Journal of Computational Physics 400 (2020) 108890

Contents lists available at ScienceDirect

Journal of Computational Physics

www.elsevier.com/locate/jcp

A finite difference discretization method for heat and mass

transfer with Robin boundary conditions on irregular domains
Min Chai, Kun Luo ∗ , Changxiao Shao, Haiou Wang, Jianren Fan
State Key Laboratory of Clean Energy Utilization, Zhejiang University, Hangzhou 310027, PR China

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

Article history: This paper proposes a finite difference discretization method for simulations of heat
Received 14 February 2019 and mass transfer with Robin boundary conditions on irregular domains. The level set
Received in revised form 9 August 2019 method is utilized to implicitly capture the irregular evolving interface, and the ghost
Accepted 9 August 2019
fluid method to address variable discontinuities on the interface. Special care has been
Available online 16 August 2019
devoted to providing ghost values that are restricted by the Robin boundary conditions.
Keywords: Specifically, it is done in two steps: 1) calculate the normal derivative in cells adjacent
Robin boundary condition to the interface by reconstructing a linear polynomial system; 2) successively extrapolate
Finite difference scheme the normal derivative and the ghost value in the normal direction using a linear partial
Irregular domain differential equation approach. This method produces second-order accurate solutions
Level set method for both the Poisson and heat equations with Robin boundary conditions, and first-
Ghost value order accurate solutions for the Stefan problems. The solution gradients are of first-order
Stefan problem
accuracy, as expected. It is easy to implement in three-dimensional configurations, and can
be straightforwardly generalized into higher-order variants. The method thus represents
a promising tool for practical heat and mass transfer problems involving Robin boundary
conditions.
© 2019 Elsevier Inc. All rights reserved.

1. Introduction

Heat and mass transfer phenomena on irregular domains or evolving boundaries are common in nature and industry.
In internal combustion engines, for example, the liquid fuel vaporizes when being injected into the combustion chamber
and significant heat transfer and mass transfer occur across the phase interface that undergoes complex topological changes.
Other examples include liquid boiling in daily life, spray cooling in medical industry, solidification in semiconductor industry
and so on. Two aspects need to be considered when solving this kind of problem: 1) the shape of the irregular or evolving
boundaries should be described accurately; and 2) different boundary conditions need to be enforced on the boundaries
(denoted as interface in the remainder of the paper) in different situations.
For irregular or evolving interface, generating boundary-fitted grids becomes difficult to conform to such complex topolo-
gies, and fixed structured grid techniques are more advantageous. In the early time, a δ -function formulation [1,2] was used
in the immersed boundary method (IBM), representing the interface as a diffusive one with a finite thickness. This strategy
is mathematically simple and easy to implement, but it smears out the sharp information near the interface, producing

* Corresponding author.
E-mail address: [email protected] (K. Luo).

https://doi.org/10.1016/j.jcp.2019.108890
0021-9991/© 2019 Elsevier Inc. All rights reserved.
2 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

unphysical continuity on the interface that certainly affects the boundary conditions. For example, the evaporation rate in
the aforementioned spray combustion process greatly depends on the local gradients of temperature and species near the
interface [3]. With such numerical smearing, one may get a wrong evaporation rate. Consequently, sharp interface methods
are more preferable to keep track of the evolving interface, including the front tracking method [4], the volume of fluid
method [5], the phase field method [6] and the level set method (LSM) [7]. Indeed, the combination of LSM and ghost fluid
method (GFM) [8] has proven to be a good and powerful choice for studying physical problems such as incompressible
multiphase flows [9–12] and compressible reacting flows [13] in the last two decades. The LSM implicitly represents the
interface as the zero level set of a continuous function, and can handle complex topological changes automatically. Note that
the discretization stencil for cells near the interface will extend to the other side, thus the local values there are certainly
invalid if variables jump across the interface. The GFM can provide ghost values instead to keep the discretization uniform
as well as to reduce parasitic currents. Besides, it is straightforward to calculate the normal and curvature of the interface
with the LSM, which may be the easiest method for parallelization and extension to three-dimensional (3D) configura-
tions [14]. Indeed, fully parallelized algorithms for the LSM have been developed by Mirzadeh et al. [15] even on adaptive
quadtree/octree Cartesian grids. Moreover, the popular fast sweeping method, which is required in the LSM community to
maintain the signed-distance property of the level set function, has also been parallelized in [16,17]. One main limitation
of the LSM, however, lies in the mass non-conservation issue. As a result, different mass-conserving strategies have been
proposed in the literature, and interested readers please refer to [18–23] and the references therein.
Once the interface is well captured, interactions between the interface and the flow field should be imposed with a
variety of boundary conditions. Many efforts have been devoted to improving the accuracy of the boundary treatments and
extending their applications. Chen et al. [24] proposed a simple LSM for solving Stefan problems with Dirichlet bound-
ary conditions. The method consists of an implicit finite difference scheme for the heat equation on uniform grids and a
quadratic extrapolation for estimating the ghost values to produce second-order accuracy. Similar work can be found in
Udaykumar et al. [25]. It is noteworthy that while benefiting from the implicit definition of the interface, the LSM has diffi-
culty in imposing boundary conditions as they are not straightforwardly imposed on the interface but in the adjacent cells
instead. Luckily, progresses have been made with the help of the GFM for the Dirichlet boundary conditions [26,27], the
Neumann boundary conditions [28] as well as the jump conditions [29,30]. In the case of jump conditions, the GFM provides
a dimension-by-dimension treatment procedure by projecting the conditions into the coordinate directions independently.
However, this treatment only produces first-order accurate solutions and the gradients do not converge in general, making
this treatment inappropriate in problems for which the gradient of the solution has more physical meaning. To fix this
issue, Guittet et al. [31] built a Voronoi tessellation local to the interface. Some recent reviews [32–34] have summarized
the numerical obstacles on how to impose boundary conditions and the state-of-art strategies to address them in the LSM
community.
However, there is little study on how to impose the more general and complicated Robin boundary conditions. The
Robin boundary conditions, known as a linear combination of the Dirichlet and the Neumann boundary conditions, are of
considerable importance to describe the inherent physics and determine the boundary fluxes in simulations involving heat
and mass transfer [35,36]. Despite of the application interests, imposing Robin boundary conditions on irregular evolving
interfaces has been a great challenge in the LSM community until Papac et al. [37] recently made progress in the framework
of finite volume method. By applying the divergence theorem, the area integrals of a diffusive term can be rewritten as the
line integrals of a normal derivative, which can be further simplified by substituting the Robin boundary conditions into the
normal derivative. This approach preserves the matrix symmetry and is efficient and straightforward to implement. Rueda
Villegas et al. [38] have successfully applied this approach to study the Leidenfrost effect. Recently, this approach has been
further generalized to solve problems involving the quadtree/octree adaptive Cartesian grids [39], the piecewise smooth
interfaces [40] and the mixing boundary conditions [41]. However, this approach does not work in the framework of finite
difference method, which represents a popular choice in computational fluid dynamics.
Difficulties of imposing Robin boundary conditions with a finite difference scheme arise in multi-dimensional configura-
tions. For example, the directions in which the boundary conditions work do not coincide with the coordinate directions,
making the implementations quite complex. When the irregular interface evolves, the implementations become more chal-
lenging and expensive. To the authors’ best knowledge, currently, there is no finite difference discretization method in the
LSM community that can effectively deal with the Robin boundary conditions on irregular evolving boundaries. To address
this issue, we present such a finite difference discretization method with the focus on the accuracy and efficiency in this
work. The method leads to second-order accuracy for both the Poisson and heat equations and first-order accuracy for the
Stefan problems, while the solution gradients are first-order accurate. It is straightforward to be generalized into 3D and
higher-order variants. The remainder of this paper is organized as follows. Section 2 presents the governing equations. The
treatments of Robin boundary conditions are presented in Section 3 while numerical validations are performed in Section 4.
In Section 5, the method is applied to more practical heat and mass transfer problems. Finally, some conclusions are drawn
in Section 6.
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 3

Fig. 1. 2D schematics of Cartesian computational domain.

2. Mathematical formulation

2.1. Level set method

Consider a Cartesian computational domain Ω with exterior boundary ∂Ω , separated into disjoint subdomains Ω − and
Ω + by a lower dimensional interface Γ as illustrated in Fig. 1. The level set method is utilized to implicitly describe the
potentially moving interface, and defines the interface as the zero level set of the signed distance function:
 
φ(x) = min |x − xΓ |, (1)

where x corresponds to the target grid location, and xΓ is the closest point on the interface from x. That is to say, φ(x) < 0
in Ω − , φ(x) > 0 in Ω + and φ(x) = 0 on the interface. It is convenient to calculate the normal vector and curvature of the
interface:

∇φ
=
n and κ = −∇ · n. (2)
|∇φ|
The evolution of the interface obeys the following transport equation:

∂φ
+ u · ∇φ = 0, (3)
∂t
where u is the velocity generated by the external flow.
This formulation is mathematically simple and elegant for immiscible multiphase flows. As the interface evolves and
topology changes, however, the level set function may lose its signed distance function property, and a re-initialization
procedure [42] is thus necessary to retain the signed distance function definition:

∂φ  
+ S |∇φ| − 1 = 0, (4)
∂τ
where S is a modified sign function and τ represents a pseudo-time.

2.2. Governing equation with Robin boundary conditions

For heat and mass transfer processes, the governing equation of the discrete variable ζ has the following form:

∂ζ
+ u · ∇ζ = k ζ, (5)
∂t
where k is the diffusion coefficient. The Robin boundary condition reads

 = c,
aζ + b∇ζ · n (6)

and works on the moving interface Γ , whose velocity is usually related to ζ . For simplification, the interface velocity can
be determined by

 I = D [∇ζ · n]Γ ,
u (7)

where D is a velocity coefficient and [ A ]Γ = A + − A − is an interface jump symbol. As a weighted combination of ζ and its
normal derivative, the Robin boundary condition is more complex than the Dirichlet and the Neumann boundary conditions.
4 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

Fig. 2. 1D schematic of the flux discretization form.

2.3. Numerical implementation

2.3.1. Spatial discretization

The present work utilizes uniform, staggered Cartesian grids in the whole computational domain, wherein scalars are
defined in the cell center and the velocity components on the cell faces. The spatial discretization of Eq. (5) utilizes the
second-order central differential scheme for diffusive terms and a fifth-order High Order Upstream Central (HOUC5) scheme
[20] for, if necessary, convective terms. Consider in the x direction of cell i as depicted in Fig. 2, the discretization is set in
the flux form as
F X i +1 − F X i
f = , (8)
x
where f is the right hand side of Eq. (5), and F X is the mass flux across the cell face. For diffusive terms, F X can be
computed as

ζ i − ζ i −1
F Xi = k . (9)
x
For convective terms, F X can be evaluated as

  
1
  
2
F X i = −0.5 u i + |u i | ηxp ζi−xp − 0.5 u i − |u i | ηxm ζi−xm , (10)
xp =−3 xm=−2

where η denotes the HOUC5 coefficients, xp and xm are the stencil points for cases of u > 0 and u < 0, respectively.
Ghost values are required by Eqs. (9) and (10) in the following situations:

1) The grid cells swept by the moving interface from t n to t n+1 no longer have valid values. That is to say, these cells were
outside the target subdomain at t n , but are inside at t n+1 .
2) Even for a still-interface case, the discretization stencil passes across the interface and extends to the other-side subdo-
main.

In addition, the ghost values should consider the restriction of the Robin boundary conditions, and the related treatments
will be introduced in Section 3.

2.3.2. Time integration

The temporal discretization is performed using a second-order, semi-implicit Crank–Nicolson iterative approach, which is
numerically stable and convergent. The residual form [43] is expressed as
   
1 ∂ f  n +1 n +1
 1 n n +1

1− t ζk+1 − ζk n
= ζ + tf ζ + ζk − ζkn+1 , (11)
2 ∂ζ 2
with an initial guess

ζ0n+1 = ζ n , (12)

where the low index notation k means the kth Newton–Raphson sub-iterative step, and ∂ f /∂ζ is the Jacobian. For more
details, one could refer to [44].

2.3.3. Calculating the velocity field

Once ζ is advanced in a new time step and the ghost values are extrapolated, two extended copies of smooth ζ -field are
valid for both Ω + and Ω − to calculate ∇ζ . That is, the velocity can be successfully determined near the interface according
to the jump condition of Eq. (7). Then we extrapolate the velocity in the normal direction from Γ towards both Ω + and
Ω − to define the essential velocity field for the level set evolution, by using the constant PDE approach of Aslam [45].
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 5

Fig. 3. Illustration of the polynomial reconstruction.

2.3.4. Full procedure

The full solution procedure at a new time step n + 1 can be summarized here:

 into a narrow band near the interface.

(1) Calculate the interface velocity from the jump condition of Eq. (7), and extend u
(2) Advance the LSM equation of Eq. (3) as well as the re-initialization equation of Eq. (4) (if necessary) to locate the
interface at t n+1 .
(3) Update the interface normal vector according to Eq. (2).
(4) Address the Robin boundary condition of Eq. (6) using the method described later in Section 3.
(5) Solve the convection-diffusion equation of Eq. (5) until the Newton–Raphson sub-iterative procedure is finished.
(6) Repeat from step (1) for the next time advancement.

3. Imposing Robin boundary conditions

As introduced earlier, imposing the Robin boundary conditions has been very challenging in the LSM community. Most
previous efforts were made in the context of finite volume method. In this work, we present a finite difference discretization
method for the Robin boundary conditions. Special attentions should be paid to extrapolating ghost values that are restricted
by the boundary conditions. In particular, the combination of polynomial reconstruction and partial differential equation
extrapolation is applied to improve the accuracy and efficiency of the simulations. For the sake of convenience, this section
introduces the discretization of the Poisson equation:

ζ = g, (13)

which is a simplification of Eq. (5) since the unsteady and convective terms are neglected. Note that with the ghost values,
discretization of Eq. (5) is straightforward.

3.1. Discretization of 1D Poisson equation

The standard second-order discretization of the 1D Poisson equation, i.e. ζxx = g, can be expressed as

ζ i +1 − 2ζ i + ζ i −1
(ζxx )i = . (14)
x2
If the interface Γ locates between xi −1 and xi as Fig. 3 depicts, the value ζi −1 for the cell of xi −1 is certainly not valid.
Therefore, a ghost value ζiG−1 should be provided there to keep the discretization of Eq. (14) uniform and take into consid-
eration of the restriction from the Robin boundary condition.
The polynomial reconstruction scheme from Luo et al. [46] is utilized. The scheme builds up a polynomial function
system between the variable and its position. For example, the linear reconstruction takes

ζ̃ (x̃) = mx̃ + n, (15)

with a system
⎧
⎨ ζ̃ (θ x) = mθ x + n = ζi
ζ̃ (0) = n = ζ I , (16)
⎩ 
ζ̃ (0) = m = (c − aζ I )/b
where
|φi |
θ= (17)
|φi −1 | + |φi |
6 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

is a parameter accounting for the sub-cell interface location, and x̃ is the signed distance from the reference point as
depicted in Fig. 3. It is noteworthy that the coefficients m and n are essentially identical to the derivate and the value on
the interface, respectively. One can compute the coefficients easily as

c − a ζi b ζi − c θ x
m= and n = . (18)
b − aθ x b − aθ x
Thus

  (aζi − c ) x
ζiG−1 = ζ̃ (θ − 1) x = ζi + (19)
b − aθ x
is the ghost value one desires. If one uses a higher-order spatial discretization such as the fourth-order central difference,
longer stencils are required. That is, the ghost values should be provided within a sufficiently wide band of xi − p , whose
value is decided from Eq. (15) as

 
ζiG− p = ζ̃ (θ − p ) x . (20)

Moreover, this linear reconstruction can be analogously generalized into an arbitrary order formula [27]. In particular,
one can use a qth-order unary polynomial


q
ζ̃ (x̃) = lk x̃k + mx̃ + n, (21)
k =2

with a system
⎧
⎪
⎪ ζ̃ ((θ + q − 1) x) = ζi +q−1
⎪
⎪ ..
⎪
⎨.
ζ̃ (θ x) = ζi (22)
⎪
⎪
⎪
⎪ ζ̃ (0) = ζ I
⎪
⎩ 
ζ̃ (0) = (c − aζ I )/b

to determine the ghost values. For instance, q = 2, 3 correspond to quadratic and cubic formulas, respectively. As the degree
gets higher, the matrix calculus is appropriate to solve the equation system. Plug the last two terms in the system of Eq. (22)
into Eq. (21), leading to


q
ζ̃ (x̃) = lk x̃k + (1 − ãx̃)n + c̃ x̃, (23)
k =2

where ã = a/b and c̃ = c /b. As a result, Eq. (22) can be rewritten into a matrix system
⎡ q ⎤ ⎡ ⎤ ⎡ ⎤
x̃i +q−1 · · · x̃2i +q−1 1 − ãx̃i +q−1 lq ζi +q−1 − c̃ x̃i +q−1
⎢ . ⎥
⎢ .. .. .. .. ⎥ ⎢ .. ⎥ ⎢ .. ⎥
⎢ . . . ⎥·⎢
⎢ . ⎥
⎥ =
⎢
⎢ . ⎥
⎥ (24)
⎢ q ⎥
⎣ x̃i +1 · · · x̃2i +1 1 − ã x̃i +1 ⎦ ⎣ l2 ⎦ ⎣ ζi +1 − c̃ x̃i +1 ⎦
q
x̃i ··· x̃2 1 − ãx̃i n ζi − c̃ x̃i
i

to solve the coefficients.

Numerical examples in Appendix A demonstrate that the reconstructions show second-, third- and fourth-order accura-
cies in the case of linear, quadratic and cubic polynomials, respectively. As expected, the reconstruction cost increases as
the polynomial order increases, and the length of the stencil to build up the matrix system also increases, requiring higher
mesh resolutions near the interface. Therefore, the linear reconstruction scheme is used in most of the literature [26].

3.2. Discretization of 2D Poisson equation

The discretization of the multi-dimensional Poisson equation becomes more challenging because the normal direction
of the interface does not coincide with the coordinate directions any more. Without loss of generality, we consider the 2D
Poisson equation, i.e. ζxx + ζ y y = g. A schematic of the 2D configuration is presented in Fig. 4.
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 7

Fig. 4. (a) Schematic and (b) local enlarged diagram of a 2D configuration, cell centers in Ω − (2), in Ω + (◦) and on the interface (•).

3.2.1. Why treat in the normal direction?

If the numerical discretization of ζxx can be treated independently from that of ζ y y , then the procedure of Section 3.1 can
be used straightforwardly in multi-dimensions. The key point of the discretization, therefore, lies in separating the Robin
boundary condition into two terms corresponding to the coordinate directions. In particular, the normal and tangential
derivatives of ζ are defined as

∇ζ · n = ζxnx + ζ y n y and ∇ζ · t = ζxn y − ζ y nx , (25)

respectively, where t = (n y , −nx ) is the tangential vector. The directional derivatives can be rewritten as

ζx = nx ∇ζ · n + n y ∇ζ · t and ζ y = n y ∇ζ · n − nx ∇ζ · t (26)

from Eq. (25). Similar to Nguyen et al. [30], one can use

ζx = nx ∇ζ · n and ζ y = n y ∇ζ · n (27)

to replace Eq. (26). In this way, the normal derivative is correctly solved when plugging Eq. (27) into Eq. (25), although the
tangential derivative is smeared out. Anyhow, it allows us to identify the two terms of the Robin boundary condition:
(c − aζ I )nx  (c − aζ I )n y
ζ x |Γ = and ζ y Γ = , (28)
b b
respectively, and make the discretization of ζxx and ζ y y independent to each other. As such, the procedure of aforementioned
polynomial reconstruction system is generalized into multi-dimensions. For example, the linear reconstruction gives the
ghost value of ζiG−1, j :

(aζi , j − c )nx x
ζiG−1, j = ζi , j + . (29)
b − aθ x n x x
Plugging Eq. (29) into the second-order central difference stencil, ζxx is approximated by
ζi +1 −ζi anx (ζi −c /a)
x
− b−anx θ x x
(ζxx )i , j = . (30)
x
Similar discretization in the dimension-by-dimension pattern is widely utilized for various boundary conditions [26,30].
In those cases, the discretization is easy to implement and usually symmetric, making the fast iterative solvers workable
for large-scale simulations. However, we should emphasize that the dimension-by-dimension discretization for the Robin
boundary conditions is naturally first-order accurate and the solution gradient does not even converge, as illustrated in
Appendix B because the directional derivatives from Eq. (27) are based on certain assumptions. The drop of solution accuracy
has also been observed in Nguyen et al. [30] for treating the interface jump conditions. To address this issue, a novel
discretization method with higher-order accuracies has been proposed in this work, which is inspired by the work of IBM
[46–49], to impose the Robin boundary conditions in the normal direction.

3.2.2. How to impose the Robin boundary conditions?

In the work of [46–49], authors reconstructed the polynomial systems and therefore discretized the Robin boundary
conditions in the normal direction. This strategy can produce second-order accurate solutions with convergent gradients.
However, the normal direction of a considered cell does not necessarily coincide with the coordinate directions as illustrated
in Fig. 4(a). Consequently, these methods in [46–49] need to generate virtual mirror points and interpolate their values for
the reconstruction polynomial system in the normal direction. This procedure is required for every cell that needs a ghost
8 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

value, and certainly induces complex implementations and additional computational consumptions. Although these methods
are not suitable for irregular evolving interface, they are instructive to the present discretization method for imposing the
Robin boundary conditions in the LSM community.
Here, we present the novel discretization method in the 2D Poisson equation with the focus on the accuracy and imple-
mentation. Specifically, the method combines the linear polynomial reconstruction system in the normal direction with a
second-order PDE extrapolation approach of Aslam [45]. The PDE approach is originally proposed to extrapolate values from
the known subdomain of ζ to the unknown subdomain in the normal direction. It is done by solving a series of sequent
PDEs to the steady state:
∂ζn
= H n · ∇ζn , (31)
∂t
∂ζ
= H (
n · ∇ζ − ζn ), (32)
∂t
where ζn = n  · ∇ζ corresponds to the normal derivative and H is a unit Heaviside function, which is used simply to not
disturb the values in the known subdomain with H = 0 in the known subdomain and H = 1 in the unknown one. Once
 · ∇ζn = 0, yielding a constant of ζn in the normal direction. The
Eq. (31) is solved to a steady state, it is reduced to n
obtained ζn -field appears as a source term in Eq. (32). Once Eq. (32) is solved to a steady state, the ghost value of ζ is
finally obtained, which has a normal derivative equal to ζn . It should be emphasized that the PDE approach is only an
extrapolation technique and cannot deal with the Robin boundary conditions. Thus, an additional procedure to deal with
the Robin boundary conditions is necessary before the extrapolation. This procedure consists of two steps:
Step 1: Build up the linear polynomial reconstruction system in the normal direction as depicted in Fig. 4(b) to impose
the Robin boundary condition:
⎧
⎨ ζ̃ (d) = md + n = ζi
ζ̃ (0) = n = ζ I . (33)
⎩ 
ζ̃ (0) = m = (c − aζ I )/b
Solving Eq. (33) gives the normal derivative ζn at cell (i, j). For the rest cells which are not adjacent to the interface, the
second-order central difference is used to calculate ζn . Consequently, ζn , as a scalar, is valid everywhere within the Ω +
subdomain.
Step 2: Successively extrapolate ζn and therefore ζ in the normal direction using the linear PDE approach of Eqs. (31)
and (32). According to Aslam [45], a second-order upwind scheme is used for spatial discretization.
Then the extrapolated ghost values are plugged into the following equation
   
ζ i +1 , j − ζ i , j ζ i , j − ζ i −1 , j 1 ζ i , j +1 − ζ i , j ζ i , j − ζ i , j −1 1
− + − = gi, j (34)
h h h h h h
to replace the invalid values in this discretization of the Poisson equation of Eq. (13), where h is the cell size.
We emphasize that in practical applications the interface always moves, and the ghost values are inherently needed for
those grid cells swept by the interface [26]. Thus, the PDE procedure is inherently required for such problems even if other
methods are utilized to impose the Robin boundary condition [37]. Besides, one does not really need to solve the PDEs to a
steady state in the whole domain, but can only advance them in a few steps because the ghost values provided in a narrow
band of the interface are already good enough during the discretization. As such, the present method consumes sufficiently
low computing resources, and provides second-order accurate ghost values with first-order accurate gradients and can be
generalized into 3D directly, as illustrated later in Sections 4.1.1 and 4.1.2.

3.3. Extension to higher order

Accuracy is always a critical concern in numerical simulations. So here comes a question that can the present method
be analogously extended to achieve higher order accuracy? The answer is yes, although building up the qth-order (q ≥ 2)
polynomial system may become very complex since additional interpolation procedure should be introduced to calculate
the values at probe points, which do not coincide with the grid points as depicted in Fig. 5. The interpolation procedure is
somewhat similar to that in [46–49], but is relatively at lower cost because cell (i , j) is utilized so that one less probe point
is required for each polynomial system and the polynomial system is reconstructed only for cells adjacent to the interface.
The overall accuracy of such an extension is comprehensively determined by the order of the polynomial reconstruction
system, the order of the interpolation procedure and the order of the PDE extrapolation approach. If one uses a cubic
polynomial system and intend to achieve fourth-order overall accuracy, for instance, then the interpolation scheme as well
as the PDE approach should be at least fourth-order accurate. Otherwise the overall accuracy can be suppressed. From this
perspective, the quadratic polynomial system needs a third-order interpolation (e.g. biquadratic interpolation for 2D case),
the cubic polynomial system needs a fourth-order interpolation (e.g. bicubic interpolation for 2D case), and so on.
To employ the PDE approach, the normal derivatives should be defined everywhere within Ω + . Fig. 6 depicts the pro-
cedure to calculate various order normal derivatives in higher-order extensions. Specifically, the whole Ω + is divided into
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 9

Fig. 5. Schematic of higher-order polynomial system in the normal direction. The superscript P corresponds to the probe points ().

Fig. 6. Illustration of calculating normal derivatives. The gray shaded area corresponds to the cells adjacent to the interface in Ω + .

two parts: cells adjacent to the interface and the rest cells. For the former part, the polynomial system can give the normal
derivatives from first-order to qth-order. Even though the proportion of this area is very small, its influence on numerical
accuracy is extremely strong. For the latter part, each successive normal derivative can be calculated from that of one lower
order by definition. Correspondingly, the initial condition for the PDE approach is valid. The qth-order normal derivative
is then extrapolated as a constant in the normal direction, and each successive normal derivative is integrated until ζ is
calculated. More information can be found in [45].
Numerical examples in Section 4.1.3 prove that it is feasible to generalize the present method into higher-order variants.
However, the difficulties of implementation severely increase as additional interpolation procedures are introduced. These
interpolations also require finer mesh resolution, which will certainly consume a large amount of computing resources and
is only affordable for small-scale simulations. In complex cases, the length scale may vary over several orders of magnitude,
the mesh resolution used is usually inadequate for small structures and can distort the interface severely. Consequently,
the surrounding points needed by the interpolation may lie outside of the known subdomain of ζ . Fortunately, the stan-
dard second-order method in Section 3.2 is efficient, satisfactory and can meet the demands for most applications. Unless
otherwise stated, the following examples are performed on this standard form.

4. Numerical validations

This section presents numerical examples for extrapolated ghost values, Poisson, heat and Stefan-type problems with
Robin boundary conditions. Each example poses its own numerical challenges. The L 1 and L ∞ norms are calculated for the
solutions and their gradients to study the convergence accuracy.

4.1. Extrapolated ghost values with Robin boundary conditions

In this subsection, we prove that the present method produces second-order accurate solutions with first-order accurate
gradients in both 2D and 3D configurations, and can be straightforwardly generalized into higher-order variants.

4.1.1. 2D configuration
The present method is first validated in a 2D configuration. We consider a circular subdomain Ω + with a radius r = 0.75,
which is placed at the center of the computational domain Ω = [−1, 1] × [−1, 1]. The exact solution of ζ = e xy is set within
Ω + and should be extrapolated into a narrow band in Ω − . The Robin boundary condition of Eq. (6) is imposed on the
interface with a = 1 and b = 1. Table 1 presents the numerical errors between the extrapolated ghost values and the exact
solution in the L 1 and L ∞ norms. For many applications, the gradient of the solution has more physical meaning rather than
the solution itself [50,51]. For instance, the Rankine–Hugoniot jump conditions of temperature and vapor concentration on
the interface in a vaporizing two-phase flow govern the vaporization rate and drive the accuracy of the problem [44]. It
10 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

Table 1
Numerical accuracy of the extrapolated ghost values in a 2D configuration.

x= y Solution Gradient
L 1 -norm Order L ∞ -norm Order L 1 -norm Order L ∞ -norm Order
2/64 1.43 × 10−3 – 1.04 × 10−2 – 3.57 × 10−2 – 1.93 × 10−1 –
2/128 3.68 × 10−4 1.96 3.00 × 10−3 1.80 1.80 × 10−2 0.99 1.06 × 10−1 0.86
2/256 9.81 × 10−5 1.91 7.49 × 10−4 2.00 9.30 × 10−3 0.95 5.37 × 10−2 0.98
2/512 2.39 × 10−5 2.04 1.86 × 10−4 2.01 4.58 × 10−3 1.02 2.63 × 10−2 1.03

Table 2
Numerical accuracy of the extrapolated ghost values in a 3D configuration.

x= y= z Solution Gradient
L 1 -norm Order L ∞ -norm Order L 1 -norm Order L ∞ -norm Order
2/32 8.59 × 10−4 – 2.14 × 10−3 – 2.74 × 10−2 – 5.68 × 10−2 –
2/64 1.85 × 10−4 2.21 6.12 × 10−4 1.80 1.25 × 10−2 1.13 3.12 × 10−2 0.86
2/128 4.59 × 10−5 2.01 1.66 × 10−4 1.88 6.26 × 10−3 1.00 1.64 × 10−2 0.93
2/256 1.10 × 10−5 2.06 4.30 × 10−5 1.95 3.04 × 10−3 1.04 8.47 × 10−3 0.95

Table 3
An example of the higher-order extension of the present method.

x= y Solution Gradient
L 1 -norm Order L ∞ -norm Order L 1 -norm Order L ∞ -norm Order
2/64 1.34 × 10−4 – 1.07 × 10−3 – 3.62 × 10−3 – 2.24 × 10−2 –
2/128 1.77 × 10−5 2.92 1.52 × 10−4 2.82 9.35 × 10−4 1.95 6.58 × 10−3 1.77
2/256 2.36 × 10−6 2.91 1.96 × 10−5 2.95 2.44 × 10−4 1.94 1.73 × 10−3 1.92
2/512 2.94 × 10−7 3.00 2.45 × 10−6 3.00 6.07 × 10−5 2.00 4.23 × 10−4 2.03

is thus of interest to report the accuracy of the solution gradient. Obviously, we observe second-order convergence for the
numerical solution and first-order convergence for its gradient.

4.1.2. 3D configuration
Next, the present method is applied to a 3D case. Consider a spheroid subdomain Ω + of radius r = 0.5, which is placed
at the center of the computational domain Ω = [−1, 1] × [−1, 1] × [−1, 1]. The exact solution of ζ = e x + y +z is set
2 2 2

within Ω + , and the Robin boundary condition of Eq. (6) is imposed on the interface with a = 1 and b = 1. The numerical
errors of the ghost values in a narrow band compared to the exact solution are presented in Table 2. It turns out that the
numerical solution and its gradient remain second-order accurate and first-order accurate, respectively.

4.1.3. Higher-order extension

A higher-order example is shown in Table 3 by using the quadratic polynomial reconstruction for the Robin boundary
conditions, the biquadratic interpolation for the probe points and the quadratic PDE extrapolation approach for the ghost
values. The formulated problem is the same as that in Section 4.1.1. Different from previous performance, the current
numerical solution converges to exact solution with third-order accuracy in the L 1 and L ∞ norms and its gradient is of
second-order accuracy. Note that this higher-order numerical scheme requires additional computation cost. The current test
on the finest grid resolution required about 20% more CPU cost due to the interpolation procedure during Step 1, and about
40% more caused by the quadratic PDE approach during Step 2 when compared to the standard second-order form used in
Section 4.1.1.

4.2. Poisson solution with Robin boundary conditions

The present method is validated in terms of applicability on irregular domains in this subsection. The formulated Poisson
problems from [37,39] are performed and the convergence results are demonstrated.

4.2.1. Over a flower-shaped domain

In this case, the target subdomain Ω + is bounded by a five-petal flower interface, which is described as the zero level set
y 5 +5x4 y −10x2 y 3
of φ = 0.5 + 3r 5
− r, where r is the distance from the center. ζ is initially specified to 0, and the Robin boundary
condition of Eq. (6) is imposed on the interface with a = 1 and b = 1. Fig. 7 presents the numerical result with a mesh
resolution of 642 . The numerical result is consistent with the exact solution of ζ = e xy , and convergence study demonstrates
that the solution and its gradient are second-order accurate and first-order accurate, respectively, in the L 1 and L ∞ norms
as depicted in Fig. 8.
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 11

Fig. 7. Numerical solution of the Poisson equation over the flower-shaped domain with a Robin boundary condition.

Fig. 8. Errors in the L 1 and L ∞ norms of (a) the Poisson solution and (b) its gradient over the flower-shaped domain with a Robin boundary condition. The
slopes of least-square fit of the L 1 and L ∞ norms are −2.07 and −2.01 for the solution while −1.74 and −0.94 for the gradient.

4.2.2. Over a star-shaped domain

In this case, an irregular domain Ω + is placed at the center and given by the zero level set of φ = 0.65 − 0.08 cos(8α ) − r,
where α ∈ [0, 2π ]. The exact solution is given by ζ = e x (sin( y ) + y 2 ), and ζ is initially set to −2. The Robin boundary
condition of Eq. (6) is imposed on the interface with a = 1 and b = 2. Fig. 9 shows the scatter plot on the numerical
solution with x = 2/64. Similarly, the numerical solution is second-order accurate in the L 1 and L ∞ norms and the
solution gradient shows first-order accuracy as depicted in Fig. 10.

4.2.3. Estimation of the condition number

The condition number for the associated linear system of a Poisson solver is also of significance to be reported because
large condition numbers can delay the convergence and lose many digits in the approximation. For irregular domains, the
choice of the extrapolated ghost values certainly affects the condition number [31,32,40]. In the present study, the ghost
values are implicitly provided by the combination of the polynomial reconstruction and the PDE extrapolation, and cannot
be further expressed as an explicit function of the Robin boundary condition and the Poisson solution. As a result, they are
preset as known values for each iteration during the Poisson discretization of Eq. (34). The entry ar ,s of the matrix from the
discretization A with cells (i r , j r ) and (i s , j s ) [52] is then given as
⎧ −2
⎪
⎪ h if i s = i r ± 1 and j s = j r
⎨
−4h−2 if s = r
ar , s = , (35)
⎪
⎪ h −2 if i s = i r and j s = j r ± 1
⎩
0 otherwise
12 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

Fig. 9. Numerical solution of the Poisson equation over the star-shaped domain with a Robin boundary condition.

Fig. 10. Errors in the L 1 and L ∞ norms of (a) the Poisson solution and (b) its gradient over the star-shaped domain with a Robin boundary condition. The
slopes of least-square fit of the L 1 and L ∞ norms are −2.01 and −1.85 for the solution while −1.23 and −0.95 for the gradient.

where 1 ≤ r , s ≤ K , and K accords to the number of cells in Ω + with the lexicographical order [53]. The 1-norm condition
numbers of A versus grid resolution for the above two Poisson examples are shown in Fig. 11. According to this estimation,
the condition number demonstrates the inverse proportionality to h2 . It is a good way to report the condition number as a
function of grid size. Indeed, Yoon and Min [52] have presented such analyses for a typical Poisson equation with Dirichlet
boundary conditions. We refer the interested reader to [52] and the references therein for more details.

4.3. Heat and Stefan problems with Robin boundary conditions

The present method is then validated and applied to the heat and Stefan problems with Robin boundary conditions.
In these conditions, the unsteady terms should be considered and the governing equation of Eq. (5) reduces to the heat
equation:

∂ζ
= k ζ. (36)
∂t
Moreover, the Stefan problems are relatively more complicated due to the evolution of the interface.

4.3.1. 2D Frank sphere solution

The 2D Frank sphere problem [54] is a classical Stefan problem with an exact solution. It describes the growing solidifi-
cation process, during which the phase interface expands outward and keeps its shape as a cylinder. The temperature field,
governed by Eq. (36) with ζ = T and k = 1, is zero in the solid phase (Ω − ) and
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 13

Fig. 11. Condition number versus grid resolution for the above Poisson problems. The slopes of least-square fit are 1.71 for the flower-shaped domain and
1.98 for the star-shaped domain.

 
E 1 ( s 2 /4 )
T = T∞ 1 − (37)
E 1 (s20 /4)
 ∞
in the liquid phase (Ω + ), where s = rt −1/2 , r = x2 + y 2 and E 1 ( z) = z
ξ −1 e −ξ dξ . Note that the initial s is related to the
far-field temperature T ∞ :
 
s20 s20 s2
0
T∞ = − E1 e 4 . (38)
4 4
The normal derivative of temperature reads as
2T ∞
e −s /4 s−1 t −1/2 .
2
∇ T · n = (39)
E 1 (s20 /4)
Thus the interface velocity of Eq. (7) is defined as
 
2T ∞
e −s /4 s−1 t −1/2 n
2
I = −
u  (40)
E 1 (s20 /4)
with D = −1. Interested readers could refer to [54] for more details.
Here, we carry out a test similar to that of Papac et al. [37]. A solid disk with an initial radius of r = 0.75 is placed at
the center of the computational domain Ω = [−3, 3] × [−3, 3]. The far-field temperature T ∞ is set to −0.246. The Robin
boundary condition of Eq. (6) is imposed on the interface with a = 1 and b = 1 while the Dirichlet boundary condition is
enforced on ∂Ω . The temporal discretization is based on the Crank–Nicolson scheme with t proportional to x, and the
simulation is performed from t = 1 to t = 2.
The temperature fields at t = 1 and t = 2 based on the grid resolution of 642 are compared in Fig. 12. The supercooled
liquid continuously turns into solid, making the interface expand outward. The temporal evolution of the radius is displayed
in Fig. 13. As the grid is refined, the numerical results approach the analytical solution. Both the temperature and the radius
converge to the exact solutions with first-order accuracy in the L 1 and L ∞ norms, as depicted in Fig. 14(a) and Fig. 15,
because the temperature gradient used in calculating the interface velocity is first-order accurate as illustrated in Fig. 14(b).
If we provide the exact interface location (which means exact interface velocity as well), the simulation essentially
reduces to a heat diffusion problem with a time varying interface. In this situation, the temperature solution achieves the
expected second-order accuracy in the L 1 and L ∞ norms while the gradient is first-order accurate as shown in Fig. 16.

4.3.2. Time-decaying solution

The Stefan problem of Eq. (36) with k = 1 is considered here. The initial domain is a circle with a radius r = π , which
is placed at the center of the computational domain Ω = [−1.5π , 1.5π ] × [−1.5π , 1.5π ]. The Robin boundary condition of
Eq. (6) is derived from the exact solution of ζ = −e −2t cos(x) cos( y ) and enforced on the interface with a = 1 and b = 1. The
interface velocity is given by u I = (2 − [∇ζ · n]Γ )
n. Temporal discretization is based on the Crank–Nicolson scheme with t
proportional to x, and the simulation is performed from t = 0 to t = 0.5.
If the interface keeps still, then the simulation essentially reduces to a heat diffusion problem. Fig. 17 compares the
ζ -fields at t = 0 and t = 0.5 in this situation. As expected, the amplitude decays with time. The numerical solution is
second-order accurate in the L 1 and L ∞ norms with first-order accurate gradient as depicted in Fig. 18.
When the interface regresses according to the velocity field, the domain becomes non-uniform as shown in Fig. 19.
The accuracy of the time-decaying Stefan solution is presented in Fig. 20, demonstrating first-order accuracy because the
gradient of ζ used to advance the interface is first-order accurate.
14 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

Fig. 12. Contour plot of the temperature fields at t = 1 and 2 of the Frank sphere solution.

Fig. 13. Temporal evolution of the numerical radius of the Frank sphere solution.

Fig. 14. Errors in the L 1 and L ∞ norms of (a) the temperature solution and (b) its gradient with a calculated interface at t = 2. The slopes of least-square
fit of the L 1 and L ∞ norms are −1.06 and −1.36 for the solution while −0.87 and −0.96 for the gradient.

5. Applications to heat and mass transfers

After the above validations, the present method is applied to study more practical heat and mass transfer problems,
which do not always have an exact solution. For each problem, we briefly discuss the physical representation to highlight
the feasibility of the present method as a promising tool for practical problems.
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 15

Fig. 15. Errors in the L 1 and L ∞ norms of the radius at t = 2. The slopes of least-square fit of the L 1 and L ∞ norms are −1.27 and −1.13.

Fig. 16. Errors in the L 1 and L ∞ norms of (a) the temperature solution and (b) its gradient with an exact interface at t = 2. The slopes of least-square fit of
the L 1 and L ∞ norms are −2.15 and −1.88 for the solution while −1.54 and −1.07 for the gradient.

Fig. 17. Contour plot of the temperature fields at t = 0 and 0.5 with a still interface.

5.1. Heat transfer in a star-shaped solid body

A simulation of heat transfer in a convectively cooled irregular body is performed. Similar to the study of Papac et
al. [37], a star-shaped solid body initially with a high temperature of 800 K is placed at the center of the computational
domain Ω = [−1, 1] × [−1, 1]. The boundary of the star is defined by the zero level set of φ = 0.5 + 0.05 cos(8α ) − r, where
16 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

Fig. 18. Errors in the L 1 and L ∞ norms of (a) the time-decaying solution and (b) its gradient with a still interface at t = 0.5. The slopes of least-square fit
of the L 1 and L ∞ norms are −2.03 and −1.99 for the solution while −1.61 and −0.99 for the gradient.

Fig. 19. Evolving interface location of the time-decaying Stefan problem at t = 0 and 0.5.

Fig. 20. Errors in the L 1 and L ∞ norms of (a) the time-decaying Stefan solution and (b) its gradient over an evolving domain at t = 0.5. The slopes of
least-square fit of the L 1 and L ∞ norms are −1.76 and −1.45 for the solution while −1.29 and −1.01 for the gradient.
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 17

Fig. 21. Contour plot of the temperature distribution in the star-shaped solid body at t = 30 s.

Fig. 22. Temporal evolution of temperature along the horizontal center line.

α ∈ [0, 2π ], and the ambient temperature is set to T ∞ = 300 K. As time evolves, the temperature of the solid body is
governed by Eq. (36). In particular, this case has ζ = T and k = λ/(ρ C p ), where λ is the thermal conduction coefficient, ρ is
the density and C p is the heat capacity at constant pressure. Applying heat conservation on the interface gives the following
Robin boundary condition:

 = hT ∞ ,
hT − λ∇ T · n (41)

where h is the convection heat transfer coefficient. For the sake of simplicity, the thermal expansion is treated by giving
a uniform interface regress velocity of 2.0 × 10−3 m/s. The physical properties of copper are used by ρ = 8954 kg/m3 ,
λ = 378 W/m K, C p = 384 J/kg K and h = 3072 W/m2 K. The simulation is conducted using a grid resolution of 256 × 256.
Fig. 21 presents the contour plot of the temperature field at t = 30 s and Fig. 22 depicts the temporal evolution of the
thermal conduction along the horizontal center line. The numerical results are consistent with those reported previously
in [37] and indicate that the temperature gradients are extremely high close to the interface especially near the concave-
interface areas.

5.2. Mass transfer via droplet evaporation

In this case, the method is applied to a droplet evaporation problem. The initial droplet radius is r = 200 μm with the
computational domain of 10r × 10r and a grid resolution of 128 × 128. The governing equation of Eq. (5) takes ζ = Y and
k = D m , where Y is the vapor mass fraction and D m is the mass diffusion coefficient. The physical properties of water and
gas are used by ρl = 1000 kg/m3 , ρ g = 1.226 kg/m3 and D m = 2 × 10−5 s−1 .
Due to evaporation, a Stefan flow occurs near the interface. Consequently, the velocity jumps across the interface as
 
1
[
u ]Γ = −ω̇ ,
n (42)
ρ Γ
18 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

Fig. 23. Plots of (a) velocity field and (b) vapor mass fraction at t = 0.05 s.

Fig. 24. Vapor mass fraction distribution along the horizontal center line at t = 0.05 s.

where ω̇ is the evaporation rate. The velocity field is computed from a velocity potential as introduced in Tanguy et al. [3],
and the Rankine–Hugoniot jump condition essentially reduces to a Robin boundary condition on the interface as

ω̇ Y + ρ g D m ∇ Y · n = ω̇. (43)
For the sake of simplicity, the evaporation rate is artificially given by ω̇ = 0.1 tanh 50t. Since the evaporation rate approaches
0.1 kg/m2 s, the evaporation gradually reaches a quasi-steady state and the vapor mass fraction field has an analytical
solution. Derivations of the analytical solution can be found in Rueda Villegas et al. [38].
Fig. 23 presents the velocity field and vapor mass fraction field around the droplet at t = 0.05 s. The velocity moves
outwards from the interface and gradually decrease, and the vapor concentration shows a uniform and symmetrical property
in circumferential direction. No obvious parasitic currents are observed. The vapor mass fraction distribution along the
horizontal center line is depicted in Fig. 24, which is largely consistent with the analytical solution, further validating the
present method.

6. Conclusions

In this work, an efficient method for solving heat and mass transfer problems with the Robin boundary conditions on
irregular domains in the framework of finite difference scheme, which is a popular choice in computational fluid dynamics,
has been developed. The method utilized the LSM to implicitly capture the potentially moving interface as well as to
automatically handle the topological changes, and the GFM to address variable discontinuities across the interface as well
to reduce parasitic currents. Special efforts have been devoted to providing ghost values that are restricted by the Robin
boundary conditions. The extrapolated ghost values are second-order accurate and the gradients are first-order accurate.
A series of numerical examples have been performed to assess the method. It proves to be second-order accurate for
both the Poisson and heat diffusion problems with Robin boundary conditions. As for the Stefan problems, the solution
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 19

drops to first-order accuracy because calculating interface location depends on the normal derivative, which is naturally
first-order accurate. Finally, the method has been successfully applied to more practical heat and mass transfer problems.
The results are highly consistent with the expected physical behavior.
The method is easy to implement in 3D and consumes sufficiently low computing resources. It should be noted that the
standard second-order form can be straightforwardly generalized into higher-order variants although finer mesh resolution
and additional computation consumption may be required.

Declaration of competing interest

The authors declare that there is no conflict of interest.

Acknowledgements

This work is financially supported by the National Natural Science Foundation of China (Nos. 91541202 and 91741203).

Appendix A. Polynomial reconstruction schemes

This part describes the polynomial reconstruction schemes in a 1D case. The test is similar to that in Gibou et al. [27].
In particular, we define φ = x − 0.5 in the 1D computational domain Ω = [0, 1]. The exact solution of ζ = x5 − x3 + 12x2 −
2.5x + 2 is given in the target subdomain Ω + and the Robin boundary condition of Eq. (6) is imposed on the interface with
a = −2 and b = 1. The errors between the extrapolated ghost values within a narrow band in Ω − and the exact solution are
presented in Tables 4–6 in the L 1 and L ∞ norms. As illustrated, the linear, quadratic and cubic polynomial reconstructions
are second-, third- and fourth-order accurate, respectively.
Computational consuming is an important concern especially in large numerical simulations. Tables 4–6 also present the
average CPU cost per time step for each polynomial scheme. In particular, the linear reconstruction is implemented directly
using Eqs. (19) and (20) while the quadratic and cubic ones should solve the matrix system of Eq. (24). As expected, the
linear reconstruction is the cheapest scheme whose cost is about one-tenth of that of the quadratic one, and one-fifteenth
of that of the cubic one.

Appendix B. Classic discretization in dimension-by-dimension pattern

The configurations here are just the same as the one in Section 4.1.1. In particular, a circular domain Ω + of radius
r = 0.75 is placed at the center of the computational domain Ω = [−1, 1] × [−1, 1]. The exact solution of ζ = e xy is set
within Ω + and the Robin boundary condition of Eq. (6) is imposed on the interface with a = 1 and b = 1. The ghost values
are extrapolated into a narrow band in Ω − using the dimension-by-dimension discretization, which separates the Robin
boundary condition into two coordinate directions as illustrated in Section 3.2.1. The errors between the extrapolated values

Table 4
Numerical accuracy and computational cost of the linear reconstruction.

x L 1 -norm Order L ∞ -norm Order CPU time/s

1/16 3.64 × 10−1 – 9.05 × 10−1 –
1/32 9.12 × 10−2 2.00 2.27 × 10−1 2.00
3.85 × 10−8
1/64 2.29 × 10−2 2.00 5.70 × 10−2 1.99
1/128 5.73 × 10−3 2.00 1.43 × 10−2 2.00

Table 5
Numerical accuracy and computational cost of the quadratic reconstruction.

x L 1 -norm Order L ∞ -norm Order CPU time/s

1/16 1.28 × 10−2 – 3.36 × 10−2 –
1/32 1.78 × 10−3 2.84 4.88 × 10−3 2.79
3.32 × 10−7
1/64 2.36 × 10−4 2.92 6.57 × 10−4 2.89
1/128 3.04 × 10−5 2.96 8.53 × 10−5 2.95

Table 6
Numerical accuracy and computational cost of the cubic reconstruction.

x L 1 -norm Order L ∞ -norm Order CPU time/s

1/16 1.09 × 10−2 – 3.33 × 10−2 –
1/32 6.78 × 10−4 4.00 2.08 × 10−3 4.00
5.16 × 10−7
1/64 4.23 × 10−5 4.00 1.30 × 10−4 4.00
1/128 2.64 × 10−6 4.00 8.16 × 10−6 4.00
20 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890

Table 7
Numerical accuracy of the classic dimension-by-dimension pattern.

x= y Solution Gradient
L 1 -norm Order L ∞ -norm Order L 1 -norm Order L ∞ -norm Order
2/64 1.34 × 10−2 – 4.76 × 10−2 – 6.78 × 10−2 – 4.44 × 10−1 –
2/128 6.58 × 10−3 1.02 1.99 × 10−2 1.26 4.81 × 10−2 0.49 4.00 × 10−1 0.15
2/256 3.25 × 10−3 1.02 9.91 × 10−3 1.01 3.78 × 10−2 0.35 4.06 × 10−1 −0.02
2/512 1.62 × 10−3 1.01 4.95 × 10−3 1.00 3.23 × 10−2 0.23 3.59 × 10−1 0.18

and the exact solution are presented in Table 7 in the L 1 and L ∞ norms. It demonstrates that the solution is suppressed to
first-order accurate and the gradient does not converge in general.

References

[1] C.S. Peskin, Numerical analysis of blood flow in the heart, J. Comput. Phys. 25 (1977) 220–252.
[2] C.S. Peskin, B.F. Printz, Improved volume conservation in the computation of flows with immersed elastic boundaries, J. Comput. Phys. 105 (1993)
33–46.
[3] S. Tanguy, T. Ménard, A. Berlemont, A level set method for vaporizing two-phase flows, J. Comput. Phys. 221 (2007) 837–853.
[4] S.O. Unverdi, G. Tryggvason, A front-tracking method for viscous, incompressible, multi-fluid flows, J. Comput. Phys. 100 (1992) 25–37.
[5] C.W. Hirt, B.D. Nichols, Volume of fluid (VOF) method for the dynamics of free boundaries, J. Comput. Phys. 39 (1981) 201–225.
[6] P. Yue, J.J. Feng, C. Liu, J. Shen, A diffuse-interface method for simulating two-phase flows of complex fluids, J. Fluid Mech. 515 (2004) 293–317.
[7] S. Osher, J. Sethian, Fronts propagating with curvature dependent speed: algorithms based on Hamilton–Jacobi formulations, J. Comput. Phys. 79 (1988)
12–49.
[8] R.P. Fedkiw, T. Aslam, B. Merriman, S. Osher, A non-oscillatory Eulerian approach to interfaces in multimaterial flows (the ghost fluid method), J.
Comput. Phys. 152 (1999) 457–492.
[9] C. Shao, K. Luo, Y. Yang, J. Fan, Detailed numerical simulation of swirling primary atomization using a mass conservative level set method, Int. J.
Multiph. Flow 89 (2017) 57–68.
[10] S. Tanguy, M. Sagan, B. Lalanne, F. Couderc, C. Colin, Benchmarks and numerical methods for the simulation of boiling flows, J. Comput. Phys. 264
(2014) 1–22.
[11] F.D.R. Gibou, L. Chen, D. Nguyen, S. Banerjee, A level set based sharp interface method for the multiphase incompressible Navier–Stokes equations with
phase change, J. Comput. Phys. 222 (2007) 536–555.
[12] G. Son, A level-set method for analysis of microdroplet evaporation on a heated surface, J. Mech. Sci. Technol. 24 (2010) 991–997.
[13] R.W. Houim, K.K. Kuo, A ghost fluid method for compressible reacting flows with phase change, J. Comput. Phys. 235 (2013) 865–900.
[14] J. Shaikh, A. Sharma, R. Bhardwaj, On sharp-interface level-set method for heat and/or mass transfer induced Stefan problem, Int. J. Heat Mass Transf.
96 (2016) 458–473.
[15] M. Mirzadeh, A. Guittet, C. Burstedde, F. Gibou, Parallel level-set methods on adaptive tree-based grids, J. Comput. Phys. 322 (2016) 345–364.
[16] M. Detrixhe, F. Gibou, C. Min, A parallel fast sweeping method for the Eikonal equation, J. Comput. Phys. 237 (2013) 46–55.
[17] M. Detrixhe, F. Gibou, Hybrid massively parallel fast sweeping method for static Hamilton–Jacobi equations, J. Comput. Phys. 322 (2016) 199–223.
[18] M. Sussman, E.G. Puckett, A coupled level set and volume-of-fluid method for computing 3D and axisymmetric incompressible two-phase flows, J.
Comput. Phys. 162 (2000) 301–337.
[19] E. Olsson, G. Kreiss, A conservative level set method for two phase flow, J. Comput. Phys. 210 (2005) 225–246.
[20] O. Desjardins, V. Moureau, H. Pitsch, An accurate conservative level set/ghost fluid method for simulating turbulent atomization, J. Comput. Phys. 227
(2008) 8395–8416.
[21] L. Zhao, X. Bai, T. Li, J.J.R. Williams, Improved conservative level set method, Int. J. Numer. Methods Fluids 75 (2014) 575–590.
[22] K. Luo, C. Shao, Y. Yang, J. Fan, A mass conserving level set method for detailed numerical simulation of liquid atomization, J. Comput. Phys. 298 (2015)
495–519.
[23] M. Chai, K. Luo, C. Shao, J. Fan, An efficient level set remedy approach for simulations of two-phase flow based on sigmoid function, Chem. Eng. Sci.
172 (2017) 335–352.
[24] S. Chen, B. Merriman, S. Osher, P. Smereka, A simple level set method for solving Stefan problems, J. Comput. Phys. 135 (1997) 8–29.
[25] H.S. Udaykumar, R. Mittal, W. Shyy, Computation of solid-liquid phase fronts in the sharp interface limit on fixed grids, J. Comput. Phys. 153 (1999)
535–574.
[26] F. Gibou, R.P. Fedkiw, L. Cheng, M. Kang, A second-order-accurate symmetric discretization of the Poisson equation on irregular domains, J. Comput.
Phys. 176 (2002) 205–227.
[27] F. Gibou, R. Fedkiw, A fourth order accurate discretization for the Laplace and heat equations on arbitrary domains, with applications to the Stefan
problem, J. Comput. Phys. 202 (2005) 577–601.
[28] Y.T. Ng, C. Min, F. Gibou, An efficient fluid–solid coupling algorithm for single-phase flows, J. Comput. Phys. 228 (2009) 8807–8829.
[29] D.Q. Nguyen, R.P. Fedkiw, M. Kang, A boundary condition capturing method for incompressible flame discontinuities, J. Comput. Phys. 172 (2001)
71–98.
[30] X.D. Liu, R.P. Fedkiw, M.J. Kang, A boundary condition capturing method for Poisson’s equation on irregular domains, J. Comput. Phys. 160 (2000)
151–178.
[31] A. Guittet, M. Lepilliez, S. Tanguy, F. Gibou, Solving elliptic problems with discontinuities on irregular domains – the Voronoi interface method, J.
Comput. Phys. 298 (2015) 747–765.
[32] F. Gibou, C. Min, R. Fedkiw, High resolution sharp computational methods for elliptic and parabolic problems in complex geometries, J. Sci. Comput.
54 (2013) 369–413.
[33] F. Gibou, R. Fedkiw, S. Osher, A review of level-set methods and some recent applications, J. Comput. Phys. 353 (2018) 82–109.
[34] K. Luo, C. Shao, M. Chai, J. Fan, Level set method for atomization and evaporation simulations, Prog. Energy Combust. Sci. 73 (2019) 65–94.
[35] C.G. Bell, H.M. Byrne, J.P. Whiteley, S.L. Waters, Heat or mass transfer at low Péclet number for Brinkman and Darcy flow round a sphere, Int. J. Heat
Mass Transf. 68 (2014) 247–258.
[36] L. Rueda Villegas, S. Tanguy, G. Castanet, O. Caballina, F. Lemoine, Direct numerical simulation of the impact of a droplet onto a hot surface above the
Leidenfrost temperature, Int. J. Heat Mass Transf. 104 (2017) 1090–1109.
 M. Chai et al. / Journal of Computational Physics 400 (2020) 108890 21

[37] J. Papac, F. Gibou, C. Ratsch, Efficient symmetric discretization for the Poisson, heat and Stefan-type problems with Robin boundary conditions, J.
Comput. Phys. 229 (2010) 875–889.
[38] L. Rueda Villegas, R. Alis, M. Lepilliez, S. Tanguy, A ghost fluid/level set method for boiling flows and liquid evaporation: application to the Leidenfrost
effect, J. Comput. Phys. 316 (2016) 789–813.
[39] J. Papac, A. Helgadottir, C. Ratsch, F. Gibou, A level set approach for diffusion and Stefan-type problems with Robin boundary conditions on quadtree/oc-
tree adaptive Cartesian grids, J. Comput. Phys. 233 (2013) 241–261.
[40] D. Bochkov, F. Gibou, Solving Poisson-type equations with Robin boundary conditions on piecewise smooth interfaces, J. Comput. Phys. 376 (2019)
1156–1198.
[41] Á. Helgadóttir, Y.T. Ng, C. Min, F. Gibou, Imposing mixed Dirichlet–Neumann–Robin boundary conditions in a level-set framework, Comput. Fluids 121
(2015) 68–80.
[42] M. Sussman, P. Smereka, S. Osher, A level set approach for computing solutions to incompressible two-phase flow, J. Comput. Phys. 114 (1994) 146–159.
[43] C.D. Pierce, Progress-Variable Approach for Large-Eddy Simulation of Turbulent Combustion, Stanford University, Palo Alto, 2001.
[44] M. Chai, K. Luo, C. Shao, H. Wang, J. Fan, A coupled vaporization model based on temperature/species gradients for detailed numerical simulations
using conservative level set method, Int. J. Heat Mass Transf. 127 (2018) 743–760.
[45] T.D. Aslam, A partial differential equation approach to multidimensional extrapolation, J. Comput. Phys. 193 (2004) 349–355.
[46] K. Luo, Z. Zhuang, J. Fan, N.E.L. Haugen, A ghost-cell immersed boundary method for simulations of heat transfer in compressible flows under different
boundary conditions, Int. J. Heat Mass Transf. 92 (2016) 708–717.
[47] G.S. Barozzi, C. Bussi, M.A. Corticelli, A fast Cartesian scheme for unsteady heat diffusion on irregular domains, Numer. Heat Transf., Part B, Fundam.
46 (2004) 59–77.
[48] D. Pan, A general boundary condition treatment in immersed boundary methods for incompressible Navier–Stokes equations with heat transfer, Numer.
Heat Transf., Part B, Fundam. 61 (2012) 279–297.
[49] A. Pacheco-Vega, J.R. Pacheco, T. Rodic, A general scheme for the boundary conditions in convective and diffusive heat transfer with immersed boundary
methods, J. Heat Transf.-Trans. ASME 129 (2007) 1506–1516.
[50] J. Seo, S. Ha, C. Min, Convergence analysis in the maximum norm of the numerical gradient of the Shortley–Weller method, J. Sci. Comput. 74 (2018)
631–639.
[51] G. Yoon, C. Min, Convergence analysis of the standard central finite difference method for Poisson equation, J. Sci. Comput. 67 (2016) 602–617.
[52] G. Yoon, C. Min, Analyses on the finite difference method by Gibou et al. for Poisson equation, J. Comput. Phys. 280 (2015) 184–194.
[53] I. Gustafsson, A class of first order factorization methods, BIT 18 (1978) 142–156.
[54] F.C. Frank, Radially symmetric phase growth controlled by diffusion, Proc. R. Soc. A 201 (1950) 586–599.