
jþ
2 2
k
2 2 2 7
6 @Rm @Rm @Rm @Rm 7
6 X2 2
.. . .. . . .. . .. . .. . .. . .. .. . . .. . .. . .. . .. .. . .. . . .. . .. . .. .. . 2 2 7
6 @ðV v Þi1 @ðV Xl Þ 1 @T v @T l 7
6 2
i
2 7
6 7
6 .. .. 7
6 . . 7
6 7
6 . .. 7
6 .. . 7
6 7
6 . .. 7
6 . 7
6 . . 7
6 . .. 7
6 . 7
6 . . 7
6 7
6 .. .. 7
6 . . 7
6 m 7
4 @R10 @Rm @Rm @Rm 10 5
@ðV Xv Þ
10
@ðV Xl Þ
                                    10
@T v @T l
i1 i1
2 2

"
#T
X ¼ a; Pi;j;k ; Pi1 ; Piþ1 ; Pj1 ; Pjþ1 ; Pk1 ; P kþ1 ; T v ; T l ; Rm m m m m m m m m m
m 1 R2 R3 R4 R5 R6 R7 R8 R9 R10 cell1
ðV v Þi1 ; ðV v Þiþ1 ; ðV xl Þi1 ; ðV xl Þiþ1 ; ðV yv Þj1 ; ðV yv Þjþ1 ; ðV yl Þj1 ;
x x Fðx Þ ¼
ð36Þ
2 2 2 2 2 2 2 .. . Rm m m m m m m m m m
1 R2 R3 R4 R5 R6 R7 R8 R9 R10 cellN
inþ1
ðV yl Þjþ1 ; ðV zv Þk1 ; ðV zv Þkþ1 ; ðV zl Þk1 ; ðV zl Þkþ1 ð33Þ
2 2 2 2 2
where R1, R2, R3, R4, R5, R6, R7, R8, R9, R10 are residual errors for the
The technique used in the solution of algebraic equations (Ort- vapor mass, liquid mass, vapor energy, liquid energy, vapor
ega and Rheinboldt, 1970) is a multidimensional extension of the momentum, and liquid momentum equations. The solving process
Newton iterative solution of algebraic equations. This choice was of this linear system is called the inner iteration. In the Newton iter-
made because a good guess guarantees convergence. The relatively ation called the outer iteration, the process is iterated to conver-
high rate of convergence will be obtained by the reduction in time gence. In general, it is a good idea to check the degree to which
step size. The linearization technique in this method is widely used both the iteration error Dxm+1 and the residual error F(xm) have con-
to numerically solve fluid dynamics equations. verged. Once either the iteration or residual error reaches machine
We linearize the old iterative accuracy, the other will not change. A highly efficient solution tech-
nique exploits this structure as follows. A major characteristic of the
xm as FðxÞ ffi Fðxmþ1 Þ ¼ Fðxm Þ þ Jðxm Þðxmþ1  xm Þ ¼ 0 ð34Þ
mass and energy equation is that the cell-centered variables, an+1,
 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99 91

2 3
Pn+1, Tn+1, are coupled only with the new time velocity components 2 3 Pmþ1 2 3
1 0 0 0 i;j;k y1 y2 y3 y4 y5 y6
on the faces of mesh cells, while each momentum equation relates 6 mþ1 7
60 1 0 6
0 76 ai;j;k 7 6
7 7 y9 y10 y11 y12 7
these velocities only to the pressure in the mesh cell and in the six 6 6 y7 y8 7
6 76 mþ1 7
þ6 7
40 0 1 0 56 7 4 y13 y14 y15 y16 y17 y18 5
surrounding cells. The momentum equations are then used to de- 4 ðT v Þi;j;k 5
rive a relationship between velocity and pressure which can be used 0 0 0 1 mþ1
ðT l Þi;j;k y19 y20 y21 y22 y23 y24
to eliminate the velocity in the mass and energy equations (Schor 2 mþ1 3
et al., 1984). First, we rearrange the momentum equations on the P i1;j;k
6 mþ1 7 ð40Þ
faces of each mesh cell into the following form: 6P 7 2 03
6 iþ1;j;k 7 y1
2 32 3 6 mþ1 7
ðx1 Þni1 ðx2 Þni1
nþ1
ðdV xv Þi1;j;k 6P 7 6 07
6 i;j1;k 7 6 y2 7
6 2 2
76 2
7 6 mþ1 7 ¼ 6 0 7
4 54 5 6P 7 4 y3 5
6 i;jþ1;k 7
ðx3 Þni1 ðx4 Þni1 nþ1
ðdV xl Þi1;j;k 6 mþ1 7 y04
2 2 2 6P 7
4 i;j;k1 5
2 3
ðy1 Þni1 ðdPnþ1 n nþ1 n P mþ1
i;j;k Þ þ ðy2 Þi1 ðdP i1;j;k Þ þ ðz1 Þi1 i;j;kþ1
6 2 2 2
7
¼4 5;
ðy3 Þni1 ðdPnþ1 n nþ1 n The resultant first line gives the seven-point differential pres-
2
i;j;k Þ þ ðy4 Þi1 ðdP i1;j;k Þ þ ðz2 Þi1
2 2
sure equation which is the nodalized form of the corresponding
2 32 nþ1 3 Poisson equation and is solved in the inner loop. This equation
ðx5 Þnj1 ðx6 Þnj1 ðdV xv Þi;j1;k
6 2 2
76 2
7 must be solved simultaneously for all mesh cells (Fig. 4). The
4 54 5 remaining three equations relate other variables (including void
ðx7 Þnj1 ðx8 Þnj1 nþ1
ðdV xl Þi;j1;k
2 2 2 fraction, liquid, and vapor temperature) to the pressure. The pres-
2 3 ð37Þ sure equations represent, in compact form, the fully coupled effects
ðy5 Þnj1 ðdP nþ1 n nþ1 n
i;j;k Þ þ ðy6 Þj1 ðdP i;j1;k Þ þ ðz3 Þj1 of sonic propagation and interphase coupling over the time step.
6 2 2 2
7
¼4 5 The pressure equation for each mesh cell can be expressed as
ðy7 Þnj1 ðdP nþ1
i;j;k Þ þ ðy8 Þnj1 ðdP nþ1
i;j1;k Þ þ ðz4 Þnj1 follows:
2 2 2

2 32 nþ1 3
ðx9 Þnk1 ðx10 Þnk1 ðdV xv Þi;j;k1 Aijk Pi;j;k þ Bijk Pi1;j;k þ C ijk Piþ1;j;k þ Dijk P i;j1;k þ Eijk P i;jþ1;k
6 2 2
76 2
7 þ F ijk Pi;j;k1 þ Gijk Pi;j;kþ1 ¼ Rijk ð41Þ
4 54 5
ðx11 Þnk1 ðx12 Þnk1 nþ1
ðdV xl Þi;j;k1
2 2 2 Number of equations is equal to the number of computational
2 3 cells. The dimension of the reduced coefficient matrix is the square
ðy9 Þnk1 ðdPnþ1
i;j;k Þ þ ðy10 Þnk1 ðdPnþ1
i;j;k1 Þ þ ðz5 Þnk1 of the number of cells. It has a block septa-diagonal structure.
6 2 2 2
7
¼4 5 A direct inversion (Gaussian elimination) procedure inside an
ðy11 Þnk1 ðdPnþ1
i;j;k Þ þ ðy12 Þnk1 ðdPnþ1
i;j;k1 Þ þ ðz6 Þnk1 iterative process was employed to solve the reduced system of
2 2 2
equations for the pressure in each mesh cell, which is to be re-
Once this system of equations is solved, each velocity is evalu-
peated for each time step (Granziera and Kazimi, 1980). This iter-
ated in terms of the pressure gradient across the momentum cell as
ation is referred to as the pressure or inner iteration. To perform
V mþ1 ¼ cm dPmþ1 þ d
m the mth iteration in the cells at level [k], the terms containing
v
ð38Þ the pressure at the bottom and top of cell (i, j, k) are passed to
V mþ1
l ¼ f m dPmþ1 þ g m
the right-hand side of Eq. (41) below:
The above equations can now be used to eliminate all velocities in
the mass and energy equations. The resulting system of equations for Aijk Pm m m m m
i;j;k þ Bijk P i1;j;k þ C ijk P iþ1;j;k þ Dijk P i;j1;k þ Eijk P i;jþ1;k
each node can be represented as follows (Kelly and Kazimi, 1979): ¼ Rijk  F ijk Pm m1
i;j;k1  Gijk P i;j;kþ1 ð42Þ
where X, X0 , X00 represent the known coefficients at the old time n. Eq.
(39) is a local matrix equation for one Newton iteration step. For each The value of Pm i;j;k1 is known because it is calculated while pres-
Newton or outer iteration step, the coefficient matrix X, X0 on the left sure iteration m is running for the cells at level k  1. Thus, the pres-
hand side and the column vector X00 on the right hand side should be sure of cells at level k is calculated using the new pressures in cells at
recalculated. The coefficient matrix of the first term on the left side is level k  1, but the old pressures are calculated in cells at level k + 1.
transformed into a [4 4] diagonal matrix. The process involves reducing the multidimensional mesh to a one-
We can put this system in the reduced form: dimensional (Cuervo et al., 2006). An iterative solution sets initial

ð39Þ
92 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99

Fig. 8. Test Series 5 bundle-averaged void fraction results – middle elevation.

estimates for the unknowns and then a solution for the pressure var- each level k, the values of the pressure will have the correct rela-
iation at each level (1D problem) is obtained by direct inversion. tionship between themselves, since the direct inversion technique
Then, the one-dimensional solution for the linear pressure variation will give the exact solution for a given right-hand side (Granziera
at each level is used as an initial guess for the linear pressure varia- and Kazimi, 1980).
tion in each mesh cell on that level in the multidimensional problem. The inner iterative procedure will stop when the maximum var-
The pressure iteration is carried out over all values of subscript k to iation of the pressure for all cells fulfills a specified convergence
obtain the new pressure vector. The global matrix equation over the criterion. Once the pressure distribution is found, void fraction, li-
entire spatial mesh can be put in the following form: quid and vapor temperature can be found via back substitution of

ð43Þ

the pressure vector in three later equations of the linear system of

NCell is the number of mesh cells in Z  direction, MCell is the Eq. (40) during the Newton iteration. The outer and inner iteration
number of mesh cells in the XY plane. convergence criteria are applied to the pressure distribution (dP)
In this technique, the new value obtained for the pressure will such that both of the specified convergence criteria (i.e., Newton
have an error because in the right-hand side of Eq. (42) the values and pressure) are satisfied. If convergence is not obtained in the
of the pressure are not the exact solution of the problem, but for specified number of iterations, the time step size is reduced, and
 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99 93

Fig. 9. Test Series 5 bundle-averaged void fraction results – upper elevation.

Table 3
Deviations of the calculated pressure data from the average of the codes (Rubin, 2011).

Deviation 5.1221 5.1222 5.2111 5.2112 5.2442 5.3441 5.3442 5.4562 5.6321 5.6322 5.6552
Lower 0.0034 0.0011 0.0011 0.00135 0.00155 0.00181 0.0153 0.0535 0.0122 0.0509 0.0746
Middle 0.0011 0.0081 0.00716 0.0001 0.026 0.015 0.0164 0.0146 0.0135 0.0288 0.0323
Upper 0.0154 0.0067 0.0022 0.0294 0.0558 0.0872 0.0704 0.209 0.0805 0.0792 0.218

the calculation is repeated. Finally, when Newton iterations con- 5. Krylov-based iterative methods
verge, the velocities are determined by using the new pressure gra-
dients, and the calculation goes onto the next time step. Note that Since the dimension of the coefficient matrix of the reduced
the wall temperature is the communication between conduction in system of equations for the pressure (Eq. (43)) is large when a large
fuel rod and fluid dynamics equations. This coupling is done via the number of mesh cells are involved, the use of a direct solution with
wall heat flux to the liquid and vapor based on Eq. (29). a Gaussian elimination method will be impractical. Then, an itera-
Thus, the overall scheme of the thermal hydraulic calculation is tive technique is recommended for solving the pressure equations.
as follows: On the other hand, the coefficient matrix arising from the linear-
ized hydrodynamic equations is weakly diagonally dominant;
Step 1: At the beginning of each time step, the heat transfer therefore, using stationary iterative methods such as Gauss–Seidel
regime and the associated heat transfer coefficients are evalu- or successive over-relaxation (SOR) are not generally efficient. Re-
ated by means of the variables which are set to their old time cently, considerable research has been performed on a nonstation-
values. ary class of techniques collectively known as Krylov subspace
Step 2: The problem of heat conduction in fuel rod is set up for methods. Nonstationary methods differ from stationary methods
each fluid cell connected to the corresponding fuel cell at a cer- in that the computations involve information that changes at each
tain axial level. Then, the forward elimination of Tridiagonal iteration. Typically, constants are computed by taking inner prod-
matrix Eq. (30), where the elements of coefficient matrix A, B ucts of residuals or other vectors arising from the iterative method.
are evaluated at the old time step, is performed. This procedure These methods are called Krylov methods because they are based
gives us the wall temperature Tw as a linear function of the old on building a solution vector from the Krylov subspace [r0,
time variables. Ar0, . . . , An1r0] where r0 is the residual of the initial solution,
Step 3: The linear expression for Tw is substituted into Eq. (29), and A is the coefficient matrix. The convergence rate of iterative
and then Qwv, and Qwl are formed. Now that the wall heat flux is methods depends on spectral properties of the coefficient matrix.
defined, the fluid dynamics equations can be solved. Hence, one may attempt to transform the linear system into an
Step 4: The new wall temperature corresponding to the new equivalent one with the same solution but with more favorable
fluid temperature is calculated following any Newton iteration spectral properties. A preconditioner is a matrix that effects such
m. Once the Newton iteration has converged, we can get the a transformation. Generating a preconditioner is costly, but apply-
remaining new fuel rod temperatures by the performing the ing it to the coefficient matrix speeds up the solution scheme. More
backward substitution of the heat conduction problem. importantly, numerical scheme may fail to converge without any
Table 4

94
Deviation (in percent) of the values for selected predicted variables obtained from the proposed Krylov solvers from corresponding direct inversion values based on the NUPEC test case.

Axial BCG CGNR DQGMRES FGMRES GMRES TFQMR

position
Pressure Velocity Pressure Velocity Pressure Velocity Pressure Velocity Pressure Velocity Pressure Velocity
(m)
(MPA) (MPA) (MPA) (MPA) (MPA) (MPA)
Vapor Liquid Vapor Liquid Vapor Liquid Vapor Liquid Vapor Liquid Vapor Liquid
(m/s) (m/s) (m/s) (m/s) (m/s) (m/s) (m/s) (m/s) (m/s) (m/s) (m/s) (m/s)
0.0762 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0.1524 0.001 0.075 0.075 0 0.025 0.025 0 0.025 0.025 0 0.025 0.025 0 0.025 0.025 0.001 0.025 0.025
0.2286 0 0 0 0 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0 0 0
0.3048 0 0 0 0 0.1 0.1 0 0 0 0 0 0 0 0 0 0 0.05 0.05
0.381 0.00025 0.025 0.025 0.001 0.025 0.025 0.001 0.025 0.025 0.001 0.025 0.025 0.001 0.025 0.025 0 0.025 0.025
0.4572 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0.5334 0 0 0 0 0.025 0.025 0 0 0 0 0 0 0 0 0 0 0 0
0.6096 0 0 0 0.001 0 0 0.001 0.075 0.075 0.001 0.075 0.075 0.001 0.075 0.075 0.0005 0 0
0.6858 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0.762 0 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0 0 0
0.8382 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0.9144 0.00075 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0.001 0 0

M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99

0.9907 0 0.1 0.1 0.001 0 0 0.001 0.1 0.1 0.001 0.1 0.1 0.001 0.1 0.1 0.001 0 0
1.066 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1.143 0 0 0 0 0 0 0.00025 0 0 0.00025 0 0 0.00025 0 0 0 0 0
1.219 0 0.075 0.075 0 0.025 0.025 0 0.075 0.075 0 0.075 0.075 0 0.075 0.075 0 0.025 0.025
1.295 0.001 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0 0 0
1.371 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1.447 0 0.075 0.075 0 0.075 0.075 0.001 0.025 0.025 0.001 0.025 0.025 0.001 0.025 0.025 0 0.075 0.075
1.524 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1.6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1.676 0 0 0 0 0 0 0 0.1 0.1 0 0.1 0.01 0 0.1 0.1 0 0.1 0.1
1.752 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1.828 0 0.1 0.1 0 0.1 0.1 0 0.075 0.075 0 0.075 0.075 0 0.075 0.075 0 0 0
1.905 0.001 0.05 0.05 0.001 0.05 0.05 0.001 0.05 0.05 0.001 0.05 0.05 0.001 0.05 0.05 0.001 0.05 0.05
1.981 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.001 0 0
2.057 0.001 0.1 0.1 0.001 0.05 0.05 0.001 0.05 0.05 0.001 0.05 0.05 0.001 0.05 0.05 0.001 0 0
2.133 0 0.05 0.05 0 0.1 0.1 0 0.1 0.1 0 0.1 0.1 0 0.1 0.1 0.001 0 0
2.21 0 0.075 0.075 0 0.1 0.1 0 0.1 0.1 0 0.1 0.1 0 0.1 0.1 0 0 0
2.286 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2.362 0 0.1 0.1 0 0.1 0.1 0.001 0.1 0.1 0.001 0.1 0.1 0.001 0.1 0.1 0.001 0 0
2.438 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2.514 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2.591 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.001 0 0
2.667 0.00075 0 0 0.00075 0 0 0.00075 0 0 0.00075 0 0 0.00075 0 0 0.00075 0 0
2.743 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2.819 0 0 0 0.001 0 0 0 0 0 0 0 0 0 0 0 0.001 0 0
2.895 0 0 0 0 0 0 0 0.1 0.1 0 0.1 0.1 0 0.1 0.1 0 0 0
2.972 0 0.1 0.1 0 0.1 0.1 0 0.1 0.1 0 0.1 0.1 0 0.1 0.1 0.00025 0 0
3.048 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3.124 0.001 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0.001 0 0 0.001 0 0
3.2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3.276 0 0 0 0 0.1 0.1 0 0 0 0 0 0 0 0 0 0 0 0
3.353 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3.429 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3.505 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3.581 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3.657 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3.734 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99 95

preconditioner. In our problem, because of fluid compressibility, To apply the direct inversion solution scheme to the system of
the coefficient matrix of the pressure equation becomes unsym- pressure equations, we stored the pressure matrix in the BLAS-gen-
metrical, so we concentrate on solving non-symmetric problems. eral-band mode in order to be supplied for LU factorization. From
Research into the design of Krylov subspace methods for solving the point of view of storage, banded matrix is the simplest case
sparse non-symmetric linear systems is going on, and new meth- of sparse matrices so that the straightforward techniques are
ods are still emerging. developed to handle it. Then, the blocked LU factorization algo-
A representative group of Krylov subspace methods was used in rithm was performed on the banded linear system of pressure
this research for such non-symmetric problems. These solvers in- equations using block Gaussian elimination with partial pivoting
clude GMRES, FGMRES, DQGMRES, CGNR, BCG, and TFQMR. In this scheme (Du Croz et al., 1990). Then, the solution of the system
section, we briefly introduce the concept of Krylov subspace meth- was obtained through forward and backward substitution. Next,
od which is used instead of the direct inversion solver to execute to investigate the feasibility of using the candidate Krylov solvers
the inner iteration solution scheme. Consider the linear system as the pressure solution scheme, use was made of the SPARSKIT li-
Ax = b where A is a n n real matrix. Let j and L be two m-dimen- brary Version. 2, a package for sparse matrix computations devel-
sional subspaces of Rn , and x0 be an arbitrary initial guess for the oped by Saad (SPARSKIT, 1994). The library includes the iterative
solution xm. Projection technique is an idea to extract an approxi- solver module containing most of the Krylov solvers and precondi-
mation solution xm by imposing these conditions: xm belongs to tioners. This library was built by CSRD, University of Illinois, and
the affine space x0 + jm, and the residual vector b  Axm is orthog- RIACS (NASA Ames Research Center) under the sponsorship of
onal to Lm . There are two choices for the subspace Lm . In its first NAS System Division and Department of Energy. The SPARSKIT is
class, the subspace Lm is the same as jm, which is known as an a free package available on the Internet. It can be redistributed
orthogonal projection technique. In the second class, Lm is differ- and/or modified under the terms of the GNU General Public License
ent from jm, which is known as an oblique projection technique. If as published by the Free Software Foundation.
there are several implementations of a projection method, there We first performed the transformation of the pressure matrix
will also be different algorithms, but all these algorithms are math- from BLAS-general-band storage mode into the compressed sparse
ematically equivalent. The projection methods applied to Krylov row (CSR) format, which is the basic format used in SPARSKIT. This
subspaces are known as Krylov subspace methods for which the format is probably the most popular for storing general sparse
subspace jm is the Krylov subspace jm ðA; tÞ ¼ span matrices as it often allows us to make substantial savings in the
ft; At; A2 t; . . . ; Am1 tg, where jm ðA; tÞ is denoted simply by jm if number of operations of some well-known algorithms and requires
there is no ambiguity. Various choices of the subspace Lm and less memory capacity and CPU time for treatment. The next step is
the ways in which the system is preconditioned result in various to select the best preconditioner, which is highly problem depen-
kinds of Krylov subspace methods. dent. Taking advantage of the structure of the coefficient matrix
An elementary property of Krylov subspace is that jm is the is a crucial factor in the development of an effective precondition-
subspace of all vectors in Rn which can be written as x = p(A)t, er. Another essential factor is the knowledge of the original phys-
where p is a polynomial of degree not exceeding m  1. All the ical problems from which the linear system arises. The block
techniques provide the same type of polynomial approximations, incomplete LU factorization with dual truncation (BILUT) mecha-
but the different choices of Lm imply different constraints to build nism was utilized as a preconditioner for the Krylov solvers. BILUT
these approximations. These methods, both orthogonal and obli- is a block variant of dual threshold incomplete LU factorization
que, are based on projection techniques and are applied to Krylov (ILUT) preconditioner (Chapman et al., 2000). Of all precondition-
subspaces. The first class of the Krylov subspace methods is based ing techniques, incomplete LU factorizations are the commonest
on the simple choice of Lm ¼ jm and minimum-residual variation way of constructing preconditioner for sparse linear systems and
Lm ¼ Ajm . GMRES and thus FGMRES and DQGMRES methods fall are known to give excellent results for this class of problems (Hig-
into this category which relies on some form of orthogonalization ham, 1996). Incomplete LU factorization, combined with a good
of the Krylov vectors. The second class of methods is based on Krylov subspace projection process, is often regarded as the best
defining Lm to be a Krylov subspace method associated with AT, ‘‘general-purpose’’ iterative solvers (Saad, 1994). We call a factor-
namely, Lm ¼ jm ðAT ; mÞ. These are projection methods that are ization incomplete if certain fill elements, positions that are zero
intrinsically non-orthogonal and built on the basis of a biorthogo- in the original matrix A but not in an exact factorization of A, are
nalization algorithm due to Lanczos. The Lanczos algorithm is a discarded during the factorization process (Barrett et al., 1994).
projection technique which is applied to a Krylov subspace, and Consider a general sparse matrix A whose elements are aij,
the representation of the projected matrix is tridiagonal. BCG, i, j = 1, . . . , N.
CGS, and TFQMR methods fall into this category. They have attrac- A general incomplete LU factorization process computes a
tive properties, but the theoretical analysis of this class is more dif- sparse lower triangular matrix L and a sparse upper triangular ma-
ficult (Saad, 2000; Simoncini and Szyld, 2006; Weiss, 1995). trix U such that the residual matrix R = LU  A satisfies certain con-
straints, such as having zero entries in some locations. A general
algorithm for building incomplete LU factorizations can be derived
6. Implementation of Krylov iterative solvers for the pressure by means of Gaussian elimination and by dropping some elements
equation in predetermined nondiagonal positions (Saad, 2000). The simplest
way to define a preconditioner is in the form of an incomplete fac-
It should be noted that the linearized pressure equation is a torization A = LU + E where L and U have the same nonzero struc-
large sparse system so that special techniques (stationary or non- ture as the lower and upper parts of A, respectively. This so-
stationary) can be utilized to solve it to take advantage of the large called ILU(0) incomplete factorization is rather easy and inexpen-
number of zero elements and their locations in the coefficient sive to compute. By definition, the L and U matrices in ILU (0) have
matrix of the pressure equation. Finite difference matrices on the same number of nonzero elements as the original matrix A
rectangular grids have regular structure because their nonzero en- (Saad, 1994). However, ILU factorizations depend on the imple-
tries form a regular pattern. This can be important for iterative mentation of Gaussian elimination.
solution methods since the efficiency of matrix-by-vector products Several variants of Gaussian elimination are known. These vari-
can improve significantly on high-performance computers depend- ants are commonly referred to as IKJ-based algorithms and depend
ing on whether they are regularly structured or not. on the order of the three loops (i, j, and k associated with the
96 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99

Fig. 10. Convergence plot of Krylov solvers (local relative residual).

Fig. 11. Convergence plot of Krylov solvers during the pressure iterations.

Table 5
1 – Norms of the pressure and phasic velocities.

l1 Norm BCG CGNR DQGMRES FGMRES GMRES TFQMR

Pressure 1.00E05 1.00E05 1.00E05 1.00E05 1.00E05 1.00E05
Vapor velocity 0.001 0.001 0.001 0.001 0.001 0.001
Liquid velocity 0.001 0.001 0.001 0.001 0.001 0.001

Table 6
2 – Norms of the pressure and phasic velocities.

l2 Norm BCG CGNR DQGMRES FGMRES GMRES TFQMR

Pressure 2.48747E05 3.25E05 3.55317E05 3.55317E05 3.55317E05 3.44601E05
Vapor velocity 0.002795085 0.002883141 0.003061862 0.003061862 0.003061862 0.0015
Liquid velocity 0.002795085 0.002883141 0.003061862 0.003061862 0.003061862 0.0015

control variables), with k being the stage number. A few other a dropping rule to an element will only mean the element is re-
alternative methods are based on dropping elements in the Gauss- placed with zero if it satisfies a set of criteria (Saad, 2000).
ian elimination process according to their magnitude rather than Saad (1994) proposed a dual threshold strategy which is as fol-
their locations. With these techniques, the zero pattern P is deter- lows: Fix a drop tolerance s and a number p of fill-ins to be allowed
mined dynamically. A generic ILU algorithm with threshold (ILUT) in each row of the incomplete L/U factors; at each step of the elim-
can be derived from the IKJ version of Gaussian elimination by ination process, drop all fill-ins that are smaller than s multiplied
including a set of rules for dropping small elements, but applying
 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99 97

by the 2 – norm of the current row; of all the remaining ones, keep bundle, cross section of heater rod, and assembly B5 of the PSBT
(at most) the p largest ones in magnitude (Benzi, 2002). rod bundle. The axial heated length of the test section is 3658 mm.
Block ILUT preconditioners, termed BILUT preconditioners in A detailed description of the PSBT benchmark and the test data
our discussion, are block variants of the point preconditioners, is provided as part of the US NRC/OECD benchmark specifications
meaning that every block matrix is considered a scalar matrix. This (Rubin et al., 2010).
structure can be exploited in block ILUT preconditioners where
Gaussian elimination is performed in terms of blocks, that is, small
dense submatrices (Chapman et al., 2000). Arbitrary accuracy is al- 7.1. Exercise 2 of Phase I: steady state in rod bundle
lowed in such preconditioners.
In our study, the coefficient matrix of the pressure equation in This exercise is designed for benchmarking meso-scopic numer-
Eq. (43), which was extracted previously from the three-dimen- ical approaches. We analyzed the steady-state bundle test B5 (Runs
sional two-fluid formalism of two-phase flow, has a block penta- 5.1221, 5.1222, 5.2111, 5.2112, 5.2442, 5.3441, 5.3442, 5.4562,
diagonal structure. This structure can be put in the form of a block 5.6321, 5.6322, and 5.6552). Assembly B5 has a uniform axial
tridiagonal matrix through block partitioning because a block tridi- power distribution. Radial power distribution of this configuration
agonal matrix has an incomplete block factorizations of a particu- is shown in Fig. 6. Table 2 gives the range of operating conditions
larly simple nature due to the fact that no fill can occur outside the for NUPEC PWR bundle test facility to measure the void fraction
diagonal blocks (Ai,i). Consider the block factorization distribution in the bundle under the steady-state condition.
A = (D + L)D1(D + U) = (D + L)(I + D1U), where D is the block diag- According to the measured data, calculated void fractions were
onal matrix of pivot blocks, and L and U are the lower and upper averaged over the four central subchannels of the 5 5 test bun-
triangles of the factorization. We can turn this into an incomplete dles (Channels 15, 16, 21, and 22 in Fig. 5) and reported at three
factorization by replacing the block diagonal matrix of pivots D different elevations along the heated length; lower: 2216 mm,
with the block diagonal matrix of incomplete factorization pivots middle: 2669 mm, and upper: 3177 mm.
X = diag(Xi). Then, the preconditioner can be represented as The heated part of the rod bundle was modeled using a 3-D grid,
M = (X + L)(I + X1U). For factorizations of this type, solving a linear with 6 6 cells in the x–y directions on the subchannel scale, and
system means solving smaller systems with the Xi matrices (Bar- 48 axial cells in the z direction. The location of the 17 grid spacers
rett et al., 1994). The BILUT preconditioning technique is applied, is not exactly equal to that of the spacers on the test bundle. To val-
through the above-said procedure, to the block factorization of idate the results, the void fractions predicted by the two-fluid
the pressure matrix. Then, various Krylov solvers were utilized to model for the central channels of test bundle B5 at the three eleva-
solve the linear algebraic system of pressure equations. tions (lower, middle, and upper) were compared with the ones pre-
We split the algorithm into two stages: (i) factor and (ii) solu- dicted by some computer codes. The latter values which were
tion. The former does not require the right hand side and does need taken from Rubin (2011) are given in Figs. 7–9.
not to be repeated if multiple systems are to be solved with the Table 3 summarizes the results of all 11 test cases based on
same coefficient matrix. The latter, on the other hand, consists of code-to-data comparisons. Deviation is expressed by the difference
the iterative solution of the linear system using a preconditioner. between simulated results and average void fraction predicted by
At each iteration, a preconditioned linear system is solved. The pro- the thermal hydraulic codes given in absolute. Depending on case
cess involves modifying the right hand side and obtaining the and level, the deviation varies between 0.0001 and 0.218.
reduced system solution. Once this solution is obtained, it is used No significant trend was observed in void fraction differences
for retrieving the full solution. between the test conditions. The results of simulation tended to
be overpredicted at lower elevation and underpredicted at the
upper level of the bundle. However, simulation produced satisfac-
7. Results and discussion tory results at the middle part of the bundle in comparison with
the values reported by other codes (Rubin, 2011). It should be
The three-dimensional two-fluid one-pressure scheme was noted that a comparison of the codes and the experimental data
evaluated against the OECD/NRC Benchmark based on the Pressur- performed within the OECD/NRC benchmark analysis shows the
ized Water Reactor (PWR) subchannel and bundle tests (PSBT) un- same results as the simulation performed in this study: over pre-
der the auspices of the Nuclear Power Engineering Corporation diction of the void fraction at the lower elevation in the bundle
(NUPEC) (OECD/NEA Benchmark Based on NU, 2009). PSBT Bench- and under prediction at the upper part of the bundle (Rubin, 2011).
mark was designed to provide an important database of void frac- The OECD/NRC benchmark analysis attributes this over predic-
tion and critical heat flux measurements in steady-state and tion to the fact that the employed method of measurement, which
transient conditions for the assessment of the key models and cor- uses X-ray densitometer, performs void fraction measurements at
relations in the wide range of flow conditions at high pressure on a the center of the subchannel (Rubin, 2011). Under subcooled boil-
prototypical PWR rod bundle. This is a very good opportunity to ing conditions, which play a major role in predicting void fraction
examine the capabilities of physical models to simulate boiling at the lower and middle elevations, the voids accumulate on
flows in PWR core geometry. heated surfaces and are therefore not seen at the center of the sub-
This research used Phase I/Exercise 2 from the PSBT benchmark channel. Thus, the experimentally-determined void fractions will
exercises for the steady-state void distributions in bundle to inves- be lower than the actual void fraction.
tigate the validity of numerical models. This section compares the At the upper elevation, the fluid is generally saturated, and tur-
predictions of the two-fluid model in steady-state conditions with bulent mixing and void drift model becomes important. This study
the PSBT void data. It also presents the efficiency studies of the did not consider vapor and turbulent diffusion enhancement due to
Krylov subspace solvers as modeling options for pressure equation. the presence of mixing grid spacers. We also ignored the lateral
The analysis only considered assembly B5, a 5 5 bundle which flow patterns within the bundle created by specific configurations
simulates the behavior of a 17 17 PWR. The other tests are geo- of the spacers’ structural elements. These phenomena are signifi-
metrically like B5, but they are different in two ways: having dif- cant when the void fraction increases since those determine the
ferent axial power distributions and/or having one of their heat flatness of the void distribution profile in the bundle. Generally,
rods replaced with a thimble. Fig. 5 shows the bundle test facility the majority of the predictions of our simulation were in accept-
of PSBT benchmark for measuring the void distribution of the rod able agreement with the data provided by other codes.
98 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99

7.2. Numerical experiments on the implemented Krylov solvers numerical solution of the two-phase mass, energy, and momentum
equations in rod bundle evaluated by simulating the NUPEC PSBT
To investigate the convergency, stability, and accuracy of the Benchmark. The second objective was investigation of the feasibil-
candidate solvers, a combination of BILUT preconditioner and some ity of implementing preconditioned Krylov methods to solve the
Krylov algorithms was implemented in a two-fluid model and was two-phase flow equations which are compatible with the phasic
then applied to the prescribed PWR steady-state problem of the transport process models. This objective was accomplished by
NUPEC PSBT database. The results obtained from applying Krylov employing the GMRES, FGMRES, DQGMRES, CGNR, BCG, and
solvers to the pressure equations were pitted against the direct TFQMR solvers in combination with a BILUT preconditioner as
inversion method as an accurate technique in a typical case from the advanced pressure iteration schemes. The main conclusions
test series B5. Specifically, six popular Krylov subspace methods are the following:
were studied: GMRES, FGMRES, DQGMRES, CGNR, BCG, and
TFQMR. There are other modern Krylov solvers such as BiCGSTAB (1) The numerical scheme capability is found to be reasonably
which were not explored in this article, but were studied in the accurate for sophisticated physical modeling in order to
previous ones. It has been found that improvement is achieved cover the steady state spectrum of conceivable bundle
when it is feasible to use good enough preconditioning techniques situations.
(Joo and Downar, 1996; Downar and Joo, 2001; Kastanya, 2002; (2) The results presented for steady-state test bundles show
Cuervo et al., 2006). that the constitutive relationships are acceptable, but it
Our test problem included general sparse and non-symmetric seems that further investigation is needed because the com-
matrices arising from two-phase flow in nuclear applications. We bined macroscopic and microscopic approaches should be
used incomplete-LU factorization (not threaded) with a drop toler- compatible with two-phase flow simulation.
ance of 104 and a fill-in of 2. We also set an upper bound of 100 (3) The presence of grid spacers in the flow channels has a sig-
for the number of preconditioned iterations as the maximum num- nificant effect on the following mechanisms: turbulent mix-
ber of matrix–vector multiplications, meaning that the iteration ing, diversion crossflow, void drift, and directed crossflow
will continue until convergence test is satisfied. The stopping crite- (Avramova, 2007). Therefore, in order to accurately simulate
rion which is applied to the relative residual for the Krylov itera- the vapor phase acceleration caused by actual spacer config-
tions is set at 106. Maximum number of inner iterations urations, the grid spacer model must be revised and sub-
(pressure iterations) is set to 50. The axial distributions of the pres- stantially improved.
sure and phasic velocities were obtained using the Krylov subspace (4) The idea of using different Krylov subspace solvers (i.e.,
methods and the direct inversion method. Table 4 shows how the GMRES, FGMRES, DQGMRES, CGNR, BCG, and TFQMR) which
results from the Krylov solvers deviate from those obtained from were introduced for non-symmetric linear systems was
direct inversion. The results are presented for central channels in examined by the NUPEC PWR PSBT Test. Within each New-
test bundle B5. It can be seen that the deviations are negligible ton iteration, an appropriately preconditioned Krylov solver,
in all comparisons. This proves that implementing the Krylov solv- with a BILUT preconditioner, was utilized for solving the lin-
ers does not introduce deviations to the two-phase results. ear system of pressure equations. The Krylov solvers were
In Fig. 10, the six mentioned Krylov schemes are compared in compared in terms of accuracy with the results obtained
terms of convergence behavior in calculating the pressure field from the direct inversion method. Our study shows that
during the Krylov iterations. The convergence plot in Fig. 10, called implementing the Krylov solvers does not impose deviations
Local Relative Residual in this paper, is calculated with the pres- to the two-phase results. We observed a more efficient and
sure iteration being fixed. stable convergence performance when performing GMRES,
Interestingly, the convergence behavior of GMRES, FGMRES, and FGMRES, and DQGMRES as the inner iterative scheme. More-
DQGMRES are very similar, and their plots coincide completely. On over, the convergence behaviors of all the three solvers were
the other hand, GMRES, FGMRES, and DQGMRES converged faster completely identical because both FGMRES and DQGMRES
than did TFQMR, BCG, and CGNR. Moreover, the convergence are flexible variants of GMRES.
behavior of GMRES, FGMRES, DQGMRES, and CGNR is smoother (5) The work cooperatively performed by Pennsylvania State
than that of BCG and TFQMR. Fig. 11 presents the plot of the conver- University (PSU) and University Polytechnic of Madrid
gence results for various solvers during the inner iterative process. (UPM) (Cuervo et al., 2004, 2005, 2006) recommended the
This plot, termed Relative Global Error in this article, is obtained in a Krylov subspace methods for large cases. Accordingly, we
selected Newton iteration during the pressure iterations. are intent on conducting another research to substantiate
All Krylov solvers show a stable convergence history after a few the superiority of the proposed two-fluid algorithm based
iterations. However, BCG, CGNR, and TFQMR exhibit abrupt jumps on Krylov methods over the same algorithm based on direct
and downs in the primary iterations, but GMRES, FGMRES, and inversion methods by performing PWR core analysis. How-
DQGMRES are stagnated. Once again, the convergence plots of ever, with the current research we have actually taken steps
GMRES, FGMRES, and DQGMRES coincide completely. Comparison toward this goal. Above all, performing the three-dimen-
of error norms of the two selected variables, pressure and phasic sional, two-fluid algorithm based on Krylov methods in the
velocities, are given in Tables 5 and 6. The Krylov solvers have PWR core analysis case requires the following:
equal 1 – norms, and their 2 – norm differences are small. i. Extended computer resources.
Our experiments generally indicate that GMRES, FGMRES, and ii. Availability of an extensive large-scale PWR database to
DQGMRES show a more efficient and stable convergence perfor- benchmark the results of the simulations. Up to now, the
mance than the BCG, CGNR, and TFQMR methods. However, it is hard above-mentioned requirements have not been available
to determine which Krylov subspace solver is the absolute winner. to us.
(6) Since the preconditioners play a crucial role in the precondi-
tioned iterative solver and usually represent the major por-
8. Conclusions tion of the memory cost, quality of the preconditioner is a
more serious problem than the choice of the Krylov subspace
This work had two objectives. The first one was the implemen- accelerator in designing a preconditioned iterative solver for
tation of a fully three-dimensional two-fluid model to the large scale CFD applications (Zhang, 2000). Thus, we propose
 M. Mohitpour et al. / Annals of Nuclear Energy 63 (2014) 83–99 99

a more sophisticated and accurate preconditioner for paral- Joo, H., Barber, D., Jiang, G., Downar, T., 1998. PARCS: A Multi-Dimensional Two-
Group Reactor Kinetics Code based on the Nonlinear Analytic Nodal Method,
lel computations because we think it would lead to a more
Tech. Rep. PU/NE-98-26, Purdue University.
robust and efficient solver in large-scale problems in nuclear Kastanya, D.F., 2002.Imolementation of a Newton–Krylov Iterative Method to
reactor safety analysis. It is clear that a comparative analysis Address Strong Nonlinear Feedback Effects in FORMOSA-B BWR Core Simulator.
that includes robustness, efficiency, speedup, and parallel- Ph.D. Thesis, North Carolina State University.
Kelly, J.E., 1980. Development of a Two-Fluid Model for Light Water Reactor
ism studies is needed to perform this task more efficiently. Subchannel Analysis. Ph.D. Thesis, Massachusetts Institute of Technology (MIT).
Kelly, J.E., Kazimi, M.S., 1979. Development and Testing of the Three Dimensional,
Two-Fluid Code THERMIT for LWR Core and Subchannel Applications. MIT
Energy Laboratory Electric Utility Program. Report No. MIT-EL-79-046.
Acknowledgement Kelly, J.E., Kazimi, M.S., 1981. Interfacial Exchange Relations for Two-Fluid Vapor–
Liquid Flow, a Simplified Regime Map Approach. MIT Energy Laboratory Electric
The authors wish to thank the Sparskit library developer Yousef Utility Program. Report No. MIT-EL-81-024.
Kelly, J.E., Kao, S.P., Kazimi, M.S., 1981.THERMIT-2: A Two-Fluid Model for Light
Saad for the work accomplished and the free distribution of the Water Reactor Subchannel Transient Analysis. MIT Energy Laboratory Electric
library. Utility Program. Report No. MIT-EL-81-014.
Liles, D.R., Reed, Wm.H., 1978. A semi-implicit method for two-phase fluid
dynamics. J. Comput. Phys. 26, 390–407.
References Mousseau, V.A., 2004. Implicitly balanced solution of the two-phase flow equations
coupled to nonlinear heat conduction. J. Comput. Phys. 200, 104–132.
Allaire, G., 1995. Solving linear systems in FLICA-4, a Thermal–hydraulic code for 3- OECD/NEA Benchmark Based on NUPEC PWR Subchannel and Bundle Tests (PSBT),
D transient computations. In: Proceedings of the International Math. Comp. 2009. JNES.
Mtg., Portland, OR, pp. 284–288. Ortega, T., Rheinboldt, W., 1970. Iterative Solution of Nonlinear Equations in Several
Avramova, M., 2007. Development of an Innovative Spacer Grid Model Utilizing Variables. Academic Press.
Computational Fluid Dynamics within a Subchannel Analysis Tool. Ph.D. Thesis, Reed, Wm.H., Stewart, H.B., Wolf, L., 1979. Applications of the THERMIT code to 3D
Pennsylvania State University. Thermal Hydraulic Analysis of LWR Cores. Brookhaven National Laboratory.
Barrett, R., Berry, M., Chan, T.F., Demmel, J., Donato, J., Dongarra, J., Eijkhout, V., Report No. BNL-25484.
Pozo, R., Romine, C., vander Vorst, H., 1994. Templates for the Solution of Linear RELAP5/MOD3.3 Code Manual, 2001. Code Structure, System Models, and Solution
Systems: Building Blocks for Iterative Methods. SIAM, Philadelphia, PA. Methods, NUREG/CR-5535 Edition, vol. I.
Benzi, M., 2002. Preconditioning techniques for large linear systems: a survey. J. Rubin, A., 2011. OECD/NEA Benchmark Based on NUPEC PWR Subchannel and
Comput. Phys. 182, 418–477. Bundle Tests (PSBT), Analysis using Subchannel Code CTF and System Code
Chapman, A., Saad, Y., Wigton, L., 2000. High order ILU preconditioners for CFD TRACE. M.S. Thesis, Pennsylvania State University.
problems. Int. J. Numer. Method Fluids 33, 767. Rubin, A., Schoedel, A., Avramova, M., Utsuno, H., Bajorek, S., Velasquez-Lozada, A.,
Cuervo, D., Avramova, M., Ivanov, K., 2004. Improving the computation efficiency of 2010. OECD/NRC Benchmark Based on NUPEC PWR Subchannel and Bundle
COBRA-TF for LWR safety analysis of large problems. In: Proceedings: PHYSOR Tests (PSBT). Experimental Database and Final Problem Specifications, Tech.
2004, Chicago, USA. Rep. NEA/NSC/DOC, vol. I.
Cuervo, D., Avramova, M., Ivanov, K., 2005. Implementation and performance of Saad, Y., 1994. ILUT: a dual threshold incomplete LU factorization. Numer. Linear
Krylov methods for the solution of two-fluid hydrodynamics equations in the Algebra Appl. 1, 387.
COBRA-TF Code. In: Proceedings: M&C 2005, Avignon, France. Saad, Y., 2000. Iterative Methods for Sparse Linear Systems, second ed. PWS
Cuervo, D., Avramova, M., Ivanov, K., Miró, R., 2006. Evaluation and enhancement of Publishing Company, Boston, MA.
COBRA-TF efficiency for LWR calculations. Ann. Nucl. Energy 33, 837–847. Schor, A.L., Kazimi, M.S., Todreas, N.E., 1984. Advances in two-phase flow modeling
Delhaye, J.G., Achard, J.L., 1976. On the averaging operators introduced in two-phase for LMFBR applications. Nucl. Eng. Des. 82, 127–155.
flow modeling. In: Specialists’ Meeting on Transient Two-Phase Flow, Toronto, Simoncini, V., Szyld, D.B., 2006. Recent computational developments in Krylov
Canada. subspace methods for linear systems. Numer. Linear Algebra Appl. 14, 1–59.
Downar, T., Joo, H., 2001. A pre-conditioned Krylov method for solution of the SPARSKIT(Version 2), 1994. A basic tool-kit for sparse matrix computations.
multi-dimensional two-fluid hydrodynamics equations. Ann. Nucl. Energy 28, doi:10.1.1.131.4808.
1251–1267. Stewart, H.B., 1979. J. Comput. Phys. 33, 259–270.
Du Croz, J., Mayes, P., Radicati, G., 1990. Factorization of band matrices using Level- Sugawara, S., Miyamoto, Y., 1990. FIDAS: Detailed subchannel analysis code based
3 BLAS. Proceedings of CONPAR 90 VAPP IV Lecture Notes in Computer Science. on the three-fluid and three-field model. Nucl. Eng. Des. 120, 147–161.
Springer, Berlin, p. 222. TRAC-M/FORTRAN 90 (Version 3.0) Theory Manual, 2000. LA-UR-00-910.
Emonot, P., Souyri, A., Gandrille, J.L., Barré, F., 2011. CATHARE 3: A new system code Turinsky, P.J., et al., 1994. NESTLE: A Few-Group Neutron Diffusion Equation Solver
for thermal–hydraulic in the context of the NEPTUNE project. Nucl. Eng. Des. Utilizing the Nodal Expansion Method for Eigenvalue, adjoint, fixed-source
241, 4476–4481. steady-state and transient problems. EGG-NRE-11406.
Frepoli, C., Mahaffy, J., Ohkawa, K., 2003. Notes on the implementation of a fully- Turner, J., Doster, J., 1991. Performance of the conjugate gradient-like algorithms in
implicit numerical scheme for a two-phase three-field flow model. Nucl. Eng. transient two-phase subchannel analysis, ANS Winter Meeting, San Francisco.
Des. 225, 191–217. VIPRE-02, 1994. A Two-Fluid Thermal–Hydraulic Code for Reactor Core and Vessel
Gan, J., Xu, Y., Downar, T., 2003. A matrix-free Newton method for coupled Analysis, vol. 1. Mathematical Modeling and Solution Methods. EPRI TR-
neutronics thermal–hydraulics reactor analysis. Nuclear mathematical and 103931-V1, Final report.
Computational Sciences: A century in Review. American Nuclear Society. Weiss, R., 1995. A theoretical overview of Krylov subspace methods. Appl. Numer.
Granziera, M.R., Kazimi, M.S., 1980. A two-dimensional, two-fluid model for sodium Math. 19, 207–233.
boiling in LMFBR fuel assemblies. Energy Laboratory Report No. MIT-EL 80-011. Yadigaroglu, G., Andreani, M., Dreier, J., Coddington, P., 2003. Trends and needs in
Higham, N.J., 1996. Accuracy and Stability of Numerical Algorithms. Ch12. SIAM, experimentation and numerical simulation for LWR safety. Nucl. Eng. Des. 221,
Philadelphia, PA. 205–223.
Hotta, A., Takeuchi, H., Takeda, S., 1998. Three dimensional evaluation of two-phase Zhang, J., 2000. Preconditioned Krylov subspace methods for solving non-
flow in BWR fuel bundles based on compressible two-fluid one-pressure and k– symmetric matrices from CFD applications. Comput. Methods Appl. Mech.
e turbulence models. Ann. Nucl. Energy 25 (7), 438–463. Eng. 189, 825–840.
Joo, H., Downar, T., 1996. An incomplete domain decomposition preconditioning
method for nonlinear nodal kinetics calculating. Nuc. Sci. Eng. 123, 403–414.