International Journal of Heat and Mass Transfer 167 (2021) 120822

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

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

Laminar to turbulent flow transition inside the boundary layer

adjacent to isothermal wall of natural convection flow in a cubical
cavity
Xin Wen a, Lian-Ping Wang a,b,c,d,e,∗, Zhaoli Guo f, Dauren B. Zhakebayev g
a
Department of Mechanical Engineering, 126 Spencer Laboratory, University of Delaware, Newark, Delaware 19716-3140, USA
b
Guangdong Provincial Key Laboratory of Turbulence Research and Applications, Center for Complex Flows and Soft Matter Research and Department of
Mechanics and Aerospace Engineering, Southern University of Science and Technology, Shenzhen 518055, Guangdong, China
c
Center for Complex Flows and Soft Matter Research, Southern University of Science and Technology, Shenzhen 518055, Guangdong, China
d
Guangdong Provincial Key Laboratory of Turbulence Research and Applications, Southern University of Science and Technology, Shenzhen 518055,
Guangdong, China
e
Guangdong-Hong Kong-Macao Joint Laboratory for Data-Driven Fluid Mechanics and Engineering Applications, Southern University of Science and
Technology, Shenzhen 518055, China
f
State Key Laboratory of Coal Combustion, Huazhong University of Science and Technology, Wuhan, P.R. China
g
Al Farabi Kazskh National University, Almaty, Kazakhstan

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

Article history: We investigate three-dimensional natural convection flow in an air-filled, differentially heated cubical
Received 20 April 2020 cavity. The vertical wall on the left is heated and the vertical wall on the right is cooled, with the re-
Revised 4 December 2020
maining four walls being adiabatic. We performed direct numerical simulations of the natural convection
Accepted 10 December 2020
flow using discrete unified gas-kinetic scheme (DUGKS), with an improved implementation of boundary
Available online 28 December 2020
conditions. Thin boundary layers are developed along the two isothermal walls. The laminar to turbulent
Keywords: flow transition inside the boundary layers is studied in this paper. The simulations are conducted at three
Natural convection Rayleigh numbers of 1.5 × 109 , 1.0 × 1010 , 1.0 × 1011 using nonuniform grids with resolution up to 3203 .
Flow transition The Prandtl number is fixed at 0.71. We provide a detailed analysis of the transition from laminar to tur-
DUGKS bulent flow inside the vertical boundary layers and its influence on the rate of heat transfer. Time traces
Accurate implementation of boundary of temperature and velocity, time-averaged flow field, statistics of fluctuation fields are presented to il-
conditions
lustrate distinct behaviors in the laminar and turbulent thermal boundary layer, as well as to determine
the transition location at different Ra numbers. The average Nusselt numbers for different Ra numbers
are compiled and compared to previous results. A guideline of the resolution requirement is suggested
based on the Ra scaling of laminar thermal boundary layer.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction and these differentially heated from the sides. The Rayleigh-Bénard
convection has been widely studied [2,3], while the latter case is
Natural convection in an enclosure is a classical configuration relatively less studied and is considered in this paper. Namely, here
in heat transfer research due to many applications such as such as the left vertical wall of the cavity is heated and the right vertical
cooling of the electronic device, energy storage system and climate wall is cooled, with the remaining four walls being kept adiabatic.
conditioning of rooms [1]. Using convection to enhance heat trans- Researchers have studied this latter configuration for several
fer in a compact space is a highly effective approach. Based on the decades, focusing on different aspects such as the instability mech-
direction of the applied temperature gradient, it can be classified anism of the convection and the Nusselt - Rayleigh number cor-
into two categories: cavity heated from below (Rayleigh-Bénard) relation, where the Rayleigh number is the key input parameter
and the Nusselt number measures the dimensionless heat trans-
fer rate, both will be defined in Section 2. Despite its simple con-
∗
Corresponding author at: Guangdong Provincial Key Laboratory of Turbulence figuration, this problem is extremely complex due to the strong
Research and Applications, Center for Complex Flows and Soft Matter Research and
Department of Mechanics and Aerospace Engineering, Southern University of Sci-
coupling between velocity field and temperature field and a large
ence and Technology, Shenzhen 518055, Guangdong, China. range of system parameters such as the cavity aspect ratio, Prandtl
E-mail address: [email protected] (L.-P. Wang). number, temperature difference, etc. When the temperature differ-

https://doi.org/10.1016/j.ijheatmasstransfer.2020.120822
0017-9310/© 2020 Elsevier Ltd. All rights reserved.
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

ence is small, the density change is negligible except in the buoy- Mallinson and de Vahl Davis [4] performed an early simulation
ancy term. And under the Boussinesq assumption the fluid prop- of three-dimensional natural convection using the finite-difference
erties such as viscosity ν, heat conductivity k, are treated as con- method with relatively coarse meshes for the Rayleigh number
stants. The viscous dissipation in the energy equation is typically range 104 ≤ Ra ≤ 106 . Fusegi et al. [16] reported a finite-difference
neglected. Led by the pioneering work of de Vahl Davis and co- numerical study of three-dimensional natural convection in a cubi-
workers [4,5], researchers used different methods to study this cal cavity for Rayleigh number range of 103 ≤ Ra ≤ 106 . Tric et al.
problem numerically and provided benchmark solutions for both [17] provided a set of benchmark solutions for 103 ≤ Ra ≤ 107 us-
two- and three-dimensional cavities; and the simulation results ing pseudo-spectral Chebyshev algorithm based on the projection-
were compared to earlier experimental results [6–8]. A summary diffusion method. Fusegi et al. [18] reported benchmark results
of previous direct numerical simulations of natural convection in a for Ra = 1010 using 62 × 122 × 62 meshes. With the introduction
differentially heated cavity related to our study is presented below. of the lateral walls in the third direction, we expect the critical
Due to the restriction of the computational resource, the early Rayleigh number for the onset of unsteady convection will be af-
numerical studies focused on steady laminar convection flow. For fected by these lateral walls. Janssen and Henkes [19] studied the
flow in a two-dimensional square cavity, de Vahl Davis and Jones transition to unsteady flow in the three-dimensional cavity with
[5] provided a set of benchmark solutions for the laminar regime adiabatic horizontal walls, and obtained the critical Ra between
(103 ≤ Ra ≤ 106 ) using the finite-difference method. Later, Quéré 2.5 × 108 and 3 × 108 by assuming symmetry so that they only
[9] used the pseudo-spectral method to solve the problem with performed simulation on a quarter of the cavity. However, Labrosse
finer meshes up to 128 × 128. He provided solutions for the full et al. [20] claimed that transition happens at Ra = 3.19 × 107 with-
range of two-dimensional steady-state flow (Ra ≤ 108 ). Beyond a out assuming symmetry.
critical Ra, the two-dimensional natural convection flow becomes In the current study, we perform three-dimensional simula-
time-dependent, the onset of the unsteadiness is affected by sev- tions in a cubical cavity with a Rayleigh number range far be-
eral controlling parameters. Using the finite-difference method, yond the critical Rayleigh number. This allows us to investigate
Paolucci and Chenoweth [10] studied two-dimensional natural con- the flow transition insides the boundary layer which is the sec-
vection with Rayleigh number up to 1010 in cavities of aspect ond type of instability for this specific configuration. Another im-
ratio near unity. They claimed that for cavity with aspect ratio portant research goal for natural convection is to predict the Nus-
(A=height/length) less than 12 or larger than 3, the primary in- selt number of the heated wall which is desired for engineering
stability takes place insides the boundary layer along the isother- applications. Utilizing the numerical results for a large range of
mal wall. However, for 12 < A < 3 the instability first happens near Rayleigh number, researchers proposed Rayleigh - Nusselt corre-
the departing corners. Janssen and Henkes [11] performed two- lations for the three-dimensional cubical enclosure. Fusegi et al.
dimensional simulations in a differentially heated square cavity [18] proposed an empirical correlation for the overall Nusselt
with Prandtl number between 0.25 and 7.0. They found that for number Nuoverall = 0.163Ra0.282 for 103 ≤ Ra ≤ 1010 with relatively
Prandtl number between 0.25 and 2.0, the flow exhibits peri- coarse mesh. Wang [21], Wang et al. [22] reported the correlation
odic, quasi-periodic behaviors before becoming turbulent eventu- Nuoverall = 0.127Ra0.3052 for steady flow regime 103 ≤ Ra < 107 and
ally. While for larger Prandtl number, flow goes from steady to tur- Nuoverall = 0.3408Ra0.241 for unsteady flow regime 107 ≤ Ra ≤ 1010 .
bulent without intermediate transition. While for the steady flow regime, the average Nusselt number re-
For an air-filled differentially heated square cavity with adia- sults are in good agreement, the results have noticeable differences
batic horizontal walls, Quéré and Behnia [12] concluded that the in the unsteady flow regime. One reason for this disagreement is
critical Rayleigh number for transition from steady to unsteady that the high Rayleigh number convection requires fine mesh, and
flow is Racr = 1.82 ± 0.01 × 108 , using a relatively coarse mesh res- Fusegi used a relatively coarse mesh. For unsteady flow regime, the
olution at 72 × 72. They studied time-dependent flow up to Ra = Nusselt - Rayleigh correlation still needs to be improved.
1.0 × 1010 in 2D, based on the pseudo-spectral Chebyshev algo- In general, three-dimensional differentially-heated natural con-
rithm. Two-dimensional natural convection flow in a square cav- vection flow in a cavity has a unique flow structure rather differ-
ity can be a good approximation of the flow at the mid-plane ent from the Rayleigh-Bénard problem. Driven by the buoyancy
of the three-dimensional cubical cavity. As stated by Janssen and force β g(T − T0 ) along the vertical direction, thin boundary lay-
Henkes [11], the boundary layer along the heated wall of the cav- ers develop along the isothermal walls, outside the boundary layer,
ity resembles those along the isothermal vertical plate. However, the core region is quiescent and there is a vertical temperature
the introduction of the top wall of the cavity changes the flow stratification in the cavity center. In the previous studies, most re-
structure. The vertical boundary layer is turned horizontal and cre- searchers focused on general flow feature of the natural convection
ates a jet-like fluid layer which induces the first instability. When and provided a set of time-averaged statistics [6,17,18]. Only a few
the Rayleigh number is further increased, the second instability of them paid attention to the transition from laminar to turbulent
takes place insides the boundary layer. Although there are numer- flow insides the boundary layer. Trias et al. [23–25] performed di-
ous studies about the boundary layer adjacent to an isothermally rect numerical simulations to a differentially heated cavity of as-
heated vertical surface [13,14], the introduction of the horizontal pect ratio 4 up to Ra = 1011 and provided a comparison between
walls restrains the free development of the vertical boundary layer. two- and three-dimensional results for this problem. They claimed
As the heat transfer is highly affected by the flow regime of the that the thermal boundary layers of the three-dimensional flow
thermal boundary layer, in this study we will focus on absolute in- remain laminar or quasi-laminar in the upstream parts, up to a
stability of thermal boundary layer, specifically, the flow transition point where the eddies are ejected. This transition location moves
insides the boundary layers and its influence on the heat trans- further upstream for 3D simulations, compared to 2D simulations.
fer. In passing we note that Janssen and Armfield [15] studied con- However, for the cubical cavity, the influence of the top and bot-
vective instability due to externally imposed perturbations within tom wall is more significant. And based on the analysis of Paolucci
the thermal boundary layer. The convective instability occurs at a and Chenoweth [10], the aspect ratio plays an important role in
much smaller Rayleigh number than those at which the absolute the development of flow instability, the instability mechanism of
instability happens. the cubic cavity is different from the tall cavity.
As the turbulent convection flow must be essentially three- The primary goal of the current work is to improve our un-
dimensional in nature, researchers extended DNS to the three- derstanding of the transition from laminar to turbulent flow in-
dimensional cavities based on previous two-dimensional studies. side the vertical boundary layer in a differentially heated cubi-

2
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

cal cavity with adiabatic horizontal and lateral walls. In recent

years, the kinetic method (or mesoscopic CFD method based on the
model Boltzmann equation) has been developed and become a re-
liable tool for thermal flow simulations [26–28]. The discrete uni-
fied gas-kinetic scheme [29] is used for the current study. Compar-
isons of performance in terms of accuracy, stability, and efficiency
have been done previously in Wang et al. [30,31] and Bo et al.
[32]. Wang et al. [30] compared DUGKS and LBM in the simula-
tion of lid-driven cavity flow, laminar flow past a square cylinder.
Wang et al. [31] compared DUGKS, LBM, and spectral method in
the simulation of decaying turbulent flows, Bo et al. [32] compared
DUGKS, LBM, and spectral method in the simulation of 3D Taylor- Fig. 1. The geometry under consideration.
Green vortex flow and wall-bounded channel flows. In these stud-
ies, the second-order accuracy has been demonstrated. In terms of
efficiency, DUGKS is slower than LBE in terms of updating the dis- ∂T
tribution functions at the lattice nodes due to its additional evalua- + (u · ∇ )T = α∇ 2 T , (1c)
∂t
tion of interface fluxes. However, as a finite volume scheme DUGKS
can use a non-uniform mesh. As a result, if the non-uniform mesh where u is the flow velocity, T is the temperature, ρ is the density
can be clustered in the region with a large flow gradient, the ef- at the mean temperature T0 , g is the gravitational acceleration, p is
ficiency of DUGKS can be better than LBM. The relative efficiency a modified pressure including the effect of mean body force ρ g, ν
of DUGKS over LBM was discussed for the simulation of the flow is the kinematic viscosity, β is the thermal expansion coefficient,
passing the square cylinder in Wang et al. [30] and for turbulent and α is the thermal diffusivity. Under the above Boussinesq as-
channel flow in Bo et al. [32]. Non-uniform meshes can be easily sumption, the temperature field is treated as a scalar field which
implemented in this scheme and the boundary condition can be is advected by the macroscopic flow with a constant diffusion co-
applied right at the cell interface on the wall. However, accurate efficient. Nevertheless, the velocity field and the temperature field
implementation of temperature and velocity boundary conditions are closely coupled: the flow is driven by the temperature field
in the kinetic method remains a research topic, and here we will through the buoyancy force, and at the same time the temperature
propose an improved implementation that is consistent with the field is advected by the flow velocity. This coupling introduces an-
Chapman-Enskog analysis. Different statistics will be used to spec- other level of nonlinearity in addition to the inertial term in the
ify the transition location, and study the influence of flow transi- momentum equation. The coordinate system used is: x is the ver-
tion on the rate of heat transfer. tical direction, y is the horizontal direction perpendicular to the
The present paper is organized as follows. The physical problem heated wall, and z is the spanwise direction.
description and numerical method (DUGKS) are first presented in With the above setup and the simplified governing equations,
Section 2. Numerical results are presented to address several physi- the system is governed by the geometrical parameters (cavity
cal aspects. In Section 3.3 we discuss the time evolution of velocity height h,cavity length l, cavity depth w, see Fig. 1), fluid properties
and temperature of a set of selected monitoring points and their (ν, (β g), α ), the reference temperature T0 , and driving temperature
power spectra. Averaged flow fields are presented in Section 3.4, difference T . Note that the mean density can be absorbed into
followed by secondary statistics in Section 3.5. The effect of flow the pressure term thus it is not listed. Also (β g) is viewed as a sin-
transition on heat transfer is presented in Section 3.6. Finally, a gle parameter as they appear together in the momentum equation.
summary is given in Section 4. By dimensional analysis, it follows that the system is governed by
the following five independent dimensional parameters
2. Governing equations and numerical methods gβ T h3 ν T h h
Ra ≡ , Pr ≡ , , , (2)
να α T0 l w
2.1. Problem description and governing equations
Furthermore, it is assumed that T /T0 << 1. With the two aspect
We consider an air-filled cubical cavity of height h, with an ratios h/l and h/w fixed to one here and the Prandtl number for
isothermal vertical hot wall at temperature Th = T0 + 0.5T on the air given as P r = 0.71, the only governing parameter remaining is
left, and a cold wall at temperature Tc = T0 − 0.5T on the right; the Rayleigh number Ra, which can be interpreted as Ra = Re2eff P r,
where the temperature difference T = Th − Tc and the mean tem- where the effective Reynolds number is defined as Reeff ≡ u0 h/ν,
perature T0 = (Th + Tc )/2. The other four walls (i.e., two other ver- with the buoyancy velocity defined as u0 = β T gh.
tical walls and two horizontal walls) are assumed to be adiabatic. We can now nondimensionalize all quantities in the governing
No-slip velocity boundary conditions are imposed on all walls. Un- equations by the length scale h, velocity scale u0 , time scale h/u0 ,
der the Boussinesq assumption, the fluid density is treated as a pressure scale ρ u20 and temperature scale T . Since the tempera-
constant except with two considerations: (1) the buoyancy force ture equation is linear, a reference value can be subtracted, so we
is retained due to density variation caused by temperature change define the normalized temperature as θ = (T − T0 )/T . The gov-
and (2) the volume expansion work in the thermal energy equation erning equations in the nondimensional form can then be written
is retained under the ideal gas assumption Kundu et al. [33]. Fur- as
thermore, the viscous dissipation in the thermal energy equation ∇ · uˆ = 0, (3a)
is neglected as it is typically very small for the case of air under
natural convection. The governing equations can then be written
as ∂ uˆ   P r 1/2
+ uˆ · ∇ uˆ = −∇ pˆ + 1/2 ∇ 2 uˆ + θ ex , (3b)
∇ · u = 0, (1a) ∂t ˆ Ra

∂u ∂θ   1
ρ + ρ (u · ∇ )u = −∇ p + ρν∇ 2 u − ρβ (T − T0 )g, (1b) + uˆ · ∇ θ = 1/2 1/2 ∇ 2 θ , (3c)
∂t ∂t
ˆ P r Ra

3
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

where uˆ, pˆ and tˆ denote the normalized velocity, pressure and where φ = g or h, Fext = 0 for h distribution function. For the con-
time, respectively, ex is the unit vector pointing vertically upward. tinuum flow regime, the Chapman-Enskog analysis implies that the
The initial state of the flow is set to be quiescent u ˆ = 0 and external body force term Fext for f can be approximated as [26]:
isothermal θ = 0. The velocity boundary condition is u ˆ = 0 on all  
a· ξ−u
walls, the thermal boundary conditions are: θ (y = 0 ) = 0.5, θ (y = Fext = −a · ∇ξ f ≈ −a · ∇ξ f eq
= f eq . (9)
1 ) = −0.5, and zero normal temperature gradient on the other four RT1
walls. As the DUGKS is a finite-volume scheme, the computational
domain is divided into a set of control volumes. Integrating
2.2. Numerical method Eq. (8) over a control volume V j centered at x j from time tn to
tn+1 . The divergence theorem is applied to the convection term to
In the present work, we perform numerical simulations us- convert the cell-volume integral to the cell-surface integral, then
ing discrete unified gas-kinetic scheme (DUGKS), which is a re- the midpoint rule is used to integrate the surface flux.
cently developed finite-volume formulation of the model Boltz-   tn+1
1
mann equation by Guo et al. [29], Wang et al. [34]. In this section, ξ · ∇x φ (x j , ξ , tn )dtdV
Vj tn
we introduce the gas-kinetic model based on the Bhatnagar-Gross-  tn+1 
1
Krook (BGK) collision model in the incompressible limit and pro- = (ξ · n )φ (x j , ξ , tn )dsdt
vide a brief description of DUGKS algorithm. Corresponding to the Vj tn ∂V j

governing equations, the gas-kinetic equations can be constructed t
as [35]: = (ξ · n )φ (x j , ξ , tn+1/2 )ds (10)
Vj ∂V j
∂f f eq − f
+ ξ · ∇x f + a · ∇ξ f = f ≡ , (4a) The collision term is treated by the trapezoidal rule. The evolution
∂t τν equation for the velocity distribution can be written as:
t
∂g geq − g φ˜ (x j , ξ , tn+1 ) = φ˜ + (x j , ξ , tn ) − F (x , ξ , tn+1/2 ), (11)
+ ξ · ∇ x g = g ≡ , (4b) |V j | φ b
∂t τc
where φ˜ (x j , ξ , tn+1 ) and φ˜ + (x j , ξ , tn ) are auxiliary distribution
where f = f (x, ξ , t ) and g = g(x, ξ , t ) are the distribution functions functions introduced to remove implicity:
for velocity and temperature, respectively. Both functions are for
particles moving with velocity ξ at position x and time t. The t t
φ˜ = φ − m , φ˜ + = φ + m . (12)
two equilibrium distributions f eq and geq take the form of the 2 2
Maxwellian equilibrium: The microscopic flux Fφ (xb , ξ , tn+1/2 ) across the cell interface xb
  at the half time step tn+1/2 is defined as:
ρ (ξ − u ) 2

f eq = exp − , (5a)
(2π RT1 ) D/2 2RT1 Fφ (xb , ξ , tn+1/2 ) = (ξ · n )φ (xb , ξ , tn+1/2 )dS, (13)

  where φ (xb , ξ , tn+1/2 ) is distribution function at cell interface xb

T ( ξ − u )2 and half time step tn+1/2 . Since the auxiliary distributions are re-
geq = exp − , (5b)
(2π RT2 ) D/2 2RT2 lated to the original distribution function φ and equilibrium φ eq ,
the conservative flow variables can be computed from φ˜ directly.
where R is gas constant, D is the spatial dimension, T1 and T2 are Thus we can track the evolution of the distribution function φ˜ in-
constants determining the effective speed of the sound. The two stead of original distribution function φ .
relaxation times τν and τc are related to the viscosity ν = τν RT1 The key point of updating evolution equation Eq. (11) is to eval-
and thermal conductivity κ = τc RT2 . By modifying these two re- uate the net flux Fφ (xb , ξ , tn+1/2 ) properly, which requires the re-
laxation times, the Prandtl number P r = ν /κ can be adjusted. Un- construction of the original distribution function φ (xb , ξ , tn+1/2 ).
der the Boussinesq approximation, the external force is given as Eq. (8) is integrated along characteristic line with the end point
a = gβ [T − T0 ]ex . (xb ) from tn to tn + h (h = t/2 ), and the trapezoidal rule is ap-
Inour simulation,
√ in lattice units we set RT1 = RT2 = 10
√ and plied to the collision term,
u0 = β gT h = 0.1 so that the Mach number Ma = u0 / RT is
small enough to satisfy the incompressible limit. The maximum φ (xb , ξ , tn + h ) − φ (xb − ξ h, ξ , tn )
√ lo-
cal Mach number in the simulation is around Mamax = umax / RT ≈ h

0.25 for the high Rayleigh number cases. The time step in DUGKS
= m (xb , ξ , tn + h ) + m (xb − ξ h, ξ , tn ) . (14)
2
is determined by the Courant-Friedrichs-Lewy (CFL) condition:
In order to remove the implicity of Eq. (14), a second set of
x auxiliary distribution functions are defined as:
t = CF L min (6)
ξmax h h
φ̄ = φ − m , φ̄ + = φ + m . (15)
where CFL number is set to be 0.5 for all simulations, xmin is 2 2
the minimal grid spacing and ξmax is the maximal discrete particle With these two auxiliary distribution functions φ̄ and φ̄ + , we
velocity. The hydrodynamic variables are computed by: apply Taylor expansion around the cell interface xb and at time tn ,
   the Eq. (14) can be rewritten as:
ρ= f dξ , ρu = ξ f dξ , T = gdξ . (7)
φ̄ (xb , ξ , tn + h ) = φ̄ + (xb , ξ , tn ) − hξ · ∇x φ̄ + (xb , ξ , tn ). (16)
One way to include the external force term is to merge the Eq. (16) is totally explicit, b , ξ , tn ) and ∇x φ̄ + (x
n ) can φ̄ + (x, ξ , t
force term into the collision term. The DUGKS can be constructed be expressed in terms of the values at the cell center by appro-
based on two kinetic equations, we rewrite Eq. (4) as: priate reconstruction method. The density and velocity at cell in-
∂φ terface at half time step can be obtained from φ̄ (xb , ξ , tn + h ), then
+ ξ · ∇x φ = m ≡ φ + Fext , (8) the equilibrium function φ eq (xb , ξ , tn + h ) can be evaluated. Finally,
∂t
4
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

based on Eqs. (15) and (16), the original distribution function at conditions in our problem: adiabatic walls with zero heat flux and
cell interface xb at time tn + h ca be obtained by: isothermal walls with a fixed temperature. For adiabatic walls, the
2 τφ Neumann boundary condition can be achieved by:
φ (xb , ξ , tn + h ) = φ̄ (xb , ξ , tn + h )    
2 τφ + h g xw , ξ α , tn + h = g xw , −ξ α , tn + h
h τ h  
+ φ eq (xb , ξ , tn + h ) + φ Fext . (17) T1 Tw ∂ρw ai Tw ∂ Tw
2 τφ + h 2 τφ + h + 2τcWα ξa,i − −
T2 ρw ∂ xi RT2 ∂ xi
This original distribution function at half time step φ (xb , ξ , tn + + O (τc2 , Ma3 ), ξ α · n > 0. (20)
h ) can be used to evaluate the net flux Fφ (xb , ξ , tn+1/2 ) in Eq. (13).
Then the net flux can be used in Eq. (11) to update φ˜ . In the where ρ , T , and temperature gradients in the tangential directions
formulations above, we have one original distribution function φ can be approximated by their respective values from the last time
and four auxiliary distribution functions φ˜ , φ˜ + , φ̄ and φ̄ + . Con- step and the temperature gradient in the wall normal direction is
servative flow variables can be computed from any one of them, set to zero. The error introduced by the approximation is of the
the most convenient way is to track the evolution of φ˜ . The hy- order O (τc t ), which is typically less than O (t 2 ) and thus is
 
drodynamic variable can be obtained as ρ = f˜dξ , ρ u = ξ f˜dξ + acceptable.

t ρ a, T = g˜d ξ . While for the isothermal walls, the distribution function leaving
2
It is worth noting that, as pointed out by Wang et al. [34], for the wall should be constructed as:
DUGKS the non-uniform meshes can be easily employed without    
g xw , ξ α , tn + h = −g xw , −ξ α , tn + h
additional effort. Non-uniform meshes allow us to use finer mesh   
near the boundary to resolve the steep gradients of temperature ∂ u j ξa,i ξa, j ∂ ui
+ 2Wα Tw 1 + τc −
and velocity especially near the cavity walls. Our 3D flow code has ∂xj RT2 ∂ x j
been parallelized using 2D domain decomposition in the y and z
directions and MPI (Message Passing Interface) [32]. In the present + O (τc2 , Ma3 ), ξ α · n > 0. (21)
study, the D3Q19 model is employed for the discretization of the
where again the velocity at the last time step can be used to evalu-
particle velocity space,
ate the velocity gradient. Most previous works neglected all O (τc )
⎧
⎨(0, 0, 0 )c, α = 0, terms for the thermal boundary conditions, this could result in in-
accurate heat fluxes at the wall, especially in view of large vari-
ξ α = (±1, 0, 0 )c, (0, ±1, 0 )c, (0, 0, ±1 )c, α = 1 − 6,
⎩ ations of velocity and temperature near the wall and around the
(±1, ±1, 0 )c, (±1, 0 ± 1 )c, (0 ± 1, ±1 )c, α = 7 − 18, corners in our problem. The above Chapman-Enskog approxima-
(18) tions of the thermal boundary conditions are derived in detail in
√ Section Appendix A.
where c = 3RT , and the corresponding weighting coefficient are
W0 = 1/3, W1−6 = 1/18 and W7−18 = 1/36.
3. Results and discussions
2.3. Boundary treatment
This section will be divided into six parts. We first verify the
For a wall-bounded thermal flow, it is crucial to implement the code and compare our results with results from the literature. In
appropriate kinetic boundary conditions for the discrete distribu- order to elucidate the flow transition inside the vertical boundary
tion functions at the solid wall. One advantage of DUGKS is that layer, we simulate the natural convection flow at three different
we can apply boundary treatment right at the wall interface nodes Rayleigh numbers: Ra = 1.5 × 109 , Ra = 1.0 × 1010 , Ra = 1.0 × 1011 .
which lie exactly on the solid wall. In DUGKS, the microscopic flux These Rayleigh numbers are beyond the critical Rayleigh number
is calculated at the cell interface xb at the half time step tn + h. Racr = 3.19 × 107 for the current configuration [20]. Beyond the
We need to apply the  boundary condition for the distribution func- critical Rayleigh number, the flow becomes time-dependent and
tion f¯ xw , ξ , tn + h , where xw is the location of the solid wall. It the first instability starts to appear in the detached regions (left
is important to point out that the boundary treatment should be upper corner and right bottom corner). The high Rayleigh num-
consistent with the Chapman-Enskog approximation which is the bers we used allow us to observe the second instability inside
basis for deriving the hydrodynamic equations. The incompressible the boundary layers along the isothermal walls. The time-averaged
Navier-Stokes equation can be recovered with the Chapman-Enskog isotherms and instantaneous fluctuation velocity contours are used
expansion of the distribution function to the order of O (τν ). To be to display the overall flow structures. The time traces of velocity
consistent with the Navier-Stokes equations, the boundary treat- and temperature inside the vertical boundary layer are used to il-
ment should also retain the terms of the order O (τν ). For the fixed lustrate the nature of fluctuations. Then we examine the averaged
no-slip wall, we apply the bounce-back rule which assuming that flow field at different vertical locations of the cavity. Statistics of
the particle just reverses its velocity after hitting the wall: fluctuation fields are also presented. Finally, The heat transfer rates
    at upstream and downstream in the thermal boundary layer are
f xw , ξ α , tn + h = f xw , −ξ α , tn + h + O (τν2 , Ma3 ), ξ α · n > 0, analyzed.
(19)
where n is the unit vector normal to the wall pointing
 to thefluid 3.1. Code verification and computational cases
cell. The corresponding boundary condition for f¯ xw , ξ , tn + h can
be derived through Eq. (15). The ‘bounce-back’ expression above is Before considering high Rayleigh number cases, the three-
widely used for the fixed no-slip wall, and it can be shown that it dimensional code is tested at Ra = 1.0 × 107 by comparing with
is fully consistent with the Chapman-Enskog analysis, namely, all three-dimensional spectral results of Tric et al. [17]. We apply non-
terms of the order O (τν ) disappear due to the zero velocity on the uniform grids in all three directions to guarantee that the steep
wall. gradient near the wall can be resolved. Table 1 summarizes the re-
However, the situation for the thermal boundary conditions is sults of time-averaged overall Nusselt number Nuo at hot wall and
somewhat more complicated. There are two types of boundary maximum flow velocities in three directions at Ra = 1.0 × 107 . It

5
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Table 1
Comparison of our simulation results with the reference results for a cubical cavity
at Ra = 1.0 × 107 .

Ra = 107 Spatial resolution Nuo uˆmax vˆ max ˆ max

w

Present 128 × 128 × 128 16.234 0.2863 0.1428 0.0308

Tric et al. [17] 81 × 81 × 81 16.342 0.2882 0.1440 0.0313
Rel. error (%) 0.661 0.601 0.379 0.158

Fig. 4. Flow monitor for Ra = 1.5 × 109 .

where S is the parameter used to alter the degree of non-

uniformity, S is set to be 3 for all simulation cases. Then the lo-
cation of the cell center can be obtained by x(i ) = [xb (i ) + xb (i +
1 )]/2. To guarantee the resolution is fine enough to resolve the
steep gradient near the wall, we use two-dimensional convection
flow to perform the convergence study. As claimed by Trias et al.
[23], two-dimensional natural convection in a square cavity is a
Fig. 2. Time-averaged overall Nusselt number at hot wall as a function of resolution good approximation of the flow at the mid-plane of the three-
for three Rayleigh numbers: Ra = 1.5 × 109 , Ra = 1.0 × 1010 , Ra = 1.0 × 1011 . dimensional cubical cavity. Fig. 2 displays the overall Nusselt num-
ber at the hot wall as a function of resolution. In each case, the
time-averaged overall Nusselt number converges to a certain value
can be observed that the present results agree well with the refer- with the increase of the resolution. As the three-dimensional sim-
ence results. The largest relative error |uuˆmax | is less than 1%. More ulations require large computational resources, the resolutions we
0
code verification simulation cases are shown in Appendix A. employ represent a reasonable compromise between accuracy and
Regarding the numerical setup, the entire cubic cavity con- computational cost. In particular, the convergence for the Ra =
stitutes the full computational domain, Table 2 tabulates the 1.0 × 1011 case is not fully achieved. Besides, an increased spa-
stretched meshes and the non-dimensional time step size of three tial resolution will result in a smaller time step due to the CFL
x
cases. In order to resolve steep gradient of velocity and tempera- condition t = CF L ξ min , this will further increase the computa-
max
ture near the wall, non-uniform meshes are introduced in three di- tional cost. Even though the resolution we used is a compromise,
rections. For a set of stretched meshes with N grid points in each there are over 10 nodes insides the temperature and the velocity
direction, the location of the cell interface xb (i ) is given by: boundary layers. This satisfies the requirement of grid resolution
  for capturing the universal structure of the boundary layer accord-
1 tanh[S(i/N − 0.5 )] ing to Grötzbach [36] but not yet reaches the standard proposed
xb ( i ) = 1+ , i = 0, 1, 2, ..., N, (22)
2 tanh(S/2 ) by Shishkina et al. [37], both for the Rayleigh-Bénard setting. Ac-

Fig. 3. The temperature θ and the vertical velocity uˆ profiles at the mid-plane z = 0.5h and mid height x = 0.5h at Ra = 1.0 × 1010 with the resolutions N = 200 (I) and
N = 320 (II).

6
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Fig. 5. Time-averaged temperature field θ̄ for: (a )Ra = 1.5 × 109 ; (b)Ra = 1.0 × 1010 .

Fig. 6. Contour plot of instantaneous fluctuating vertical velocity uˆ = u−u

u0
(top) and horizontal velocity vˆ  = vu−0v (bottom) for three Rayleigh numbers: (a )(b) Ra = 1.5 × 109 ,
(c )(d ) Ra = 1.0 × 10 , (e )( f ) Ra = 1.0 × 10 .
10 11

7
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Fig. 7. Time traces of temperature at monitoring points in 9 vertical locations, x/h = 0.1 to 0.9 with increment of 0.1, at (a) Ra = 1.5 × 109 , (b) Ra = 1.0 × 1010 , (c) Ra =
1.0 × 1011 .

cording to the scaling of laminar thermal boundary layer thickness, to reach the stationary stage. After tˆ = 300, all the monitoring
we propose that the value of (Ra1/4 · xmin /l) should be less than statistics reach the stationary stage. The statistics used for analy-
a given threshold. Table 2 tabulates the value Ra1/4 · xmin /l for sis are averaged over 75 eddy turn-over times after tˆ = 300. The
each case. Since the case of Ra = 1.5 × 109 is shown to be well code was run on the National Center for Atmospheric Research’s
resolved, we could use the condition of Ra1/4 · xmin /l ≤ 0.30 as (NCAR-Wyoming) Supercomputer, known as Cheyenne, equipped
the resolution requirement for Ra = 1.0 × 1010 and Ra = 1.0 × 1011 with 2.3-GHz Intel Xeon E5-2697V4 processors. The computational
cases. Two resolutions (N = 200 and N = 320) are used to simu- domain are decomposed in the y and z direction, 400 processors
late the Ra = 1.0 × 1010 case for some additional grid sensitivity are employed for the case Ra = 1.5 × 109 . The wall clock time per
study. Fig. 3 shows the time-averaged temperature θ and vertical step is 9.56 × 10−2 s, and it takes 9.5 × 106 time steps to reach
velocity uˆ profiles for the case Ra = 1.0 × 1010 with two resolutions tˆ ≈ 300. For the case Ra = 1.0 × 1010 with N = 200 resolution, the
N = 200 and 320. The result shows that the boundary layer thick- wall clock time per step is the same with the case Ra = 1.5 × 109 .
ness is nicely captured by both resolutions. In the following sec- For the case Ra = 1.0 × 1011 , 1024 processors are used and the wall
tions, the results of the case Ra = 1.0 × 1010 are obtained with the clock time per step is 1.83 × 10−1 s, and 1.5 × 107 time steps are
N = 200 resolution unless otherwise mentioned. The resolution of needed to reach tˆ ≈ 300. Fig. 5 shows the time-averaged temper-
the highest Rayleigh number case (Ra = 1.0 × 1011 ) does not meet ature field θ̄ at the mid-plane (z = 0.5 ) of the cavity, when the
the requirement, and as such the accuracy of results for this case flow reaches the statistically stationary stage. Generally speaking,
should be taken with caution. the time-averaged isotherms is asymmetric about the cavity center,
extremely thin boundary layers are developed along the isother-
3.2. Overall flow structures mal wall. With the increase of the Rayleigh number, the boundary
thickness becomes smaller.
Before a detailed analysis of the boundary layer, we first For these high Rayleigh numbers, the flow is unsteady and fluc-
present the overall structures of the flow. All statistics used in tuates in time. Fig. 6 shows the instantaneous velocity fluctuation

the following analysis are obtained after the flow reaches the field u ˆ = uu−ū at three Rayleigh numbers after flow reaches the
0
quasi-steady or statistically stationary stage. The development of stationary stage. The cavity center is no longer steady and now
the flow is tracked by a few spatially-averaged flow monitors fluctuates in time. The vortices are generated at the downstream
(uˆ2 , vˆ 2 , w
ˆ 2 , p, prms , T , Trms ). For example, Fig. 4 shows the spatial- of the boundary layers and are shed from the upper corner on the
averaged flow statistics for the case Ra = 1.5 × 109 , as a function heated wall. The size of the vortices becomes smaller with the in-
of time. This shows that different quantities took different times crease of the Rayleigh numbers.

8
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Fig. 8. Time trace of velocity at monitoring points x = 0.1h ∼ x = 0.9h at (a) Ra = 1.5 × 109 , (b) Ra = 1.0 × 1010 , (c) Ra = 1.0 × 1011 .

3.3. Time traces of the temperature and velocity ate over time, and the fluctuation amplitude and the frequency
range increase with Ra. Similar fluctuation behaviors of the ve-
Driven by the buoyancy force, the flow gradually develops from locity and temperature in the downstream region are reported by
laminar to turbulent flow along the vertical boundary layer adja- Xin and Quéré [38] for vertical boundary layer at the hot wall of
cent to two isothermal walls. The upstream part is usually lami- a two-dimensional differentially heated cavity of aspect ratio at
nar, and up to a point after which the flow becomes turbulent. As 4. Based on the time traces of uˆ velocity for Ra = 1.0 × 1010 and
the velocity signals of laminar and turbulent flow are evidentially Ra = 1.0 × 1011 , Table 3 tabulates the partition of energy in the
different, we track the time traces of the temperature and veloc- mean and in the fluctuating motion. The ratio is negligible in the
ity inside the vertical boundary layer along the hot wall. Firstly, upstream for both Rayleigh numbers, while in the downstream re-
we need to ensure the monitoring points are located inside the gion more kinetic energy participates in fluctuation. Especially for
boundary layer. Based on the time-averaged temperature and ve- the Rayleigh number Ra = 1.0 × 1011 , the relative partition of fluc-
locity profile at nine different vertical locations (x = 0.1h − 0.9h), tuation energy at the height x = 0.9h is more than 14%.
for temperature evolution, a set of monitoring points are cho- To investigate the fluctuation frequency of the velocity and tem-
sen at the mid-plane (z = 0.5h), as a y location where T (x, y, z = perature insides boundary layer, we provide the spectra of velocity
T +T (x,z=0.5h )  ˆ
0.5h ) = h ∞ 2 , T∞ (x, z = 0.5h ) is the temperature outside fluctuation u ( f ) (u = u − ū ) and temperature fluctuation θ  ( fˆ) in
u0
the boundary layer at the given x level. Due to temperature strati- the Figs. 9 and 10. The spectra are obtained from time trace signals
fication in the core region, T∞ (x ) changes with the vertical location spanning over a time duration of approximately 70h/u0 for the two
x. For the velocity trace, the monitoring points are at the location lower Rayleigh number cases, and 100h/u0 for the Ra = 1.0 × 1011 .
where u = umax at that height. In this way, we can guarantee that Three locations are selected to show the transition of the velocity:
all the monitoring points are located inside the boundary layer. x = 0.2h near the starting point of the upstream, x = 0.5h at the
The time traces of temperature and velocity at 9 monitoring mid-height, and x = 0.9h at the downstream. As the temperature
points for each of the three different Rayleigh numbers are shown time trace is similar to the velocity time trace, we only provide
in Figs. 7 and 8. All the time traces are obtained after the flow the spectra of temperature θ  ( fˆ) for the location x = 0.9h. For two
reaches the stationary stage. From the plots, we can observe that locations at the upstream, the spectral magnitudes are relatively
the upstream region with time-independent behavior is from x = 0 small and the velocity spectra are dominated by low-frequency os-
to x = 0.7h for the cases Ra = 1.5 × 109 and 1.0 × 1010 ; but for cillations. It is also noted that the magnitude of the spectra for
Ra = 1.0 × 1011 the upstream region is reduced to the region from the location x = 0.2h is very small, especially for the cases Ra =
x = 0 to x = 0.6h, roughly speaking. For the cases Ra = 1.5 × 109 1.5 × 109 and Ra = 1.0 × 1010 , the oscillations in the upstream are
and 1.0 × 1010 the flow is laminar and steady in the upstream negligible. While for the upstream, energy concentrates near the
region, while for Ra = 1.0 × 1011 it is laminar and unsteady. In mean mode ( fˆ = 0), the downstream velocity fluctuation spreads
contrast, the temperature and velocity of the downstream fluctu- out into the high-frequency region. As pointed out by Janssen and

9
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Fig. 9. Spectra of velocity fluctuation uˆ ( fˆ) at three monitoring points. From top to bottom: (a) (b) (c) for Ra = 1.5 × 109 ; (d) (e) (f) for Ra = 1.0 × 1010 ; (g) (h) (i) for
Ra = 1.0 × 1011 .

Table 2
The non-uniform meshes used for the simulations.
xmin
Ra 4 · xlmin xmax u0 t
1
Ra Spatial resolution l xmin h

1.5 × 109 200 × 200 × 200 1.52 × 10−3 0.299 5.4587 3.1623 × 10−5
1.0 × 1010 [I] 200 × 200 × 200 1.52 × 10−3 0.481 5.4587 3.1623 × 10−5
1.0 × 1010 [II] 320 × 320 × 320 9.44 × 10−4 0.299 5.4869 1.9764 × 10−5
1.0 × 1011 320 × 320 × 320 9.44 × 10−4 0.531 5.4869 1.9764 × 10−5

Table 3
uˆ 2
Relative partition of energy in the mean and in the fluctuation (%).
uˆ 2

Ra 0.1h 0.2h 0.3h 0.4h 0.5h 0.6h 0.7h 0.8h 0.9h

1.0 × 10 10
1.1088e-4 2.0790e-4 8.6940e-4 7.2612e-4 5.0661e-4 7.1279e-4 0.0286 0.7535 0.9667
1.0 × 1011 0.0022 0.0053 0.0087 0.0123 0.0403 0.0689 0.6112 3.9789 14.1829

Henkes [11], the time-periodic flow is characterized by the phe- fˆ = 0.8, this agrees well with the two-dimensional results of the
nomenon that the power spectra of the velocity show a peak at Janssen and Henkes [11]. They obtain the peak frequency around
a single frequency; and the quasi-periodic flow shows at lease fˆ = 0.8 at Ra = 7.5 × 108 from the temperature time trace at the
two spikes (two fundamental frequencies and the linear combina- same height. This non-dimensional peak frequency of fˆ = 0.8 com-
tion of the fundamental spike) at two frequencies in their power pares well with the period of about 1.21 observed in time trace of
spectra; while turbulent flow has a broadband power spectra. The the temperature and the velocity as shown in Figs. 7(a ) and 8(a ).
peak frequency at the x = 0.9h of the case Ra = 1.5 × 109 is around From the spectra at the location x = 0.9h for three Rayleigh num-

10
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Fig. 10. Spectra of temperature fluctuation θ  ( fˆ) at monitoring points x = 0.9h. For (a) Ra = 1.5 × 109 , (b) Ra = 1.0 × 1010 , (c) Ra = 1.0 × 1011 .

Fig. 11. (a) Time-averaged temperature, (b) Vertical velocity profiles in the mid-plane z = 0.5h at x = 0.1h ∼ x = 0.9h, Ra = 1.5 × 109 (black line), Ra = 1.0 × 1010 (red dash
line), Ra = 1.0 × 1011 (blue dash line). The purple squares in (b) are experimental data of Salat et al. [6] at Ra = 1.5 × 109 . (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

11
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Fig. 12. (a) Temperature profiles with abscissa scaled by Ra1/4 , and (b) vertical velocity profiles with abscissa scaled by Ra1/4 in the mid-plane z = 0.5h at x = 0.1h ∼ x = 0.9h,
Ra = 1.5 × 109 (black line), Ra = 1.0 × 1010 (red dash line), Ra = 1.0 × 1011 (blue dash line). (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

bers, we can observe that the two lower Ra cases are characterized shed vortices to produce quasi-periodic or turbulent flow down-
by spectra with spikes at two fundamental frequencies plus spikes stream. With the increase of the Rayleigh number, the downstream
at linear combinations of the fundamental frequencies which in- of the boundary layer becomes turbulent eventually.
dicate the flow is quasi-periodic; and the case Ra = 1.0 × 1011 has
broadband spectra instead of the specific peaks at a certain fre- 3.4. Averaged flow field
quency. The spectra of the location x = 0.5h for Ra = 1.0 × 1011
is characterized by a single spike at low-frequency and this phe- The time-averaged temperature and vertical velocity profiles at
nomenon shows the flow in this location is periodic. In the case the mid-plane are displayed in Fig. 11. As shown in Fig. 11(a ) and
of Ra = 1.0 × 1011 , the frequency spectrum of the velocity fluctu- (b), both the thicknesses of velocity and thermal boundary in-
ations Fig. 9(i) appears to decay exponentially, similar to the dis- crease with increasing distance from the starting location of the
sipation range of high-Reynolds number turbulence, the frequency boundary layer. At the same height, with the increase of Ra, the
spectrum of the temperature fluctuations Fig. 10(c) exhibits a more thickness of boundary layer becomes smaller and the vertical ve-
complex shape implying perhaps different structures of tempera- locity peak moves towards the wall. And at mid-height x = 0.5h,
ture fluctuations compared to the velocity fluctuations. From the our results agree well with the experimental results of Salat et al.
analysis above we can observe that insides the boundary layer, the [6]. Corresponding to the thinner boundary layer, the temperature
upstream part remains laminar until when the local flow starts to gradient near the boundary becomes larger at higher Ra. Fig. 12(a)

12
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Fig. 13. Averaged turbulent fluctuation statistics from top to bottom: uˆrms , vˆ rms , −uˆ vˆ  × 104 . Left: Ra = 1.5 × 109 , middle: Ra = 1.0 × 1010 , right: Ra = 1.0 × 1011 .

and (b) show the time averaged temperature and velocity profiles by vertical velocity uˆ and horizontal velocity vˆ , we measure the
with the abscissa scaled by the laminar scaling factor Ra1/4 [1]. For Reynolds stress −uˆ vˆ  . It should be noted that the velocity w
ˆ in the
the upstream part (x = 0 to x = 0.7h), almost identical temperature lateral direction is almost one order of magnitude lower than uˆ, vˆ .
and velocity profiles are obtained after the scaling. Similar results For Ra = 1.5 × 109 and 1.0 × 1010 , the active turbulence is located
are reported by Trias et al. [23]. Discrepancies occur only in down- at left top and bottom right corners indicating regions for the first
stream part, and the differences start to appear at the point where occurrence of turbulence. For these two Rayleigh numbers, the tur-
temperature and velocity start to fluctuate and reach its maximum bulent fluctuation statistics start to be significant at the height x =
around x = 0.8h. This observation again confirms that for the Ra 0.7h. Outside the boundary layer, all turbulent statistics are almost
range being studied, the major part (x = 0 − 0.7h) of the vertical zero. For Ra = 1.0 × 1011 , the distribution of turbulent fluctuation
boundary layer is still laminar, and the transition from the lami- statistics are more complex than the lower Ra cases. Small fluctu-
nar to time-dependent turbulent flow happens around the height ations exist in the core region, the high value region of turbulent
x = 0. 7h. fluctuation statistics still concentrates in the downstream corners.
And the transition point moves upstream, the significant turbulent
3.5. Turbulent fluctuation statistics fluctuations appear around the height x = 0.65h. From this local
distribution of the turbulence statistics, we can clearly observe that
Turbulent fluctuation statistics (uˆrms , vˆ rms , −uˆ vˆ  ) of the mid- for the lower two Ra cases, the core region and upstream of the
plane z = 0.5h are displayed in Fig. 13. As the flow is dominated vertical boundary are still laminar, the turbulent fluctuations are

13
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Table 4
Time-averaged mean Nusselt number and Nu − Ra correlations.

Nuo Nuo Nuup Nudown Nudown

Ra Nuo Ra1/4 Ra1/3
Nuup Ra1/4
Nudown Ra1/4 Ra1/3

1.5 × 109 58.08 0.2953 0.0508 73.30 0.3725 22.65 0.1151 0.0198
Fusegi et al. [18] 63.07 0.3205 0.0551
Wang et al. [22] (1.0 × 109 ) 52.08 0.2929 0.0521
1.0 × 1010 [I] 89.29 0.2824 0.0414 112.1 0.3545 36.05 0.1140 0.0167
1.0 × 1010 [II] 94.21 0.2979 0.0437 134.1 0.4240 35.91 0.1135 0.0167
Fusegi et al. [18] 107.7 0.3406 0.0500
Wang et al. [22] 89.2 0.2821 0.0414
1.0 × 1011 154.6 0.2750 0.0333 217.6 0.3870 62.59 0.1113 0.0135

only significant in the downstream corners. It is suggested that for

Ra = 1.5 × 109 and 1.0 × 1010 , the flow is only weakly turbulent as
pointed out by Salat et al. [6]. For Ra = 1.0 × 1011 , the fluctuations
exist in the whole domain, but the high value of turbulent statis-
tics concentrate in the downstream corners, and the flow becomes
turbulent. To the authors’ knowledge, Ra = 1.0 × 1011 is the highest
Ra reported for three dimensional natural convection in a differen-
tially heated cubical cavity with adiabatic horizontal walls.

3.6. Heat transfer

From the averaged flow field we can clearly observe that the
thickness of the boundary layer becomes smaller with the increase
of Ra, and thinner boundary layer leads to larger temperature gra-
dient near the boundary. Steep temperature gradient intensifies the
heat transfer near the boundary. Mean Nusselt number Nuo at hot
wall is defined as:
    1 1 
1 h h
∂ T (x, z )  ∂θ (xˆ, zˆ) 
Nuo =
∂ y y=0  dxdz =  dxˆdzˆ Fig. 14. Time averaged overall Nusselt number for the Rayleigh number range:
h T 0 0 0 0 ∂ yˆ yˆ=0 103 ≤ Ra ≤ 1011 . The two data points at Ra = 1.0 × 1010 case represent results from
(23) two resolutions, as shown in Table 2, with the higher resolution giving a higher Nu
value.
Table 4 shows time-averaged mean Nusselt number Nuo at the
hot wall for three different Rayleigh numbers. Our results are in
excellent agreement with Wang et al. [22]. It should be noted As pointed out in the previous sections, the flow insides the
that the Nu − Ra correlation reported by Wang et al. [22] was vertical boundary layer transitions from steady to time-dependent
given in two Ra ranges respectively (Nuo = 0.127Ra0.3052 for 103 ≤ flows at the height x = 0.7h for Ra = 1.5 × 109 and Ra = 1.0 × 1010 .
Ra ≤ 107 ; Nuo = 0.3408Ra0.241 for 107 ≤ Ra ≤ 1010 ), while Fusegi For Ra = 1.0 × 1011 the transition point is moved to x = 0.65h.
et al. [18] proposed Rayleigh-Nusselt dependence in one correla- To distinguish the heat transfer behavior of upstream and down-
tion (Nuo = 0.163Ra0.282 for 103 ≤ Ra ≤ 1010 ). And the resolution stream, we compute the mean Nusselt number for upstream and
used by Wang et al. [22] is relatively finer than Fusegi et al. [18]. downstream respectively. Table 4 shows Nusselt number for up-
In the current study, we evaluate the time-averaged overall Nus- stream Nuup which is integrate from x = 0 to x = 0.7h for two
selt number for the Rayleigh range 103 ≤ Ra ≤ 1011 as shown in lower Rayleigh numbers, and integrate from x = 0 to x = 0.65h for
Fig. 14. In the steady regime, our results agree well with both ref- Ra = 1.0 × 1011 . Even with this domain separation, both Nuup and
erence fitting. In the unsteady flow regime, the thermal bound- Nudown are close to classical laminar scaling of Ra1/4 . This implies
ary layer becomes thinner with the increase of the Rayleigh num- that the local transition to turbulent flow here has a negligible ef-
ber. The maximum number of grid points used by Fusegi et al. fect on the heat transfer rate, due to a very limited domain and
[18] is 122 × 62 × 62. Wang et al. [22] employed 2003 non-uniform the constraint of the top and bottom walls.
meshes for Ra = 1010 . In the current study, the resolution we used
for the highest Rayleigh number (Ra = 1.0 × 1011 ) may not be ad- 4. Conclusion
equate, as such the average Nu number for this case is not shown
in the figure. A set of direct numerical simulations for natural convection
From Table 4, we can observe that the Nu − Ra correlation for flow in a differentially heated cubical cavity are performed, with
mean Nusselt number at the hot wall is much closer to laminar the goal to improve our understanding of flow transition insides
scaling factor Ra1/4 than turbulent scaling Ra1/3 expected from nat- the vertical boundary layer adjacent to isothermal walls and its
ural convection over an unbounded flat plate [1]. When the con- influence on heat transfer rate. The Rayleigh number range con-
vection flow reaches the stationary stage, due to temperature strat- sidered (Ra = 1.5 × 109 to Ra = 1.0 × 1011 ) extends these studied
ification, the cold fluid is distributed in the bottom while hot fluid in the literature. A three dimensional DUGKS code using domain
in the top, thus the temperature gradient in the upstream of the decomposition and MPI is created and verified by comparing re-
vertical boundary would be larger than in the downstream. The sults with reference works at Ra = 107 . Specifically, an improved
reason for Nu − Ra correlation closer to laminar scaling factor is treatment of temperature and velocity boundary conditions is pro-
that most heat transfer happens in the upstream of the vertical posed, based on a consistency consideration with the Chapman-
boundary layer where it is almost a laminar flow. Enskog approximation.

14
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

To specify the transition location, several statistics are mea- Appendix A. Chapman-Enskog analysis of boundary condition
sured. Discernible different behaviors of the upstream flow and
the downstream flow are observed in the time traces of temper- For the mesoscopic method used here, it is crucial to apply ap-
ature and velocity. While the temperature and velocity remain es- propriate boundary condition to the unknown distribution func-
sentially constant in the upstream, fluctuations are observed in the tions near a wall. For the current double distribution function
downstream region. The spectra of velocity time traces also show model, velocity and temperature boundary condition should be ap-
that the velocity fluctuations spread out to high frequencies in the plied to f (xw , ξ α , t ) and g(xw , ξ α , t ), respectively. It is important to
downstream region. The upstream insides the boundary layer re- point out that the boundary condition should be fully consistent
mains laminar until a transition location, and the vortices start to with the Chapman-Enskog approximation. Based on the Chapman-
eject in the downstream yielding quasi-periodic or turbulent flows. Enskog expansion, the distribution functions can be written ap-
With the increase of the Rayleigh number, the downstream of the proximately as:
thermal boundary layer becomes turbulent eventually.  
Time-averaged velocity and temperature profiles are obtained. ∂ f eq ∂ a (ξ − u j ) eq
f = f eq − τν + (ξ j f eq ) + τν j j f + O (τν2 ),
When scaled by laminar thermal boundary thickness scaling, al- ∂t ∂xj RT1
most identical profiles are obtained for the upstream region, and
(A.1a)
the discrepancies appear downstream of the transition location
(x = 0.65h to 0.7h depending on Ra). The turbulent fluctuation  
statistics in the mid-plane also confirm that the transition loca- ∂ geq ∂
g = geq − τc + (ξ j geq ) + O (τc2 ). (A.1b)
tion for two lower Rayleigh numbers is at x = 0.7h, while for ∂t ∂xj
Ra = 1.0 × 1011 the transition location is approximately x = 0.65h.
The scalings of heat transfer rate in the upstream and the down- Based on the Chapman-Enskog analysis we realize that the

stream are also different. As the main part of the vertical boundary third-order moments of equilibrium ξi ξ j ξk f eq dξ is needed for
layer is laminar, the time-average mean Nusselt number for the hot momentum flux evaluation, which requires third-order Hermite
wall is closer to laminar scaling Ra1/4 . The local transition to tur- expansion of f eq,N=3 and Gauss-Hermite quadrature with sixth-
bulent flow observed at high Ra numbers observed in the simula- degree of precision. For energy equation, the second-order mo-

tion was found to have a negligible effect on the heat transfer rate, ments of equilibrium ξi ξ j geq dξ is needed for the evaluation of
perhaps due to a very limited domain and the constraint of the top energy flux. As the Mach number is very small in the current work,
and bottom walls. the third-order terms can be neglected. The equilibrium distribu-
Despite the efforts we made here for simulating natural convec- tion is expanded to the second-order as:
tion flows at high Rayleigh numbers in the three-dimensional cav-  
ity, three-dimensional flow structures are far from fully explored. ξα · u ( ξ α · u )2 u2
eq,N=2
fα = Wα ρ 1 + + − + O (Ma3 ),
Especially for turbulent natural convection flow in a cavity, it re- RT1 2(RT1 )2 2RT1
quires the significant computational resources to obtain accurate (A.2a)
statistics. Further research is thus needed in this direction.

 
Declaration of Competing Interest ξ · u ( ξ α · u )2 u2
geq,N=2
α = Wα T 1 + α + − + O (Ma3 ).
RT2 2(RT2 ) 2 2RT2
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to (A.2b)
influence the work reported in this paper. For temperature and velocity boundary conditions, we can de-
rive the following equations:
CRediT authorship contribution statement
a ju j ∂
f (ξ α ) − f (ξ ᾱ ) = (1 − τν )( f eq+ − f eq− ) − τν ( f eq+ − f eq− )
Xin Wen: Conceptualization, Data curation, Formal analysis, RT1 ∂t
Writing - original draft, Writing - review & editing. Lian-Ping ∂ eq+ aξ
− τν ξa, j ( f + f eq− ) + τν j a, j ( f eq+ + f eq− ), (A.3a)
Wang: Conceptualization, Methodology, Writing - review & edit- ∂xj RT1
ing, Resources, Funding acquisition. Zhaoli Guo: Conceptualization,
Writing - review & editing. Dauren B. Zhakebayev: Funding acqui-
∂ eq+ eq−
sition. g(ξ α ) − g(ξ ᾱ ) = (geq+ − geq− ) − τc (g − g )
∂t
∂ eq+ eq−
Acknowledgments − τc ξa, j ( g + g ), (A.3b)
∂xj
This work has been supported by the U.S. National Sci-
ence Foundation (CNS-1513031, CBET-1706130), the National Nat- ∂ eq+ eq−
g(ξ α ) + g(ξ ᾱ ) = (geq+ + geq− ) − τc (g + g )
ural Science Foundation of China (91852205, 11988102, 91741101 ∂t
& 11961131006), Kazakhstan Ministry of Education and Science ∂ eq+ eq−
(AP05132121), the National Numerical Wind Tunnel program, − τc ξa, j ( g − g ). (A.3c)
∂xj
Guangdong Provincial Key Laboratory of Turbulence Research and
Applications (2019B21203001), Guangdong-Hong Kong-Macao Joint For simplicity, we use superscript + to represent the particles ξα
Laboratory for Data-Driven Fluid Mechanics and Engineering Ap- bouncing back from the wall, while the superscript - represents the
plications(2020B1212030 0 01), Shenzhen Science & Technology Pro- particles moving in the opposite direction to ξα . geq+ + geq− , geq+ −
gram (Grant No. KQTD20180411143441009), and Guangdong Sci- geq− , heq+ + heq− and heq+ − heq− can be derived from the equilib-
ence & Technology Program (2020B1212030 0 01). Computing re- rium distributions:
sources are provided by the Center for Computational Science and  
( ξ α · u )2 u2
Engineering of Southern University of Science and Technology and f eq+ + f eq− = 2Wα ρ 1 + − , (A.4a)
by National Center for Atmospheric Research (CISL-UDEL0 0 01).
2(RT1 )2 2RT1

15
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

 
ξα · u
f eq+ − f eq− = 2Wα ρ , (A.4b)
RT1
 
( ξ α · u )2 u2
geq+ + geq− = 2Wα T 1 + − , (A.4c)
2(RT2 )2 2RT2
 
eq+ eq− ξα · u
g −g = 2Wα T . (A.4d)
RT2
∂u
It is important to point out that the ∂∂ut and ∂ x i operation will
j
reduce the Mach order by one order. To keep the error term to
O (τν2 , Ma3 ), the equilibriums need to be expanded to one-order
higher, namely, to the fourth-order for f eq and the third-order
for geq . However, all the high order terms contain u3 and u4 ,
which can be canceled under the no-slip boundary condition due
to derivative by parts. Thus, we did not explicitly write out the
high order terms.
Substituting these expressions into Eq. (A.3a), as all walls are
assumed to be no-slip in velocity, we can set u = 0 to terms not
related to time- or spatial- derivatives. The u2 and u3 terms are Fig. A.15. The comparison of mean Nusselt number Num (zˆ) distribution among
also eliminated due to derivative by parts, then the following re- present DUGKS results and reference results Peng et al. [27]; Wang et al. [39]; Wang
[21]).
sults can be obtained:
ξa,i ∂ρ ui ∂ρ
f (ξ α ) − f (ξ ᾱ ) = −2τν Wα − 2τν Wα ξa, j
∂tRT1 ∂xj Leading order energy equation is used to replace the time
∂Tu
a j ξa, j derivatives ∂∂Tt = − ∂ x j + O (τ ). Then, the final result for isother-
+ 2τν Wα ρ . (A.5) j
RT1 mal wall is
 
The Euler equations can be used to replace the time derivative ∂ u j τc ξa,i ξa, j ∂ ui
ρ ∂∂uti = −ρ u j ∂∂ xui − ∂∂xp + ai ρ + O (τ ), therefore g(ξ α ) + g(ξ ᾱ ) = 2Wα T 1 + τc −
j i ∂xj RT2 ∂ x j
 
ξa,i ∂ρ + O (τc2 , Ma3 ). (A.12)
f (ξ α ) − f (ξ ᾱ ) = −2τν Wα −RT1 + ρ ai
RT1 ∂ xi In summary, for no-slip boundary condition, the standard
∂ρ a j ξa, j “bounce-back” is consistent with the Chapman-Enskog analysis.
− 2τν Wα ξa, j + 2τν Wα ρ + O (τν2 , Ma3 ).
∂xj RT1 However, for temperature boundary conditions, most literature
works only keep the O (1 ) term. In fact, the O (τc ) terms have in-
(A.6)
fluence on the heat flux at the boundary. In actual implementation,
Finally, we have the mesoscopic representation of the no-slip the derivatives of hydrodynamic variables can be evaluated approx-
boundary condition as imately using the values from last time step. To show that the
f (ξ α ) − f (ξ ᾱ ) = 0 + O (τν2 , Ma3 ). (A.7) temperature boundary condition can be accurately implemented
by the current kinetic boundary condition, we performed a sim-
Substituting the geq
equilibrium into Eq. (A.3b) with uw = 0, the ulation of natural convection in an air-filled cubical cavity with
adiabatic boundary condition can be written as: Ra = 1.0 × 103 . Fig. A.15 shows the mean Nusselt number distribu-
 
∂ ξ ·u ∂ tion along the z-direction, where Num (zˆ) is defined as:
g(ξ α ) − g(ξ ᾱ ) = −τc 2Wα T α − τc ξa, j (2Wα T )  
∂t RT2 ∂xj 1
∂θ (xˆ, zˆ) 
Num (zˆ) =  dxˆ. (A.13)
+ O (τc2 , Ma3 ). (A.8) 0 ∂ yˆ yˆ=0
Again the Euler equations are used to replace the time deriva- We compared the mean Nusselt distribution of the current
∂u ∂u
tive ρ ∂ ti = −ρ u j ∂ x i − ∂∂xp + ai ρ + O (τ ), to obtain DUGKS results and those from the literature [21,27,39]. Our re-
j i
  sults agree well with the conventional CFD results of Wang et al.
ξa,i T 1∂ρ RT1 [39] and LBM results of Peng et al. [27]. The results provided by
g(ξ α ) − g(ξ ᾱ ) = −2τcWα − + ai
RT2 ρ ∂ xi Wang [21] is obtained using the DUGKS scheme with only O (1 )
boundary condition implementation. It is obvious that the bound-
∂T
− 2τcWα ξa, j + O (τc2 , Ma3 ). (A.9) ary condition is improved using our current kinetic boundary con-
∂xj dition including the O (τc ) terms.
The final result of mesoscopic representation of the adiabatic To further confirm the convergence order of the DUGKS, we
boundary condition becomes performed a series of two-dimensional simulations of natural con-
  vection in a square cavity at Ra = 1.0 × 103 with different meshes.
T1 T ∂ρ ai T ∂ T
g(ξ α ) − g(ξ ᾱ ) = 2τcWα ξa,i − − Simulations are conducted with meshes 20 × 20, 40 × 40, 80 ×
T2 ρ ∂ xi RT2 ∂ xi 80, 160 × 160, 320 × 320 and the CFL numbers are adjusted to keep
+ O (τc2 , Ma3 ). (A.10) the time step constant accordingly. The L2 erros in temperature
and velocity field are measured in Table A.5, where the L2 error
For isothermal boundary condition, we can derive the following
is defined as:
expression from Eq. (A.3c): 
∂T ξa,i ξa, j ∂ ui x,y |φ (x, y, t ) − φe (x, y, t )|2
g(ξ α ) + g(ξ ᾱ ) = 2Wα T − 2Wα τc − 2Wα τc T . (A.11) E (φ ) =  (A.14)
∂t RT2 ∂ x j x,y |φe (x, y, t )|
2

16
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Fig. A.16. Time trace of temperature θ at location (xˆ, yˆ ) = (0.88, 0.15 ) for: (a) Ra = 1.6 × 108 ;(b) Ra = 1.8 × 108 ; (c) Ra = 1.82 × 108 ; (d) Ra = 1.84 × 108 ; (e) Ra = 2.0 × 108 .

where φ = θ or uˆ, and φe is the benchmark value with the mesh formed. The non-uniformity parameter is set to be S = 3, and the
320 × 320. At least second-order convergence of the DUGKS is con- minimum grid spacing is xmin = 6.0219 × 10−4 L. Fig. A.16 shows
firmed. time trace of temperature at monitoring point for the simulation
Besides the steady natural convection simulation, we also vali- with five different Rayleigh numbers. The red line represents the
date our code by calculating the critical Rayleigh number of tran- time integration was started from the steady flow of the case
sition to unsteady flow regime for two-dimensional natural con- Ra = 1.8 × 108 and the blue dash line represents the simulation
vection in a square cavity. A set of simulations of two-dimensional was started from the unsteady solution of the case Ra = 2.0 × 108 .
natural convection with non-uniform 500 × 500 meshes are per- Steady results are obtained for the cases Ra = 1.6 × 108 and Ra =

17
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

Table A.5 Table B.6

Error and convergence order in velocity and temperature. L2 Errors and convergence order in velocity and pressure with different
mesh sizes (dt = 1.0 × 10−5 ).
N 20 40 80 160
N 20 40 80 160
E (θ ) 5.3583 × 10−3 9.2810 × 10−4 1.5528 × 10−4 2.3295 × 10−5 U0 dt
dx
2 × 10−4 4 × 10−4 8 × 10−4 1.6 × 10−3
order - 2.5294 2.5794 2.7368 cs dt
2 × 10−3 4 × 10−3 8 × 10−3 1.6 × 10−2
E (uˆ ) 4.3024 × 10−2 7.2083 × 10−3 1.2445 × 10−3 2.1210 × 10−4 dx

order - 2.5774 2.5341 2.5527 E (u ) 9.343971E-3 2.187037E-3 5.718380E-4 1.371459E-4

order - 2.0951 1.9353 2.0599
E ( p) 2.562765E-2 7.614341E-3 1.927406E-3 4.033651E-4
order - 1.7509 1.9821 2.2565
1.8 × 108 , for both initial conditions. For the cases Ra = 1.82 × 108
and Ra = 1.84 × 108 , it shows that the solution first experiences
damped oscillations and then reaches periodic oscillations. With Table B.7
L2 Errors and convergence order in velocity and pressure with different time
these high resolution results, the critical Rayleigh number would
step sizes (fixed CFL number).
be between 1.80 × 108 and 1.82 × 108 , which is consistent with
the the Le Quéré and Behnia result for Rac , Rac = (1.82 ± 0.01 ) × dt 2.0 × 10−3 1.0 × 10−3 5.0 × 10−4 2.5 × 10−4
N 20 40 80 160
108 . We also note that the amplitude of oscillations increases with U0 dt
0.04 0.04 0.04 0.04
Ra for Ra ≥ 1.82 × 108 . dx
cs dt
0.4 0.4 0.4 0.4
dx

E (u ) 5.798385E-1 1.059667E-1 1.384611E-2 1.317480E-3

Appendix B. Order of accuracy test
order - 2.4520 2.9361 3.3936
E ( p) 8.176620E-1 1.914017E-1 2.716178E-2 2.573582E-3
To validate the second-order accuracy of our scheme in both order - 2.0949 2.8170 3.3997
space and time, we simulated the unsteady 2D Taylor-Green vor-
tex flow in a square domain. The Taylor-Green vortex flow is an
exact solution of the Navier-Stokes equation in a two-dimensional  
∂ ui ci c j ∂ ui
where fα(1 ) = fα
eq
periodic domain, representing the viscous decay process of a vorti- ∂ xi − RT1 ∂ x j . Then we can construct a new
cal flow. The velocity and pressure fields of this unsteady flow are set of the initial distribution as:
given as:
 2π x   2π y  2
τν fα(1) (x, 0 ) ≈ fα (x, t ) − fαeq [ρ (x, t ), u(x, t )] + O (τν2 ),
− 8π 2 ν t
u(x, y, t ) = −U0 cos sin e L , (B.1a) (B.5a)
L L
 2π y   2π x  8π 2 ν t fα (x, 0 ) = τν fα(1) (x, 0 ) + fαeq [ρ (x, t ), u(x, 0 )]. (B.5b)
v(x, y, t ) = U0 cos sin e− L2 , (B.1b)
L L 4. Repeat step 2 and 3 until the pressure and stress fields at the
    initial time converge.
1 2π 2π 16π 2 ν t
p(x, y, t ) = − U02 cos (x − y ) cos ( x + y ) e − L2 , The parameters used in the simulations are: Re = 10 0 0, ν =
2 L L
(B.1c) 0.001, L = 1.0, U0 = 1.0, RT = 100. We compare the velocity pro-
files and pressure profiles at two dimensionless times (not shown)
where ν is the shear viscosity. In order to use the analytical solu-
with the analytical solution, and find that the numerical solution
tion as the rigorous benchmark, we must initialize the distribution
agrees well with the theoretical solution. To confirm the conver-
functions carefully to be fully consistent with the hydrodynamic
gence order of DUGKS, a set of simulations with different resolu-
velocity and pressure fields. In the DUGKS simulation, the initial
tions and time steps are performed. The L2 errors
distributions are generated by iteration as follows: 
1. Begin with the initial distribution function defined as i, j |VN (i, j ) − VT (i, j )|2
L2 = (B.6)
f (x, ξ α , 0 ) = fα [ρ (x, 0 ), u(x, 0 )], where ρ (x, 0 ) = p(x, 0 )/cs2 .
eq N2
  where V represents velocity or pressure. The L2 errors are shown
ξ · u ( ξ α · u )2 u2
fαeq,N=2 = Wα ρ 1 + α + − + O(Ma3 ) in Tables B.6 and B.7, along with the resulting order of accuracy.
RT1 2(RT1 ) 2 2RT1 In Table B.6, time step size (as viewed by the conventional CFL
(B.2) number and the kinetic CFL number cs dt/dx) is kept very small
so the L2 errors represent mainly the space discretization errors,
2. Evolve the distribution function for one time step with DUGKS. the order of accuracy is around 2. In Table B.7, we fixed the ki-
Update the hydrodynamic variable and denote them as ρ (x, t ) netic CFL number to 0.4 so both the time discretization error and
and u(x, t ). space discretization error are present, the realized order of accu-
3. Based on the Chapman-Enskog expansion, the distribution racy is between 2 to 3. Since the requirement that cs dt/dx < 1
function can be written as follows: implies that we cannot isolate space discretization error from the
 
∂ fαeq ∂ a (ξ − u j ) eq time discretization error in DUGKS. Overall, Tables B.6 and B.7 to-
f α = f α − τν
eq
+ ( ξ f ) + τν j j
eq
fα + O (τν2 ) gether demonstrate the second-order accuracy in both space and
∂t ∂xj j α RT1
time.
(B.3)
With the unexpanded form of the equilibrium distribution and References
the Euler equation, the distribution function can be written
[1] J.H. Lienhard, A Heat Transfer Textbook, fourth edition, Phlogiston Press, 2015.
as:
   [2] E. Bodenschatz, W. Pesch, G. Ahlers, Recent developments in Rayleigh-Bénard
∂ u i ci c j ∂ u i convection, Annu. Rev. Fluid Mech. 32 (1) (20 0 0) 709–778.
fα = fαeq 1 + τν − + O(τν2 ) [3] D. Lohse, K.Q. Xia, Small-scale properties of turbulent Rayleigh-Bénard convec-
∂ xi RT1 ∂ x j tion, Annu. Rev. Fluid Mech. 42 (1) (2010) 335–364.
[4] G.D. Mallinson, G. de Vahl Davis, Three-dimensional natural convection in a
= fαeq + τν fα(1) + O (τν2 ). (B.4) box: a numerical study, J. Fluid Mech. 83 (1977) 1–31.

18
X. Wen, L.-P. Wang, Z. Guo et al. International Journal of Heat and Mass Transfer 167 (2021) 120822

[5] G. de Vahl Davis, I.P. Jones, Natural convection in a square cavity: a comparison [23] F.X. Trias, M. Soria, A. Oliva, C.D. Pérez-segarra, Direct numerical simulations
exercise, Int. J. Numer. Methods Fluids 3 (1983) 227–248. of two- and three-dimensional turbulent natural convection flows turbulent
[6] J. Salat, S. Xin, P. Joubert, A. Sergent, F. Penot, P.L. Quéré, Experimental and nu- natural flows in a differentially heated cavity of aspect ratio 4, J. Fluid Mech.
merical investigation of turbulent natural convection in a large air-filled cavity, 586 (1) (2007) 259–293.
Int. J. Heat Fluid Flow 25 (2004) 824–832. [24] F.X. Trias, A. Gorobets, M. Soria, A. Oliva, Direct numerical simulation of a dif-
[7] Y.S. Tian, T.G. Karayiannis, Low turbulence natural convection in an air filled ferentially heated cavity of aspect ratio 4 with Rayleigh numbers up to 1011 -
square cavity part 1: the thermal and fluid flow fields, Int. J Heat Mass Transf. Part 1: numerical methods, Int. J. Heat Mass Transf. 53 (2010) 665–673.
43 (20 0 0) 849–866. [25] F.X. Trias, A. Gorobets, M. Soria, A. Oliva, Direct numerical simulation of a dif-
[8] Y.S. Tian, T.G. Karayiannis, Low turbulence natural convection in an air filled ferentially heated cavity of aspect ratio 4 with Rayleigh numbers up to 1011
square cavity Part 2: the turbulence quantities, Int. J. Heat Mass Transf. 43 - Part 2: Heat transfer and flow dynamics, Int. J. Heat Mass Transf. 53 (2010)
(20 0 0) 867–884. 674–683.
[9] P.L. Quéré, Accurate solutions to the square thermally driven cavity at high [26] X. Shan, Simulation of Rayleigh-Bénard convection using a lattice Boltzmann
Rayleigh number, Comput. Fluids 20 (1991) 29–41. method, Phys. Rev. E 55(3) (1997) 2780–2788.
[10] S. Paolucci, D.R. Chenoweth, Transition to chaos in a differentially heated ver- [27] Y. Peng, C. Shu, Y.T. Chew, A 3D incompressible thermal lattice Boltzmann
tical cavity, J. Fluid Mech. 201 (1989) 379–410. model and its application to simulate natural convection in a cubic cavity, J.
[11] R.J.A. Janssen, R.A.W.M. Henkes, Influence of Prandtl number on instability Comput. Phys. 193 (2003) 260–274.
mechanism and transition in a differentially heated square cavity, J. Fluid [28] A. Xu, L. Shi, H.D. Xi, Lattice Boltzmann simulations of three-dimensional ther-
Mech. 290 (1995) 319–344. mal convective flows at high Rayleigh number, Int. J. Heat Mass Transf. 140
[12] P.L. Quéré, M. Behnia, From onset of unsteadiness to chaos in a differentially (2019) 359–370.
heated square cavity, J. Fluid Mech. 359 (1998) 81–107. [29] Z. Guo, K. Xu, R. Wang, Discrete unified gas kinetic scheme for all Knudsen
[13] Y.L. Zhao, C.W. Lei, J.C. Patterson, Transition of natural convection boundary number flows: low-speed isothermal case, Phys. Rev. E 88 (2013) 033305.
layer-a revisit by Bicoherence analysis, Int. Commun. Heat Mass Transf. 58 [30] P. Wang, L.H. Zhu, Z.L. Guo, K. Xu, A comparative study of LBE and DUGKS
(2014) 147–155. methods for nearly incompressible flows, Commun. Comput. Phys. 17 (2015)
[14] Y.L. Zhao, C.W. Lei, J.C. Patterson, Natural transition in natural convection 657–681.
boundary layers, Int. Commun. Heat Mass Transf. 76 (2016) 366–375. [31] P. Wang, L.P. Wang, Z.L. Guo, Comparison of the LBE and DUGKS methods for
[15] R. Janssen, S. Armfield, Stability properties of the vertical boundary layers in DNS of decaying homogeneous isotropic turbulence, Phys. Rev. E 94 (2016)
differentially heated cavities, Int. J. Heat Fluid Flow 17 (1996) 547–556. 043304.
[16] T. Fusegi, J.M. Hyun, K. Kuwahara, B. FAROUK, A numerical study of three-di- [32] Y.T. Bo, P. Wang, Z.L. Guo, L.P. Wang, DUGKS simulations of three-dimensional
mensional natural convection in a differentially heated cubical enclosure, Int. Taylor-Green vortex flow and turbulent channel flow, Comput. Fluids 155 (1)
J. Heat Mass Transf. 34 (1991) 1543–1557. (2017) 9–21.
[17] E. Tric, G. Labrosse, M. Betrouni, A first incursion into the three-dimensional [33] P.K. Kundu, I.M. Cohen, D.R. Dowling, Fluid Mechanics, fourth edition, Aca-
structure of natural convection of air in a differentially heated cubic cav- demic Press, 2008, pp. 124–128.
ity, from accurate numerical solutions, Int. J. Heat Mass Transf. 43 (20 0 0) [34] P. Wang, S. Tao, Z. Guo, A coupled discrete unified gas-kinetic scheme for
4043–4056. Boussinesq flows, Comput. Fluids 120 (2015) 70–81.
[18] T. Fusegi, J.M. Hyun, K. Kuwahara, Three-dimensional simulations of natural [35] K. Xu, S. Liu, Rayleigh-Bénard simulation using the gas-kinetic Bhatnagar–
convection in a sidewall-heated cube, Int. J. Numer. Meth. Fluids 13 (1991) Gross-Krook scheme in the incompressible limit, Phys. Rev. E 60(1) (1999)
857–867. 464–470.
[19] R.J.A. Janssen, R.A.W.M. Henkes, Instabilities in three-dimensional differential- [36] G. Grötzbach, Spatial resolution requirement for direct numerical simulation of
ly-heated cavities with adiabatic horizontal walls, Phys. Fluids 8 (1996) 62–74. the Rayleigh-Bénard convection, Int. J. Numer. Meth. Fluids 49 (1983) 241–264.
[20] G. Labrosse, E. Tric, H. Khallouf, M. Betrouni, A direct(pseudo-spectral) solver [37] O. Shishkina, R.J. Stevens, S. Grossmann, D. Lohse, Boundary layer structure in
of the two-/three-dimensional stokes problem: transition to unsteadiness of turbulent thermal convection and its consequences for the required numerical
natural-convection flow in a differentially heated cubical cavity, Numer. Heat resolution, N. J. Phys. 12 (2010) 075022.
Transf. 31 (1997) 261–276. [38] S. Xin, P.L. Quéré, Direct numerical simulations of two-dimensional chaotic
[21] P. Wang, Discrete unified gas-kinetic scheme for incompressible flows and its natural convection in a differentially heated cavity of aspect ratio 4, J. Fluid
applications, Huazhong University of Science and Technology, 2016 phd thesis. Mech. 304 (1995) 87–118.
[22] P. Wang, Y. Zhang, Z. Guo, Numerical study of three-dimensional natural con- [39] C.A. Wang, H. Sadat, C. Prax, A new meshless approach for three dimen-
vection in a cubical cavity at high Rayleigh numbers, Int. J. Heat Mass Transf. sional fluid flow and related heat transfer problems, Comput. Fluids 69 (2012)
113 (2017) 217–228. 136–146.

19