Available online at www.sciencedirect.

com
ScienceDirect

Mathematics and Computers in Simulation 180 (2021) 328–353

www.elsevier.com/locate/matcom

Original articles

Introducing compressibility with SIMPLE algorithm

Jian Qin, Huachen Pan, M.M. Rahman ∗, Xiaoqing Tian, Zefei Zhu
Hangzhou Dianzi University, School of Mechanical Engineering, 310018 Hangzhou, China
Received 13 June 2020; received in revised form 7 September 2020; accepted 9 September 2020
Available online 11 September 2020

Abstract
A comparative assessment between SIMPLE and its variant SIMPLE-C is conducted based on two-dimensional (2-D)
incompressible flows, using a cell-centered finite-volume ∆-formulation on a collocated grid. The SIMPLE-C (SIMPLE
Compressibility) scheme additionally combines the concept of artificial compressibility (AC) with the pressure Poisson equation,
provoking the diagonal dominance of influence coefficients. An improved nonlinear momentum interpolation scheme is
employed at the cell face in discretizing the continuity equation to suppress pressure oscillations. The pseudo-time step ∆ti
remains the same in both schemes to conserve an analogous scaling with momentum and scalar nodal influence coefficients.
Numerical experiments in reference to buoyancy-driven cavity flow dictate that both contrivances execute a residual smoothing
enhancement, facilitating an avoidance of the velocity/pressure under-relaxation (UR). However, compared with the SIMPLE
approach, included benefits of the SIMPLE-C method are the use of larger Courant numbers, enhanced robustness and
convergence. Excellent consistency is obtained between results available in the literature and numerical solutions obtained
by both SIMPLE and SIMPLE-C solvers. The segregated SIMPLE algorithm is finally reformulated in conjunction with a
new slope/flux limiter function to predict fluid flow at all speeds. Numerical results show that the compressible variant of
SIMPLE replicates correct shock speed, well-resolved shock front, contact discontinuity and rarefaction waves when compared
with analytical solutions. Compressible laminar flows are computed to further support the accuracy and robustness of proposed
algorithm.
⃝c 2020 The Author(s). Published by Elsevier B.V. on behalf of International Association for Mathematics and Computers in
Simulation (IMACS). This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Keywords: SIMPLE algorithm; Artificial compressibility; Under-relaxation factor; Convergence and robustness; Limiter function

1. Introduction
The dual role of pressure comprising strongly coupled to velocity and density, respectively at low-Mach and
high-Mach number regimes, facilitates the success of pressure-based methods in solving Navier–Stokes (NS)
equations [1,7,51] wherein the concept of artificial compressibility (AC) is exploited [9,38,40,42,49,54]. In principle,
the AC is an exceptional aspect related to the preconditioned density-based compressible flow method, confronting
a significant coupling of both pressure–velocity and pressure–density simultaneously [49]. However, a consistently
formulated pressure-based algorithm can resolve stability and convergence issues due to the strong coupling and
nonlinearity of governing equations [6,11,31,57].
The preconditioning of compressible code with an AC parameter is an eminent means of simulating nearly
incompressible flows, associating conventionally a time-dependent pressure term to the continuity equation together
∗ Corresponding author.
E-mail address: [email protected] (M.M. Rahman).

https://doi.org/10.1016/j.matcom.2020.09.010
0378-4754/⃝ c 2020 The Author(s). Published by Elsevier B.V. on behalf of International Association for Mathematics and Computers in
Simulation (IMACS). This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Nomenclature
C Sound speed
CFL Courant number
c p,v Specific heat capacities
E Total internal energy
e Specific internal energy
F, G Flux vectors in x and y directions
Gr Grashof number
H Total enthalpy
L Characteristic length
Ma Mach number
Ṁ ∗ Fictitious mass source
p Pressure
p′ Pressure correction
p∗ Tentative pressure
Pr Prandtl number
Q Source term
Ra Rayleigh number
S Cell-face area
t Time
∆t ′ Weighting factor
T Temperature
UR Under-relaxation
u, v Velocity components in x and y directions
u ′, v′ Velocity corrections
u ∗, v∗ Tentative velocities
U, V Contravariant or dimensionless velocity components
W Conservative variable vector
x, y Cartesian coordinates
Γ Diffusion coefficient
γ Specific heat capacity ratio
ζ Cell-face damping factor
θ Dimensionless temperature
ν Kinematic viscosity
ρ Density
ρ∗ Tentative density
ρ′ Density perturbation or correction
φ Scalar variable or slope limiter
Ψ Reference velocity or flux limiter
∀ Cell volume
Subscript
r Reference condition
nb Neighboring grid point
w Wall condition

329
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

with a multiplicative variable recognized as an artificial sound speed C [9]. This approach allows the mass
continuity to be advanced in a time-marching manner, analogous to the momentum equations and hence, the AC
term plays a crucial role in replicating a compatible accuracy, stability and convergence adhering to a numerical
method [38,40,42,49,54]. In particular, considerable research has been devoted to formulating pressure-based
algorithm SIMPLE (Semi-Implicit Method for Pressure Linked Equation) and its variants for predicting flows at all
Mach numbers wherein the sound speed C is significantly influential to determine a credible success [10,20,31];
incompressibility is recovered as a limiting case of this formulation (e.g., setting density as constant).
The SIMPLE algorithm [32,34] with the pressure–velocity coupling for incompressible flows assembles a number
of consecutively progressive algorithms, merely aims at enhancing the robustness and convergence [2,3,8,12,16–
18,27,28,32,33,35,37,39,41,43,44,48,50,55]. The SIMPLER (SIMPLE Revised) of Patankar [32,33] employs a
pressure-correction equation to adjust the velocity field, resembling the SIMPLE; however, it uses a separate
pressure equation to correct the pressure field, requiring much more CPU time when compared with the SIMPLE.
The PRIME (Pressure Implicit Momentum Explicit) algorithm of Maliska and Raithby [27] solves momentum
equations explicitly; no under-relaxation (UR) is required and the pressure field represents the most important
factor in the overall convergence. Spalding [48], and Markatos and Pericleous [28] have developed the SIMPLEST
(SIMPLE ShorTened) algorithm where the implicit treatment of diffusion terms in discretized momentum equations
are combined with an explicit treatment of the convection terms. The SIMPLEST algorithm can be found to be
a compromise between SIMPLE and PRIME [31]. The SIMPLEX (X appended to SIMPLE indicating the use of
extrapolation) algorithm of Van Doormaal and Raithby [50] ensures that the rate of convergence will not degrade
with grid refinement. This concern is addressed in SIMPLEX by considering the influence of pressure differences
outside but in the vicinity of pressure-difference, local to the velocity. The SIMPLEM (SIMPLE-Modified) algorithm
of Acharya and Moukalled [2] consists of a predictor and a corrector stage wherein disadvantages and advantages
of SIMPLE are interchanged [31]. Van Doormaal and Raithby [12] have provided several suggestions for the
enhancements of SIMPLE method in terms of boundary conditions, solution techniques, convergence criteria and
a new variant SIMPLEC (SIMPLE Consistent), which partially accounts for the influence of surrounding velocity
nodes, and thus obtaining the improved velocity correction equation. Issa et al. [17] and Issa [16] have proposed
PISO (Pressure Implicit with Split Operator) algorithm, which includes one prediction step and two corrector steps.
Rahman and Siikonen [39], and Rahman et al. [37,43] have also modified the pressure-correction method, enhancing
the robustness and convergence when compared with the traditional SIMPLE algorithm.
Above-mentioned analyses dictate that a family of SIMPLE variants has been developed, preserving undoubtedly
a creditable success. Unfortunately, it seems likely that no consensus of exact preferences is available as to
which variant of SIMPLE is the best for general scientific and engineering applications. The methodology of
SIMPLE is to devise a second-order Poisson equation for the pressure-correction by manipulating the continuity and
momentum equations, updating afterward the pressure and velocity fields until the discrete continuity is satisfied.
This deterministic coupling together with the UR factor is possibly the major reason, leading to slow convergence
when compared with its variants. Nevertheless, the SIMPLE algorithm may remain competitive with its developed
variants when the actual computational time rather than the number of iterations is considered as the measure of
merit. Consequently, the SIMPLE can be improved by avoiding the involved UR factors, preventing the divergence
of iterative solution for strongly nonlinear problems; this artifact is implicitly introduced with the coefficients of
SIMPLEC pressure-correction equation [12,31]. In this way, the SIMPLE without UR factors eventually achieves an
enhanced competency with a higher convergence-rate. However, on a collocated grid, the physical requirement of
SIMPLE is to muster assertively a precluded decoupling of pressure–velocity fields. A lot of innovative formulations
based on primitive variables have been constructed to eliminate the destabilizing effect arising from the pressure
checker-boarding on collocated grid arrangements [37,41,44].
In particular, combining the pressure Poisson equation of SIMPLE method with an AC parameter is conducive
to magnifying the diagonal dominance of influence coefficients. In addition, the convergence to the steady state
can be considerably accelerated with a higher CFL number as the pseudo-time step is parameterized with diffusive,
convective and AC-generated sound speeds. With the finite-volume ∆-formulation, the implicit stage conserves the
explicit density residual ∆ρ (e.g., corrected mass imbalance in the SIMPLE method) due to the use of primitive
rather than conservative variables; ∆ρ is explored to smooth momentum and scalar residuals which can reduce the
well-recognized slow convergence aspect associated with the SIMPLE approach. In the present study, the cell-face
velocity interpolation technique devised by Rahman et al. [37] is modified to suppress the local extrema into the
330
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

tentative cell-face velocity, entailing essentially a prerequisite (since the fictitious mass imbalance is responsible for
estimating an appropriate pressure correction) to controlling the velocity or pressure UR involved in the SIMPLE
method.
To this end, it is worth pointing out that the only difference between SIMPLE and SIMPLEC is in the
definition of coefficients (e.g., UR factor is implicitly introduced) in the pressure-correction equation [12,31],
resulting in an avoidance of velocity or pressure UR. Nevertheless, this aspect is not fully supported by the
SIMPLEC when numerically computing flows with dominant source terms (and where the pressure undergoes a
very large change) in the interest of convergence and robustness [18,55]. In the current research, the SIMPLEC-like
competency is integrated with the SIMPLE in conjunction with an AC parameter entitled as the SIMPLE-
C (SIMPLE Compressibility); however, compared to the SIMPLEC, the SIMPLE-C neglects the influence of
surrounding velocity nodes. Equations of motion are solved in a segregated fashion depending on the diagonal
dominance for convergence. The overall artifact dramatically expedites a residual smoothing enhancement, thereby
facilitating an avoidance of velocity or pressure UR. Viscosity-driven and buoyancy-driven cavity flows wherein
the flow-regime ranges from highly diffusive to convection dominated, could be appropriate numerical experiments
[5,13,22,25,26,29,52,53] to validate the efficacy and accuracy of proposed SIMPLE incompressible and compressible
algorithms. Finally, it must be stressed herein that the SIMPLE-C inherits a relevant root to extend its capability
toward predicting fluid flow at all speeds due to the inclusion of an AC parameter; this is, in principle, accomplished
in the present study together with a newly designed slope/flux limiter. The compressible version of SIMPLE is
extensively examined with various typical benchmark tests of one- and two-dimensional compressible flows.

2. Governing equation
Two-dimensional (2-D) convection–diffusion equations with an inclusion of a scalar variable φ can be represented
in the following form:
∂W ∂(Finv − Fvis ) ∂(G inv − G vis )
+ + =Q (1)
∂t ∂x ∂y
where W = (ρ, ρu, ρv, ρφ)T and Q signifies the source term. The inviscid fluxes are:
ρu ρv
⎛ ⎞ ⎛ ⎞
⎜ ρu 2 + p ⎟ ⎜ ρuv ⎟
⎝ ρuv ⎠ ,
Finv = ⎜ G inv = ⎜
⎝ ρv 2 + p ⎠ (2)
⎟ ⎟

ρuφ ρvφ
In Eq. (2), ρ is the density, u and v are the Cartesian velocity components and p implies the pressure. The viscous
fluxes can be given as:
⎛ ⎞ ⎛ ⎞
0 0
⎜ Γ (∂u/∂ x) ⎜ Γ (∂u/∂ y) ⎟
⎟,
⎟
Fvis = ⎜⎝ Γ (∂v/∂ x) ⎠ G vis =⎜ ⎟
⎝ Γ (∂v/∂ y) ⎠ (3)
Γ (∂φ/∂ x) Γ (∂φ/∂ y)
Γ denotes the diffusion coefficient, e.g., either viscosity or thermal conductivity of the fluid.

3. Finite-volume discretization with ∆-formulation

The governing equations are discretized using a cell-centered finite-volume scheme, having an integral form
∫ ∫ ∫
d
W d∀ + F(W ) · d S = Qd∀ (4)
dt ∀ s ∀
for an arbitrary region ∀ with a boundary S. Carrying out the integration for a computational cell i yields:
d Wi ∑
∀i = −S F̂ + ∀i Q i (5)
dt f aces

where the summation is taken over the faces of computational cell, each surface has a unit normal vector n, defined
by
Sx Sy
n = nx i + n y j = i + j (6)
S S
331
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

with the corresponding cell-face flux:

F̂ = n x F + n y G (7)
where fluxes F and G are defined by Eqs. (2) and (3), respectively. To account for the directional impact on
the upwinding process, the inviscid flux at the cell face (i + 1/2) is evaluated using a rotational matrix [39].
The rotational matrix transforms contravariant momentum components to primitive components, representing an
analogous flux-expression to that obtained in the Cartesian coordinate system. On a curvilinear grid, a second-order
central differencing scheme together with a thin-layer approximation is used to evaluate the viscous fluxes.

3.1. Tentative velocity field

In principle, this is the predictor step of SIMPLE-type algorithm; the transformation from conservative to
primitive variables is performed before the implicit stage, providing fluid constitutive explicit relations which can
be presented as a matrix form:
ρ 0 0 0
⎛ ⎞ ⎛ ⎞⎛ ⎞
∆ρ ∆ρ
⎜ ∆u ⎟ 1⎜⎜ −u 1 0 0 ⎟ ⎜ ∆ρu ⎟
⎟⎜ ⎟
⎝ ∆v ⎠ = ρ ⎝ −v 0 1 0 ⎠ ⎝ ∆ρv ⎠
⎜ ⎟ (8)
∆φ −φ 0 0 1 ∆ρφ
The conservative explicit changes ∆W can be obtained from:
∆ti
∆W = (∆ρ, ∆ρu, ∆ρv, ∆ρφ)T = Ri (9)
∀i
where ∆ti is the spatially varying pseudo-time step. Residuals Ri are evaluated by the right-hand side of Eq. (5);
a fully upwinded second-order (FUS) difference is utilized to approximate the convective flux residual since this
scheme uses only upwinded terms to extrapolate the cell face value, maintaining a stable solution [52]. It is worth
mentioning that the operation confronts an explicit density residual ∆ρ, resembling the mass imbalance generated
by corrected velocity fields; linearizing momentum and scalar residuals with ∆ρ is presumably beneficial to predict
a realistic pressure field with an avoidance of UR factor.
Allowing a first-order backward difference to the temporal variation of W̃i , the implicit pseudo-time integration
pertaining to Eq. (5) yields a discretized system of algebraic equations [39]:
[ ]
∆t ∗ ∑ ∆t ∗ ∑
1+ Anb ∆W̃i = Anb ∆W̃nb + ∆W̃i∗ (10)
∀i nb ∀i nb
where ∆t ∗ = ξ ∆ti and ξ = 1.5, a factor governing the stability of implicitness. The subscript nb stands for a
run over the neighboring nodes (i + 1), (i − 1), ( j + 1), and ( j − 1). The quantity ∆W̃i∗ = (∆u, ∆v, ∆φ)T
represents primitive explicit residuals of dependent variables as defined by the left hand side of Eq. (8). The
influence coefficient A consists of diffusion and mass fluxes which are discretized using central and first-order
upwind differencings, respectively.
Primitive variables W̃ = (u ∗ , v ∗ , φ)T are updated from the following relation after the implicit stage:
W̃ n+1 = W̃ n + (∆u, ∆v, ∆φ)T (11)
where W̃ n+1 indicate new values at time level (n + 1)∆t. Eq. (11) provides tentative velocity fields with an arbitrary
pressure field. Unlike the commonly used SIMPLE-like algorithms, velocity and scalar residuals are calculated
simultaneously during an iteration cycle in the same manner as in density-based methods; however, a mass imbalance
is allowed to form with tentative velocity fields for the pressure-correction step (e.g., corrector step of SIMPLE
algorithm).

3.2. Pressure-correction equation

Fundamentally, this is recognized as the corrector step of SIMPLE algorithm in which mass and momentum
equations are satisfied with tentative velocity and pressure fields after repeated iterations. The velocity and pressure
fields are corrected as (for convenience, only u and p are considered):
u = u∗ + u′ (12)
332
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

p = p∗ + p′ (13)
′ ′ ∗ ∗
where u and p signify the incremental velocity and pressure, respectively. Quantities u and p represent tentative
values. The velocity-correction can be linked to the pressure-correction on a curvilinear coordinate system by [39]:
∆ti′ (Sn x ∂ p ′ )i + (Sn x ∂ p ′ ) j
u i = u i∗ − (14)
ρ ∀i
with
(Sn x ∂ p ′ )i (n x p ′ )i+1/2 − (n x p ′ )i−1/2
≈ (15)
∀i ∆n i
(Sn x ∂ p ′ ) j (n x p ′ ) j+1/2 − (n x p ′ ) j−1/2
≈ (16)
∀i ∆n j
where ∆ti′ is a weighting factor; ∆n i and ∆n j are the control-volume cell thicknesses in i and j directions,
respectively. The cell face value of pressure-correction p ′ is obtained as an average from two adjacent nodal points.
Within the framework of traditional AC method, the continuity equation is modified by introducing a time-
derivative of p ′ , relaxing the incompressibility constraint [9]; applying the chain rule yields:
∂ρ ′ ∂ p ′ ∂u ∂v
( )
+ρ + =0 (17)
∂ p ′ ∂t ∂x ∂y
Rearranging Eq. (17) to obtain:
1 ∂ p′ ∂u ∂v
( )
+ρ + =0 (18)
C 2 ∂t ∂x ∂y
where C is recognized as the artificial sound speed/AC parameter:
1 ∂ρ ′
= (19)
C2 ∂ p′
and is optimized numerically as:
√ [ ]
1
C = β max (u 2 + v 2 ); Ur2 (20)
2
where Ur implies a reference velocity and β indicates a compressibility parameter. Ostensibly, C depends on β,
having a significant impact on the convergence and stability of solver when C has the involvement in evaluating
the pseudo-time step ∆ti . Usually, values of β stay in the range 1–10 to comply with better convergence to the
steady-state.
The finite-volume discretization of Eq. (18) can be written as:
∀i i+1/2 j+1/2
p ′ + (ρU S)i−1/2 + (ρV S) j−1/2 = 0 (21)
Ci ∆ti i
2

Applying relations like Eqs. (14) and (21) results in a truncated pressure-correction relation:
( )
∀i ∑ ∑
2
+ Bnb pi′ = ′
Bnb pnb − Ṁi∗ (22)
Ci ∆ti nb nb
where
∑
Bnb = Bi+1 + Bi−1 + B j+1 + B j−1 (23)
nb
∆t ′ S ∆t ′ S
( ) ( )
Bi±1 = , B j±1 =
∆n i±1/2 ∆n j±1/2

The fictitious mass imbalance Ṁi∗ in Eq. (22) can be given by:
Ṁi∗ = Ṁi+1/2
∗ ∗
− Ṁi−1/2 + Ṁ ∗j+1/2 − Ṁ ∗j−1/2 (24)
∗
Ṁi±1/2 = (ρU S)i±1/2 ,
∗
Ṁ ∗j±1/2 ∗
= (ρU S) j±1/2
333
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

It must be stressed herein that cross-diffusion fluxes in Eq. (22) are neglected using a thin-layer approximation in
order to maintain diagonally dominant coefficient matrices [39,43]. Physically, the first term on the left-hand side
in Eq. (22) signifies an addition of more weight on the diagonal nodal coefficient. It is worth noting that Eq. (22)
represents the SIMPLE algorithm with an AC parameter identified by SIMPLE-C as mentioned earlier and the
parent SIMPLE algorithm can be recovered with an exclusion of the first term on the left-hand side from Eq. (22).
However, the pseudo-time step ∆ti remains the same with both SIMPLE and SIMPLE-C algorithms to benefit the
overall numerical scheme.

3.3. Time step and weighting factor

The spatially varying pseudo-time step ∆ti is determined by employing the relation [39]:
⎡ ⎤
C F L∀i λ 2
∆ti = min ⎣ √ , ⎦ λ = min(∆n i , ∆n j ) (25)
(|U |S)i (|V |S) j + Ci Si2 + S 2j 2kvi
where k = 2 (and for three-dimensional case k = 3), νi denotes the kinematic viscosity; U and V are the
contravariant velocity components in i and j directions, respectively. Grid properties at (i, j) are interpolated from
cell faces. In both SIMPLE and SIMPLE-C algorithms, the pseudo-time step ∆ti remains the same in order to keep
an analogous scaling with nodal influence coefficients involved in discretized momentum and scalar equations.
Relying on the species of a given geometry and flow field, values of CFL in the range 1-30 are recommended for
better convergence to the steady state at which mass and momentum conservation laws are enforced.
The weighting factor ∆ti′ involved in the cell-face dissipation scheme and velocity–pressure correction equation
may be evaluated in many ways [43]. To accommodate better controlling and scaling of both phenomena
(e.g., dissipation scheme and velocity–pressure correction), ∆ti′ is determined from Eq. (10) as:
∆t ∗ ∑
∆ti′ = min(1, C F L) Ai = Anb (26)
Ai nb
where the coefficient Ai emerges naturally from a consequence of discretized momentum equations. The choice of
∆t ′ is an additional criterion to augment stability and convergence in SIMPLE-like algorithms.

3.4. Cell-face velocity

On a collocated grid, the elimination of odd–even point decoupling is accelerated by employing an improved
Rhie–Chow [44] interpolation method on the control-volume cell-face developed by Rahman and Siikonen [37,41]
is adopted. In the present work, the dissipation scheme is modified such that it surmounts local extrema into the
cell-face velocity. This is appropriately preserved by the following explicit reconstruction; referring to a boundary
cell face (i + 1/2):
{ ( }
′
1 ∆ti+1 ∆ti′ ∆t ′
) ( )
∗
Ui+1/2 = U i+1/2 − ζi+1/2
∗
( R L
)
− ( f i+1 − f i ) − p −p (27)
4 ρi+1 ρi 2 ρ ∆n i+1/2
∗
( )
a2
( )
∆n U
with ζi+1/2 = , ai+1/2 =
1 + a 2 i+1/2 ν
i+1/2
∗
where the overbar denotes an averaging between grid nodes, U signifies a contravariant tentative velocity field
and f = Q u n x + Q v n y . The dissipation limiter ζ is presumably well suited to stagnation/highly-diffusive flow
regions, preventing local extrema into the cell-face velocity. The FUS scheme is used to determine state variables
p ∗L and p ∗R on the left and right sides of cell-face. Eq. (27) ensures a strong coupling between pressure and velocity
fields since both variables are mutually dependent on adjacent nodal values, constituting a nonlinear cell-face
interpolation scheme. Eq. (27) is used to calculate cell-face velocity to form a fictitious mass source for the pressure-
correction equation. Apparently, the current dissipation approach has several desirable attributes when compared
with the Rhie–Chow [44] cell-face interpolation scheme: (a) it is not dependent on UR factors; (b) it incorporates a
compact formulation; (c) the included non-pressure gradient source term influences the solution stability [37] and
(d) local extremums into the cell-face velocity are prevented. Thus, the formulation has a compatible competency
in suppressing non-physical oscillations in fluid flow and heat transfer problems with strong source terms.
334
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

3.5. Boundary conditions

Boundary conditions are very sensitive to accuracy and stability of any numerical scheme. Two layers of ghost
cells are set to boundaries; ghost cells retain boundary conditions such that actual boundary conditions are recovered
at the cell-face when a central difference scheme is applied. On a boundary cell face (i −1/2), Dirichlet and Neumann
conditions are formulated as follows:
( )
∂ W̃
W̃i−1 = 2W̃i−1/2 − W̃i , W̃i−1 = W̃i − ∆n (28)
∂n
i−1/2

where (i − 1) represents a ghost cell nodal point and W̃i−1/2 or (∂ W̃ /∂n) is the specified boundary value. On the
solid surface, the central expression of viscous flux is replaced by a second-order one-sided formula. A second-order
extrapolation from the computational domain is applied to evaluate the wall pressure. In the initial iteration step,
∆W̃i−1 = 0. In the pressure-correction equation, the boundary conditions are treated implicitly as follows. In the
ghost cell, Eq. (22) becomes:
′
Bi−1 pi−1 = Bi pi′ (29)
where Dirichlet and Neumann boundary conditions assume that Bi = 0 and Bi−1 = Bi , respectively. This strategy
enhances the convergence-rate.

3.6. Steps of algorithm

Equations are solved by using a tridiagonal matrix algorithm (TDMA), The different steps of solver can be
summarized as follows:
1. Guess pressure, velocity, and scalar fields.
2. Solve momentum and scalar equations using Eq. (10) and update tentative velocity and scalar fields with
relation (11).
∗
3. Calculate Ũi±1/2 , using relations like Eq. (27) to form the mass imbalance Ṁi∗ and solve Eq. (22) for p ′ .
4. Correct velocity and pressure fields.
5. Repeat steps 2–4 until convergence is attained.

4. SIMPLE for compressible flow

To facilitate the subsequent development, the calculation advances with a recourse to the 1-D Euler equations,
governing an unsteady compressible inviscid flow which can be expressed in the conservative form as:
∂W ∂ F(W )
+ =0 (30)
∂t ∂x
with unknown conservative variables W and inviscid fluxes F:
ρ ρu
⎛ ⎞ ⎛ ⎞

W = ⎝ ρu ⎠ , F = ⎝ ρu 2 + p ⎠ (31)
ρE ρu H
where E and H denote the total specific internal energy and total specific enthalpy, respectively. It is worth
noting that an incompressible flow usually utilizes a pressure-difference from an ambient pressure; however, in the
compressible flow an absolute pressure must be used in the energy flux which is obtained by adding the constant
ambient pressure to the interpolated pressure-difference. This set of equations is completed with the constitutive
law of the gas, i.e., equation of state (EOS). Considering the perfect gas model, the following state equations are
obtained:
C2 u2
√
p p p
C= γ , ρ= , H=E+ = + (32)
ρ (γ − 1)e ρ γ −1 2
where e = (E − u 2 /2) is the specific internal energy, C is the local sound speed and γ = c p /cv is the ratio of
specific heat coefficients at constant pressure and volume, respectively; typically γ = 1.4 for air. For simplicity,
but without loss of generality, viscous stresses, heat conduction and external forces are neglected in Eq. (30).
335
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

With the SIMPLE method, velocity and density fields must be corrected to satisfy the mass conservation;
therefore, linearzing the mass flux ρu yields:
(ρ ∗ + ρ ′ )(u ∗ + u ′ ) = ρ ∗ u ∗ + ρ ∗ u ′ + ρ ′ u ∗ + ρ ′ u ′ (33)
where ρ ′ is the density correction. The second order correction term (ρ ′ u ′ ) is usually ignored since it is smaller
than other terms and it does not influence the convergence-rate or final solution field. Using the chain rule and
relation (19), the modified 1-D mass continuity equation together with the weighting factor ∆t ′ defined in Eq. (26),
becomes:
∂ p′ ∗ ∂p
′
′ 2 ∂ p
2 ′
2 ∂ρ ∂ρ ∗ u ∗
( ∗ )
+u − ∆t C = −C + (34)
∂t ∂x ∂x2 ∂t ∂x
representing a convection–diffusion equation. However, the driving force is the time derivative of density and mass
imbalance on the right-hand side of Eq. (34). The parameter C is redefined as a bounded sound speed to identify
the nature of flow field:
C = min C ∗ ; Cac
( )
(35)
where C ∗ is calculated using tentative values, as both momentum and energy equations are solved in the predictor
step of SIMPLE:
√ √
p∗ p∗
( )
1
ρ =
∗
, C = γ ∗ , Cac = β max u 2 ; Ur2
∗
(36)
(γ − 1)e ρ 2
As a thumb-rule, compressible flows preserve the limit of incompressibility when the local Mach number Ma =
|u|/C becomes lower than around 0.3 [46]. Therefore, the magnitude of β is recommended as β ≈ 3.0 for better
convergence to the steady state regarding incompressible flow computations. It is worth pointing out that in some
compressible flow cases, a reference velocity is not available, however, a reference sound speed is attainable; in
this situation it is relevant to set Ur = Cr . In connection with the relation (36), Eq. (35) is based on the assumption
that Ma ≥ 0.3. For 2-D case, the sound speed due to AC, Cac is equivalent to Eq. (20) with β = 3.0.
A finite-volume discretized algebraic relation of Eq. (34) can be approximated as:
[ ( )]
∆t ∗ ∑ ∆t ∗ ∑ ∆ti Ci2 ρi∗ − ρin
( )
∗ ′ ′ ∗
1+ Bnb + ṁ i pi = Bnb pnb − ∀i + Ṁi (37)
∀i nb
∀i nb ∀i ∆ti
where ∆t ∗ is defined earlier and ρ n is the density from the previous iteration cycle. For 1-D case ∆n = ∆x,
S = 1.0, ∀i = ∆x and
∑
Bnb = Bi+1 + Bi−1 ,
⏐ ⏐
ṁ i∗ = ⏐(u ∗ S)i+1/2 − (u ∗ S)i−1/2 ⏐
nb
C 2 ∆t ′ S C 2 ∆t ′ S
( ) ( )
Bi−1 = ∗
+ max[u S, 0] , Bi+1 = + max[−u ∗ S, 0]
∆n i−1/2 ∆n i+1/2

The fictitious mass imbalance Ṁi∗ in Eq. (35) can be given by:
Ṁi∗ = Ṁi+1/2
∗ ∗
− Ṁi−1/2 , ∗
Ṁi±1/2 = (ρ̃ ∗ ũ ∗ S)i±1/2 ,
where the cell-face velocity ũ ∗ uses the modified Rhie–Chow scheme, Eq. (27) and ρ̃ ∗ is the cell-face value
computed using the MUSCL approach with a TVD (total variation diminishing) [14] limiter. The first-order upwing
differencing mass flux ṁ i∗ is added with the left-hand side of Eq. (37) to guarantee diagonal dominance since
volumetric flows in and out are not generally equal.
Note-worthily, popular choices are the minmod slope limiter [45]
φ1 (r ) = max (0, min(1, r ))
which is the most diffusive limiter; it corresponds to the lower boundary of second-order TVD region, and the
Superbee limiter [45]
φ2 (r ) = max (0, min(1, 2r ), min(2, r ))
336
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 1. Admissible limiter region for second-order TVD scheme.

which is the least diffusive limiter; it corresponds to the upper boundary of second-order TVD region. The ratio of
consecutive slopes r gives an indication of the local smoothness of discrete solution. When r < 0, the numerical
solver faces the presence of an extremum and necessarily delimits that anything different from φ = 0 will produce
oscillations. Above all, convergence issues may become a problem when an iterative scheme is used due to abrupt
changes in the slope of φ(r ) function for minmod and superbee limiters. A differentiable/smoother limiter function
can alleviate such problems. A somewhat better choice is obtained with a currently developed cubic slope limiter
as shown in Fig. 1:
( )
2r r 2 + 1
φ3 (r ) = 3 ; lim φ3 (r ) = 2
r +r +2 r →∞

It seems likely that the φ3 (r ) limiter maintains good correspondence with the minmod φ1 (r ) (i.e., 0 ≤ φ ≤ 1)
and superbee φ2 (r ) (i.e., 1 ≤ φ ≤ 2) limiters for 0 ≤ r ≤ 1 and r > 1, respectively. Apparently, φ3 (r ) is a
combination of φ1 (r ) and φ2 (r ) limiters, however, the devised slope limiter is somewhat diffusive when compared
with the superbee limiter (see Fig. 1). Indeed, more diffusivity results in better convergence properties, indicating
an appropriate conservative choice for a limiter function when governing equations are non-linear and being solved
on poor quality meshes. The flux limiter Ψ (r ) for the left and right interfaces can be conveniently given for the
MUSCL interpolation scheme as:
1
Ψ (r ) = [(1 − κ) + r (1 + κ)] φ (r ) ≤ 2 (38)
2
where −1 ≤ κ ≤ 1). Flux limiter functions can be applied when evaluating explicit momentum and energy residuals
with compressible flows.
The primitive explicit residuals can be calculated as
1
∆u = (∆ρu − u∆ρ)
ρ
1 (39)
∆E = (∆ρ E − E∆ρ)
ρ
∆e = ∆E − u∆u
where ∆ρ is the explicit density residual produced by the corrected velocity field (as mentioned earlier) and ∆e is
the explicit residual of specific internal energy e = cv T (i.e., T signifies the absolute temperature). Similar to the
momentum residual, Eq. (10) can be used to smooth the explicit residual ∆e (or ∆T = ∆e/cv ) for the implicit
stage and the specific internal energy can be updated as en+1 = en + ∆e. Tentative variables are updated in the
corrector step of SIMPLE as follows:
∆t ′ ∂ p ′
u = u∗ − ∗ , p = p∗ + p′ (40)
ρ ∂x
337
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353
{
1 if C ∗ ≤ Cac
ρ = ρ ∗ + ρ ′ = ρ ∗ + αρ p ′ /C ∗ 2 , αρ = (41)
0 otherwise
Since the specific internal energy is updated in the predictor step of SIMPLE method, the total internal energy and
total enthalpy can be updated using new values of density, velocity and pressure:
E = e + u 2 /2, H = E + p/ρ (42)
Thermodynamic properties can also be renewed using an EOS at this stage. Evidently, αρ = 0 in Eq. (41) implies
that the flow is weakly compressible or nearly incompressible with ρ = ρr (e.g., preferably constant or could be kept
variable); an application of this constraint to Eq. (37) sets the time derivative of density and convection term to zero,
resulting in a recovery of the incompressible SIMPLE-C algorithm. Therefore, the above-mentioned development
is essentially an extension of the SIMPLE-C formulation, entailing to be equally applicable for incompressible and
compressible flows with a symmetric stencil for the pressure gradient [23]. In principle, Eq. (37) evidently works
at all flow speeds without switching off additional compressibility terms; actual numerical coding for 2-D and 3-D
cases pertaining to the current solver is quite simple and straightforward.

5. Incompressible numerical tests

To demonstrate the effectiveness and generality of present formulations, a few applications to laminar flows
consisting of a buoyancy-driven flow in a square cavity and a buoyancy-driven flow in a half-concentric annulus
are addressed. These test cases are probably the most common computational experiments to validate the predictive
performance of a solver [5,22,25,26,29,53]. Since no analytically exact solutions to these problems are available,
literature data [25,26,29,53] and results of control-volume-based finite-element method (CVFEM) with vorticity
stream function formulation [21,36] are considered as standards for estimating the accuracy of computations
obtained by proposed methods. It is believed that a realistic assessment between SIMPLE and SIMPLE-C algorithms
for incompressible flows could be achieved based on these numerical experiments, containing both highly diffusive
central and convective near-wall flows.
It is worth mentioning that the SIMPLE method is recovered with an exclusion of the first term on the left-
hand side in Eq. (22); however, both SIMPLE and SIMPLE-C algorithms evaluate the pseudo-time step ∆ti from
Eq. (25) and weighting factor ∆ti′ from Eq. (26). An important parameter to judge the convergence of SIMPLE-like
algorithms is the root-mean-square of mass imbalance in the computational domain, defined as:
√
∗ 2
∑N P
i=1 | Ṁi |
M∗ = ≤ ηt (43)
NP
where NP is the number of computational cells and ηt is the user-defined tolerance limit. The choice of a compatible
mesh system for buoyancy-driven cavity flows is obvious since the wall-proximity occupies most of the high-
gradient regions. Consequently, each computational test case employs a non-uniform grid spacing to capture the
sharp gradients in near-wall regions. The maximum possible CFL number is approximated for each computational
grid, enhancing the fastest convergence. Grid densities (not shown) are varied to guarantee the grid independence
of numerical results; numerical errors due to the grid size is evaluated as less than 2%, based on grid dependence
tests for both velocity and temperature fields. Calculations that follow, use the compressibility parameter β = 2.

5.1. Buoyancy-driven flow in square cavity

Fig. 2 shows a schematic diagram of the natural convection in a square cavity. The system has a characteristic
dimension of L with insulated top and bottom walls. Both hot and cold vertical walls are isothermal having
temperatures Th and Tc , respectively. A non-dimensional form of governing equations for two-dimensional case
can be obtained from Reference [37]. Two dimensionless parameters dominating the test problem, spring from the
normalization: Grashof number, Gr = gγ (Th − Tc )L 3 /ν 2 (ratio of buoyancy to viscous forces, controlling natural
convection) and Prandtl number, Pr = ν/α (ratio of momentum to thermal diffusivity, governing temperature
field and its relationship with fluid flow characteristics); quantities g, α and γ are the gravitational acceleration,
thermal diffusivity and thermal-expansion coefficient, respectively. Present computations set Pr to 1.0 and results
are plotted in the form of V = vL/ν and θ = (T − Tc )/(Th − Tc ) versus X = x/L. A dimensional reference velocity
338
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 2. Physical model of buoyancy-driven cavity flow.

√ √
can be formulated√as Ψ = gγ (Th − Tc )L = ν Gr /L, yielding a dimensionless reference velocity amounted to
Ur e f = Ψ L/ν = Gr . Another non-dimensional important parameter for this buoyancy-driven cavity flow is the
Rayleigh number Ra defined by Ra = Gr Pr ; however, Ra = Gr since Pr = 1.0.
No-slip boundary conditions (U = V = 0) are applied to all walls. Temperature field enforces Dirichlet boundary
conditions θh = 1.0 and θc = 0.0 on the left and right vertical walls, respectively. A Neumann boundary condition
∂θ/∂Y is imposed on bottom and top horizontal walls. In principle, the buoyancy-driven flow in a plane square
cavity with the clarity in boundary conditions, highly convective near-wall and diffusive central regions serves as a
relevant test-case to validate the competency of new computational algorithms.
A 40 × 40 non-uniform grid for Ra = 105 , 50 × 50 for Ra = 106 , 60 × 60 for Ra = 107 and 80 × 80 for
Ra = 108 , having finer grid resolutions in the near-wall region than in the core, are employed for computations.
These grid arrangements are presumably to be adequate for producing consistent and reliable simulations by the
present numerical formulations. The use of literature data [25,26,29,53] and CVFEM [21,36] solutions further
supports the current investigations. According to Reference [15], the laminar-flow region of natural convection at
a local Ra ≥ 109 may promote the transition from laminar to turbulence in the vertical boundary layer; therefore,
present simulations are limited to Ra = 108 .
Plots of vertical velocity and temperature profiles at the horizontal centerline of cavity are presented at different
Ra numbers in Figs. 3 and 4, respectively. Plots also include semi-implicit solutions from a characteristic-based
split algorithm by Massarrotti et al. [29] together with CVFEM computational data. As is observed, comparisons
exhibit an encouraging qualitative agreement. The velocity and temperature distributions indicate that the boundary
layers move more closer to vertical walls with an increase in Ra number; this is a physically expected phenomenon
and both algorithms produce identical results.
Fig. 5 elucidates local Nusselt number distributions along the hot vertical wall for various Rayleigh numbers;
results of CVFEM [21,36] are also included for comparisons. A careful inspection shows that both SIMPLE and
SIMPLE-C replicate analogous results and they maintain a fairly good agreement with CVFEM computations. A
slight variation in the profiles is observed due to the differences in methodological phenomena; this aspect could
be naturally resolved to an extent using a finer grid-resolution at the close proximity of wall.
Fig. 6 displays the convergence histories of fictitious mass residuals with buoyancy-driven flows for different Ra
numbers, starting from the same initial conditions. The variation in mass imbalance M ∗ against iteration counter
is regarded as the convergence history. The trends regarding the performances with CFL numbers indicate that the
SIMPLE-C outperforms the SIMPLE with increasing Ra number. Interestingly, the buoyancy-driven cavity flow
339
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 3. Vertical velocity profile on horizontal midplane for buoyancy-driven cavity flow.

compromises almost twofold CFL number with the SIMPLE-C relative to the SIMPLE; this is due to the fact that
the former algorithm incorporates an additional AC term in the pressure-correction equation, enhancing the diagonal
dominance of influence nodal coefficients associated with the discretized pressure-correction equation.
Quantitative comparisons are illustrated in Table 1, representing the magnitudes of maximum horizontal Umax
and vertical Vmax velocities at the mid-sections and average Nusselt number on the hot wall of buoyancy-driven
cavity flow. As can be noticed, both SIMPLE and SIMPLE-C algorithms have a reasonable agreement with
available numerical data from the literature [25,26,29,53]. Such a tabular comparison appears to be encouraging
for enthusiastic researchers, reflecting a good congruence among all six independent numerical schemes. In
addition, discrepancies in predictions adhering to SIMPLE and SIMPLE-C models are indistinguishable. Tabular
investigations advocate that these schemes can be faithfully used to obtain a reliable simulation of industrial-scale
fluid flow and heat transfer problems.

5.2. Buoyancy-driven flow in an annulus

The proposed scheme is further validated against a buoyancy-driven flow in a half-concentric annulus shown
schematically in Fig. 7; the radii of inner and outer surfaces are Ri and Ro , respectively, with (Ro /Ri ) = 3. The
Cartesian-coordinate origin of the system is located at the center of circles with a characteristic length L, setting
to (Ro − Ri ) having insulated straight top and bottom walls. The inner curved surface is hot with a temperature of
Th and the outer one is cold with a temperature of Tc . Governing equations and boundary conditions are identical
to those of the buoyancy-driven flow in a square cavity, containing both highly diffusive central and convective
340
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 4. Temperature profiles on horizontal midplane for buoyancy-driven cavity flow.

Table 1
Buoyancy-driven cavity flow: comparisons of average Nusselt number, maximum horizontal and
vertical velocity components at mid-sections (X = 0.5 and Y = 0.5) with literature data.
. Ref. [25] Ref. [53] Ref. [26] Ref. [29] SIMPLE SIMPLE-C
Ra = 105
Umax 34.0 33.39 34.95 33.18 33.64 33.64
Vmax 68.64 70.81 68.66 70.33 69.38 69.38
Nu 4.52 4.60 4.57 4.63 4.78 4.78
Ra = 106
Umax 64.83 65.40 65.08 65.30 64.98 64.98
Vmax 220.60 228.05 222.70 226.6 221.70 221.70
Nu 8.83 8.98 9.02 9.11 9.44 9.44
Ra = 107
Umax 148.80 143.56 141.20 143.02 152.0 152.0
Vmax 699.30 720.54 706.30 706.91 687.50 687.50
Nu 16.50 16.66 16.80 17.03 17.74 17.74
Ra = 108
Umax −− 296.71 246.80 289.48 300.80 300.80
Vmax −− 2291.05 1935.0 2202.87 2207.0 2207.0
Nu −− 31.49 31.0 30.71 32.11 32.11

341
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 5. Local Nusselt number along hot wall for buoyancy-driven cavity flow.

near-wall flow regimes. From the literature review it can be concluded that benchmark data are not available
for this test-case; however, the flow geometry demonstrates the application of present schemes on a curvilinear
grid. Additional supports are received by the results of CVFEM. For this configuration, the buoyancy-driven flow
is stable at low Ra numbers but unstable at a higher Ra number. The flow transits from laminar to turbulence
when Ra ≥ 5 × 106 with Pr = 1.0 [58]; therefore, the present investigation is confined to Ra = 106 at
Pr = 1.0.
The non-uniform computational grid used in this test-case is composed of 40 radial and 30 circumferential
line segments; this grid refinement is considered to be sufficiently accurate to describe the flow and transport
characteristics. Figs. 8(a) and 8(b) depict the vertical velocity and temperature profiles, respectively, at the horizontal
midplane for Ra = 106 ; X is counted exactly from the inner surface. Results of CVFEM are also plotted
for comparisons. Remarkably, the correspondence of both algorithms with the CVFEM is quite satisfactory and
reproduces fairly good agreement with CVFEM data; both SIMPLE and SIMPLE-C provide an identical solution.
Articulately, local Nusselt number distributions in Fig. 8(c) compare favorably with CVFEM data. Nevertheless,
some discrepancies between both algorithms and CVFEM appear at the lower portion of inner wall. The probable
inconsistency is due to the reasoning that the CVFEM estimates the temperature by computing the weighted nodal
temperature. This incongruity may be reduced to some extent using a finer grid resolution and a higher-order
formulation with current methods. Fig. 8(d) represents the convergence history of mass imbalance; it is once more
ascertained that the performance of SIMPLE-C is better than that of SIMPLE in smoothing mass residuals.
342
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 6. Convergence of mass imbalance for buoyancy-driven cavity flow.

Fig. 7. Physical model of buoyancy-driven flow in annulus.

343
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 8. Buoyancy-driven flow in annulus: (a) velocity profiles on horizontal midplane; (b) temperature profiles on horizontal midplane; (c)
local Nusselt number along hot wall and (d) convergence of mass residuals.

6. Compressible numerical experiment

Three 1-D test cases, namely Sod [47], Lax [24] and Mach 3 shock [4] problems , focusing on highly non-
linear compressible flows with shocks, contacts and rarefactions, are considered to analyze the performance of
compressible SIMPLE solver. Due to their conceptual simplicity and well-defined theoretical solutions, they are
frequently used for validation and comparison of a numerical method. Traditionally, these types of problems are
generally solved with an explicit time integration and the explicit component is an upwind scheme. Air is used
as the working fluid, covering the Mach number Ma = |u|/C in the range 0.9 ≥ Ma ≤ 3 with γ = 1.4 and
cv = 718 J kg −1 K −1 . Computations are performed with equidistant 400 cells and a CFL number of C F L ≤ 0.2
based on the maximum magnitude of flow velocity for 1-D Euler equations. An explicit first-order backward Euler
method is applied and inviscid fluxes are calculated using the TVD MUSCL scheme. The weighting factor ∆t ′ = ∆t
in the SIMPLE algorithm, indicating that cell-face dissipation (with damping factor ζ = 1.0) and influence nodal
coefficients are governed by a constant explicit time-step ∆t. In the present investigation, the limited MUSCL
scheme uses κ = −1 such that a fully upwinded second-order difference is achieved and at the same time,
an equivalence exists between slope and flux limiters. Since considered tests cases are solved using an explicit
predictor step and an implicit corrector step in the framework of SIMPLE algorithm, the overall contrivance can be
designated as a semi-implicit method. In addition, compressible 2-D viscosity-driven cavity flow and flow around
an NACA-0012 airfoil are computed to investigate the efficacy of proposed compressible SIMPLE solver.
344
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 9. Solutions to Sod’s shock tube problem after 0.1 s with TVD MUSCL scheme for different limiters: (a) density; (b) velocity; (c)
pressure and (d) total internal energy.

6.1. Sod’s shock tube

The performance of proposed compressible flow variant of SIMPLE algorithm is contrasted with the well-known
Sod’s shock tube problem [47], representing the Riemann problem with discontinuous initial conditions given by
if 0.0 < x ≤ 0.5
{
(1.0, 0.0, 1.0)
(ρ, u, p)t=0 =
(0.125, 0.0, 0.1) if 0.5 < x < 1.0
The solution is evolved over the interval x ∈ [0, 1], from t = 0 to t = 0.1. Dirichlet boundary conditions
(i.e., primitive quantities from specified initial conditions) are imposed at either boundary.
The considered shock tube features a shock wave, a rarefaction fan and a contact discontinuity and hence,
provides a comprehensive test-case to validate the predictive performance of a TVD solver in conjunction with an
implicit SIMPLE pressure-correction method. To evaluate the reliability and accuracy of the present formulation,
computational results are compared with those obtained from exact solutions of the shock-tube problem.
Fig. 9 illustrates the performance of TVD MUSCL scheme and its adaptive dissipative actions in capturing
accurately shocks and other discontinuities without oscillations for the shock-tube problem. In the framework of
SIMPLE algorithm together with flux limiters minmod, superbee and cubic (e.g, developed in this research), it
345
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 10. Solutions to Lax’s shock tube problem after 0.12 s with TVD MUSCL scheme for different limiters: (a) density; (b) velocity; (c)
pressure and (d) total internal energy.

is seen that the shock is sharply captured, the contact discontinuity is better resolved and no oscillations are
produced. In fact, the TVD limiter functions provide insight into the manner by which the present solver is
susceptible to controlling the degree of biasing the dissipation. Computational results compare very favorably with
the exact solution and are particularly striking when compared to those obtained with the superbee limiter. A careful
investigation shows that a slight difference (naturally at contact discontinuities) in the predicted profiles of Sod’s
parameters is observed among different flux limiters. Therefore, it is a rather difficult task to judge the superiority of
one limiter function to another for this test case. However, the satisfaction is that each of them when embedded with
the current solver, provides a marked improvement in producing oscillation-free solutions. Results are comparable
with References [19,56] for a given mesh resolution and time step.

6.2. Lax’s shock tube

The performance of proposed solver is further validated against the well-known Lax’s shock tube problem [24]
which has similar characteristics to those of Sod’s problem, except that the initial condition has a discontinuity in
the velocity field; Lax’s shock tube problem considers the Riemann problem with following initial data:
(0.445, 0.698, 3.528) if 0.0 < x ≤ 0.5
{
(ρ, u, p)t=0 =
(0.5, 0.0, 0.571) if 0.5 < x < 1.0
Fig. 10 demonstrates well resolved shock and contact solutions. It seems likely that the physical process functioning
herein is the advection which generates a linear contact wave with a discontinuous jump in density. Articulately,
346
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 11. Solutions to Mach 3 shock problem after 0.09 s with TVD MUSCL scheme for different limiters: (a) density; (b) velocity; (c)
pressure and (d) total internal energy.

the present solver shows a merit for handling this type of linear contact waves. Three independent limiters produce
identical solutions and fairly comparable with literature data [19,56] for the same grid resolution.

6.3. Mach 3 shock

The competency of proposed solver is again validated against a Mach 3 shock tube problem which is
characterized by a strong expansion fan containing sonic point. The configuration is created by the following initial
conditions:
(3.857, 0.92, 10.333) if 0.0 < x ≤ 0.5
{
(ρ, u, p)t=0 =
(1.0, 3.55, 1.0) if 0.5 < x < 1.0
Numerical results shown in Fig. 11 dictate that a well-resolved expansion wave is obtained. The right moving shock
recognized by a small jump in this system, is also visible in the numerical solution. However, it can be observed that
the superbee limiter scheme produces some tiny oscillations in the velocity field (e.g., after the end of expansion
wave) probably due to its less diffusive nature when compared with that of minmod and cubic limiters as mentioned
earlier. It can be ensured that present computations are compatible to those of [19,56].

6.4. Compressible viscosity-driven cavity flow

The 2-D test problem is shown schematically in Reference [39] with the top wall moving to the right at a
velocity û r while the three sides are at rest. The viscosity-driven cavity flow problem is a typical low-Mach
347
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

number test-case to justify the Mach-uniform efficiency of a compressible solver. The cavity is assumed to be
filled with air and the flow-field inside the structure is isothermal. Selected reference values are: ρ̂r = 1.18 kg/m3 ,
µ̂r = 18.46 × 10−6 N-s/m2 , T̂r = 300 K, Pr = 0.71 and L̂ = 0.001 m. The dimensional form of compressible
equations is not favorable because the disparity of physical variables may introduce undesirable errors. Therefore,
all variables are non-dimensionalized using dimensionless parameters (dimensional quantities are indicated by a
‘hat’ sign), defined as:
x̂ ŷ û v̂ tˆû r
x = ; y = ; u = ; v = ; t =
L̂ L̂ û r û r L̂
(44)
ρ̂ µ̂ T̂ µ̂r R̂ T̂r
ρ = ; µ = ; T = ; Pr = ; R =
ρ̂r µ̂r T̂r ρ̂r α̂r û r2
where α̂r is the thermal diffusivity. The main dimensionless group evolving is the Reynolds number Re = ρ̂r û r L
/µ̂r ; the reference velocity û r can be determined from a given Re. The non-dimensional pressure and other
thermodynamic parameters can be evaluated using the EOS p = ρ RT and cv = R/(γ − 1). The governing
non-dimensional equations are given as:
1 ( vis
Wt + Fxinv + G inv F + G vis
)
y = (45)
Re x y

where W = (ρ, ρu, ρv, ρ E)T . The inviscid fluxes are:

ρu ρv
⎛ ⎞ ⎛ ⎞
⎜ ρu 2 + p ⎟ ⎜ ρuv ⎟
F inv = ⎜ ⎟, G inv
= (46)
⎝ ρuv ⎠ ⎝ ρv 2 + p ⎠
⎜ ⎟

ρu H ρv H
The viscous fluxes can be given as:
⎛ ⎞ ⎛ ⎞
0 0
τ τ12
F vis = ⎜ 11 ⎟, G vis
⎜ ⎟ ⎜ ⎟
=⎜ (47)
τ12 τ22
⎟
⎝ ⎠ ⎝ ⎠
uτ11 + vτ12 + c p Tx /Pr uτ12 + vτ22 + c p Ty /Pr
with
τ11 = 2[u x − (u x + v y )/3], τ22 = 2[vx − (u x + v y )/3], τ12 = vx + u y (48)
The velocity boundary conditions are u = v = 0 on the bottom and two vertical walls, and Ur = u w = 1.0, v = 0
on the top wall. Since the temperatures of moving and stationary walls are the same as the initial temperature, the
temperature boundary conditions on all walls can be set to Tw = 1.0 (or adiabatic boundary conditions can be
imposed on all walls). The viscosity µ = 1/Re, thermal conductivity Γ = µc p /Pr and specific heats of the gas
are assumed to be independent of temperature; the density ρ is computed from an EOS.
A 32 × 32 non-uniform grid for Re = 400, 40 × 40 for Re = 1000, 60 × 60 for Re = 5000 and
80 × 80 for Re = 10000 with finer grid points near the walls than in the core, are employed for computations. A
comprehensive grid resolution and convergence study have been conducted; however, are not shown here for brevity.
This range of flows is chosen in order to evaluate the performance of proposed scheme on flow regimes ranging
from highly diffusive to convection dominated. Inviscid fluxes are calculated using a fully upwinded second-order
scheme without a flux limiter and β = 3 is used. This test-case encloses reference Mach numbers in the range
1.8 × 10−2 < Ma < 0.5.
Plots of horizontal velocity profiles at the vertical centerline of cavity are presented at different Re’s in Fig. 12.
Results of Ghia et al. [13] and Malan et al. [26] for incompressible flows are also included for comparisons. As
is observed, the agreement of SIMPLE solutions with those of Ghia et al. and Malan et al. is quite striking at
Re = (400, 1000, 5000) with Ma = (0.0181, 0.046, 0.23) in the incompressible limit. However, discrepancies in
predictions start to appear (especially at the upper part of vertical centerline) between incompressible (e.g., Ghia
et al. and Malan et al.) and compressible (e.g., SIMPLE) solutions at Re = 10000 with Ma = 0.452. It is once
more ascertained that the flow transits from incompressible to compressible when Ma ≥ 0.3, predicted by the
thumb-rule as mentioned earlier.
348
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 12. Velocity profiles on vertical midplane for viscosity-driven cavity flow.

Fig. 13 displays the convergence histories of mass residuals with viscosity-driven flows for different Re numbers
(with previously mentioned grids), starting from the same initial conditions. Numerical computations recommend
CFL numbers such that the ’best convergence’ could be obtained. The trends regarding the performances with CFL
numbers indicate that retaining the compressibility term in the pressure-correction equation accommodates better
convergence in all computations with a higher CFL. It is worth mentioning that the pseudo-time ∆ti parameterized
with diffusive, convective and AC-generated sound speeds further assists in using a higher CFL number when the
full compressible solver is used to compute incompressible flow.

6.5. Compressible NACA-0012 airfoil

To further evaluate the performance, the flow over a well-documented 2-D NACA-0012 airfoil at a zero angle
of attack (AOA) with a free-stream Mach number of Ma = 0.5 and a Reynolds number of Re = 5000 [30] is
simulated. As the NACA-0012 airfoil is symmetrical, the theoretical lift at AO A = 0 is zero. The free-stream
temperature is 300 K; the thermally insulated wall boundary condition is applied at the airfoil surface. The Re for
this test-case can be labeled as the upper limit of steady laminar flows prior to the onset of turbulence [30]. The
representative O-type structured grid is displayed in Fig. 14(a); the grid size is 126 × 92 and heavily clustered
near the solid wall. The utilized O-type grid avoids rapid changes in the spacing of grid-cell near the solid surface.
This study reveals that a non-uniform O-type 126 × 92 grid would be sufficient to establish a grid independent
solution (not shown). The far-field boundary is located 15 chord lengths out from the airfoil. Governing equations
and reference conditions are identical to those of the compressible viscosity-driven flow in a square cavity. The
349
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 13. Convergence of mass imbalance for viscosity-driven cavity flow.

thermal conductivity Γ = µc p /Pr = µ(γ − 1)−1 Ma −2 /Pr with viscosity µ = 1/Re; a fully second-order upwind
scheme without the flux limiter is used to calculate the inviscid fluxes.
The pressure distribution around the airfoil is an important parameter, both in terms of aerodynamic charac-
teristics (e.g., lift and drag coefficients of airfoil) and in terms of boundary layer behavior since the pressure
gradient is known to affect the development of boundary layers. Fig. 14(b) represents the pressure coefficient C p
curves of the airfoil, plotted against the non-dimensional chord length x/c; surface pressure coefficients match the
literature data [30] very well. Fig. 14(c) exhibits the convergence histories of mass residuals with various CFL
values. Evidently, faster convergence is achieved at a higher CFL number; the smooth convergence unambiguously
confirms the conclusion that the present compressible SIMPLE method works effectively toward eliminating the
non-physical oscillations.

7. Conclusions
A comparative assessment is conducted between SIMPLE and SIMPLE-C algorithms within the framework of a
cell-centered finite-volume ∆-approximation on collocated grids for incompressible flows. The SIMPLE-C includes
an AC parameter with the pressure Poisson equation, amplifying the diagonal dominance of influence coefficients.
Relevant phenomena associated with both algorithms alleviate the need for UR factor even in the presence of
dominant source terms in momentum equations. Both SIMPLE and SIMPLE-C schemes have been additionally
sensitized to enhanced convergence and robustness by including the pseudo-time step ∆ti , parameterized with
diffusive, convection and artificial sound speeds. Results demonstrate that both methods compare favorably with
literature and CVFEM data for a specified level of grid resolution. Compared with the SIMPLE formula, flow
quantities namely the velocity, temperature and heat transfer-rate remain almost unaffected by the additional of AC
term embedded with the SIMPLE-C; however, convergence studies essentially reveal that the SIMPLE-C scheme
is benefited with enhanced convergence and robustness characteristics when compared with the SIMPLE.
The capability of SIMPLE-C algorithm is extended toward predicting fluid flow at all speeds. Furthermore, a
new cubic limiter function is developed which inherits the characteristics of both minmod and superbee limiters
in giving rise to oscillation-free numerical solutions. To substantiate the validity of the proposed SIMPLE solver
accompanied by the limited MUSCL scheme, computations have been carried out for 1-D Euler equations with the
well-known shock-tube problems having sharp discontinuities. The current formulation compares favorably with
the analytical solutions. In addition, 2-D compressible viscosity-driven flow and flow around an NACA-0012 airfoil
350
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

Fig. 14. NACA-0012 airfoil: (a) computational structured grid; (b) surface pressure coefficient and (c) convergence of mass imbalance.

are calculated to gain more confidence in the methodology. Computations agree well with the literature data in
the incompressible limit where the full compressible equations are solved without switching off the compressible
SIMPLE algorithm to the incompressible one. Obviously, the present compressible variant of SIMPLE procedure
can readily be extended to the 3-D case and is expected to be stable for a wide range of compressible flows in all
Mach number regimes without the need for UR factor.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could
have appeared to influence the work reported in this paper.

Acknowledgments
We are grateful to the National Natural Science Foundation of China (No. 51709070), Natural Science Foundation
of Zhejiang Province (LQ17E090007), Key R& D Program of Zhejiang Province (2018C04002).

References
[1] S. Acharya, B.R. Baliga, K. Karki, J.Y. Murthy, C. Prakash, S.P. Vanka, Pressure–based finite–volume methods in computational fluid
dynamics, J. Heat Transfer 129 (2007) 407–424.
351
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

[2] S. Acharya, F. Moukalled, Improvements to incompressible flow calculation on a non-staggered curvilinear grid, Numer. Heat Transf.
B 15 (1989) 131–152.
[3] S.W. Armfield, Finite difference solutions of the Navier–Stokes equations on staggered and nonstaggered Grids, Comput. & Fluids 20
(1991) 1–17.
[4] M. Arora, P. Roe, A well-behaved TVD limiter for high resolution calculations of unsteady flow, J. Comput. Phys. 132 (1) (1997)
3–11.
[5] A. Castrejon, D.B. Spalding, Experimental and theoretical study of transient free-convection flow between horizontal convection
cylinders, Int. J. Heat Mass Transfer 31 (1998) 273–284.
[6] Z. Chen, A.J. Przekwas, A coupled pressure-based computational method for incompressible/compressible flows, J. Comput. Phys. 229
(2010) 9150–9165.
[7] K.-H. Chen, Pletcher R., Primitive variable, strongly implicit calculation procedure for viscous flows at all speeds, AIAA J. 29 (1991)
1241–1249.
[8] Y.P. Cheng, T.S. Lee, H.T. Low, W.Q. Tao, Improvement of SIMPLER algorithm for incompressible flow on collocated grid system,
Numer. Heat Transf. B 51 (5) (2007) 463–486.
[9] A.J. Chorin, A numerieal method for solving incompressible viscous flow problems, J. Comput. Phys. 2 (1967) 12–26.
[10] I. Demirzi, V. Lilek, M. Peri, A collocated finite volume method for predicting flows at all speeds, Internat. J. Numer. Methods Fluids
16 (1993) 1029–1050.
[11] F. Dnner, Fully-coupled pressure-based algorithm for compressible flows: linearisation and iterative solution strategies, Comput. &
Fluids 175 (2018) 53–65.
[12] J.P. van Doormaal, G.D. Raithby, Enhancement of the SIMPLE method for predicting incompressible fluid flow, Numer. Heat Transfer
7 (1984) 147–163.
[13] U. Ghia, K. Ghia, C. Shin, High-Re solutions for incompressible flow using Navier–Stokes equations and a multigrid method, J.
Comput. Phys. 48 (1996) 387–411.
[14] A. Harten, High resolution schemes for hyperbolic conservation laws, J. Comput. Phys. 135 (1997) 260–278, 1997.
[15] F.P. Incropera, D.P. DeWitt, Fundamentals of Heat and Mass Transfer, forth ed., Wiley, New York, 1996.
[16] R.I. Issa, Solution of the implicit discretized fluid flow equations by operator splitting, J. Comput. Phys. 62 (1) (1986) 40–65.
[17] R.I. Issa, A.D. Gosman, A.P. Watkins, The computation of compressible and incompressible recirculating flows by a non-iterative
scheme, J. Comput. Phys. 62 (1986) 66–82.
[18] P. Johansson, L. Davidson, Modified collocated SIMPLEC algorithm applied to buoyancy-affected turbulent flow using a multigrid
solution procedure, Numer. Heat Transf. B 28 (1995) 39–57.
[19] S.Y. Kadioglu, M. Sussman, M.S. Osher, J. Wright, M. Kang, A second order primitive preconditioner for solving all speed multi-phase
flows, J. Comput. Phys. 209 (2005) 477–503.
[20] K.C. Karki, S.V. Patankar, Pressure based calculation procedure for viscous flows at all speeds in arbitrary configurations, AIAA J.
27 (1989) 1167–1174.
[21] C.F. Kettleborough, S.R. Husain, C. Prakash, Solution of fluid flow problems with the vorticity-stream function formulation and the
control volume based finite element method, Numer. Heat Transf. B 16 (1989) 31–58.
[22] T.H. Kuehn, R.J. Goldstein, Experimental and theoretical study of natural convection in the annulus between horizontal convection
cylinders, J. Fluid Mech. 74 (1976) 695–719.
[23] J.W. Kurzrock, Exact Numerical Solution of the Time-Dependent Compressible Navier–Stokes Equations (Ph.D. thesis), Cornell
University, Ithaca, New York, 1966.
[24] P.D. Lax, Weak solutions of nonlinear hyperbolic equations and their numerical approximations, Comm. Pure Appl. Math. 7 (1965)
159–193.
[25] P. Le Quere, T. Alziary de Roquefort, Computation on natural convection in two-dimensional cavities with Chebyshev polynomial, J.
Comput. Phys. 57 (1985) 210–228.
[26] A.G. Malan, R.W. Lewis, P. Nithiarasu, An improved unsteady, unstructured, artificial compressibility, finite volume method for viscous
incompressible flows: part II. Application, Internat. J. Numer. Methods Engrg. 54 (2002) 715–729.
[27] C.R. Maliska, G.D. Raithby, Calculating 3-D fluid Flows Using non-orthogonal Grid, in: Proc. Third Int. Conf. on Numerical Methods
in Laminar and Turbulent Flows, Seattle, 1983, pp. 656–666.
[28] N.C. Markatos, K.A. Pericleous, Laminar and turbulent natural convection in an enclosed cavity, Int. J. Heat Mass Transf. 27 (1985)
755–772.
[29] N. Massarrotti, F. Arpino, R.W. Lewis, P. Nithiarasu, Explicit and semi-implicit CBS procedures for incompressible viscous flows,
Internat. J. Numer. Methods Engrg. 66 (2006) 1618–1640.
[30] D.J. Mavriplis, A. Jameson, Multigrid olution of the Navier–Stokes equations on triangular meshes, AIAA J. 28 (8) (1990) 1415–1424,
http://dx.doi.org/10.2514/3.25233.
[31] F. Moukalled, M. Darwish, A unified formulation of the segregated class of algorithms for fluid flow at all speeds, Numer. Heat Transf.
B 37 (2000) 103–139.
[32] S.V. Patankar, Numerical Heat Transfer and Fluid Flow, Mc Graw-Hill, New York, 1980, pp. 131–134.
[33] S.V. Patankar, A calculation procedure for two-dimensional elliptic situations, Numer. Heat Transfer 4 (1981) 409–425.
[34] S.V. Patankar, D.B. Spalding, A calculation procedure for heat, mass and momentum transfer in three–dimensional parabolic flows,
Int. J. Heat Mass Transfer 15 (1972) 1787–1806.
[35] M. Peric, Analysis of pressure velocity coupling on nonorthogonal grids, Numer. Heat Transf. B 17 (1990) 63–82.
[36] M.M. Rahman, M.J. Lampinen, Numerical study of natural convection from a vertical surface due to combined buoyancies, Numer.
Heat Transf. A 28 (1995) 409–429.
352
J. Qin, H. Pan, M.M. Rahman et al. Mathematics and Computers in Simulation 180 (2021) 328–353

[37] M.M. Rahman, A. Miettinen, T. Siikonen, Modified SIMPLE formulation on a collocated grid with an assessment of the simplified
quick scheme, Numer. Heat Transf. B 30 (1996) 291–314.
[38] M.M. Rahman, P. Rautaheimo, T. Siikonen, Numerical Study of turbulent heat transfer from a confined impinging jet using a pseudo-
compressibility method, in: 2nd International Symposium on Turbulence, Heat and Mass Transfer, Delft, The Netherlands, 1997,
511–520.
[39] M.M. Rahman, T. Siikonen, An improved SIMPLE method on a collocated grid, Numer. Heat Transf. B 38 (2000) 177–201.
[40] M.M. Rahman, T. Siikonen, An artificial compressibility method for incompressible flows, Numer. Heat Transf. B 40 (2001) 391–409.
[41] M.M. Rahman, T. Siikonen, A dual-dissipation scheme for pressure-velocity coupling, Numer. Heat Transf. B 42 (2002) 231–242.
[42] M.M. Rahman, T. Siikonen, An artificial compressibility method for viscous incompressible and low Mach number flows, Int. J. Numer.
Method Eng. 75 (2008) 1320–1340.
[43] M.M. Rahman, T. Siikonen, A. Miettinen, A pressure correction method for solving fluid Flow Problems on a Collocated Grid, Numer.
Heat Transf. B 32 (1997) 63–84.
[44] C.M. Rhie, W.L. Chow, Numerical study of the turbulent flow past an airfoil with trailing edge separation, AIAA J. 21 (1983)
1525–1532.
[45] P.L. Roe, Some Contributions To the Modeling of Discontinuous Flows, in: Lect. Notes Appl. Math., vol. 22, 1985, pp. 162–193.
[46] S. Roller, C.D. Munz, A low Mach number scheme based on multi-scale asymptotics, Comput. Vis. Sci. 3 (2000) 85–91.
[47] G.A. Sod, A survey of several finite difference methods for systems of nonlinear hyperbolic conservation laws, J. Comput. Phys. 27
(1978) 1–31.
[48] D.B. Spalding, A general purpose computer program for multi-dimensional one-and-two phase flow, Math. Comput. Simul. 23 (1981)
267–276.
[49] E. Turkel, R. Radespiel, N. Kroll, Assessment of preconditioning methods for Multi-dimensional Aerodynamics, Comput. & Fluids 26
(1997) 613–634.
[50] J.P. Van Doormaal, G.D. Raithby, An Evaluation of the segregated approach for predicting incompressible fluid flows, in: ASME Paper
85-HT-9, Presented at the National Heat Transfer Conference, Denver, Colorado, 1985, pp. 4–7, August.
[51] J. Van Doormaal, G. Raithby, McDonald B., The segregated approach to predicting viscous compressible fluid flows, ASME J.
Turbomach. 109 (1987) 268–277.
[52] W. W. Shyy, A study of finite difference approximations to steady-state convection dominated flow problems, J. Comput. Phys. 15
(1985) 415–438.
[53] D.C. Wan, B.S.V. Patnaik, G.W. Wei, A new benchmark quality solution for the buoyancy driven cavity by discrete singular convolution,
Numer. Heat Transf. B 40 (2001) 199–228.
[54] J.M. Weiss, Smith W.A., Preconditioning applied to variable and constant density flows, AIAA J. 33 (1995) 2050–2057.
[55] X. Wen, D.B. Ingham, A new method for accelerating the rate of convergence of SIMPLE-like algorithm, Int. J. Numer. Methods
Fluids 17 (1993) 385–400.
[56] F. Xiao, Unified formulation for compressible and incompressible flows by using multi-integrated moments I: one-dimensional inviscid
compressible flow, J. Comput. Phys. 195 (2004) 629–5654.
[57] C.-N. Xiao, F. Denner, B. van Wachem, Fully–coupled pressure–based finite–volume framework for the simulation of fluid flows at
all speeds in complex geometries, J. Comput. Phys. 346 (2017) 91–130.
[58] X. Yang, S.-C. Kong, Numerical study of natural convection in a horizontal concentric annulus using smoothed particle hydrodynamics,
Eng. Anal. Bound. Elem. 102 (2019) 11–20.

353