International Journal of Multiphase Flow 162 (2023) 104409

Contents lists available at ScienceDirect

International Journal of Multiphase Flow

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

An accurate and efficient multiphase solver based on THINC scheme and

adaptive mesh refinement
Dezhu Chen a , Xin Tong a , Bin Xie a ,∗, Feng Xiao b , Ye Li a
a
School of Naval Architecture, Department of Ocean and Civil Engineering, Shanghai Jiaotong University, Shanghai, 200240, China
b
School of Engineering, Department of Mechanical Engineering, Tokyo Institute of Technology, Tokyo, 152-8550, Japan

ARTICLE INFO ABSTRACT

Keywords: This paper presents an accurate and efficient multiphase solver which is developed based on THINC/QQ
Multiphase flow (THINC method with quadratic surface representation and Gauss quadrature) scheme and AMR (adaptive mesh
Free surface refinement) technique in the framework of OpenFOAM. In order to make consistent interpolation after the
THINC/QQ
mesh refinement, a novel algorithm is proposed for the algebraic mapping of the volume fraction field in the
Interface capturing
child cells, which significantly improves the geometric faithfulness of mapped interface particularly for curved
Adaptive mesh refinement
surface. Second-order schemes are used for the spatial and temporal discretization of momentum equations
which further enhances the solution accuracy compared with interFoam. Using THINC/QQ scheme with AMR
technique in the multiphase simulation, the so-called thincFoam delivers high fidelity solutions in the vicinity
of interface while largely improving the computational efficiency. As verified by various benchmark tests in
2D and 3D, such as bubble rising, dam-break and propagation of solitary wave, it convinces that the numerical
results predicted by thincFoam are in excellent agreement with experimental data and other reference solutions.
Thus, the present solver possesses a promising prospect in capturing complicated interfaces undergoing large
deformations and topological changes for the simulation of nonlinear interaction between multiphase flows
with solid structures.

1. Introduction function in a fixed grid which needs to be further investigated to resolve

violent motions with high fidelity and meanwhile ensure the finer grid
It is universally found for incompressible immiscible two-phase resolution in the vicinity of interface.
flows in both nature and engineering, such as the propagation of wave The key to solving two-phase flows on the Eulerian grid is how to
in sea, jet flow, oil extraction and wave energy conversion (Xu et al., deal with the discrete representation of interface. For decades, various
2019). In general, two-phase flows feature the large-density ratio, and
interface capturing methods have been developed greatly and used to
tend to generate large deformations and geometric topological changes
simulate the two-phase flows numerically, in which volume of fluid
of interface, which makes the Computational Fluid Dynamics (CFD)
(VOF) method is particularly appealing. The VOF method represents
simulations of two-phase flow meet great challenges, particularly for
the complex flows with a large range of temporal and spatial scales. the interface by an indicator function defined by the proportion of one
Furthermore, it is noted that not only the apparent phenomena but also fluid. Furthermore, it is classified into algebraic and geometric VOF
small-scale structures, for example breakup in bubble flow, should be methods according to whether with explicit interface reconstruction or
investigated and studied by using numerical methods in detail. The is- not. The geometric VOF methods, represented by piecewise linear in-
sue mentioned above can be addressed separately from the perspective terface calculation (PLIC), reconstruct the interface by a linear (2D) and
of Lagrangian and Eulerian methods. In the past decades, it has made planar (3D) segment in target cell, in which the direction of outward
significant progresses for particle methods (Luo et al., 2015, 2021; normal vector of interface is considered to improve the numerical ac-
Zhang et al., 2022) which are free from meshes and numerical diffusion curacy (Youngs, 1984; Puckett, 1991; Pilliod and Puckett, 2004; Rider
of advection terms. Moreover, Adaptive Particle Refinement (APR)
and Kothe, 1998; Benson, 2002; Aulisa et al., 2003). Apart from the
approaches (Chiron et al., 2018; Sun et al., 2018) are also proposed to
traditional linear reconstruction, the piecewise parabolic reconstruction
improve upon the error minimization and uniform error distribution.
of VOF method has been proposed by Renardy and Renardy (2002) to
In the Eulerian framework, the interface is captured by solving a field

∗ Corresponding author.
E-mail address: [email protected] (B. Xie).

https://doi.org/10.1016/j.ijmultiphaseflow.2023.104409
Received 22 September 2022; Received in revised form 9 January 2023; Accepted 3 February 2023
Available online 9 February 2023
0301-9322/© 2023 Elsevier Ltd. All rights reserved.
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

calculate the surface tension term accurately. In addition to the struc- et al. (2020a) proposed a consistent and balanced-force model for in-
tured algorithms mentioned above, some geometric VOF methods are compressible multiphase flows by combining FVMS3 (Xie et al., 2019)
proposed for unstructured grids (López et al., 2016; Marić et al., 2020) and THINC/QQ scheme. The numerical results demonstrate remarkable
and polygonal/polyhedral grids (Dai and Tong, 2018, 2019; Roenby capability to predict complex flows with largely deformed interface
et al., 2016; Scheufler and Roenby, 2019), despite that it substantially and high density ratio, which outperforms interFoam solver with much
increases difficulties due to the complicated geometrical operations. less mesh dependency and numerical dissipation. This solver was then
On the other hand, the algebraic VOF methods have been devised in used by Huang et al. (2021) to simulate the propagation of solitary
a different strategy by adding an artificial compression or anti-diffuse wave, which significantly improves the solution quality by reducing
source term into VOF equation (Ubbink and Issa, 1999; Deshpande undesirable numerical effects, including energy loss, wave decay, phase
et al., 2012) which does not require geometric reconstruction like PLIC shift, and overestimation of the velocity profile even after long du-
methods and thus largely improves the computational efficiency and al- ration computation. More recently, Yin et al. (2021) quantitatively
gorithmic simplicity. The major disadvantage of algebraic VOF method investigated the flow pattern and hydrodynamic parameters of pile
is that the quality of numerical solutions will deteriorate owing to breakwater under solitary wave, where the interface was captured by
the large skewness and non-orthogonality on the unstructured meshes, THINC/QQ scheme as well. Xiao et al. (2022) added a mass transport
which results in a variety of variant of CICSAM (compressive interface model into VOF equation to simulate the cavitation flows, detailed
capturing scheme for arbitrary meshes) scheme such as the work of structures of which were resolved by high accuracy interface capturing
Denner and van Wachem (2014) and Zhang et al. (2014). scheme. Despite considerable success gained above, THINC schemes
Being an alternative candidate of VOF method, Tangent of Hy- should be further developed for practical applications with other ad-
perbola INterface Capturing (THINC) scheme was first proposed on vanced techniques such as arbitrary Lagrangian–Eulerian (ALE) and
one-dimensional building block by Xiao et al. (2005), who represent adaptive mesh refinement (AMR) methods.
the interface jump by using a step-like hyperbolic tangent function AMR technique is particularly attractive for the multiphase simula-
in each control volume. Neither explicit geometric manipulation nor tions of multi-scale dynamics such as spray atomization, which contains
special numerical treatment, THINC scheme effectively maintains the both larger and smaller flow structures. It provides a favorable ap-
transition zone of interface jump by adopting a steepness parame- proach to maximize the grid resolution in the vicinity of interface while
ter to eliminate the numerical diffusion, which delivers competitive simultaneously keeping the number of grids relatively small so as to
lessen the computational burden. For decades, a variety of multiphase
numerical accuracy than geometric VOF method without loss of sim-
solvers have been developed by combining AMR strategy and either
plicity. Via multi-dimensional hyperbolic tangent function, it has been
geometric or algebraic scheme for interface capturing. For example,
developed by Ii et al. (2012, 2014) and Xie et al. (2014b) to re-
the well known flow solvers ‘‘Gerris’’ and ‘‘Basilisk’’ were developed
construct interfaces on multi-dimensional unstructured meshes which
by Popinet (Popinet, 2003, 2009) and widely used for the simulations
includes geometric information such as orientation and curvature of
of two-phase DNS flows, which were implemented based on operator
interface, into the reconstruction of THINC function. As the up-to-
splitting approach and thus not directly applicable to unstructured
date version of THINC scheme, THINC method with quadratic surface
meshes. For unstructured grids, interFoam solver was combined with
representation and Gauss quadrature (THINC/QQ) was proposed by
AMR technique and used to simulate various multiphase flows, such
Xie and Xiao (2017) to improve the geometrical faithfulness of curved
as isothermal film flow (Cooke et al., 2014), cavitation flow around
surface by using quadratic function in the THINC formulation, which
Clark-Y hydrofoil (Wang et al., 2020) or cylindrical configuration with
can be straightforwardly applied for 2D and 3D standard elements
a hemispherical head (Li et al., 2020) as well as spray problems (Kuo
of unstructured grids including triangles, quadrilaterals, tetrahedrons,
and Trujillo, 2021). As the interface capturing method of interFoam,
hexahedrons and pyramids. Different from the previous work based
MULES (Damián and Nigro, 2014) scheme, however is found to suffer
on local coordinate, the THINC/QQ scheme on the global coordinate
from severe distortions and ripples which significantly degrades the
is developed by Xie et al. (2020a) and further improved by Chen
solution quality particularly for unstructured grids (Deshpande et al.,
et al. (2022) for polyhedral unstructured grids, which demonstrates the 2012). Moreover, due to the absence of geometrical information of
higher accuracy and geometric faithfulness compared with other VOF interface, the mapping of volume fraction in the child cells after mesh
methods for polyhedral grids. refinement is hard to be realized more accurately other than taking the
Given the advantages of THINC formulation for interface capturing, average values of parent cells straightforwardly. This approach ensures
a variety of multiphase solvers have been developed based on THINC- mass conservation while brings large interface deformation at each time
type schemes in the framework of OpenFOAM, which are verified of mapping, which will accumulate with large numerical errors after
and validated by the simulations of some complicated flows such long duration computation. On the other hand, it is easier to handle
as interaction between wave and structures, breakup of wave and the mapping of volume fraction for geometric VOF method by consid-
dam-break with obstacle. Nakayama with two of our authors, i.e. , ering position of reconstructed interface. Marić et al. (2013) proposed
Nakayama et al. (2015) first developed a new solver and named it an advanced geometric VOF algorithm with support of unstructured
thincFoam by using the UMTHINC scheme (Xie et al., 2014b) on the meshes and AMR strategy, which renders the execution of geometric
unstructured grids with quadrilateral and hexahedral elements. It has mapping of volume fraction by geometrical intersection in child cells.
been validated by two test cases including low speed liquid jet and The work done by Marić et al. (2013) demonstrates high accuracy
Plateau–Rayleigh instability, which shows the good capability for the and reliability on unstructured grids even for large CFL number and
gas–liquid two-phase flow simulation. Then, Kamra and Hu combined complicated velocity field. At the same time, Dai and Tong (2022) sim-
it with Reynolds averaging approach for turbulence model and per- ulated cavitation flows by making use of adaptive PLIC-VOF method,
formed numerical simulations about violent sloshing flows (Hu and where similar as Marić et al. (2013), geometric mapping algorithm
Kamra, 2020) and dam-break flow(Kamra et al., 2018). The numerical based on PLIC scheme (Dai and Tong, 2018, 2019) was used to map
results demonstrate good agreement with experimental records, which the volume fraction after mesh refinement. Moreover interflow was
verifies the excellent performance of thincFoam on 3D unstructured developed by Roenby et al. (2016) based on isoAdvector scheme with
grids. In addition, Zhang et al. (2019) build a numerical tank based support for any types of grids including polyhedral. The interflow is
on Navier–Stokes equations by using VPM (Xie et al., 2014a; Xie and capable of simulating multiphase flow with AMR, mapping for volume
Xiao, 2014) and THINC/QQ (Xie and Xiao, 2017) schemes. It is found fraction of which is accomplished by considering the correct position
that this model significantly improves the dissipation properties of the of PLIC interface in parent cell (Scheufler, 2019). Simultaneously as
propagating wave even with larger time steps and coarse grids. Xie well, Laurila et al. (2019) simulated a pressure-swirl atomizer and

2
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

adequately capture the motion of liquid interface near the nozzle by and momentum equation,
making use of isoAdvector scheme and AMR, where the mapping error 𝜕(𝜌𝐮)
was minimized by applying remeshing strategy. Compared with alge- + ∇ ⋅ (𝜌𝐮𝐮) = −∇𝑝 + ∇ ⋅ (𝜇∇𝐮) + ∇𝐮 ⋅ ∇𝜇 + 𝜌𝐠 + 𝜎𝜅∇𝜙, (3)
𝜕𝑡
braic mapping with piecewise constant, geometric mapping schemes where volume fraction 𝜙 = 𝑉𝑤𝑎𝑡𝑒𝑟 ∕𝑉 ∈ [0, 1] denotes the proportion of
improves geometrical faithfulness of mapped interface at the expense water in control volume, 𝐮 is velocity field, 𝐠 is acceleration of gravity,
of larger algorithmic complexity and computational cost. In addition, 𝜅 is the interface curvature and 𝜎 is surface tension coefficient.
the linear reconstruction of PLIC-type scheme makes it intrinsically In general, the piezometric pressure is defined as 𝑝rgh = 𝑝 − 𝜌𝐠 ⋅ 𝐱 by
difficult to represent the curved surfaces with large curvature. For subtracting hydrostatic pressure from the total one, which results in
example, the mapped interface of sphere looks like polyhedron which 𝜕(𝜌𝐮)
will bring substantial numerical errors for the approximation of surface + ∇ ⋅ (𝜌𝐮𝐮) = −∇𝑝rgh + ∇ ⋅ (𝜇∇𝐮) + ∇𝐮 ⋅ ∇𝜇 − 𝐠 ⋅ 𝐱∇𝜌 + 𝜎𝜅∇𝜙. (4)
𝜕𝑡
tension. Being the hybrid geometric/algebraic type VOF approach,
The density and viscosity of fluid in the vicinity of interface between
THINC formulation provides an attractive choice to devise a more water and gas are computed from the volume fraction, that is
sophisticated mapping scheme by making use of quadratic interface
function. Along this direction, Hu and Liu (2018) proposed a multi- 𝜌 = 𝜌H 𝜙 + 𝜌L (1 − 𝜙), 𝜇 = 𝜇H 𝜙 + 𝜇L (1 − 𝜙). (5)
phase solver with AMR method based on THINC/SW scheme by using
where subscripts H and L represent heavy and light fluid respectively.
dimensional splitting strategy which cannot be extended to curved
surfaces and unstructured grids due to its feature of 1D formulation 2.2. THINC/QQ scheme with normalization on global coordinate
of hyperbolic tangent function.
In this paper, towards arriving at higher accuracy and efficiency, For water–gas two-phase flows, the indicator function of volume
we present a novel multiphase solver called thincFoam based on the fraction can be described as
state-of-the-art THINC/QQ scheme and AMR technique which can be ⎧1 water,
seen as the successive work of Nakayama et al. (2015). In this solver, ⎪
 (𝐱, 𝑡) = ⎨0 gas, (6)
THINC/QQ scheme is implemented on the global coordinate so as to
⎪
facilitate the computation of numerical fluxes on the created faces ⎩𝜖 (0 < 𝜖 < 1) interface.
during refinement. In order to make consistent interpolation after The computational domain is divided into a set of non-overlapping
mesh refinement, we propose a novel algebraic mapping algorithm of polyhedral cells with arbitrary number of faces and vertices. According
volume fraction field by taking advantage of quadratic polynomial of to finite volume method, the volume integration average of volume
THINC/QQ scheme. The major merits of this scheme is to preserve fraction on the target cell 𝛺𝑖 is defined as
quadratic interface during the mapping process which outperforms
1
geometric mapping approach of PLIC-type VOF schemes in terms of 𝜙̄𝑖 (𝑡) =  (𝐱, 𝑡) d𝐱, (7)
|𝛺𝑖 | ∫𝛺𝑖
solution fidelity particularly for curved surface. The volume fraction
of each child cell is determined by simply computing integration of where |𝛺𝑖 | is the volume of target cell. Considering the discontinu-
hyperbolic tangent function in its parent cell, which makes it as ef- ous property of  (𝐱, 𝑡), we piecewisely approximate it by using the
ficient as algebraic mapping scheme without geometric operations. multi-dimensional hyperbolic tangent function which can be written by
As a result, the proposed mapping scheme significantly improves the
1( ( ( )))
geometrical faithfulness of mapping interface while retains algorith- ̃ 𝑖 (𝐱, 𝑡) = 1 + tanh 𝛽 𝑃𝑖 (𝐱) + 𝑑𝑖 . (8)
2
mic simplicity and computational efficiency. Besides, the second-order
temporal and spatial discretization is used to further improve the The jump steepness of transition region is controlled by the parameter 𝛽
which is given as 6.0∕𝛥 with 𝛥 being the characteristic length of target
solution accuracy. By means of AMR method, the grid in the vicinity
cell by default. 𝑃𝑖 (𝐱) is a quadratic polynomial and used to represent
of interface is automatically refined while unchanged for those regions
interface by
that are less concerned about, which largely saves the computational
resource. Verified by various benchmark tests, the present solver is able 𝑃𝑖 (𝐱) = 𝐶200 𝑋 2 + 𝐶020 𝑌 2 + 𝐶002 𝑍 2 + 𝐶110 𝑋𝑌 + 𝐶101 𝑋𝑍 + 𝐶011 𝑌 𝑍
(9)
to ensure the higher accuracy and geometric fidelity while improving + 𝐶100 𝑋 + 𝐶010 𝑌 + 𝐶001 𝑍,
the computational efficiency significantly in multiphase simulations,
in which 𝑋, 𝑌 , 𝑍 denotes 𝑥 − 𝑥𝑖𝑐 , 𝑦 − 𝑦𝑖𝑐 and 𝑧 − 𝑧𝑖𝑐 with (𝑥𝑖𝑐 , 𝑦𝑖𝑐 , 𝑧𝑖𝑐 )
especially for 3D large-scale computation.
being the cell center. The coefficients 𝐶𝑖𝑗𝑘 can be approximated from
The rest of this paper is organized as follows. Firstly the numer-
the unit normal vector and curvature of interface based on the least-
ical methods of present solver, such as THINC/QQ scheme, mapping square approach. The unique unknown 𝑑𝑖 is determined from the given
algorithm and solution procedure, are described in Section 2. Then volume fraction following constraint condition
Section 3 presents a set of benchmark tests that are carried out to
1
examine the performance of proposed mapping scheme and numerical ̃ (𝐱, 𝑡)d𝐱 = 𝜙̄ 𝑖 . (10)
|𝛺𝑖 | ∫𝛺 (𝐱) 𝑖
solver. At last, some concluding remarks are drawn in Section 4 to end | | 𝑖

this paper. Once the THINC function is piecewisely reconstructed in each cell,
we update the volume fraction by solving Eq. (1) which can be recast
into
2. Numerical methods
( )
1 ∑ ∑
𝐽 𝐺
𝜕 𝜙̄ 𝑖 1
2.1. Governing equations =− ̃ 𝑖𝑢𝑝 𝐮 ⋅ d𝐒 = − 𝑢𝑛𝑖𝑗 𝛤𝑖𝑗 𝜔𝑔𝑖𝑗 ̃ 𝑖𝑢𝑝 (𝐱𝑔𝑖𝑗 ) ,
𝜕𝑡 |𝛺𝑖 | ∫𝛤𝑖𝑗 |𝛺𝑖 | 𝑗=1 𝑔=1

For the incompressible two-phase flows, based on one-fluid model, (11)

the governing equations include the VOF equation, where 𝑢𝑛𝑖𝑗 is velocity projected onto the outward normal vector of
𝜕𝜙 surface segment 𝛤𝑖𝑗 . The numerical flux in Eq. (11) is computed by
+ 𝐮 ⋅ ∇𝜙 = 0, (1)
𝜕𝑡 making integration on the THINC function in upstream cell 𝛺𝑖𝑢𝑝 , in
the continuity equation, which 𝜔𝑔𝑖𝑗 and 𝐱𝑔𝑖𝑗 denote weights and coordinates of quadrature
points on boundary surface belonging to 𝛺𝑖𝑢𝑝 . The interested reader can
∇ ⋅ 𝐮 = 0, (2) refer to Chen et al. (2022) for more details.

3
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

2.3. Algebraic mapping of volume fraction field for AMR for the mapping of curved interface. Therefore, we propose an algebraic
mapping algorithm by making use of the quadratic reconstruction of
2.3.1. Adaptive mesh refinement with load balancing interface based on THINC/QQ formulation. The mapping algorithms
To increase the spatial resolution where necessary, we apply h- mentioned above are denoted by algebraic mapping with piecewise
adaptive approach in the vicinity of interface which is implemented constant (amPC), geometric mapping with linear interpolation (gmLI),
in OpenFOAM via the AMR library developed by Rettenmaier et al. and algebraic mapping with quadratic interface (amQI) hereafter, and
(2019). It is adaptive for mesh refinement including unstructured grids plotted in Fig. 1 for refinement of a quadrilateral cell in 2D. It is
of 2D quadrilateral and 3D hexahedral elements, which can deal with evident that the present amQI algorithm substantially improves geo-
flow domain of arbitrary geometrical complexity by changing overall metrical fidelity of interface mapping in particular for curved surfaces,
mesh connectivity information. It provides some new features to fa- without involving geometric operations and thus ensuing algorithmic
cilitate the implementation on the unstructured grids, such as mesh simplicity.
splitting of cell-type other than hexahedron, the multi-criterion of mesh The core idea of amQI algorithm is to map the volume fraction
refinement as well as dynamic load balancing in parallel computation, from the THINC function with quadratic surface representation. Given
which will be described in brief as follows. the volume fraction 𝜙𝑖 in each parent cell, the interface is implicitly
(1) The mesh splitting is realized by the class of hexRef with three retrieved from THINC function and expressed as 𝑃 (𝐱) + 𝑑 = 0 given
derived classes, i.e. hexRef8, hexRef4 and hexRef4Axi, which adds in Eq. (9), in which its coefficients 𝐶𝑖𝑗𝑘 (𝑖, 𝑗, 𝑘 ∈ 0, 1, 2) and interface
the function to handle quadrilateral and prismatic elements for 2D and position 𝑑 are determined from algebraic calculation and stored for
axisymmetric cases by using quad-tree refinement respectively. It is reuse after mesh refinement. Then the volume fraction of child cell (𝛺𝑖𝑠 )
called by the class of dynamicRefineFvMesh which selects appro- is computed by making numerical integration of THINC function on the
priate quad/octree refinement based on the number of dimensions of sub-cell as
solution.
(2) The multi-criterion refinement class is included as a modular 1 ∑𝐺
1( ( ( )))
𝜙𝑠𝑖 = ̃ 𝑖 (𝐱, 𝑡) d𝐱 = 1 + tanh 𝛼𝛽 𝑃 (𝐱𝑔𝑖𝑠 ) + 𝑑 , (13)
plugin to dynamicRefineFvMesh, which combines different criteria |𝛺𝑖 | ∫𝛺𝑠
𝑠
𝑔=1
2
𝑖
such as selectable fields, their gradients and curls, interface and geo- ( )1∕𝑚
metrical features as well as other individual criterion. Compared with where 𝑠 = 4 for 2D and 𝑠 = 8 for 3D, 𝛼 = 𝛺𝑖 ∕𝛺𝑖𝑠 is a scaling factor
single criterion in the official version, it offers higher flexibility to select to control the width of interface with 𝑚 being dimensionless, in which
cells for refinement and coarsening for complicated flow problems. 𝛺𝑖 and 𝛺𝑖𝑠 are volume of parent and child cells respectively. In our
(3) The derived class of dynamicRefineBalancedFvMesh is added work, we set 𝑚 as 2 in 2D or 3 in 3D based on numerical feedback. The
to realize the dynamic load balancing among each processors in parallel mapping algorithm of amQI is described in Algorithm 1, which includes
for a transient problem. This approach redistributes the number of ele- two central steps: identification of refined cells and computation of
ments in each processor according to the evaluation of load imbalance volume fraction in the child cell. It is noted that OpenFOAM provides
rate which is computed by a special class mapPolyMesh to contain mesh-to-mesh mapping infor-
|𝑁𝑟𝑝 − 𝑁𝑖𝑝 | mation after local dynamic refinement, which facilitates the execution
𝑁𝑒
𝐿= , 𝑁𝑖𝑝 = , (12) of algebraic mapping of volume fraction based on THINC/QQ scheme.
𝑁𝑖𝑝 𝑁𝑝
In mapPolyMesh class, cellMap() accesses to information from current
where 𝑁𝑖𝑝 and 𝑁𝑟𝑝 are the ideal and realistic number of elements in mesh back to previous mesh, instead reverseCellMap() has an opposite
each processor, 𝑁𝑒 is total number of element, and 𝑁𝑝 is the number function.
of processors in parallel.
Algorithm 1 Algebraic mapping of volume fraction using THINC/QQ
2.3.2. Algebraic mapping algorithm based on THINC/QQ scheme scheme for AMR
In this paper, we realize the high-fidelity computation of moving Input: polynomial coefficients 𝐶𝑖𝑗𝑘 , interface position 𝑑, center
interface by making the combination of THINC/QQ scheme and AMR coordinate 𝐱𝑐 , VOF field 𝜙𝑜𝑙𝑑
approach, which refines grids in the vicinity of interface automatically 1: for cellI ∈ reverseCellMap() do
to improve the numerical accuracy and solution resolution. Part of cells 2: oldCellI ← cellMap()[cellI]
( ) ( )
after refinement will be identified as polygonal/polyhedral cell in the 3: if |𝑉 [cellI] − 𝑉𝑜𝑙𝑑 [cellI]| > 10−15 ∩ 𝜖 ≤ 𝜙𝑜𝑙𝑑 ≤ 1 − 𝜖 then
OpenFOAM which can be readily handled by the THINC/QQ scheme on 4: Remark isRefinedCell[cellI] ← True
the global coordinate without using any extra treatment like Li et al. 5: Recalculate Gaussian quadrature points 𝐱𝑔 and weights 𝜔𝑔
(2022). Since the connectivity information and mesh data structures are 6: Compute volume fraction of sub-grids 𝜙𝑠𝑖 by Eq. (13)
changed after each refinement cycle, it is necessary to re-allocate the 7: end if
memory to store the mesh data and re-calculate the coefficient matrices 8: end for
that are used in the THINC formulation. Despite that these operations Output: volume fraction of sub-grids 𝜙𝑠𝑖
bring some extra computational cost, it only occupies small portion of
the overall time involved in the flow simulations considering a huge
number of static grids with the same resolution. 2.4. Solution procedure of flow simulation
Another critical issue associated with THINC/QQ scheme in AMR
is how to make consistent interpolation of volume fraction between In the framework of finite volume method, given the volume frac-
parent and refined child cells. For algebraic VOF scheme that does not tion 𝜙𝑛𝑖 , velocity 𝐮𝑛𝑖 in each cell at time level 𝑛, we predict the solution
contain geometric information of interface, volume fraction in the child of volume fraction 𝜙𝑛+1 and velocity 𝐮𝑛+1 at time level 𝑛 + 1 by solving
cell simply receives the cell-centered value of the parent cell which is Navier–Stokes equation with pressure-implicit with splitting of opera-
efficient but largely degrades the solution accuracy. As shown later, the tor (PISO) algorithm (Issa, 1985) in which second-order Runge–Kutta
mapped interface after several cycles of refinement is found with large (RK2) method is used for the temporal discretization of all prognostic
deformations due to the mapping errors of this approach. On the other variables like Sussman (2003). Numerical procedures for the solution
hand, the geometrical mapping is presented to calculate the volume of Eqs. (1)∼(3) from 𝑡 = 𝑛 to 𝑡 = 𝑛 + 1 are summarized as follows.
fraction in the child cells according to the interface which is recon-
structed by a piecewise straight line (2D) or a plane (3D) segment. This 1. Initialize the solution variables 𝐮𝑛 and 𝜙𝑛 , and refine the mesh in
algorithm improves the geometric faithfulness but still have downside the vicinity of interface.

4
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 1. Diagram of different mapping of volume fraction for AMR in 2D. It should be noted that more quadrature points are arranged in child cells for the present mapping
algorithm (amQI) in practice.

2. Map the volume fraction in child cell by amQI in Algorithm 1, 5. PISO algorithm is performed to solve pressure 𝑝𝑛+1,𝑟 and velocity
and other variables by amPC for the sake of simplicity. 𝐮𝑛+1,𝑟 in each stage of RK2 iteration. At first, we predict the
3. Start the RK2 iteration with 𝑟 = 0 and compute the volume intermediate velocity 𝐮𝑛+1,∗ by solving momentum equation ne-
fraction 𝜙𝑛+1,𝑟 at the sub-stage by glecting the pressure gradient, surface tension and gravity terms
( ) as written by
𝛥𝑡 ∑ ∑
𝐽 𝐺
𝜙 𝑛+1,𝑟 𝑛
=𝜙 − 𝑢 𝛤 ̃
𝜔  (𝐱 ) , (14) ∑
|𝛺𝑖 | 𝑗=1 𝑛𝑖𝑗 𝑖𝑗 𝑔=1 𝑔𝑖𝑗 𝑖𝑢𝑝 𝑔𝑖𝑗 𝑎𝑃 𝐮∗𝑃 = 𝑎𝑁 𝐮𝑁 + 𝑆𝑝𝑛 = 𝐻(𝐮), (18)

where 𝑟 equals either 0 or 1 which denotes the sub-iteration of where subscript 𝑃 and 𝑁 represent the cell in question and its
RK2 scheme. neighbors respectively. 𝑎𝑃 is the diagonal part of the momentum
4. Update the density 𝜌 and viscosity 𝜇 from the volume fraction by coefficient matrix, i.e. the diagonal entries of velocity, conversely,
𝑎𝑁 the off-diagonal entries of velocity. The operator 𝐻(𝐮) denotes
the source terms including extra source from time derivative 𝑆𝑃𝑛 .
𝜌𝑛+1,𝑟 = 𝜌H 𝜙𝑛+1,𝑟 + 𝜌L (1 − 𝜙𝑛+1,𝑟 ), 𝜇𝑛+1,𝑟 = 𝜇H 𝜙𝑛+1,𝑟 + 𝜇L (1 − 𝜙𝑛+1,𝑟 ), Given the pressure gradient, surface tension and gravity terms
neglected in Eq. (18), then it arrives at
(15)
( )
𝑎𝑃 𝐮∗𝑃 = 𝐻(𝐮) + 𝜎𝜅∇𝜙𝑛+1,𝑟 − 𝐠 ⋅ 𝐱∇𝜌𝑛+1,𝑟 − ∇𝑝rgh , (19)
𝜌𝑛+1,𝑟
𝑖𝑗 = 𝜌H 𝜙𝑛+1 𝑛+1,𝑟
𝑖𝑗 + 𝜌L (1 − 𝜙𝑖𝑗 ), 𝜇𝑖𝑗𝑛+1,𝑟 = 𝜇H 𝜙𝑛+1,𝑟
𝑖𝑗 + 𝜇L (1 − 𝜙𝑛+1,𝑟
𝑖𝑗 ). which can be solved either explicitly or implicitly. Enforcing
(16) divergence-free condition on the both side of Eq. (19), then the
pressure Poisson equation is obtained as
The volume fraction at surface 𝜙𝑛+1,𝑟
is computed by
𝑖𝑗 ( ) ( )
(𝐺 ) 1 𝐻(𝐮∗ ) 1 ( )
∇⋅ ∇𝑝∗rgh = ∇ ⋅ + 𝜎𝜅∇𝜙𝑛+1,𝑟 − 𝐠 ⋅ 𝐱∇𝜌𝑛+1,𝑟 .
𝑛+1,𝑟 1 ∑ ( ) ∑ 𝐺
( ) 𝑎𝑝 𝑎𝑝 𝑎𝑝
𝜙𝑖𝑗 = ̃
𝑖 𝐱𝑔 + ̃
𝑗 𝐱𝑔 , (17)
| |
2 |𝛤𝑖𝑗 | 𝑔=1 𝑔=1 (20)
| |
where ̃ 𝑖 and ̃ 𝑗 denote the THINC function in the neighboring which should be repeated for two or more steps to couple the
cells of 𝛤𝑖𝑗 . velocity and pressure. Finally, the velocity is corrected to satisfy

5
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 2. A diagram for the whole solution procedure of thincFoam.

the continuity equation by using the pressure obtained above 3. Numerical results and discussions
( )
1
𝐮∗∗ = 𝐻(𝐮) − ∇𝑝∗rgh − 𝐠 ⋅ 𝐱∇𝜌𝑛+1,𝑟 + 𝜎𝜅∇𝜙𝑛+1,𝑟 , (21) In this section, we present a series of benchmark tests to investigate
𝑎𝑝
the proposed solver with AMR technique and evaluate its performance
After the PISO iteration loop, velocity and pressure are put for-
by making comparisons with other existing results.
ward to the converged solutions at 𝑟 stage, which can be ex-
At the first step of computation, we initialize the volume fraction
pressed as 𝐮𝑛+1,𝑟 = 𝐮∗∗ , 𝑝𝑛+1,𝑟 = 𝑝∗ .
field by making use of level-set function so as to take the width of
6. Repeat the step 3∼5 for RK2 iteration and store the solution fields
interface into account, which makes it more consistent with THINC
of volume fraction 𝜙𝑛+1,𝑟 , velocity 𝐮𝑛+1,𝑟 and pressure 𝑝𝑛+1,𝑟 for
formulation. A sophisticated procedure is described as follows. The
reuse. ( )
level-set function 𝛹 𝐱𝑠 is first obtained by calculating the minimum
7. Average the volume fraction, velocity and pressure in each RK2
distance between the centroid of target cell and the interface. Then it
time stage, hence
is used to compute the volume fraction in each cell by taking average
1 ( 𝑛+1,0 ) 1 ( 𝑛+1,0 ) of THINC function with numerical integration as
𝜙𝑛+1 = 𝜙 + 𝜙𝑛+1,1 , 𝐮𝑛+1 = 𝐮 + 𝐮𝑛+1,1 ,
2 2
(22)
1 ∑1(
𝑆
1 ( 𝑛+1,0 ) ( ( )))
𝑝𝑛+1 = 𝑝 + 𝑝𝑛+1,1 . 𝜙̄𝑖 = 1 + tanh 𝛽𝛹 𝐱𝑠 , (23)
2 𝑁𝑠 𝑠=1 2

in which 𝛽 is parameter used in solution of VOF function before, 𝐱𝑠 is

Remark.
coordinates of sample points. It is noted that more sample points should
(1) In velocity–pressure solution of step 5, it should be noted
be used to refine the volume fraction in the vicinity of interface.
that Gauss linear and Gauss linear corrected schemes are adopted for
gradient term and Laplacian term respectively, and vanLeer is applied
3.1. Interface reconstruction of circle and sphere after mesh refinement
for the advection term, which provides second-order accuracy for the
numerical solution of momentum equations.
We reconstruct 2D circle and 3D sphere to verify the algebraic
(2) The moving interface is captured by THINC/QQ scheme in thinc-
mapping of volume fraction in AMR that has been described in Sec-
Foam, which is superior to the MULES scheme used in interFoam solver
tion 2.3.2. A circle with radius of 0.25 is centered at (0.5, 0.5) in unit
on numerical accuracy and geometric fidelity of capturing moving
square computational domain [0, 1]2 . The initial grid configuration is
interface with large-distortions and topological changes, especially for
10 × 10, and then grids in the vicinity of interface (𝜙 ∈ [10−3 , 1 − 10−3 ])
those grids with large skewness.
are refined by adaptive mesh refinement. Fig. 3 shows the interface
(3) thincFoam solver is developed with second-order RK2 time
reconstructed by amQI algorithm for refinement level 𝑅 = 1 by using
integration scheme, which has distinct advantages on the solution accu-
THINC formulation in the child cells. As a comparison, the interface
racy for prognostic variables of velocity, pressure and volume fraction
reconstructed with amPC and gmLI are displayed as well. It is seen that
obtained by interFoam solver when the first-order Euler scheme is used.
the interface mapped with gmLI and amQI algorithm all prevent inter-
In order to make readers comprehensible, we also elaborate a flow face from worsening after mesh refinement since these two algorithms
chart for the whole solution procedure of thincFoam for each time step take the geometrical information of interface such as normal vector
in Fig. 2. at the previous time step into full consideration. Then, we increase

6
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 3. Visualized interface represented by 𝜙 = 0.5 using algebraic mapping with piecewise constant (amPC), geometric mapping with linear interpolation (gmLI) and algebraic
mapping with quadratic interface (amQI) of volume fraction for refinement level 𝑅 = 1. The red and black line denote reference solution and numerical solution respectively. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 4. Volume fraction contour calculated by algebraic mapping with piecewise constant (amPC), geometric mapping with linear interpolation (gmLI) and algebraic mapping with
quadratic interface (amQI) of volume fraction for refinement level 𝑅 = 4. The black line represents the analytical interface. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

the refinement level to 𝑅 = 4 and plot the volume fraction contour Table 1
A comparison of numerical errors 𝐸𝑠 for various mapping algorithms of volume fraction
mapped with amPC, gmLI and amQI algorithms after mesh refinement
and refinement level 𝑅 in 3D sphere test.
as shown in Fig. 4. The results indicate that amPC algorithm produces
Algorithm 𝑅=1 𝑅=2 𝑅=3 𝑅=4
extra numerical errors to degrade the solution fidelity as the refinement
amPC 6.141 × 10−3 7.639 × 10−3 8.330 × 10−3 8.676 × 10−3
level increases, while amQI scheme can well preserve the geometrical
gmLI 1.121 × 10−3 1.141 × 10−3 1.070 × 10−3 1.869 × 10−3
fidelity by making use of quadratic reconstruction of interface in THINC amQI 1.876 × 10−4 1.829 × 10−4 1.753 × 10−4 1.895 × 10−4
formulation, which yields a more proper distribution of volume fraction
than amPC. Another point should be noted that gmLI algorithm makes
an remarkable improvement in the geometrical faithfulness of interface
shape compared with amPC. On the other hand, the mapped interface deteriorating as refinement level increases, which can also be proved
is not as smooth as that of amQI, which will bring in extra numerical by the numerical errors in Table 1 defined as
errors for the curvature estimation particularly in the case of curved ∑
𝑁
surfaces. Throughout the reconstruction of circle, it is deemed that the 𝐸𝑠 = |𝜙𝑛𝑖 − 𝜙𝑒𝑖 ||𝛺𝑖 |, (24)
interface mapped with amQI algorithm shows better agreement with 𝑖=1
the reference solution than amPC and gmLI. in which |𝛺𝑖 | denotes the area (2D) or volume (3D) of target cell.
We carry out similar test in 3D domain for a 𝑅 = 0.25 sphere Compared with amPC algorithm, the interface mapped with gmLI
centered at (0.5, 0.5, 0.5). Various refine levels and mapping algo- algorithm shows improved geometrical faithfulness, where the given
rithms are applied for this case and the mapped interface of sphere numerical errors are approximately one seventh of amPC mapping
is visualized in Fig. 5. More quantitatively, Table 1 makes a rigorous algorithm. However, it is still inferior to the results computed by the
comparison about numerical errors for different mapping algorithms proposed amQI algorithm with respect to numerical errors and geo-
and refinement levels. It is seen that the interface calculated by amPC metrical fidelity. Due to quadratic reconstruction of interface in THINC
algorithm almost loses its original shape for the larger refinement formulation, the amQI scheme shows remarkable capability to preserve
level. Additionally, the interface shape mapped with amPC will keep the shape and solution smoothness of the curved surfaces, in which the

7
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 5. The visualized sphere after mesh refinement using algebraic mapping with piecewise constant (amPC), geometric mapping with linear interpolation (gmLI) and algebraic
mapping with quadratic interface (amQI) of volume fraction for different refinement levels.

level of numerical errors nearly reduce by one magnitude against amPC initial volume fraction is given as a circle with radius of 0.15, center of
and gmLI mapping algorithm. It convinces that algebraic mapping which is (0.5, 0.75). The volume fraction is set as unit if the coordinates
throughout THINC function with quadratic interface representation is reside into the circle. The transported velocity field is expressed by
equipped with higher numerical accuracy and geometrical fidelity, as ( )
well as algorithmic simplicity. In a word, all numerical results demon- ⎧ 2 2 𝜋𝑡
⎪𝑢 = 𝜋 sin (𝜋𝑥)sin(𝜋𝑦)cos(𝜋𝑦) cos 𝑇
strate the validity of present mapping algorithm of volume fraction in ⎨ ( ) (25)
2D and 3D computation, which guarantees the accuracy of numerical ⎪𝑣 = − 𝜋2 sin(𝜋𝑥)cos(𝜋𝑥)sin2 (𝜋𝑦) cos 𝜋𝑡 ,
⎩ 𝑇
solution for two-phase flows using THINC scheme and AMR method.
where 𝑇 = 8 is one computational period. The circle is stretched by the
3.2. 2D shear vortex flow given velocity field and become a spiral curve with a thin tail in the
first half of cycle. Then under the action of reversed velocity, the circle
2D shear vortex flow (Rider and Kothe, 1998) is widely used to will return back to its initial shape theoretically, which is used as the
evaluate the performance of different interface capturing methods. The exact solution.

8
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 6. Two kinds of computational grids (left: gridA, right: gridB) for 2D shear vortex flow. It should be noted that these grids are illustrative and supposed to refined to 1282 ∕2562
in practice.

Fig. 7. The interface of 2D shear vortex flow by THINC/QQ (left) and MULES (right) schemes on gridA with resolution of 1282 , in which black line is analytical solution, red
(blue) line is numerical solution at 𝑡 = 4 (𝑡 = 8). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

3.2.1. Comparison between THINC/QQ and MULES results computed by MULES scheme. The results demonstrate that
In order to verify the capability of interface capturing, we first THINC/QQ scheme has higher numerical stability on unstructured grids
carried out the 2D shear flow by using the THNC/QQ scheme on two and adequate resolvability for tiny fluid structure. At time instants of
types of unstructured grids which are featured with large skewness and 𝑡 = 𝑇 , the interface predicted by THINC/QQ scheme agrees with the
non-orthogonality as shown in Fig. 6. The resolution of triangular grid initial shape very well, which delivers higher numerical accuracy and
is 1282 , and the polygonal grid is transformed from triangular grid with geometric fidelity than MULES scheme even on the unstructured grids
2562 in configuration. For comparison, this case is reproduced under with large skewness and non-orthogonality.
the same configuration with MULES scheme which is the standard Quantitatively, we evaluate the numerical errors related to the
choice used by multi-phase solver in OpenFOAM. The RK42 method is shape deformation and volume conservation, which are defined as 𝐸𝑠
used to update the VOF equation for time integration with maximum in Eq. (24) and 𝐸𝑣 as
Courant number given as 0.15. Additionally, the steepness parameter
|∑ |
𝛽 of THINC function is set to 6.0∕𝛥 and artificial compression factor of |𝑁 ( ) |
𝐸𝑣 = || 𝜙𝑛𝑖 − 𝜙𝑒𝑖 𝛺𝑖 || . (26)
MULES is specified as 1.0 according to the recommendation of Cifani | 𝑖=1 |
et al. (2016) and Klostermann et al. (2013). | |
Figs. 7 and 8 display the interface on the non-orthogonal triangular The upper and lower bound of volume fraction are also included for
and polygonal grids. At the time instants of 𝑡 = 𝑇 ∕2, the interface THINC/QQ and MULES schemes on gridA. From the results in Table 2,
is largely stretched into a spiral with a thin tail which may break we can see that the shape error of THINC/QQ scheme is only one
into small droplets since the distorted tail becomes too thin to be twentieth that of MULES. It is also convinced that THINC/QQ scheme
resolvable by finite resolution of the grids. It is seen that the interface is capable to capture the large-deformed interface more accurately than
computed MULES scheme is observed with severe oscillations which algebraic MULES scheme on poor quality mesh. Apart from higher
is caused by the numerical instability on the non-orthogonal meshes. numerical accuracy, THINC/QQ can effectively control the solution
On the contrary, the interface is predicted with excellent solution boundedness without overshoot and undershoot, which is superior to
quality by THINC/QQ scheme which is apparently superior to the the results computed by the MULES scheme.

9
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 8. Same as Fig. 7, but on gridB with resolution of 2562 .

Table 2
The numerical errors computed by THINC/QQ and MULES on gridA for 2D shear flow. Table 3
Algorithm 𝐸𝑠 𝐸𝑣 min 𝜙 − 1 max 𝜙 The numerical errors computed by THINC/QQ on static and AMR grids.
NOE denotes the number of elements.
MULES 4.22 × 10−2 2.58 × 10−15 −2.03 × 10−4 2.00 × 10−4
THINC/QQ 2.15 × 10−3 4.38 × 10−10 0.0 0.0 Types of grids Grid resolution max(NOE) 𝐸𝑠
AMR grid 1/64 11389 1.33 × 10−3
static grid 1/256 65536 1.33 × 10−3

3.2.2. 2D shear vortex flow with AMR

In this section, we continue to make use of 2D shear flow to verify
the capability of interface capturing of THINC/QQ scheme with AMR cube. And the sphere is stretched by the following reversed velocity
approach, which is denoted by THINC/QQ-AMR in short. The time 𝐔 = (𝑢, 𝑣, 𝑤), in which
step is adjusted by maximum Courant number 0.15 and the grid is
automatically refined according to the refinement criteria that volume ⎧𝑢 = 2sin2 (𝜋𝑥)sin(2𝜋𝑦)sin(2𝜋𝑧) cos(𝜋𝑡∕𝑇 ),
⎪ 2
fraction fall into the range of [𝜖, 1 − 𝜖] with 𝜖 being 10−6 . Furthermore, ⎨𝑣 = −sin(2𝜋𝑥)sin (𝜋𝑦)sin(2𝜋𝑧) cos(𝜋𝑡∕𝑇 ), (27)
the refinement level is set as 2 and amQI algorithm is use to map the ⎪ 2
⎩𝑤 = −sin(2𝜋𝑥)sin(2𝜋𝑦)sin (𝜋𝑧) cos(𝜋𝑡∕𝑇 ).
volume fraction field after each mesh refinement.
The computational domain is divided into structured hexahedral
The interface is predicted by THINC/QQ-AMR for 2D shear flow and
grids with 323 resolution at 𝑡 = 0 instant and refinement level of
plotted at four time instants in Fig. 9. The resolution of background
grid equals 2. The refined criterion is the volume fraction falling in
grid is 642 , and the grid size is same as that of static grid with 2562
[10−6 , 1 − 10−6 ]. In the whole computation, the time step is adjusted
resolution after twice refinement. It is seen that the grid resolution is
by maximum Courant number of 0.5. Additionally, the 3D deforma-
refined in the vicinity of interface and remain unchanged otherwise,
tion flow is simulated in parallel with four cores. Fig. 11 displays
which largely improves the solution quality without decreasing the
the interface computed by THINC/QQ-AMR scheme along with the
numerical efficiency. There is no distinct difference on the numerical
refined grids for 3D deformation flow. At time 𝑡 = 𝑇 ∕2, the sphere
solution of interface shape between the AMR and finer static grids (see
is stretched into a very thin film with largest deformation which is
Fig. 10). The interface eventually converges to a shape of circle at the
challenging to be resolved by nearly all interface capturing schemes. As
final instant, which not only shows the strong robustness of THINC/QQ
shown in Fig. 11(b), THINC/QQ-AMR scheme is capable of resolving
scheme on the different types of grids, but also demonstrates the
the extremely thin film by locally refining the grid in the vicinity of
excellent capturing ability and geometric faithfulness for THINC/QQ-
interface, which arrives at adequate resolution with much less grid cells
AMR approach. Quantitatively, Table 3 provides the numerical errors of than the computation with static mesh.
2D shear flow on the static and AMR grids with different resolution. The In 3D case, the simulation is carried out with AMR in parallel based
maximum number of elements for AMR grid is only one sixth of that on domain decomposition, in which the changeable load on proces-
for static grid. In spite of less number of elements, the THINC/QQ-AMR sors may considerably reduce the numerical efficiency. The processor
delivers almost the same numerical error as the static grid. All above with largest load becomes a bottleneck when the data communica-
numerical results convince that the THINC/QQ-AMR scheme attains tion is required by other waiting processors. Therefore, we combine
adequate numerical resolution and solution accuracy of interface cap- the THINC/QQ-AMR with a load balancing algorithm as described in
turing while improving the computational efficiency, which has more Section 2.3.1 which dynamically redistributes the number of grid cells
advantages in the 3D computation. in each core so as to improve the numerical efficiency in the parallel
computation. Table 4 elaborates the numerical errors and CPU time for
3.2.3. 3D deformation flow with AMR different algorithms in 3D deformation flow where allowable imbalance
In order to further evaluate the capability of THINC/QQ-AMR rate is given as 10%. It is noted that the number of grids using AMR
scheme for interface capturing, we choose the representative 3D de- is decreased to 1.6 × 105 which is only 7% of number of static grid
formation flow (Enright et al., 2002) as another test case. Initially, a with 1283 resolution. The computational time is dropped to 75% of
sphere with radius of 0.15 is centered at (0.35, 0.35, 0.35) in a unit that without using load balancing algorithm. It is realized that the

10
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 9. The visualized interface computed by THINC/QQ with AMR at different instants for 2D shear flow.

Fig. 10. Enlarged interface of 2D shear flow at 𝑡 = 8 on static (left) and AMR (right) grids.

11
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 11. The visualized interface computed by THINC/QQ with AMR for 3D deformation flow.

Fig. 12. The evolution of Rayleigh–Taylor instability without surface tension at different time instants.

Table 4
The comparison of numerical errors and computational efficiency for different algorithms in 3D deformation flow.
The minimum grid size is 1/128, and maximum Courant number is 0.5. Symbol √ denotes AMR grid, but symbol 𝛥
denotes static grid. LB in this table illustrates that the dynamic load balancing strategy is adopted.
Algorithms AMR Number of processors Computational time (s) 𝐸𝑠
voFoam (Marić et al., √ – – 1.303 × 10−3
2013)
THINC/QQ 𝛥 64 778 1.110 × 10−3
THINC/QQ √ 4 2162 1.057 × 10−3
THINC/QQ+LB √ 4 1594 1.061 × 10−3

computational cost is largely reduced without loss of numerical accu- is over the lighter one. The balance between two phases is interrupted
racy. For comparison, we also include the numerical errors computed by some little interfacial turbulence. Then the interface is deformed
by voFoam (Marić et al., 2013) for this case, where geometric PLIC extremely due to the gravity and then develops into the feature of
scheme is used to capture the moving interface. The numerical error ‘‘mushroom cloud’’. This is a canonical test to evaluate the capability
of THINC/QQ-AMR is reduced by 20% compared with that of voFoam, of numerical solver in resolving complex flows with moving interface
which further demonstrates the strong capability of proposed scheme that undergoes large deformation and topological change.
to handle complex interfaces with adequate accuracy and efficiency. We set up our simulation in a rectangular computational region
of size 1 m × 4 m with center of (𝑥𝑐 , 𝑦𝑐 ) being (0.5, 2.0). The interface
3.3. Rayleigh–Taylor instability without surface tension between two fluids with different properties is placed in the middle
of the domain,( which )is imposed by an initial perturbation as 𝑦 =
As the first case of two-phase flow, Rayleigh–Taylor instability 𝑦𝑐 + 0.05 cos 2𝜋(𝑥 − 𝑥𝑐 ) . The density of the heavier fluid above the
problem is considered under certain circumstance that the heavier fluid interface 𝜌ℎ is 1.225 kg∕m3 , and dynamic viscosity 𝜇ℎ is 3.13×10−3 Pa ⋅ s;

12
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 13. The evolution of Rayleigh–Taylor instability without surface tension at different grids. The left part of subfigure is computed on AMR grid, and the other is on static grid.

For the lighter fluid, the density 𝜌𝑙 is 0.1694 kg∕m3 and dynamic viscos- step is determined by maximum Courant number 0.2 and the upper
ity 𝜇𝑙 is also 3.13 × 10−3 Pa ⋅ s. The gravity is given as 9.81 m∕s2 along limit of 1 × 10−4 .
the negative direction of 𝑦-axis and the effect of surface tension is not In this test, the mesh is refined for those cells whose volume fraction
considered here. The top and bottom sides of computational region are belong to [10−3 , 1 − 10−3 ]. It should be noted that this refined criterion
is applicable to the following numerical simulations of two-phase flows
considered as non-slip boundary condition, but free slip wall boundary
in 2D and 3D if no special statement.
condition is applied for the other. Both static and AMR grids are used The evolution of interface for Rayleigh–Taylor flow without surface
for simulation in this case. The resolution of static grid is 100 × 400, tension force at different time instants is shown in Fig. 12. It is seen that
and the background grid of AMR is 50 × 200, which is refined into the heavier fluid gradually penetrates into the lighter fluid under the
200 × 800 in the vicinity of interface after twice refinement. The time effect of gravity as time goes by. The interface is extremely stretched

13
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 14. The time histories of 𝑦 coordinate of the peak and trough interfacial points for Rayleigh–Taylor instability without surface tension.

efficiency. It is further demonstrated that AMR technique apparently

enhances the numerical solutions to capture the interfaces of film that
becomes too thin to be resolvable by the finite resolution of the grids.
Fig. 14 plots the time histories of 𝑦 coordinate of the peak and
trough interfacial points to evaluate the growth rate of interface in-
stability, which also includes the results of other numerical methods
for comparisons. The solid and dashed lines correspond the numerical
results computed by thincFoam on the uniform and AMR mesh, which
are well coincident with each other as well as reference solutions
for the interface height of peak in Fig. 14(a). With respect to the
trough interfacial point in Fig. 14(b), the numerical results obtained
by thincFoam are in good agreement with those of Haghshenas et al.
(2017) and Oevermann et al. (2000), and show lower interfacial height
than the results in Zuzio and Estivalezes (2011) as well as Popinet
(2009), in which the relative errors are evaluated at time 𝑡 = 0.9 s
below 0.1% and 0.8%, respectively. It is convinced that the present
solver that combines THINC/QQ scheme and AMR method, is capable
Fig. 15. Configuration of 2D bubble rising benchmark.
of simulating more complex two-phase flow with large deformations
and small-scale fluid structures accurately.

3.4. Bubble rising benchmark

and develop into a shape of ‘‘mushroom cloud’’ when the time instant
exceeds 0.6 s. Besides, some breakup of filament are observed on the This section presents a benchmark test of 2D bubble rising to
static mesh of 100 × 400 since the distorted tail becomes too thin evaluate the performance of thincFoam by comparing with reference
to be resolvable by finite resolution of the grids. It is predictable solutions given in Hysing et al. (2009) and Gamet et al. (2020). As
that the interface will become vague and messy as two phases are shown in Fig. 15, a circular bubble of gas with diameter of 0.5 is
mixed together progressively. Fig. 13 makes a thorough comparison initially centered at (0.5, 0.5) in a rectangular domain [0, 1] × [0, 2], in
about the evolution of interface predicted on the static and AMR grids. which remaining region is occupied by the liquid. We impose non-
The left part of subfigure plots the interface computed by thincFoam slip condition on the top and bottom boundaries, and slip boundary
with AMR method, while the counterpart depicts the interface on the conditions on the vertical sides. Two types of cases are carried out
static grid, which should be symmetric with respect to the 𝑦-axis at under the effect of gravity, viscosity and surface tension. A minor
the location of 𝑥 = 0.5. It is concluded that the interface profiles surface tension coefficient 𝜎 = 1.96 N∕m is applied for case I. We specify
computed on the AMR mesh show good agreement with those on the the density 𝜌𝑙 = 1000 kg∕m3 , dynamic viscosity 𝜇𝑙 = 10 Pa ⋅ s for liquid
static mesh, which demonstrates the excellent capability of present and 𝜌𝑔 = 1 kg∕m3 , 𝜇𝑔 = 0.1 Pa ⋅ s for gas respectively. The physical
solver in predicting the complex interfaces with a thin tail. However, properties of gas are considered as 100 kg∕m3 and 1 Pa ⋅ s for case II,
some irregularities are observed on the AMR grid in Fig. 13(a) and (b), where the surface tension coefficient is 24.5 N∕m. The acceleration of
which results from the mapping errors of volume fraction during the gravity is 0.98 m∕s2 for two cases.
mesh refinement. Simultaneously the mapping error will be magnified Following Hysing et al. (2009), we define the position and velocity
with computation continuing especially for Rayleigh–Taylor instability of centroid of bubble in which the position of centroid (𝑋𝑐 , 𝑌𝑐 ) is
featured with sensitivity for small interfacial perturbation. As shown in computed by
Fig. 13(d), small-scale fluid structure is well captured by present solver ∑ 𝑁𝑒 ( ) ∑ 𝑁𝑒 ( )
with AMR approach since the grid is refined in the vicinity of interface, 𝑖=1 𝑖
𝑥 1 − 𝜙𝑖 𝑦 1 − 𝜙𝑖
𝑖=1 𝑖
𝑋𝑐 = ∑𝑁 ( ) , 𝑌𝑐 = ∑𝑁𝑒 ( ) . (28)
which largely improve the solution resolution without loss of numerical 𝑒
1 − 𝜙𝑖 1 − 𝜙𝑖
𝑖=1 𝑖=1

14
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 16. The temporal evolution of bubble shape predicted on the AMR grid for case I.

Fig. 17. Vertical position and rising velocity on different grids for case I: (a) position of centroid and (b) rising velocity.

15
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 18. The evolution of bubble shape for case II. The left part of subfigure is on AMR grid, and the other is on static grid.

Fig. 19. Configuration of 3D dam-break benchmark.

3.4.1. Case I : bubble rising with weaker surface tension force

We perform the computation on the uniform and AMR grids for case
Table 5
The information of grids and time steps for case I. Symbol √ : AMR grid; Symbol 𝛥: I until 𝑡 = 3 s. Table 5 presents the resolution of grids and setup about
uniform grid. NOE : Number of Elements. time steps for case I of bubble rising test.
Initial resolution AMR max(NOE) Minimum grid size Time step Fig. 16 displays the temporal evolution of bubble shape for case
ℎ = 1∕160 𝛥 51200 1/160 4 × 10−4
I on the AMR mesh with grid resolution of 1/160 near the interface.
ℎ = 1∕20 √ 1706 1/80 5 × 10−4 It is seen that an inward decompression emerges from the bottom of
ℎ = 1∕40 √ 4997 1/160 2.5 × 10−4 bubble due to relatively larger velocity at the beginning. Eventually,
ℎ = 1∕80 √ 16541 1/320 1.25 × 10−4 this bubble lies the state that features skirted and dimpled ellipsoidal-
cap regimes. From the visualization of bubble shape in Fig. 16, the
present solver demonstrates an excellent agreement on the evolution of
bubble shape at different time instants with other published literature
The velocity of centroid of bubble (𝑢𝑐 , 𝑣𝑐 ) is expressed as
(Zuzio and Estivalezes, 2011; Hysing et al., 2009; Gamet et al., 2020).
∑𝑁𝑒 ( ) ∑ 𝑁𝑒 ( ) It is notable that the shape of main bulk of bubble predicted by
𝑢 1 − 𝜙𝑖
𝑖=1 𝑖
𝑣 1 − 𝜙𝑖
𝑖=1 𝑖
𝑢𝑐 = , 𝑣𝑐 = , (29) the present solver shows negligible differences compared with other
∑ 𝑁𝑒 ( ) ∑ 𝑁𝑒 ( )
𝑖=1
1 − 𝜙𝑖 𝑖=1
1 − 𝜙𝑖 numerical methods, and the shape of tails and breakup is captured with
more details due to the finer resolution of AMR grids.
where 𝑁𝑒 is total number of elements in the whole computational The time histories of position and velocity are plotted in Fig. 17
region, and 𝜙𝑖 denotes the volume fraction of control volume 𝛺𝑖 . for the centroid of bubble on the grid with various resolutions. The

16
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 20. The time histories of interfacial height for different positions in 3D dam-break benchmark.

profiles computed by thincFoam with AMR approach on the different range of 1.88 s∼2.06 s. The results indicate that the present solver
background grids converge to the reference solutions obtained on the can resolve fluid dynamics and interface deformation with adequate
finer static grids. It is seen that the vertical position of centroid is solution accuracy and robustness for 2D bubble rising with weaker
nearly linear to the time which convinces that the skirted region with surface tension force.
thin filament and breakup has less influence on the overall movement
of bubble. As shown in Fig. 17(b), it portrays the time histories of 3.4.2. Case II : bubble rising with stronger surface tension force
rising velocity which has two maximum points during the rising of Same as case I, we carry out this simulation on the static and AMR
bubble, corresponding to the time 𝑡 = 0.75 s and 𝑡 = 2 s respectively. grids with fixed time step 𝛥𝑡 = 2.5 × 10−4 s. The resolution of static grid
Table 6 elaborates the final position and maximum vertical velocity is 160 × 320, while the background grid configuration of AMR is only
as well as its corresponding time instants of rising bubble, which 20 × 40 and refined to 160 × 320 near the interface after fourth times
includes the results computed by other numerical methods such as mesh refinement.
TP2D, FreeLIFE and MooNMD based on FEM method (Hysing et al., Fig. 18 plots the bubble shapes at different time instants after mesh
2009), and geometric isoAdvetor-plicRDF and algebraic MULES inter- refinement. The circular bubble evolves into ellipsoidal shape that
face capturing methods (Gamet et al., 2020) for comparisons. The is distinctly different from case I accounting for the dominant effect
position of centroid predicted by the present solver at time 𝑡 = 3 s of surface tension. It reveals that the surface tension force plays a
shows slightly lower height than the numerical results in Hysing et al. vital role in keeping the interface from deforming and stretching into
(2009), in which the relative error is controlled less than 1%. It is tails or breakup. As shown in Fig. 18, the shapes of bubble solved
seen that the magnitude of maximum velocity at the first appearance by thincFoam on static (blue line) and AMR (red line) grid have a
is about 0.25 m∕s, corresponding to time instant around 0.73 s as given negligible difference which looks in satisfactory agreement with other
by other numerical solvers. The maximum velocity of centroid at the results (Hysing et al., 2009; Balcázar et al., 2016; Xie et al., 2020a). To
second appearance is about 0.23 m∕s corresponding to the time in the evaluate the computational cost, we measure the CPU time of present

17
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 21. The time histories of flow velocity at different positions for 3D dam-break benchmark, subfigure(a) for experiment is cited from Nguyen and Park (2018). Every 0.05 m
denotes a level from 𝑦 = 0.05 m to 𝑦 = 0.4 m in subfigure(b).

solver for the case with static and AMR grids and tabulate the results in the bottom of domain is set as non-slip wall boundary condition while
Table 7 for comparisons. It is noted that the elapsed time on AMR grid is the water can flow away from other three open boundaries. In our
only one tenth of the static one. This is attributed to the computational simulation, two inner faces at 𝑥 = 0 are represented as boundary faces
resources saved by less grid elements which is approximately only 5% with zero thickness by using createBaffles tool in OpenFOAM.
of static mesh. Another attention should be paid is that the elapsed time The simulation is performed on the structured hexahedral grid with
of THINC reconstruction (𝑇thinc ) is nearly the same for AMR and static resolution of 300 × 100 × 100 using automatic time step adjusted by
grids, which justifies that the same level of grid resolution is obtained in maximum Courant number of 0.2. We also reproduce the simulation
the vicinity of interface due to application of AMR approach. It gives a with the AMR grid in which the initial resolution and refinement level
convincing illustration that the present solver that combined with AMR are set as 75 × 25 × 25 and 2, respectively. To investigate the interfacial
approach has excellent performance in terms of numerical accuracy and height and flow velocity quantitatively, a set of gauging points are
efficiency which is of great importance for the simulation of two-phase located at the 𝑥 − 𝑧 cross-section to measure interfacial height and flow
flows with diverse scales of interface structures. velocity following Fraccarollo and Toro (1995), positions of which are
presented in Fig. 19.
3.5. 3D dam-break Fig. 20 plots the time histories of interfacial height for different
probes. It is seen that the height predicted by the present solver,
In order to evaluate the capability of present solver in 3D simula- including on the static and AMR grids, has fairly good agreement with
tions, we first perform the simulation with AMR approach for a dam- experimental data, which is superior to numerical results computed by
break flow with experiments given by Fraccarollo and Toro (1995). As Nguyen and Park (2018). In terms of position 0, the extreme point
shown in Fig. 19, the rectangular domain of [−1, 0] × [0, 1] × [−1, 1] is of interfacial height at 𝑡 = 0.8 s calculated by a variety of numerical
initially filled with water of height 0.6 m in which a spout with width methods differs from the experiment, which tends to deliver a higher
of 0.4 m (𝑧 ∈ [−0.2, 0.2]) is located on the boundary 𝑥 = 0. The top of interface since the gate cannot be open immediately and ideally at 𝑡 =
tank is connected with atmosphere and other sides are specified with 0 s as pointed out by Nguyen and Park (2018). Remarkably, the present
non-slip boundaries. In the downstream area of [0, 2] × [0, 1] × [−1, 1], solver performs well in fitting the tendency and height of interface for

18
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 22. The evolution of visualized interface in 3D dam-break benchmark, and different colors denote different heights of interface. (For interpretation of the references to color
in this figure legend, the reader is referred to the web version of this article.)

19
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 23. Schematic view of dam-break against a vertical wall, where all dimensions are in cm.

in Research Institute of Applied Mechanics, Kyushu University (Kamra

et al., 2018), the schematic view of which is shown in Fig. 23. Ini-
tially, a water column with height of 0.2 m is closed by a door in
its right side, which is waiting to flow into the dry horizontal bed
under the impact of gravity of 9.81 m∕s2 . To measure impact pressure
on the wall boundaries, several small pressure sensors are placed on
the right-hand vertical impact wall, the precise locations of which
are also depicted in Fig. 23. Non-slip velocity conditions are imposed
in the bottom and surrounding wall boundaries, correspondingly, the
pressure boundary conditions are specified as fixedFluxPressure
in simulation. The density and viscosity of water are set as 997 kg∕m3
and 8.9 × 10−3 kg∕(m ⋅ s) and those of air are given as 1.185 kg∕m3 and
1.7894 × 10−5 kg∕(m ⋅ s), respectively.
The simulation is carried out by the present solver on both 2D
and 3D grids and compared with the experimental results. For 2D
static grid, computational domain is divided into 9600 and 48000
non-uniform quadrilateral elements concentrating on the floor and
right-hand vertical wall, which are denoted as gridA and gridB. At
meanwhile, gridA is set as the background grid of AMR simulation and
then refined with refinement level 1 and 2 respectively. Fig. 24 plots the
time evolution of wave front on different 2D grids, which demonstrates
the solution independence on the computational grids. According to
Fig. 24. Wave front evolution for different grids. the numerical results, gridB and AMR grid with refinement level 2
are adopted in the subsequent numerical tests in order to attain better
grid resolution in the vicinity of interface. On the other hand, gridB is
position 8A and 10A. However, the numerical results given by Nguyen extruded into 100 layers in 𝑧 direction to generate 3D static grids, but
and Park (2018) underestimates the height of interface compared with gridA is extruded into 25 layers to create 3D AMR background grid,
experimental data. Besides, it is evident that irregular and oscillatory refinement level of which is set to 2.
behavior is observed in Nguyen and Park (2018) at 𝑡 = 0.4 s for Fig. 25 shows the contour line and iso-surface (𝜙 = 0.5) calculated
both position 8A and 10A, which is absent for our simulation and on 2D and 3D grids for different dimensionless time instants 𝑡∗ , which
experiment. For position -5A inside the container, the height predicted equals to 𝑡 ⋅ (𝑔∕𝐻)1∕2 . It is seen that the surface profiles computed
by the present solver is smooth and coincides with the experiment by present solver coincide well with experimental snapshots in Kamra
and simulation results in Nguyen and Park (2018) except for the et al. (2018) despite that some differences can be found due to the
shallow water method. Furthermore, the velocity curves predicted by viscous effect of boundary wall. From Fig. 25(d), it demonstrates that
the present solver are shown in Fig. 21 for various height level of the resolution of grid in the vicinity of surface is automatically refined,
position -3A and -3D, which shows good agreement with experimental which significantly improves the capability of capturing for small fluid
results. Fig. 22 delivers a visible result about the shape and height of structures, such as liquid droplets and air bubbles. Same as Fig. 25,
interface at different time instants. It reveals that the present two-phase Fig. 26 gives the iso-metric view of 3D dam-break simulation at the
solver is able to predict the evolution of interface and simulate complex selected instants, in which the free surfaces are predicted in better
two-phase flows accurately and faithfully in 3D computation. agreement with experiment as well. More quantitatively, time histories
of impact pressure for two sensors on the vertical wall is measured
3.6. Dam-break flow against a vertical wall and shown in Fig. 27. It is seen that the present solver delivers an
overall agreement on the position and magnitude of the pressure peaks.
In this section another dam-break problem is also presented to Compared with 2D results, the present 3D simulation shows a better
evaluate the performance of present solver in resolving violent motions match with experimental data in P1 as well as results in Kamra et al.
of moving interfaces in the two-phase flows simulation. Numerical (2018). With respect to P5, the magnitude and tendency around the
setup is determined according to the laboratory experiment conducted first and second pressure peak predicted by thincFoam agree better

20
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 25. Comparisons on free surface profiles between experiment and numerical simulation. The left column is 2D solution, middle column is front view of 3D solution, right
column is experimental snapshots (Kamra et al., 2018).

with experiment than those of Kamra et al. (2018). Throughout this water. To generate the wave accurately, the solitary wave is created
case, it convinces that the present solver is capable of simulating two- based on the analytical solution of solitary wave theory, which can be
phase flows accurately with severe distortions and deformations of free expressed as follows (Lee et al., 1982)
surface while only grids around the interface are refined automatically. 𝐴
𝜂= , (30)
cosh2 (𝐾𝑋)
3.7. Propagation of solitary wave around a vertical cylinder ( ( ))
√ 1 ℎ 3𝑦2 d2 𝜂
𝑢 = 𝜂𝑟 𝑔ℎ 1 − 𝜂𝑟 + 1− , (31)
4 3𝜂𝑟 2ℎ2 d𝑋 2
As the final test, we simulate a propagation of solitary wave around (( ( ))
𝑦√ ) d𝜂 𝑦2 d3 𝜂
a vertical cylinder in 3D domain in which the non-linear effect is highly 1 ℎ2
𝑣=− 𝑔ℎ 1 − 𝜂𝑟 + 1− , (32)
associated with the design of coastal and offshore structures. Since ℎ 2 d𝑋 3 2ℎ2 d𝑋 3
the experimental study of Yates and Wang (1994), it has been widely where 𝜂 is the wave elevation, 𝐴 is the wave height, 𝑢 and 𝑣 denotes
investigated by various numerical methods (Zhao et al., 2007; Jiang the velocity along streamwise and vertical directions, relative
√ position
et al., 2019; Xie et al., 2020b) for decades. We set up the simulation 𝑋 = 𝑥 − 𝑥0 − 𝑐𝑡, 𝑥0 is the offset √
of wave peak initially, 𝑐 = 𝑔 (ℎ + 𝐴)
according to the schematic view shown in Fig. 28. A cylinder with is the wave celerity and 𝐾 = 3𝐴∕4ℎ3 , 𝜂𝑟 = 𝜂∕ℎ. In addition, the
radius 𝑅 of 0.0635 m is located at the center of the whole domain, ratio between wave height and water depth 𝐴∕ℎ is considered as 0.4
lengths of which is 32𝑅, 12𝑅 and 2ℎ along the streamwise, spanwise here. In order to evaluate the free surface elevation, 12 guaging points
and vertical directions respectively, where ℎ denotes the depth of still are evenly divided into three groups and placed in different radial

21
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 26. Experimental and numerical interface (iso-metric view) calculated by thincFoam on 3D. The experiment snapshots come from Kamra et al. (2018).

Fig. 27. The time histories of pressure for different sensors on the vertical wall.

Fig. 28. Schematic view for propagation of solitary wave around a vertical cylinder.

directions 𝜃 of 0◦ , 100◦ and 180◦ clockwise as well as distances 𝑟 from non-uniform hexahedral elements, which are respectively labeled as
the center of cylinder to measuring points, where 𝜃 = 0◦ is the opposite coarse, middle and fine grid. From Fig. 30(a), it demonstrates that
direction of propagation of solitary wave as displayed in Fig. 28. the elevations of free surface calculated on different grids make no
In order to investigate the mesh dependence of numerical solutions, disparity in this case. Therefore, the middle grid is chosen to complete
the computational domain is divided into 0.22, 1.09 and 13.14 million the simulation of static grid. In terms of simulation of AMR grid, the

22
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 29. A comparison of grid in the vicinity of interface before and after mesh refinement.

Table 6
The final position and maximal velocity of centroid, corresponding time instants for case I of bubble rising
simulation.
Algorithms 𝑌𝑐 (𝑡 = 3) 𝑣𝑐,max 1 𝑡𝑣𝑐 ,max 1 𝑣𝑐,max 2 𝑡𝑣𝑐 ,max 2
ℎ = 1∕80
thincFoam+AMR 1.1084 0.2534 0.7375 0.2316 1.9410
FreeLIFE (Hysing et al., 2009) 1.1099 0.2518 0.7188 0.2384 1.9062
TP2D (Hysing et al., 2009) 1.1370 0.2638 0.7250 0.2597 1.9688
isoAdvector-plicRDF (Gamet – 0.2507 0.7268 0.2317 1.9127
et al., 2020)
MULES (Gamet et al., 2020) – 0.2493 0.7308 0.2261 1.8890
ℎ = 1∕160
thincFoam 1.1145 0.2478 0.7272 0.2300 2.0000
thincFoam+AMR 1.1114 0.2498 0.7293 0.2319 1.9778
FreeLIFE (Hysing et al., 2009) 1.1249 0.2514 0.7281 0.2440 1.9844
TP2D (Hysing et al., 2009) 1.1377 0.2570 0.7430 0.2522 2.0234
isoAdvector-plicRDF (Gamet – 0.2509 0.7268 0.2351 1.9990
et al., 2020)
MULES (Gamet et al., 2020) – 0.2494 0.7308 0.2318 1.9617
ℎ = 1∕320
thincFoam+AMR 1.1137 0.2489 0.7256 0.2301 1.9594
TP2D (Hysing et al., 2009) 1.1387 0.2538 0.7340 0.2467 2.0553
isoAdvector-plicRDF (Gamet – 0.2509 0.7367 0.2374 2.0163
et al., 2020)
MULES (Gamet et al., 2020) – 0.2497 0.7332 0.2357 2.0101
MooNMD (Hysing et al., 2009) 1.1376 0.2502 0.7317 0.2393 2.0600

coarse grid is regarded as initial grid, then refined twice in the vicinity Fig. 30(b)∼(d) displays the evolution of the water surface elevation
of wavefront, just as shown in Fig. 29. It is obvious that the resolution along different 𝜃 predicted by the present solver on static and AMR
grids, in which the wave elevation 𝜂 and computational time 𝑡 are
of grid is significantly improved near the interface while remains to √
normalized by ℎ and ℎ∕𝑔 respectively. The experimental data (Yates
be unchanged in the other region so that the total number of grid ele-
and Wang, 1994) and numerical results computed by Xie et al. (2020b)
ments amounts to only 1.8 million, whereas 13.14 million hexahedral
are also included to facilitate comparisons. It reveals that the tendency
elements are required to arrive at the same level of grid resolution if and magnitude of 𝜂∕ℎ computed by present solver matches fairly well
static mesh is used. The time step is limited by the maximum Courant with experimental data for different 𝜃 and relative distances 𝑟∕𝑅 except
number 0.5 and a upper limit of 2 × 10−4 s. from 𝑟∕𝑅 = 1.03 with 𝜃 = 0◦ and 𝜃 = 180◦ . Furthermore, the present

23
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Fig. 30. The evolution of the free surface elevation along different 𝜃 as predicted numerically together with experimental data.

Table 7 static and AMR grid. Besides we also plot the contour of interface
Computational efficiency of present solver with/without AMR strategy for case II of
height at different incident times in Fig. 31. As shown in Fig. 31(a),
bubble rising simulation. 𝑇total denotes elapsed time in total, 𝑇thinc the elapsed time for
interface reconstruction by THINC function, 𝑇U the elapsed time for the solution of the solitary wave gradually approaches to the cylinder at first and runs
√
momentum equation, 𝑇AMR the elapsed time for updating grids and THINC matrices. up to the highest elevation around 𝑡∕ (ℎ∕𝑔) = 8.61, which corresponds
NOE is the number of elements.
to the climbing segment of curve in Fig. 30(b) 𝑟∕𝑅 = 1.03. Then, the
Types of grid max(NOE) 𝑇total /s 𝑇thinc ∕s 𝑇U ∕s 𝑇AMR ∕s
solitary wave surpasses the cylinder and is decomposed into two groups
AMR 2774 935.46 342.07 171.1 415.26
static 51200 9089.72 323.48 8689.01 – of waves. Passing the cylinder for a while, two groups of waves are
integrated to a new wave to propagate downstream and a scattering
wave starts to emerge around the cylinder at the same time. Eventually,
solver delivers a better agreement with experimental study than the the scattering wave spreads to the whole computational region, which
previous work done by Xie et al. (2020b) under some circumstances, can be observed in Fig. 31(c) and (d). It reasonably accounts for
such as 𝑟∕𝑅 = 2.61 with 𝜃 = 0◦ as well as 𝑟∕𝑅 = 1.03 with 𝜃 = 100◦ , the apparent declination of free surface elevation in Fig. 30(b) and
which can be contributed to the higher solution accuracy of THINC/QQ (c). Moreover, the local refinement in the vicinity of interface can
scheme for interface capturing and adequate grid resolution provided be visualized at surrounding boundaries. It is evident that the grid
by AMR strategy in the vicinity of interface. Overall, it demonstrates resolution of interface is finer than regions away from the interface,
that thincFoam possesses the excellent capability in capturing largely which improves the numerical accuracy for two-phase flows definitely
deformed interface with sufficient accuracy and efficiency on the both while remaining the less number of elements for 3D simulation.

24
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

√
Fig. 31. Visualized results about the free surface height for different dimensionless instants 𝑡∕ (ℎ∕𝑔), and the grid configuration is also plotted to observe the refined procedure.

4. Conclusion remarks experimental data as well as reference solutions obtained by other

numerical methods. It is convincing that present solver delivers higher
In this paper, we present an accurate and efficient numerical solver numerical accuracy and stronger robustness for 2D and 3D multi-
for incompressible multiphase flows by using the new developed THINC phase flows with large deformations and topological changes, while
/QQ scheme (Chen et al., 2022) and AMR technique (Rettenmaier et al., improving the computational efficiency significantly. Therefore, the
2019) in the framework of OpenFOAM. Compared with the existing present solver is deemed to be particularly appealing for the numerical
interFoam solver, we have made the following major new efforts in simulations of complex multiphase flows that involve diverse flow
this work: (1) The latest version of THINC/QQ scheme is developed scales.
with AMR technique for interface capturing which realizes quadratic
representation of curved surfaces on unstructured polyhedral grids. CRediT authorship contribution statement
Without any geometric reconstruction, it demonstrates higher accuracy
of interface capturing than algebraic MULES scheme in interFoam, Dezhu Chen: Writing – original draft, Formal analysis, Validation,
even if on the large-skewness grids; (2) An algebraic algorithm with Methodology. Xin Tong: Methodology, Visualization, Software. Bin
quadratic interface (amQI) is proposed for mapping the volume fraction Xie: Conceptualization, Methodology, Reviewing and editing, Project
field based on THINC formulation which largely improves the geomet- administration. Feng Xiao: Reviewing and editing, Supervision, Project
ric fidelity of mapped interface in the child cells after mesh refinement; administration. Ye Li: Project administration, Supervision, Funding
(3) Second-order schemes of the spatial and temporal discretization are acquisition.
adopted with PISO algorithm for the solution of momentum equations
which further enhances numerical accuracy of both velocity and pres-
Declaration of competing interest
sure variables. The so-called thincFoam solver adaptively refines the
grid resolution in the vicinity of interface which significantly improve
The authors declare that they have no known competing finan-
the numerical accuracy and efficiency for multiphase simulations that
cial interests or personal relationships that could have appeared to
involve complicated interface deformations.
influence the work reported in this paper.
The algebraic mapping algorithm called amQI is first evaluated
quantitatively by interface reconstruction of circle and sphere. Com-
pared with other existing algorithms, it is found with higher numerical Data availability
accuracy and geometric fidelity after mesh refinement, in particular for
curved surfaces. Then the interface capturing method that combines Data will be made available on request.
THINC/QQ scheme and AMR technique is verified by 2D shear flow and
3D deformation flow, which demonstrates higher numerical accuracy Acknowledgments
and robustness than MULES scheme particularly for unstructured grids
with non-orthogonality. Furthermore, a series of benchmark tests of This work was partially supported by the National Key Research and
two-phase flow, such as Rayleigh–Taylor instability, bubble rising, Development Program of China (Grant No. 2019YFE0102500) and the
dam-break and propagation of solitary wave, are carried out to assess National Natural Science Foundation of China (Grant No. 11802178)
the performance of thincFoam solver which adaptively refines the mesh as well as the fund from Shanghai Pilot Program for Basic Research -
in the vicinity of interface. All numerical results coincide well with Shanghai Jiao Tong University, China (No. 21TQ1400202).

25
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

References Lee, J.J., Skjelbreia, J.E., Raichlen, F., 1982. Measurement of velocities in solitary
waves. J. Waterw. Port Coast. Ocean Div. 108 (2), 200–218.
Aulisa, E., Manservisi, S., Scardovelli, R., Zaleski, S., 2003. A geometrical Li, L.M., Hu, D.Q., Liu, Y.C., Wang, B.T., Shi, C., Shi, J.J., Xu, C., 2020. Large
area-preserving volume-of-fluid advection method. J. Comput. Phys. 192 (1), eddy simulation of cavitating flows with dynamic adaptive mesh refinement using
355–364. OpenFOAM. J. Hydrodyn. 32 (2), 398–409.
Balcázar, N., Lehmkuhl, O., Jofre, L., Rigola, J., Oliva, A., 2016. A coupled volume-of- Li, C., w, Duan., Zhao, B., 2022. Breaking wave simulations for a high-speed surface
fluid/level-set method for simulation of two-phase flows on unstructured meshes. vessel with hybrid THINC and HRIC schemes. Appl. Ocean Res. 125, 103257.
Comput. & Fluids 124, 12–29. López, J., Hernández, J., Gómez, P., Faura, F., 2016. A new volume conservation
Benson, D.J., 2002. Volume of fluid interface reconstruction methods for multi-material enforcement method for PLIC reconstruction in general convex grids. J. Comput.
problems. Appl. Mech. Rev. 55 (2), 151–165. Phys. 316, 338–359.
Chen, D., Xie, B., Xiao, F., 2022. Revisit to the THINC/QQ scheme: Recent progress Luo, M., Khayyer, A., Lin, P., 2021. Particle methods in ocean and coastal engineering.
to improve accuracy and robustness. Internat. J. Numer. Methods Fluids 94 (7), Appl. Ocean Res. 114, 102734.
719–755. Luo, M., Koh, C.G., Gao, M., Bai, W., 2015. A particle method for two-phase flows with
Chiron, L., Oger, G., De Leffe, M., Le Touzé, D., 2018. Analysis and improvements of large density difference. Internat. J. Numer. Methods Engrg. 103 (4), 235–255.
Adaptive Particle Refinement (APR) through CPU time, accuracy and robustness Marić, T., Kothe, D.B., Bothe, D., 2020. Unstructured un-split geometrical
considerations. J. Comput. Phys. 354, 552–575. volume-of-fluid methods–a review. J. Comput. Phys. 420, 109695.
Cifani, P., Michalek, W.R., Priems, G.J.M., Kuerten, J.G., van der Geld, C.W.M., Marić, T., Marschall, H., Bothe, D., 2013. voFoam-A geometrical Volume of Fluid
Geurts, B.J., 2016. A comparison between the surface compression method and algorithm on arbitrary unstructured meshes with local dynamic adaptive mesh
an interface reconstruction method for the VOF approach. Comput. & Fluids 136, refinement using OpenFOAM. arXiv preprint arXiv:1305.3417.
421–435. Nakayama, H., Xie, B., Xiao, F., 2015. Numerical simulations of liquid breakup by
Cooke, J.J., Armstrong, L.M., Luo, K.H., Gu, S., 2014. Adaptive mesh refinement of thincfoam-a numerical model for interfacial multiphase flows. In: Proceedings of
gas–liquid flow on an inclined plane. Comput. Chem. Eng. 60, 297–306. 29th Symposium on Computational Fluid Dynamics. C04-3 (in Japanese).
Dai, D., Tong, A.Y., 2018. An analytical interface reconstruction algorithm in the PLIC- Nguyen, V.T., Park, W.G., 2018. Enhancement of Navier–Stokes solver based on an
VOF method for 2D polygonal unstructured meshes. Internat. J. Numer. Methods improved volume-of-fluid method for complex interfacial-flow simulations. Appl.
Fluids 88 (6), 265–276. Ocean Res. 72, 92–109.
Dai, D., Tong, A.Y., 2019. Analytical interface reconstruction algorithms in the PLIC- Oevermann, M., Klein, R., Berger, M., Goodman, J., 2000. A projection method for
VOF method for 3D polyhedral unstructured meshes. Internat. J. Numer. Methods two-phase incompressible flow with surface tension and sharp interface resolution.
Fluids 91 (5), 213–227. Zuse Inst. Berlin-Rep. 17, 1–23.
Dai, D., Tong, A.Y., 2022. The adaptive plic-vof method in cavitating flow simulations. Pilliod, Jr., J.E., Puckett, E.G., 2004. Second-order accurate volume-of-fluid algorithms
Comput. Therm. Sci. Int. J. 14 (4), 75–90. for tracking material interfaces. J. Comput. Phys. 199 (2), 465–502.
Damián, S.M., Nigro, N.M., 2014. An extended mixture model for the simultaneous Popinet, S., 2003. Gerris: a tree-based adaptive solver for the incompressible Euler
treatment of small-scale and large-scale interfaces. Internat. J. Numer. Methods equations in complex geometries. J. Comput. Phys. 190 (2), 572–600.
Fluids 75, 547–574. Popinet, S., 2009. An accurate adaptive solver for surface-tension-driven interfacial
Denner, F., van Wachem, B.G., 2014. Compressive VOF method with skewness cor- flows. J. Comput. Phys. 228 (16), 5838–5866.
rection to capture sharp interfaces on arbitrary meshes. J. Comput. Phys. 279, Puckett, E.G., 1991. A volume-of-fluid interface tracking algorithm with applications
127–144. to computing shock wave refraction. In: Proceedings of the Fourth International
Deshpande, S.S., Anumolu, L., Trujillo, M.F., 2012. Evaluating the performance of the Symposium on Computational Fluid Dynamics. pp. 933–938.
two-phase flow solver interFoam. Comput. Sci. Discov. 5 (1), 014016.
Renardy, Y., Renardy, M., 2002. PROST: a parabolic reconstruction of surface tension
Enright, D., Fedkiw, R., Ferziger, J., Mitchell, I., 2002. A hybrid particle level set
for the volume-of-fluid method. J. Comput. Phys. 183 (2), 400–421.
method for improved interface capturing. J. Comput. Phys. 183 (1), 83–116.
Rettenmaier, D., Deising, D., Ouedraogo, Y., Gjonaj, E., Gersem, H.De., Bothe, D.,
Fraccarollo, L., Toro, E.F., 1995. Experimental and numerical assessment of the shallow
Tropea, C., Marschall, H., 2019. Load balanced 2D and 3D adaptive mesh
water model for two-dimensional dam-break type problems. J. Hydraul. Res. 33 (6),
refinement in OpenFOAM. SoftwareX 10, 100317.
843–864.
Rider, W.J., Kothe, D.B., 1998. Reconstructing volume tracking. J. Comput. Phys. 141
Gamet, L., Scala, M., Roenby, J., Scheufler, H., Pierson, J.L., 2020. Validation of
(2), 112–152.
volume-of-fluid OpenFOAM® isoAdvector solvers using single bubble benchmarks.
Roenby, J., Bredmose, H., Jasak, H., 2016. A computational method for sharp interface
Comput. & Fluids 213, 104722.
advection. R. Soc. Open Sci. 3 (11), 160405.
Haghshenas, M., Wilson, J.A., Kumar, R., 2017. Algebraic coupled level set-volume of
Scheufler, H., 2019. DLR-RY/VoFLibrary. Accessed from https://github.com/DLR-RY/
fluid method for surface tension dominant two-phase flows. Int. J. Multiph. Flow.
VoFLibrary.
90, 13–28.
Scheufler, H., Roenby, J., 2019. Accurate and efficient surface reconstruction from
Hu, C., Kamra, M.M., 2020. An unstructured mesh method for numerical simulation of
volume fraction data on general meshes. J. Comput. Phys. 383, 1–23.
violent sloshing flows. J. Hydrodyn. 32 (2), 259–266.
Hu, C., Liu, C., 2018. Simulation of violent free surface flow by AMR method. J. Sun, P.N., Colagrossi, A., Marrone, S., Antuono, M., Zhang, A.M., 2018. Multi-resolution
Hydrodyn. 30 (3), 384–389. Delta-plus-SPH with tensile instability control: Towards high Reynolds number
Huang, Y., Li, J., Xie, B., Lin, Z., Borthwick, A.G., 2021. High-fidelity numerical flows. Comput. Phys. Comm. 224, 63–80.
simulation of solitary wave propagation. Ocean Eng. 224, 108698. Sussman, M., 2003. A second order coupled level set and volume-of-fluid method for
Hysing, S.R., Turek, S., Kuzmin, D., Parolini, N., Burman, E., Ganesan, S., Tobiska, L., computing growth and collapse of vapor bubbles. J. Comput. Phys. 187, 110–136.
2009. Quantitative benchmark computations of two-dimensional bubble dynamics. Ubbink, O., Issa, R.I., 1999. A method for capturing sharp fluid interfaces on arbitrary
Internat. J. Numer. Methods Fluids 60 (11), 1259–1288. meshes. J. Comput. Phys. 153 (1), 26–50.
Ii, S., Sugiyama, K., Takeuchi, S., Takagi, S., Matsumoto, Y., Xiao, F., 2012. An Wang, Z., Li, L., Cheng, H., Ji, B., 2020. Numerical investigation of unsteady cloud
interface capturing method with a continuous function: the THINC method with cavitating flow around the Clark-Y hydrofoil with adaptive mesh refinement using
multi-dimensional reconstruction. J. Comput. Phys. 231 (5), 2328–2358. OpenFOAM. Ocean Eng. 206, 107349.
Ii, S., Xie, B., Xiao, F., 2014. An interface capturing method with a continuous function: Xiao, F., Honma, Y., Kono, T., 2005. A simple algebraic interface capturing scheme
The THINC method on unstructured triangular and tetrahedral meshes. J. Comput. using hyperbolic tangent function. Internat. J. Numer. Methods Fluids 48 (9),
Phys. 259, 260–269. 1023–1040.
Issa, R.I., 1985. Solution of the implicitly discretised fluid flow equations by Xiao, T., Xie, B., Deng, X., Du, Y., 2022. Numerical simulations for incompressible
operator-splitting. J. Comput. Phys. 62, 40–65. turbulence cavitation flows with tangent of hyperbola interface capturing (THINC)
Jiang, C., Liu, X., Yao, Y., Deng, B., 2019. Numerical investigation of solitary wave scheme. Phys. Fluids 34 (2), 022108.
interaction with a row of vertical slotted piles on a sloping beach. Int. J. Nav. Xie, B., Deng, X., Liao, S., 2019. High-fidelity solver on polyhedral unstructured grids
Archit. Ocean Eng. 11 (1), 530–541. for low-Mach number compressible viscous flow. Comput. Methods Appl. Mech.
Kamra, M.M., Mohd, N., Liu, C., Sueyoshi, M., Hu, C., 2018. Numerical and experimen- Engrg. 357, 112584.
tal investigation of three-dimensionality in the dam-break flow against a vertical Xie, B., Ii, S., Ikebata, A., Xiao, F., 2014a. A multi-moment finite volume
wall. J. Hydrodyn. 30 (4), 682–693. method for incompressible Navier–Stokes equations on unstructured grids:
Klostermann, J., Schaake, K., Schwarze, R., 2013. Numerical simulation of a single Volume-average/point-value formulation. J. Comput. Phys. 277, 138–162.
rising bubble by VOF with surface compression. Internat. J. Numer. Methods Fluids Xie, B., Ii, S., Xiao, F., 2014b. An efficient and accurate algebraic interface capturing
71 (8), 960–982. method for unstructured grids in 2 and 3 dimensions: The THINC method with
Kuo, C.W., Trujillo, M.F., 2021. An analysis of the performance enhancement with quadratic surface representation. Internat. J. Numer. Methods Fluids 76 (12),
adaptive mesh refinement for spray problems. Int. J. Multiph. Flow. 140, 103615. 1025–1042.
Laurila, E., Roenby, J., Maakala, V., Peltonen, P., Kahila, H., Vuorinen, V., 2019. Xie, B., Jin, P., Du, Y., Liao, S., 2020a. A consistent and balanced-force model for
Analysis of viscous fluid flow in a pressure-swirl atomizer using large-eddy incompressible multiphase flows on polyhedral unstructured grids. Int. J. Multiph.
simulation. Int. J. Multiph. Flow. 113, 371–388. Flow. 122, 103125.

26
D. Chen et al. International Journal of Multiphase Flow 162 (2023) 104409

Xie, Z., Stoesser, T., Yan, S., Ma, Q., Lin, P., 2020b. A Cartesian cut-cell based mul- Youngs, D.L., 1984. An Interface Tracking Method for a 3D Eulerian Hydrodynamics
tiphase flow model for large-eddy simulation of three-dimensional wave-structure Code. Vol. 44. No. 92. Atomic Weapons Research Establishment (AWRE) Technical
interaction. Comput. & Fluids 213, 104747. Report, p. 35.
Xie, B., Xiao, F., 2014. Two and three dimensional multi-moment finite volume solver Zhang, D., Jiang, C., Liang, D., Chen, Z., Yang, Y., Shi, Y., 2014. A refined volume-of-
for incompressible Navier–Stokes equations on unstructured grids with arbitrary fluid algorithm for capturing sharp fluid interfaces on arbitrary meshes. J. Comput.
quadrilateral and hexahedral elements. Comput. & Fluids 104, 40–54. Phys. 274, 709–736.
Xie, B., Xiao, F., 2017. Toward efficient and accurate interface capturing on arbitrary Zhang, Z., Zhao, X., Xie, B., Nie, L., 2019. High-fidelity simulation of regular waves
hybrid unstructured grids: The THINC method with quadratic surface representation based on multi-moment finite T volume formulation and THINC method. Appl.
and Gaussian quadrature. J. Comput. Phys. 349, 415–440. Ocean Res. 87, 81–94.
Xu, Q., Li, Y., Yu, Y.H., Ding, B., Jiang, Z., Lin, Z., Cazzolato, B., 2019. Experimental Zhang, C., Zhu, Y., Wu, D., Hu, X., 2022. Review on smoothed particle hydrodynamics:
and numerical investigations of a two-body floating-point absorber wave energy methodology development and recent achievement. arXiv preprint arXiv:2205.
converter in regular waves. J. Fluids Struct. 91, 102613. 03074.
Yates, G.T., Wang, K.H., 1994. Solitary wave scattering by a vertical cylinder: Zhao, M., Cheng, L., Teng, B., 2007. Numerical simulation of solitary wave scattering
experimental study. In: The Fourth International Offshore and Polar Engineering by a circular cylinder array. Ocean Eng. 34 (3–4), 489–499.
Conference. OnePetro. Zuzio, D., Estivalezes, J., 2011. An efficient block parallel AMR method for two phase
Yin, M., Zhao, X., Luo, M., Sun, H., 2021. Flow pattern and hydrodynamic parameters interfacial flow simulations. Comput. & Fluids 44 (1), 339–357.
of pile breakwater under solitary wave using OpenFOAM. Ocean Eng. 235, 109381.

27