1. Introduction
This investigation focuses on mixing flows induced by hydrodynamic interfacial instabilities, such as the Rayleigh–Taylor (RT) (Rayleigh Reference Rayleigh1883; Taylor Reference Taylor1950) and Richtmyer–Meshkov (RM) (Richtmyer Reference Richtmyer1960; Meshkov Reference Meshkov1969) instabilities. The RT instability arises when a denser fluid is accelerated by a lighter one under constant acceleration, while the RM instability emerges if the acceleration is pulsing regardless of the shocking direction. In both cases, the misalignment between density and pressure gradients drives interfacial perturbation growth. Baroclinic vorticity plays a crucial role in amplifying interfacial perturbations while simultaneously entraining surrounding irrotational fluids into the mixing zone. Mixing is frequently intensified by the Kelvin–Helmholtz instability (Helmholtz Reference Helmholtz1868; Kelvin Reference Kelvin1871) and subsequently transitions to turbulence. This interfacial mixing phenomenon is observed in systems spanning multiple scales, from cosmic-scale supernova explosions to laboratory-scale applications in scramjet propulsion and inertial confinement fusion (ICF) (Zhou Reference Zhou2017a ). The prevalence of these mixing processes across such diverse physical systems underscores their fundamental importance in both natural phenomena and engineering applications.
The aforementioned mixing flows essentially demonstrate significant compressibility effects across numerous scenarios. Taking ICF as an example, the internal flow can be characterised as a highly nonlinear compressible system progressing through four characteristic stages: shock-transit, acceleration, deceleration and peak compression (Craxton et al. Reference Craxton2015). The first stage typically initiates with a sequence of laser pulses of increasing intensities irradiating the ICF target. These pulses generate pressure waves in the plasma, launching a sequence of shock waves that propagate into and compress the target. During this stage, surface perturbations on the shells evolve as a result of a RM-like instability, due to imperfections in fabrication and laser-beam non-uniformity. After the shock reaches the inner surface, the acceleration stage begins, characterised by an outward-moving rarefaction wave and inward acceleration of the target shells. The subsequent deceleration stage sees continued shell compression with a gradually decreasing speed through centripetal inertia, with kinetic energy converted into internal energy, resulting in increase of pressure and temperature. The compression process culminates in a peak compression phase, where the fuel reaches the maximum compression level, yielding central hot spots with extraordinarily high temperature and density due to compressibility. Throughout these stages, hydrodynamic instabilities, particularly the RT and RM instabilities, amplify surface perturbations, compromising shell integrity and compression efficiency. The resulting mixing flows dissipate substantial energy, cooling the fuel and potentially causing ignition failure. Notably, the density profiles within each shell vary radially with the radial location
$r$
, as illustrated in figure 1
$(b)$
, contrasting with the constant density profiles of incompressible cases in figure 1
$(a)$
. This ignition process underscores the critical importance of studying compressible mixing flows.
Current studies on compressible mixing flows primarily focus on those induced by RT instability, where compressibility effects stem from two distinct origins (Gauthier & Le Creurer Reference Gauthier and Le Creurer2010). The first is the static compressibility, which results from the fluid’s variable density. In an acceleration field, it causes density stratification and a gradient for light and heavy fluids. The second is dynamic compressibility, an effect fundamentally associated with the finite speed of sound. For a perfect gas, it can be characterised by the adiabatic indices
$\varGamma$
in the equation of state (EOS). The dynamic compressibility mainly depends on the inherent properties of the fluid, while existing numerical and theoretical investigations have demonstrated that compressibility in RT flows is predominantly governed by static compressibility (Livescu Reference Livescu2004; Jin et al. Reference Jin, Liu, Lu, Cheng, Glimm and Sharp2005; Xue & Ye Reference Xue and Ye2010; Gauthier Reference Gauthier2013, Reference Gauthier2017). Therefore, the present study concentrates on static compressibility, specifically density-stratified compressible RT mixing flows.
Schematic diagrams of the initial density distributions for
$(a)$
the incompressible and
$(b)$
the density-stratified compressible RT flows.

The classical incompressible RT mixing, as illustrated in figure 1
$(a)$
, maintains initial constant density distributions within both light and heavy fluid layers. In contrast, the density-stratified compressible RT mixing initially displays a significant density gradient, as depicted in figure 1
$(b)$
. When heavy and light fluids are set to the same temperature in the initial static state, density distribution exhibits exponential variation with spatial locations (Mellado, Sarkar & Zhou Reference Mellado, Sarkar and Zhou2005). The stratification strength is commonly quantified using the density stratification parameter (Luo & Wang Reference Luo and Wang2022):
with
$g$
,
$L_r$
,
$T_r$
,
$M_r$
and
$R$
representing the gravitational acceleration, characteristic length, temperature, mole mass and universal gas constant, respectively. The larger
$S_r$
corresponds to a steeper density gradient, indicating a stronger compressibility effect. This fundamental difference highlights the distinctive behaviour of compressible versus incompressible RT mixing.
Existing studies have revealed prominent differences between density-stratified compressible RT flows and their incompressible counterparts. Owing to the initial density stratification, the growth rate of RT mixing is substantially affected (Xue & Ye Reference Xue and Ye2010). George & Glimm (Reference George and Glimm2005) have emphasised that self-similar rates differ markedly between compressible and incompressible scenarios, noting that the quadratic growth rate with time (denoted as
$t$
) observed in incompressible cases cannot be directly applied to compressible flows. This perspective has been corroborated by high-fidelity (HiFi) numerical simulations conducted by Gauthier (Reference Gauthier2017) and Luo & Wang (Reference Luo and Wang2022), which confirm that the
$t^2$
-scaling law does not hold for density-stratified compressible RT mixing flows. Furthermore, density stratification enhances asymmetric evolution of the mixing layers on the light and heavy fluid sides. Additionally, these studies have revealed that spatial profiles of density and temperature are significantly influenced by initial density stratification, exhibiting complex distributions that contrast with the monotonic variations seen in incompressible cases.
Energy transport mechanisms also exhibit fundamental differences between density-stratified compressible and incompressible RT flows. For the compressible cases, kinetic energy generation stems from two primary sources: potential energy conversion and pressure-dilatation work (Zhao, Liu & Lu Reference Zhao, Liu and Lu2020; Luo & Wang Reference Luo and Wang2022). When the
$S_r$
is small, the kinetic energy derives predominantly from potential energy, which is similar to the incompressible situation. However, for larger
$S_r$
, pressure-dilatation work becomes the dominant mechanism for kinetic energy generation. Moreover, the kinetic energy is subsequently converted back into potential energy. These observations demonstrate that the underlying energy transport physics differs substantially between compressible and incompressible RT flows (Zhao, Betti & Aluie Reference Zhao, Betti and Aluie2022).
While considerable progress has been made in understanding the fundamental mechanisms of density-stratified compressible RT flows, modelling efforts remain limited. Current research primarily focuses on two approaches: the buoyancy-drag model (Jin et al. Reference Jin, Liu, Lu, Cheng, Glimm and Sharp2005; Fu et al. Reference Fu, Zhao, Xu, Wang, Liu, Wan and Lu2022) and the large-eddy simulation (LES) techniques (Luo et al. Reference Luo, Li, Yuan, Peng, Liu, Wang and Wang2024). However, significant limitations persist. The buoyancy-drag model, while computationally efficient, oversimplifies the physics and fails to capture essential flow characteristics required for engineering applications. Conversely, LES provides more detailed flow information but at prohibitively high computational costs, making it impractical for routine engineering predictions. This disparity underscores the urgent need to develop alternative modelling approaches that can provide acceptable computational accuracy at an affordable cost, which is a critical step towards practical engineering applications of compressible RT flow predictions.
Reynolds-averaged Navier–Stokes (RANS) simulation has been widely adopted in industrial applications owing to its favourable balance between computational efficiency and cost. A number of RANS mixing models have been proposed for interfacial flow predictions, such as the
$K$
–
$L$
(Dimonte & Tipton Reference Dimonte and Tipton2006; Morgan & Greenough Reference Morgan and Greenough2016; Kokkinakis, Drikakis & Youngs Reference Kokkinakis, Drikakis and Youngs2019; Zhang et al. Reference Zhang, He, Xie, Xiao and Tian2020),
$K$
–
$\epsilon$
(Morán-López & Schilling Reference Morán-López and Schilling2013; Xie et al. Reference Xie, Xiao and Zhang2021a
) and Besnard–Harlow–Rauenzahn (BHR) models (Denissen et al. Reference Denissen, Rollin, Reisner and Andrews2014; Kokkinakis et al. Reference Kokkinakis, Drikakis and Youngs2019; Xie et al. Reference Xie, Xiao and Zhang2021b
). However, most of the existing RANS models predominantly address flows where the compressible effect plays an insignificant role in mixing development. The fundamental differences between compressible and incompressible RT mixing flows, as previously discussed, render existing RANS models inadequate for the density-stratified compressible RT scenario. Indeed, compressible RANS modelling remains largely unexplored – a critical gap this study seeks to address. Our approach builds upon the recently proposed
$K$
–
$L$
–
$\gamma$
mixing transition model (Xie et al. Reference Xie, Qi, Xiao, Zhang and Zhao2025), which has been shown to have good performance for capturing the mixing transition process in flows where compressibility is not prominent. Through systematic compressibility modifications, we extend this model to accurately capture both the transition dynamics (Qi et al. Reference Qi, He, Xu and Zhang2024) and compressibility effects in compressible mixing flows.
This paper is organised as follows. Section 2.1 presents the governing equations of the baseline
$K$
–
$L$
–
$\gamma$
mixing transition model. The performance of this baseline model for density-stratified compressible RT flows is then evaluated in § 2.2. Through comparison with HiFi simulation data and application of local instability criteria, the fundamental limitations of the baseline model applied to compressible RT flows are systematically analysed. Section 2.3 details the theoretical derivation of the proposed compressibility modifications. These developments are subsequently validated in § 3, where computational results demonstrate the improved predictive capability. In § 4, the impact of the proposed corrections on RM mixing prediction and scenarios of considering these modifications are discussed. Finally, § 5 summarises the contributions of the present study.
2. Methodology
2.1. Governing equations and baseline model
For mixing flows, the multicomponent RANS equations are solved considering molecular transport and thermodynamic coefficients. The transport equations for the mean density
$\bar {\rho }$
, velocity
$\tilde {u}_{i}$
, total energy
$\tilde {E}$
of the mixture and mass fraction
$\tilde {Y}_{\alpha }$
of species
$\alpha$
are given as follows:
\begin{align}&\qquad\qquad\qquad\quad\,\,\, \frac {\partial \bar {\rho }{{{\tilde {Y}_{\alpha }}}}}{\partial t}+\frac {\partial \bar {\rho }{{{\tilde {Y}_{\alpha }}}}{{{\tilde {u}}}_{j}}}{\partial {{x}_{j}}}=D_{Y}+\frac {\partial }{\partial {{x}_{j}}}\left (\bar {\rho }\bar {D}\frac {\partial \tilde {Y}_{\alpha }}{\partial x_{j}}\right ). \end{align}
The overbar and tilde represent the Reynolds and Favre averaged fields, respectively. The subscripts
$i,j=1,2,3$
denote three spatial directions (
$x,y,z$
). The
$\bar {p}$
denotes pressure of the mean field. The heat flux
$\bar {q}_{c}$
, the interspecies diffusional heat flux
$\bar {q}_{d}$
and the viscous stress tensor
$\bar {\sigma }_{ij}$
are given as
$\bar {q}_{c}=-\bar {\kappa } \partial \tilde {T}/\partial x_{j}$
,
$\bar {q}_{d}=-\sum \bar {\rho } \bar {D}\bar {c}_{p,\alpha }\tilde {T}\partial \tilde {Y}_{\alpha }/\partial x_{j}$
,
$\bar {\sigma }_{ij}= 2\bar {{\rm \mu} }(\tilde {S}_{ij}-\tilde {S}_{kk}\delta _{ij}/3)$
, where
$\tilde {S}_{ij}=(\partial \tilde {u}_{i}/\partial x_{j}+\partial \tilde {u}_{j}/\partial x_{i})/2$
. Here,
$\bar {{\rm \mu} }$
,
$\bar {D}$
,
$\bar {\kappa }$
, and
$\bar {c}_{p,\alpha }$
represent dynamic viscosity, mass diffusivity, thermal conductivity and constant-pressure specific heat of species
$\alpha$
, respectively. In correspondence with Livescu (Reference Livescu2013), the mixture’s EOS
$\bar {p}\tilde {M}=\bar {\rho }R\tilde {T}$
is used, which is calculated with the assumptions of iso-temperature (i.e.
$T_1=T_2=\cdots =T_{\alpha }$
) and partial pressure (i.e.
$p=\sum {p}_{\alpha }$
), and the fluid properties of the mixture are determined using the species-linearly weighted assumption.
Equations (2.1)
${-}$
(2.4) are deduced based on the ensemble average operation, which results in the unclosed terms, i.e. the Reynolds stress
$\tau _{ij}$
, the turbulent diffusion terms
$D_{E}$
,
$D_{K}$
and
$D_{Y}$
of the total energy, the turbulent kinetic energy (TKE,
$\tilde {K}$
), and the mass fraction, respectively. With the widely used gradient diffusion assumption (GDA) that the turbulent flux of an arbitrary physical variable
$f$
can be proportional to its mean gradient, i.e.
where ‘
$'$
’ and ‘
$''$
’ denote the Reynolds and Favre fluctuations, respectively, as well as
$N_{f}$
being a model coefficient, these turbulent diffusion terms are modelled as
$ {D}_{E}= ({\partial }/{\partial x_{j}}) (({{\rm \mu} _{t}}/{N_{h}}) ({\partial \tilde {h}}/{\partial x_{j}}) )$
,
${D}_{K}=({\partial }/{\partial x_{j}}) ( ({{\rm \mu} _{t}}/{N_{K}})({\partial \tilde {K}}/{\partial x_{j}} ))$
,
${D}_{Y}= ({\partial }/{\partial x_{j}}) (({{\rm \mu} _{t}}/ {N_{Y}})({\partial \tilde {Y}}/{\partial x_{j}}) )$
, respectively, the enthalpy
$h=e+p/\rho =c_pT$
, where
$e$
represents the internal energy.
In the
$K$
–
$L$
–
$\gamma$
model, the TKE
$\tilde {K}$
and the turbulent length scale
$\tilde {L}$
are used to describe the turbulent viscosity
$\nu _{t}$
, which is limited by the intermittent factor
$\gamma$
to enable this model to describe mixing transition flows (Xie et al. Reference Xie, Qi, Xiao, Zhang and Zhao2025). The
${\rm \mu} _{t}$
is expressed as
${\rm \mu} _{t}=\bar {\rho }\nu _t=C_{{\rm \mu} }\bar {\rho } \gamma \tilde {L}\sqrt {2\tilde {K}}$
, where
$C_{{\rm \mu} }$
is a model coefficient. Based on the classical Boussinesq eddy viscosity hypothesis, the Reynolds stress is modelled as
$ {{{\tau }}_{ij}}={{C}_{P}}\bar {\rho }\tilde {K}{{\delta }_{ij}}-2{{{\rm \mu} }_{t}}(\tilde {S}_{ij}-\tilde {S}_{kk}{{\delta }_{ij}}/3)$
; here,
$C_{P}$
is a model coefficient and
$\delta _{ij}$
is the Kronecker symbol. To describe the spatio-temporal evolutions of the
$\tilde {K}$
,
$\tilde {L}$
and
$\gamma$
, their transport equations are introduced as follows:
\begin{align}&\frac {\partial \bar {\rho }\tilde {K}}{\partial t}+\frac {\partial \bar {\rho }\tilde {K}{{{\tilde {u}}}_{j}}}{\partial {{x}_{j}}}=-{{{\tau }}_{ij}}\frac {\partial {{{\tilde {u}}}_{i}}}{\partial {{x}_{j}}}+\frac {\partial }{\partial {{x}_{j}}}\left (\frac {{{{\rm \mu} }_{t}}}{{{N}_{K}}}\frac {\partial \tilde {K}}{\partial {{x}_{j}}}\right )+{{S}_{Kf}}-{{C}_{D}}\bar {\rho }{{\Big ( \sqrt {2\tilde {K}}\Big )}^{3}}/{\tilde {L}}, \end{align}
\begin{align}&\qquad\quad \frac {\partial \bar {\rho }\tilde {L}}{\partial t}+\frac {\partial \bar {\rho }\tilde {L}{{{\tilde {u}}}_{j}}}{\partial {{x}_{j}}}=\frac {\partial }{\partial {{x}_{j}}}\left (\frac {{{{\rm \mu} }_{t}}}{{{N}_{L}}}\frac {\partial \tilde {L}}{\partial {{x}_{j}}}\right )+{{C}_{L}}\bar {\rho }\sqrt {2\tilde {K}}+{{C}_{C}}\bar {\rho }\tilde {L}\frac {\partial {{{\tilde {u}}}_{j}}}{\partial {{x}_{j}}}, \end{align}
The
$P_{\gamma }=F_{\textit{onset}}G_r$
is the production term of the
$\gamma$
equation, the dissipation term
$\epsilon =\gamma P_{\gamma }$
ensuring the intermittent factor varying from 0 to 1. The
$F_{\textit{onset}}$
serves as a transition switch and
$G_r$
describes the growth of
$\gamma$
, which are expressed as
\begin{align} G_{r}&=C_{1}\sqrt {\gamma }\left \{ (1-\gamma ) \left [\bar {\rho }\sqrt {2\tilde {S}_{ij}\tilde {S}_{ij}}+C_{2}\sqrt {{\textit{max}} \left (0,-\frac {\partial \bar {\rho }}{\partial x_k} \frac {\partial \bar {p}}{\partial x_k}\right )}\right ]\right . \nonumber\\&\quad\left . +\, \left (\gamma S_{Kf}-{{{\tau }}_{ij}}\frac {\partial {{{\tilde {u}}}_{i}}}{\partial {{x}_{j}}}\right )/\tilde {K} \right \}. \end{align}
Here,
$C_1$
and
$C_2$
are model coefficients,
$\nu$
denotes the kinetic viscosity, and
${\textit{Re}}_{\textit{tra}}$
denotes a pre-defined key transition Reynolds number set to 10, following Xie et al. (Reference Xie, Qi, Xiao, Zhang and Zhao2025).
The buoyancy-driven source term
${S}_{Kf}$
is modelled based on the following formation (Kokkinakis et al. Reference Kokkinakis, Drikakis, Youngs and Williams2015):
\begin{equation} S_{Kf} = \begin{cases} C_{B} \bar {\rho }\sqrt {2\tilde {K}} \left | \left [(1 - \omega _{l}) A_{0i} + \omega _{l} A_{\textit{ssi}} \right ]g_i\right | &\textit{if}\; \varTheta _{\textit{gf}}\lt \varTheta _{gm}, \\[5pt] C_{B} \bar {\rho }\sqrt {2\tilde {K}} {\textit{max}} \left \{0, \left [(1 - \omega _{l}) A_{0i} + \omega _{l} A_{\textit{ssi}}\right ]g_i \right \} & \textit{otherwise}, \end{cases} \end{equation}
where
$C_B$
is a model coefficient. The mean- and turbulent-field accelerations
$\varTheta _{gm}$
and
$\varTheta _{\textit{gf}}$
are used to identify RM and RT mixings, whose expressions can be seen from Xiao, Zhang & Tian (Reference Xiao, Zhang and Tian2020), the acceleration
$g_{i}=-({1}/{\bar {\rho }})({\partial \bar {p}}/{\partial x_{i}})$
. The initial cell Atwood number
${A}_{0i}$
describes the density’s discontinuous characteristic at the interfaces between fluids and can be expressed by the reconstructed values of the densities at the cell’s faces (Kokkinakis et al. Reference Kokkinakis, Drikakis, Youngs and Williams2015). The local Atwood number
${A}_{\textit{ssi}}$
in the self-similar regime depicts the density’s gradual variation, and is calculated by
\begin{equation} {{A}_{\textit{ssi}}}=\frac {{{C}_{A}}{\tilde {L}}}{\bar {\rho }+\tilde {L}\left | \dfrac {\partial \bar {\rho }}{\partial x_{i}} \right |}\dfrac {\partial \bar {\rho }}{\partial x_{i}}, \end{equation}
where
$C_A$
is a model coefficient. The weighting factor
$\omega _{l}={\rm min}(({\tilde {L}}/{\Delta x}),1)$
(
$\Delta x$
is the grid scale) ensures that mixing evolution is solely decided by
$A_{\textit{ssi}}$
when
$\tilde {L}$
becomes sufficiently large, while initially is dominated by
$A_{0i}$
(Kokkinakis et al. Reference Kokkinakis, Drikakis, Youngs and Williams2015).
Equations (2.1)
${-}$
(2.4), (2.6)
${-}$
(2.12) constitute the complete
$K$
–
$L$
–
$\gamma$
mixing transition model, which has demonstrated a good capability for predicting RT/RM mixing transition flows with insignificant compressible effect. More details about this model can be found from Xie et al. (Reference Xie, Qi, Xiao, Zhang and Zhao2025). Given the prominent transition processes observed in density-stratified compressible RT mixing flows, the
$K$
–
$L$
–
$\gamma$
model is adopted as the baseline model in this study to be expected to capture the transition effect. In the next section, performance of the baseline model for the targeted flows will be evaluated. In addition, the widely used
$K$
–
$L$
model, which is designed for fully developed turbulent mixing flows, is also tested for comparison, highlighting the necessity of considering the transitional effects by introducing the
$\gamma$
equation.
2.2. Baseline model evaluation for compressibility effects
The evaluated cases originate from the three-dimensional (3-D) simulation of initially isothermal density-stratified compressible RT mixing implemented by Luo & Wang (Reference Luo and Wang2022). These simulations solved the multicomponent Navier–Stokes equations for cases with different density stratification parameters
$S_r$
= 0.5, 1, 2 and 3. An eighth-order central compact finite difference scheme was employed on a uniform grid of
$1024\times 512\times 512$
grid points in a rectangular box
$x\times y\times z$
of
$2L_r\times L_r\times L_r$
,
$L_r$
= 1. Time integration was performed using a third-order Runge–Kutta scheme. The simulations of Luo & Wang (Reference Luo and Wang2022) are thus highly resolved and referred to as HiFi simulations in the present study.
Due to the homogeneity of the flow in the spanwise plane (i.e. the
$y{-}z$
plane), the 3-D cases are statistically reduced to the one-dimensional (1-D) equivalents for RANS computations. The flow configuration is described as follows. The light and heavy fluids, represented by the subscript
$L$
and
$H$
, respectively, are initially placed in the computational domains
$[-L_x,0]$
cm and
$[0,L_x]$
cm, respectively, with
$L_x$
=
$L_r$
= 1. A constant gravitational acceleration
$g$
= 1 cm s
$^{-2}$
is imposed along the
$-x$
direction. Both fluids are compressible ideal gases with the same adiabatic exponent
$\varGamma$
= 5/3, and the molar masses
$M_L$
and
$M_H$
are
$0.029$
kg mol–1 and
$0.087$
kg mol–1, respectively, yielding an Atwood number
$A_t\equiv ({M_H-M_L})/({M_H+M_L})$
= 0.5, consistent with the definition by Luo & Wang (Reference Luo and Wang2022). The stratification parameter
$S_r$
is defined based on (1.1). The characteristic Reynolds number
${\textit{Re}}\equiv ({g^{1/2}{L_r}^{3/2}})/({{\rm \mu} _r/\rho _r})$
= 10 000, and the Prandtl number and the Schmidt number are 0.7 and 1.0, respectively. Reference quantities used to define these dimensionless parameter include the temperature
$T_r$
=
$T_H$
=
$T_L$
, molar mass
$M_r=(M_H+M_L)/2$
, dynamic viscosity
${\rm \mu} _r={\rm \mu} _H={\rm \mu} _L$
and density
$\rho _r=(\rho _{H,0}+\rho _{L,0})/2$
, where
$\rho _{H,0}$
and
$\rho _{L,0}$
represent the initial densities on either side of the interface. Additionally, the characteristic time scale is defined as
$\tau =\sqrt {L_r/(A_tg)}$
. All reference quantities are consistent with the HiFi simulations to ensure direct comparability between RANS and HiFi results. The system is initialised with a hydrostatic background state in thermal equilibrium. By setting
$\tilde {u}$
= 0 in the momentum equation, we can obtain the initial density and pressure distributions:
where
$p_I$
and
$T_I$
represent the initial pressure and temperature at the interface, respectively, the subscript
$m$
representing either a heavy (
$H$
) or light (
$L$
) fluid. Specifically,
$p_I$
= 1 g cm
$^{-1}$
s
$^{-2}$
and
$T_I$
equals to
$T_r$
determined based on (1.1). Figure 2 shows comparisons of the initial density and pressure fields from RANS and HiFi simulations, ensuring the correctness of the present configurations.
Initial distributions for density and pressure fields. The symbols and lines represent the initial flow fields of HiFi simulations and RANS simulations, respectively. The lines coloured in black, red, blue and green correspond to cases of
$S_r=$
0.5, 1, 2 and 3, respectively.

Performance of the baseline model for the targeted flows is tested here. The RANS calculations are implemented in the finite difference code for compressible fluid dynamics developed by Y. Zhang et al. in 2016, which has been extensively validated for RT and RM mixing studies (Li et al. Reference Li, He, Zhang and Tian2019; Zhang et al. Reference Zhang, He, Xie, Xiao and Tian2020; Xiao et al. Reference Xiao, Hu, Dai and Zhang2022; Xie, Zhao & Zhang Reference Xie, Zhao and Zhang2023; Xiao, Qi & Zhang Reference Xiao, Qi and Zhang2025; Xie et al. Reference Xie, Qi, Xiao, Zhang and Zhao2025). A combination of MUSCL5, i.e. the improved total variation diminishing scheme with a fifth-order limiter, and the Rusanov-like Riemann solver seen in Toro, Spruce & Speares (Reference Toro, Spruce and Speares1994) is used to calculate the convection terms. To reduce the numerical dissipation, low-Mach-number modification (Thornber et al. Reference Thornber, Drikakis, Williams and Youngs2008a , Reference Thornber, Mosedale, Drikakis, Youngs and Williamsb ) is used during the reconstruction of the half-point flux. The diffusion terms are calculated with the sixth-order central difference scheme. The temporal integration marches with a third-order Runge–Kutta scheme, with Courant–Friedrichs–Lewy number 0.05.
The model coefficients used in the baseline model are listed in table 1, which are determined based on the methodology proposed by Zhang et al. (Reference Zhang, He, Xie, Xiao and Tian2020). For initialisation of the turbulence model variables, we adopt the common practices: near the interface, the initial turbulent length scale
$\tilde {L}_0$
is
$0.04$
cm, approximating to the corresponding 3-D simulation’s initial perturbation width, the initial TKE
$\tilde {K}_0$
is
$0.0001A_tg\tilde {L}_0$
(cm s−1)
$^2$
, as well as the initial intermittent factor
$\gamma _0=1\times 10^{-16}$
following Xie et al. (Reference Xie, Qi, Xiao, Zhang and Zhao2025).
Model coefficients of the baseline model.

Temporal evolution of the mixing width predicted by
$(a)$
the
$K$
–
$L$
model and
$(b)$
the
$K$
–
$L$
–
$\gamma$
baseline model. The colours used to distinguish different
$S_r$
cases are consistent with figure 2.

Figure 3 presents temporal evolution of the mixing width, an important statistical quantity used to characterise the global evolution of mixing flows. The mixing width is defined as
The mole fraction
$X$
is calculated by the mass fraction (Luo & Wang Reference Luo and Wang2022). The
$W_b$
and
$W_s$
represent the bubble mixing width, i.e. distance of the light fluid penetrating the heavy fluid, and the spike mixing width, i.e. distance of the heavy fluid penetrating the light fluid, respectively. Figure 3
$(a)$
shows the prediction results of the
$K$
–
$L$
model, which gives a
$t^2$
-scaling law throughout the entire simulation for all cases, failing to depict the transition process. This is expected because the
$K$
–
$L$
model is only suitable for fully developed turbulent mixing flows with insignificant compressible effect. In contrast, the baseline model performs well during the instability development and mixing transition stages at approximately
$(t/\tau )^{2}\lt 4$
, providing predictions consistent with HiFi data for all cases, as illustrated in figure 3
$(b)$
. It is attributed to the fact that the baseline model can capture mixing transition processes effectively. Beyond the transition stage, the mixing width predicted by the baseline model exhibits a quadratic growth in time, in accordance with the late-stage self-similar evolution of incompressible RT mixing flows. Figure 3 highlights the superior capability of the baseline model in predicting mixing transition flows.
Notably, during the early stage, the baseline model gives consistent predictions with HiFi simulations across all compressibility strengths considered here, even though it does not introduce the compressibility corrections. It is attributed to the negligible compressibility effects in the early development of the density-stratified RT mixing flows, where behaviour closely resembles incompressible cases (Gauthier Reference Gauthier2017). However, the baseline model’s performance deteriorates significantly during intermediate and late stages. The discrepancies between RANS predictions and HiFi data progressively widen with increasing
$S_r$
values. It can be understood as follows. The
$S_r$
directly correlates to the compressibility strength: a smaller
$S_r$
corresponds to a weaker compressibility effect, with flow approaching to the incompressible case as
$S_r\,\rightarrow$
0. In the case of
$S_r$
= 0.5, the compressibility is not prominent so that the baseline model can give a comparatively reasonable prediction. For the cases with larger
$S_r$
, strong density stratification brings a strong compressibility effect, resulting in failures of the baseline model. The degradation in predictive accuracy directly correlates with the strength of density stratification, highlighting the need for compressibility corrections in the existing RANS modelling.
The baseline model consistently underestimates the evolution of mixing width across all cases examined here. To understand the discrepancy, the underlying modelling philosophy is carefully analysed. For the
$K$
–
$L$
–
$\gamma$
and the
$K$
–
$L$
models, the TKE transport equation plays a crucial role in governing the dynamic evolution of both mixing and turbulence. Following Chassaing et al. (Reference Chassaing, Antonia, Anselmet, Joly and Sarkar2002) and Gauthier (Reference Gauthier2017), the exact averaged TKE equation without any modelling is written as
\begin{align} \frac {\partial (\bar {\rho }\tilde {K})}{\partial t}+\frac {\partial (\bar {\rho }\tilde {K}\tilde {u}_j)}{\partial x_j}&=\underbrace {-\overline {\rho u^{\prime \prime }_iu^{\prime \prime }_j}\frac {\partial \tilde {u_i}}{\partial x_j}}_{I}\underbrace {-\overline {u^{\prime \prime }_i}\frac {\partial \bar {p}}{\partial x_i}}_{II}\underbrace {-\frac {\partial \left (\overline {\rho u^{\prime \prime }_jK}+\overline { p^{\prime }u^{\prime \prime }_j}+\overline { \sigma _{ij}u^{\prime \prime }_i}\right )}{\partial x_j}}_{III}\nonumber\\[3pt]&\quad\underbrace {+\overline {p^{\prime }\frac {u^{\prime \prime }_i}{\partial x_i}}}_{IV}\underbrace {-\overline {\sigma _{ij}\frac {\partial u^{\prime \prime }_i}{\partial x_j}}}_{V}, \end{align}
where terms
$I\text{--} V$
represent shear production, turbulent mass flux (buoyancy production), diffusion, pressure–strain correlation and dissipation, respectively. Based on the HiFi simulations, budgets of the terms
$I\text{--} V$
at
$t/\tau =$
1, 2.5 and 4 are given in figure 4, where the cases of
$S_r=$
0.5 and 3 are chosen to represent the weak and strong density stratification cases. Figure 4 reveals that the important terms on the right-hand side of the TKE equation (2.16) are the turbulent mass flux (term
$\textit{II}$
) and dissipation (term
$V$
) terms. Given the dissipation enslaved by the production, the turbulent mass flux
$-\overline {u^{\prime \prime }_i} ({\partial \bar {p}}/{\partial x_i})$
is thus considered as the dominant term.
Budgets for terms
$I\text{-} V$
of the TKE equation (2.16). Panels
$(a){-}(c)$
and
$(d){-} (f)$
correspond to moments of
$t/\tau$
= 1, 2.5 and 4 for cases of
$S_r$
= 0.5 and 3, respectively.

In the
$K$
–
$L$
–
$\gamma$
model, the turbulent mass flux
$-\overline {u^{\prime \prime }_i}({\partial \bar {p}}/{\partial x_i})$
is closed using the term
$S_{Kf}$
, as expressed by the combination of (2.11) and (2.12). Therefore, it is natural to analyse the modelling mechanism of (2.11) and (2.12) to explore underlying reasons. The key physical mechanism incorporated in
$S_{Kf}$
is the baroclinic term
$ ({\partial \bar {p}}/{\partial x_{i}})({\partial \bar {\rho }}/{\partial x_{i})}$
, which acts as an effective production source when
$({\partial \bar {p}}/{\partial x_{i}})({\partial \bar {\rho }}/{\partial x_{i}})\lt 0$
; otherwise
$S_{Kf}$
= 0. The latter condition leads to a stagnation of mixing evolution, consistent with the observations in figure 3, where the growth rate of the mixing width approaches zero for cases with
$S_r$
= 2 and 3 during late-stage development. This behaviour can be attributed to the premature disappearance of the baroclinic term in the present flows. Specifically, due to the initial density stratification, the sign of the density gradient
${\partial \bar {\rho }}/{\partial x_{i}}$
transitions from positive to negative too early in the mixing region. Since the pressure gradient
$ {\partial \bar {p}}/{\partial x_{i}}$
remains negative under gravitational field, the product
$ ({\partial \bar {p}}/{\partial x_{i}})( {\partial \bar {\rho }}/{\partial x_{i}})$
becomes positive once
$ {\partial \bar {\rho }}/{\partial x_{i}}\lt 0$
, forcing
$S_{Kf}$
= 0. Consequently, the baseline model fails to supply the necessary source term to sustain mixing and turbulence evolution, thereby deviating from the true physical behaviour.
Identification of unstable regions based on the local instability criteria for incompressible flow
$Q_1=-({\partial \bar {\rho }}/{\partial x})( {\partial \bar {p}}/{\partial x})$
and for compressible flow
$Q_2=\bar {c}^2({\partial \bar {\rho }}/{\partial x})-( {\partial \bar {p}}/{\partial x})$
. The
$Q_2$
is equivalent to the expression (2.18) for the RT cases considered here, owing to the negative pressure gradient in the gravitational field. The ordinate has been normalised by their maximum amplitudes. Panels
$(a){-} (c)$
and
$(d){-} (f)$
correspond to moments of
$t/\tau$
= 1, 2.5 and 4 for the cases of
$S_r$
= 0.5 and 3, respectively.

Furthermore, stability analysis can serve as a practical tool for investigating compressible stratified flows characterised by finite density gradients and continuous profiles (Gauthier & Le Creurer Reference Gauthier and Le Creurer2010). Notably, local instability criteria can effectively identify unstable flow regions. For incompressible flow, the necessary and sufficient condition of a stratified heterogeneous fluid being stable is
$ {\partial \rho }/{\partial x_i}\lt 0$
everywhere, whereas for any region where
${\partial \rho }/{\partial x_i}\gt 0$
, the stratification is considered unstable (Chandrasekhar Reference Chandrasekhar1961). In a gravitational field, the instability condition is expressed as
For compressible flow, the instability criterion within the linear approximation for the Euler equations is suggested as (Gamalii et al. Reference Gamalii, Rozanov, Samarskii, Tishkin, Tyurina and Favorskii1980)
where
$c$
is the isentropic speed of sound. Obviously, the compressible criterion (2.18) can revert to the incompressible one (2.17) as
$c\,\rightarrow \,\infty$
.
We further assess the differences of these criteria in identifying unstable regions by analysing their spatial distributions along the gravitational direction. These distributions are obtained through the spanwise-plane average of the HiFi data. Figure 5 reveals that for the case of
$S_r$
= 0.5, both criteria identify identical unstable regions at
$t/\tau$
= 1 and 2.5. However, at
$t/\tau$
= 4, the compressible criterion predicts a significantly wider unstable area than its incompressible counterpart, reflecting the baseline model’s under-prediction for late-stage mixing evolution. For the case of
$S_r$
= 3, both criteria initially agree at
$t/\tau$
= 1 due to negligible compressibility effects in early stages. However, remarkable differences emerge at
$t/\tau$
= 2.5 and 4, where the compressible criterion identifies substantially wider unstable regions. These observations thus indicate that the incompressible criterion fails to correctly identify instability when the compressibility effects are significant. Considering that the instability criterion is essential for RANS models, as shown in the closures (2.11) and (2.12), the preceding analysis suggests that compressibility effects must be incorporated in modelling the turbulent mass flux.
2.3. Compressibility corrections
Analysis of the TKE equation budgets from HiFi simulations confirms that the turbulent mass flux plays a dominant role. Since the seminal work of Dimonte & Tipton (Reference Dimonte and Tipton2006), this term
$-\overline {u^{\prime \prime }_i}({\partial \bar {p}}/{\partial x_i})$
has been modelled based on a simple buoyancy-drag model and has received subsequent refinements continuously, leading to the current form (2.11) or similar variants (Kokkinakis et al. Reference Kokkinakis, Drikakis and Youngs2019, Reference Kokkinakis, Drikakis and Youngs2020). However, the compressibility effect has not been considered yet. Indeed, the majority of published RANS mixing studies concentrate on flows where the influence of the compressible effect on mixing evolution is insignificant, with quite limited attention given to scenarios with a significant compressibility effect. This section aims to explore how to incorporate compressibility effects into the modelling of the turbulent mass flux term
$-\overline {u^{\prime \prime }_i} ({\partial \bar {p}}/{\partial x_i})$
, with a particular focus on the turbulent mass flux velocity
$\overline {u^{\prime \prime }_i}$
, which serves as a critical component. Using the definitions of Reynolds and Favre averaging, the turbulent mass flux velocity
$\overline {u^{\prime \prime }_i}$
can be expressed as
$-{\overline {\rho ^{\prime }u^{\prime }_{i}}}/{\bar {\rho }}$
. This expression highlights that
$\overline {u^{\prime \prime }_i}$
fundamentally characterises the turbulent transport on density fluctuations. Therefore, it is natural to explore the potential influence factors on density fluctuation in compressible flows.
To establish a generalised physical relationship, the EOS for a perfect gas is employed in the following form for the mixture:
Averaging (2.19) gives
With the Reynolds decomposition, the instantaneous variables
$p$
,
$\rho$
,
$\varGamma$
and
$e$
can be written as
$p=\bar {p}+p^{\prime }$
,
$\rho =\bar {\rho }+\rho ^{\prime }$
,
$\varGamma =\bar {\varGamma }+\varGamma ^{\prime }$
and
$e=\bar {e}+e^{\prime }$
, which are substituted into (2.19) and (2.20), then we subtract (2.20) from (2.19) to yield the linear relation between fluctuation variables as follows:
Spatial distributions of
$-\overline {p^{\prime }u_i^{\prime }}/\bar {p}$
and
$\overline {e^{\prime }u_i^{\prime }}/\bar {e}$
. Panels
$(a){-} (c)$
and
$(d){-} (f)$
correspond to moments of
$t/\tau$
= 1, 2.5 and 4 for the cases of
$S_r$
= 0.5 and 3, respectively.

where the nonlinear fluctuation correlations are discarded. The adiabatic index fluctuation
$\varGamma ^{\prime }$
can be considered neglected due to the following reasons. First, for mixtures comprising gases with identical molecular structures, the adiabatic index
$\varGamma _L$
and
$\varGamma _H$
of the light and heavy fluids are identical. Consequently,
$\varGamma$
of the mixture is independent of the species and can be treated as a constant, with
$\varGamma ^{\prime }$
becoming appreciable only at extreme temperatures which is beyond the scope of the present study. Second, for mixtures consisting of gases with different molecular structures, the variation in
$\varGamma$
of the mixture remains insignificant, even if the adiabatic index
$\varGamma _L$
and
$\varGamma _H$
of the light and heavy fluids are different and the
$\varGamma$
of the mixture is species-dependent. This is attributed to the small difference in adiabatic indices for gases with different molecular structures. For example,
$\varGamma$
= 1.4 for diatomic gas,
$\varGamma$
= 1.33 for polyatomic gas, which indicates negligible spatial gradients along the direction of mixing development. In contrast, the fluctuations in density, pressure and internal energy are substantially larger than
$\varGamma ^{\prime }$
, and as thus, the
$\varGamma ^{\prime }$
can be considered negligible. Consequently, (2.21) reduces to
The mass flux velocity is then obtained by multiplying (2.22) by
$u^{\prime }$
and then applying averaging:
The contributions of the two terms,
$-\overline {p^{\prime }u_i^{\prime }}/\bar {p}$
and
$\overline {e^{\prime }u_i^{\prime }}/\bar {e}$
, to the mass flux velocity
$\overline {u_{i}^{\prime \prime }}$
are examined using HiFi simulation results for cases with
$S_r$
= 0.5 and 3. Figure 6 presents their spatial profiles at
$t/\tau$
= 1, 2.5 and 4, revealing that
$\overline {e^{\prime }u_i^{\prime }}/\bar {e}$
dominates the mass flux velocity generation, whereas the influence of
$-\overline {p^{\prime }u_i^{\prime }}/\bar {p}$
is negligible. Consequently, the expression (2.23) is further reduced as
With the relation
$e=c_vT$
, (2.24) is written as
where a third-order fluctuation correlation term
$\overline {c^{\prime }_vT^{\prime }u_i^{\prime }}$
is ignored. These two terms in (2.25) will be modelled respectively.
Using the GDA,
$({\overline {c_v^{\prime }u^{\prime }_{i}}})/{\bar {c}_v}$
is modelled as
\begin{eqnarray} \frac {\overline {c_v^{\prime } u_i^{\prime }}}{\bar {c}_v} &=& -\frac {C_3 \nu _t}{\bar {c}_v} \frac {\partial \bar {c}_v}{\partial x_i} \nonumber \\ &=& -\frac {C_3 \nu _t}{\bar {c}_v} \dfrac {\partial }{\partial x_i} \left ( \dfrac {\bar {p}}{\bar {\rho }(\bar {\varGamma }-1)\bar {T}} \right ) \nonumber \\ &=& -\frac {C_3 \nu _t}{\bar {c}_v} \dfrac {\bar {\rho }(\bar {\varGamma }-1)\bar {T} \dfrac {\partial \bar {p}}{\partial x_i} - \bar {p}(\bar {\varGamma }-1) \dfrac {\partial (\bar {\rho }\bar {T})}{\partial x_i}}{\left[\bar {\rho }(\bar {\varGamma }-1)\bar {T}\right]^2} \nonumber \\ &=& -C_3 \nu _t \frac {\bar {\rho }\bar {T} \dfrac {\partial \bar {p}}{\partial x_i} - \bar {p} \dfrac {\partial (\bar {\rho }\bar {T})}{\partial x_i}}{\bar {c}_v \bar {\rho }^2 (\bar {\varGamma }-1)\bar {T}^2} \nonumber \\ &=& -C_3 \nu _t \left ( \frac {1}{\bar {p}} \dfrac {\partial \bar {p}}{\partial x_i} - \dfrac {1}{\bar {T}} \dfrac {\partial \bar {T}}{\partial x_i} - \dfrac {1}{\bar {\rho }} \dfrac {\partial \bar {\rho }}{\partial x_i} \right ) \nonumber \\ &=& \frac {C_3 \nu _t}{\bar {\rho }} \left [ \dfrac {\partial \bar {\rho }}{\partial x_i} - \dfrac {\bar {\varGamma }}{\bar {c}^2} \left ( \frac {\partial \bar {p}}{\partial x_i} - \frac {\bar {p}}{\bar {T}} \frac {\partial \bar {T}}{\partial x_i} \right ) \right ], \end{eqnarray}
where the second equals sign replaces the
$\bar {c}_v$
with
$\bar {p}/ [\bar {\rho }(\bar {\varGamma }-1)\bar {T} ]$
to obtain a similar expression to the instability criterion (2.18) for compressible flows,
$C_3$
representing a model coefficient to be determined.
The pure turbulent heat flux
$\overline {T^{\prime }u^{\prime }_{i}}$
is straightforwardly modelled based on the expression proposed by Cloutman (Reference Cloutman2003), reading
where
$C_4$
is a model coefficient to be determined. By combining with (2.27),
$({\overline {T^{\prime }u^{\prime }_{i}}})/{\bar {T}}$
is expressed as
It is worth emphasising that this formulation of (2.27) deviates from the conventional GDA framework by explicitly accounting for counter-gradient heat transport mechanisms. Classical gradient diffusion theories, such as Fourier’s law, posit that heat transfer occurs from high-temperature to low-temperature regions, with flux magnitude proportional to the local temperature gradient. However, turbulent flow dynamics can induce anomalous heat transport phenomena where heat is transferred against the temperature gradient – from low-temperature to high-temperature regions. Such a counter-gradient behaviour is ubiquitously observed in convection-dominated boundary layer, buoyancy-dominated turbulence, turbulent coherent structures, among other scenarios.
In the density-stratified compressible RT mixing flows investigated in this study, counter-gradient heat transfer is observed, as clearly evidenced by the flow field visualisation in figure 7. To elucidate this phenomenon, we analyse the late-stage flow field at
$t/\tau$
= 4 from the central
$x{-}z$
plane of the 3-D HiFi simulations for weak (
$S_r$
= 0.5) and strong (
$S_r$
= 3) stratification cases. The visualisation employs a background temperature colour map overlaid with instantaneous turbulent heat flux vectors
$\overline {T^{\prime }u^{\prime }_{i}}$
. Figure 7 reveals the following points. First, the heavy fluid regions generally exhibit lower temperatures compared with the light fluid regions and, moreover, density stratification promotes a wider range of temperature variations across the mixing zone. Second, strong density stratification (
$S_r$
= 3) contributes to produce organised large-scale flow structures, enhancing heat transfer efficiency. The case with weak density stratification (
$S_r$
= 0.5) displays dispersed small-scale turbulent structures, resulting in heat transfer dominated by local mixing. Last but not least, the turbulent heat flux shows significant deviation from the local temperature gradient direction, with clear evidence of heat transfer from low-temperature to high-temperature regions. These findings underscore the importance of turbulent heat flux in the flows considered here, and further necessitate explicit consideration of counter-gradient diffusion in modelling
$\overline {T^{\prime }u^{\prime }_{i}}$
. Consequently, based on the work of Cloutman (Reference Cloutman2003), we adopt the closure formulation (2.27) to address the specific requirements of modelling counter-gradient diffusion in density-stratified compressible RT flows.
Contours of instantaneous temperature in the central
$x{-}z$
plane for cases of
$(a)$
$S_r$
= 0.5 and (
$b$
)
$S_r$
= 3 at
$t/\tau$
= 4. The instantaneous turbulent heat flux is also displayed with the vector arrows.

Finally, collecting the expressions (2.28) and (2.26), the turbulent mass flux velocity is modelled as
Based on a principle of minimal modifications to the baseline model, the expression (2.29) is integrated into the baseline closure (2.12) for the local Atwood number
$A_{\textit{ssi}}$
, yielding
\begin{equation} {{A}_{\textit{ssi}}}=\frac {{\tilde {L}}}{\bar {\rho }+\tilde {L}\left | \dfrac {\partial \bar {\rho }}{\partial x_{i}} \right |}\!\left \{ C_3\!\left [\frac {\partial \bar {\rho }}{\partial x_i}-\frac {\bar {\varGamma }}{\bar {c}^2}\left (\!\frac {\partial \bar {p}}{\partial x_i}-\frac {\bar {p}}{\bar {T}}\frac {\partial \bar {T}}{\partial x_i}\right )\!\right ]-C_4\frac {1}{\bar {c}^2}\left [\frac {\bar {\varGamma }\bar {p}}{\bar {T}}\frac {\partial \bar {T}}{\partial x_{i}}-(\bar {\varGamma }-1)\frac {\partial \bar {p}}{\partial x_{i}}\right ] \right \}\!. \end{equation}
To ensure consistency with the baseline form (2.12) in the incompressible limit
$\bar {c}\,\rightarrow \,\infty$
, the model coefficient
${C}_{3}$
is required to meet the constraint
To determine the coefficient
$C_4$
, the ratio
$C_4/C_3$
is analysed using HiFi simulation results, specifically, with (2.28) and (2.26),
\begin{equation} \frac {C_4}{C_3}= \frac {\dfrac {\overline {T^{\prime }u^{\prime }_{i}}}{\bar {T}}/\left \{-\dfrac {1}{\bar {c}^2}\left [\dfrac {\bar {\varGamma }\bar {p}}{\bar {T}}\dfrac {\partial \bar {T}}{\partial x_{i}}-(\bar {\varGamma }-1)\dfrac {\partial \bar {p}}{\partial x_{i}}\right ]\right \}}{\dfrac {\overline {c_v^{\prime }u^{\prime }_{i}}}{\bar {c}_v}/ \left [\dfrac {\partial \bar {\rho }}{\partial x_i}-\dfrac {\bar {\varGamma }}{\bar {c}^2}\left (\dfrac {\partial \bar {p}}{\partial x_i}-\dfrac {\bar {p}}{\bar {T}}\dfrac {\partial \bar {T}}{\partial x_i}\right )\right ]}. \end{equation}
Figure 8 presents the spatial distributions of
$C_4/C_3$
across the mixing regions at
$t/\tau$
= 1, 2.5 and 4 for the case where
$S_r=3$
, indicating the ratio predominantly varying within the range
$[0,1]$
. To derive a representative constant value, the mean value of
$C_4/C_3$
is computed as
$ ({\int C_4/C_3\,{\rm d}x})/{H}$
, where
$H$
denotes the cutoff mixing width. The resulting mean ratios are 0.4, 0.5 and 0.55 at
$t/\tau$
= 1, 2.5 and 4, respectively. Furthermore, the influence of
$C_4$
on the predictive result is examined, as shown in figure 9, revealing that the bubble width exhibits weak sensitivity to
$C_4$
, whereas the spike width remains largely insensitive. Based on these findings, a comparatively optimal ratio of
$C_4/C_3$
= 0.4 is adopted in this study.
Spatial distributions of the ratio
$C_4/C_3$
across the mixing regions (marked by the arrows) at
$t/\tau$
= 1, 2.5 and 4.

Temporal evolutions of the mixing widths for the case of
$S_r$
= 3, based on different
$C_4$
.

Finally, the local Atwood number
$A_{\textit{ssi}}$
is expressed as
\begin{align} {{A}_{\textit{ssi}}}=\frac {{\tilde {L}}}{\bar {\rho }+\tilde {L}\left | \dfrac {\partial \bar {\rho }}{\partial x_i} \right |} \left \{ C_A\frac {\partial \bar {\rho }}{\partial x_{i}}+\underbrace {\frac {1}{{{{\bar {c}^{2}}}}}\left \{\left [C_4(\bar {\varGamma }-1)-C_A\bar {\varGamma }\right ]\frac {\partial \bar {p}}{\partial x_{i}}+(C_A-C_4)\frac {\bar {\varGamma }\bar {p} }{\bar {T}}\frac {\partial \bar {T}}{\partial x_{i}}\right \}}_{\rm compressibility \, corrections} \right \}. \end{align}
The second term in the expression (2.33) incorporates all compressibility corrections proposed here. It is worth noting that the local instability criterion (2.18) for compressible flows has been incorporated into the new closure (2.33). In the incompressible limit
$c\,\rightarrow$
0, the expression (2.33) naturally converges to its baseline counterpart given by (2.12), demonstrating consistency with the baseline formulation.
The proposed modelling strategy demonstrates that in compressible mixing flows, generation of the turbulent mass flux primarily depends on two key mechanisms: the turbulent composition and heat fluxes. Particularly, the former arises fundamentally from species fluctuations caused by turbulent mixing. It is attributed to that the mixture’s constant-volume specific heat is determined through a species-linearly weighting of the light and heavy fluid components, i.e.
$c_v$
=
$c_v^{H}Y+c_v^{L}(1-Y)$
, implying
$c_v^{\prime }$
essentially resulting from species fluctuation, while
$c_v^{H}$
and
$c_v^{L}$
can be regarded as constants under non-extreme temperature situations. The present scheme physically captures density variations induced by both turbulent composition and thermal effects, and is expected to provide a comprehensive framework for characterising compressibility effects in variable-density compressible turbulent mixing flows.
3. Model validation
To evaluate the proposed compressibility corrections, density-stratified compressible RT mixing flows with different stratification parameters
$S_r$
are tested. The flow configurations, including model coefficients and initialisation of turbulence model variables, remain consistent with the baseline model, as stated in § 2.2.
Figure 10 illustrates the temporal evolutions of bubble (
$W_b$
) and spike (
$W_s$
) mixing widths for 4 cases with density stratification parameters
$S_r$
ranging from 0.5 to 3. The proposed compressibility modifications show remarkable consistency with HiFi simulation results across all tested cases. This improved performance stems from two fundamental physical considerations: incorporation of the local instability criterion for compressible flows by considering the mixing composition effect and appropriate treatment of turbulent heat flux contribution. By incorporating these effects, the model achieves a more comprehensive representation of the turbulence production term in the TKE equation, leading to significantly improved predictions of compressible mixing evolution. It should be noted that identical initialisation of the turbulent model variables and model coefficients are maintained throughout all the cases. The model’s robustness is particularly evident in strongly stratified cases (
$S_r$
= 3), where it still delivers accurate predictions despite the significant compressibility effect. The success of the compressible closure (2.33) stems from its physically grounded formulation, which effectively captures the essential dynamics of compressible mixing flows.
Temporal evolutions of the mixing widths. Panels
$(a){-} (d)$
correspond to the cases of
$S_r$
= 0.5, 1, 2 and 3, respectively.

Spatial profiles of the mean species at three different moments:
$t/\tau =1$
(coloured in black),
$t/\tau =2.5$
(coloured in blue) and
$t/\tau =4$
(coloured in red). Panels
$(a){-} (d)$
correspond to the cases with
$S_r=$
0.5, 1, 2 and 3, respectively.

Except for the temporal evolution of the mixing width, the spatial profiles of the important variables are also examined. Figure 11 presents the mass fraction profiles at three different moments
$t/\tau$
= 1,
$t/\tau$
= 2.5 and
$t/\tau$
= 4, across stratification parameters
$S_r$
varying from 0.5 to 3. Excellent agreements with HiFi simulations are observed for all cases, demonstrating the model’s robust predictive capability. This validation is further reinforced by the density profile comparisons, as shown in figure 12, which serve as a critical metric for compressible flow characterisation.
Spatial profiles of the mean density at three different moments:
$t/\tau =1$
(coloured in black),
$t/\tau =2.5$
(coloured in blue) and
$t/\tau =4$
(coloured in red). Panels
$(a){-} (d)$
correspond to the cases with
$S_r=$
0.5, 1, 2 and 3, respectively.

The density profile variations depicted in figure 12 provide critical physical insights, offering a more direct understanding of the baseline model’s distinct performance across cases with different
$S_r$
. Specifically, for the cases with a relatively small
$S_r$
, i.e.
$S_r$
= 0.5 and 1, the sustained positive density gradient throughout the simulation ensures an effective baroclinic production term. It leads to the baseline model’s better performance for mixing width in cases with weaker stratification (
$Sr=$
0.5 and 1) compared with stronger stratification cases (
$Sr=$
2 and 3), as shown in figure 10. Additionally, it is also noted that compared with the case of
$S_r$
= 1, the baseline model shows superior accuracy for the case of
$S_r$
= 0.5, as the positive density gradient persists across the entire mixing region, providing a relatively sufficient TKE production term. However, for the case of
$S_r$
= 1, the positive density gradients become confined to the core mixing region at later stages, while negative gradients emerge at the edge of the mixing region. It is well established that the amplitude of the mixing width depends on the penetration dynamics of bubbles and spikes at the edge of the mixing zone. In RANS simulation, the strength of penetration is closely linked to the amplitude of the TKE production term. Due to the disappearance of the baroclinic term at the edge (where the density gradient becomes negative), the TKE production term diminishes to zero, significantly reducing the TKE and halting further development of the mixing evolution. Consequently, the mixing width predicted by the baseline model is substantially lower than the actual physical evolution.
The underestimate of the mixing width for cases of
$S_r$
= 2 and 3 can also be explained by the aforementioned analysis. For these cases, which exhibit strong density stratification, fundamentally different behaviour is observed. The density gradient prematurely reverses sign during the simulation – occurring at the edge of the mixing region in the intermediate stage for
$S_r$
= 2 and across the entire mixing region in the intermediate stage for
$S_r$
= 3. This gradient reversal causes a severe deficit in the turbulence production source. Since the baseline model’s closure relies solely on the baroclinic term, it fails to provide sufficient TKE production, leading to a deterioration in model performance with increasing stratification strength.
Section 2.2 establishes that flow regions where baroclinic terms vanish may maintain instability, as determined by the local instability criterion for compressible flows. This phenomenon has also been validated by HiFi simulations (Gauthier Reference Gauthier2017; Luo & Wang Reference Luo and Wang2022), which demonstrate ongoing mixing and turbulence development in density-stratified compressible RT flows despite baroclinic term disappearance. The proposed model incorporates more comprehensive physical mechanisms for compressible mixing flows and supplements the TKE production source appropriately, leading to successful predictions across all investigated cases.
4. Discussion
4.1. Impact on RM mixing prediction
Given the inherent compressibility of shock wave motions that drive RM instability, a natural question arises regarding whether the proposed compressibility corrections would significantly alter RANS predictions for RM mixing. To reveal this, we examine the inverse chevron reshocked RM mixing, a case previously validated using the baseline
$K$
–
$L$
–
$\gamma$
model without compressibility corrections (Xie et al. Reference Xie, Qi, Xiao, Zhang and Zhao2025). The flow configuration and computational set-up remain identical to those described in the earlier work (Xie et al. Reference Xie, Qi, Xiao, Zhang and Zhao2025). Figure 13 presents the temporal evolution of important mixing quantity, comparing results with and without compressibility corrections. The remarkable agreement between both sets of predictions indicates that the proposed modifications have minimal impact on RM mixing, which is attributed to two factors. First, while the shock passage itself is highly compressible, its duration is sufficiently transitory that the compressibility effect does not dominantly influence the subsequent mixing evolution. Second, during the post-shock mixing evolution stage, the flow field does not generate significant pressure and temperature gradients – the crucial terms for the present compressibility corrections.
Temporal evolution of the mixing quantity for the inverse chevron case predicted by the
$K$
–
$L$
–
$\gamma$
model. The red and blue lines represent the RANS results with and without compressibility corrections, respectively.

4.2. Scenarios of considering compressibility corrections
Given the distinct importance of the present modifications in RT and RM mixing flows, conditions under which the proposed compressibility corrections become important deserve discussion. The present RANS results across cases with different
$S_r$
demonstrate that these modifications play an increasingly prominent role in flows characterised by stronger density stratification. As discussed in § 1, density stratification is a crucial characteristic to reflect static compressibility, which is related to the expansion and compression motions and is associated with density fluctuations of fluid elements in response to change in pressure. Therefore, the compressibility corrections should be considered when there is a significant static compressibility effect.
Furthermore, the significant temperature fluctuation proves the importance of the proposed compressibility corrections. The associated turbulent heat flux exerts substantial influence on flow compressibility since temperature variations directly modify density and pressure through the equation of state. Previous implementations of turbulent heat flux corrections in RANS modelling have been quite limited, which primarily stems from the historical focus on mixing scenarios where thermodynamic effects play negligible roles in flow evolution. However, HiFi simulations, particularly for late-stage evolution at high
$S_r$
values, demonstrate pronounced temperature fluctuations that fundamentally alter mixing dynamics (Gauthier Reference Gauthier2017). Moreover, the test of discarding turbulent heat flux closure shows that there is a significant underestimation of late-time mixing evolution, especially in highly compressible cases. The incorporation of turbulent heat flux closures therefore represents not merely an incremental improvement for mixing width, but rather a physical necessity for properly modelling the strongly compressible flows.
5. Summary
Compressible mixing flow induced by hydromechanics interfacial instability play a crucial role in engineering applications, such as the ICF. RANS simulations remain the most practical approach for engineering predictions. However, most of the existing RANS mixing models focus on flows where the influence of compressibility on mixing evolution is insignificant and fail to predict mixing flows with significant compressibility influence. This study examines the density-stratified RT flows as representative cases to: (i) identify limitations in the existing models for compressible flow predictions; and (ii) develop an improved compressible RANS mixing model. Our analysis begins with a comparative study of local instability criteria for incompressible versus compressible flows, demonstrating the physical inadequacy of conventional baroclinic term modelling (
$ ({\partial \bar {p}}/{\partial x_{i}}) ({\partial \bar {\rho }}/{\partial x_{i}})$
) for compressible cases. Subsequently, based on the
$K$
–
$L$
–
$\gamma$
mixing transition model proposed by Xie et al. (Reference Xie, Qi, Xiao, Zhang and Zhao2025), we derive compressibility corrections for modelling the key turbulent mass flux term in the TKE equation. Validation studies across various density-stratified compressible RT cases confirm the model’s improved predictive capability. Compared with the baseline model, two fundamental improvements account for the proposed model’s success. One is the consideration of the turbulent composition effect, incorporating the local instability criterion for compressible flows, and the other is proper closure of turbulent heat flux contribution. Moreover, the impact of the proposed corrections on RM mixing prediction and scenarios of considering these modifications are discussed.
While elements of the present modelling formulation have appeared in previous studies, their application to RANS modelling of interfacial mixing flows remains limited. The present closure shares some similar modelling with formulations dating back nearly four decades in stellar evolution studies (Cloutman Reference Cloutman1987). Furthermore, the kinetic closure model for turbulent counter-gradient diffusion developed by Cloutman (Reference Cloutman2003) provides theoretical support for modelling the turbulent heat flux in this study. In the existing studies, several researchers have employed partially similar expressions of the current compressibility corrections. For instance, the
$K$
–
$\epsilon$
model by Morán-López & Schilling (Reference Morán-López and Schilling2013) uses the form
$({\partial \bar {\rho }}/{\partial x_{i}})-({\bar {\rho }}/{\bar {p}}) ({\partial \bar {p}}/{\partial x_{i}})$
to close the turbulent mass flux velocity
$\overline {u^{\prime \prime }_{i}}$
. Zhou (Reference Zhou2017b
) have also emphasised the importance of such an expression in modelling the TKE equation’s buoyancy source term. However, these pioneering contributions have received relatively little attention in modelling interfacial mixing flows. This may be attributed to the fact that the influence of compressibility on mixing evolution is insufficiently pronounced to demonstrate the models’ full potential.
Contributions of this study include the following points. First, we present a systematic derivation of comprehensive compressibility closures based on the EOS of a perfect gas. This represents a significant advance over previous sporadic and incomplete formulations found in the literature. The developed closures maintain both physical fidelity and computational tractability. Second, our results demonstrate the crucial roles of turbulent composition and heat fluxes in compressible flows, while the latter is usually ignored in previous studies. Through detailed analysis, it is revealed that turbulent heat flux: (i) significantly modifies buoyancy effects by mediating density fluctuations, especially for highly compressible cases; and (ii) exhibits counter-gradient diffusion behaviour in density-stratified compressible RT flows, which necessitates specific modifications, while the traditional GDA fails to describe that. Third, this work provides the first successful RANS-based framework for predicting density-stratified compressible RT flows, representing a significant step forward in turbulence modelling for practical engineering applications.
The present study primarily establishes a foundational modelling strategy for compressible mixing flows and provides potential modelling implications for other RANS mixing models (see the Appendix), with comprehensive verifications left for future investigations.
Acknowledgements
The authors sincerely thank the reviewers for the constructive suggestions, which significantly improve the quality of this paper.
Funding
This work was supported by the National Natural Science Foundation of China (Grant Nos. 12222203, 12588301, 12532013, 92152202 and 12432010) and the National Key Laboratory of Computational Physics (Grant Nos. 6142A05240201, JK2024-03). Additionally, T. Luo was supported by the Research Grants Council (RGC) of the Hong Kong Special Administrative Region, China (Project Reference: AoE/P-601/23-N) and the Center for Ocean Research in Hong Kong and Macau (CORE), a joint research center between the Laoshan Laboratory and the Hong Kong University of Science and Technology (HKUST).
Declaration of interests
The authors report no conflict of interest.
Appendix Potential modelling implications for other RANS mixing models
In addition to the
$K$
–
$L$
model examined in this study, alternative RANS mixing models exist, such as the BHR-type formulations, e.g. BHR-2 and
$K$
–
$L$
–
$a$
–
$b$
models (Kokkinakis et al. Reference Kokkinakis, Drikakis and Youngs2019). The proposed compressibility corrections may offer potential enhancements for these models.
Temporal evolutions of the mixing widths predicted by the
$K$
–
$L$
–
$a$
–
$b$
model, whose physical model and model coefficients can be found from Zhao et al. (Reference Zhao, Xie, Xiao, Wang and Zhang2025). Panels
$(a){-} (d)$
correspond to the cases of
$S_r$
= 0.5, 1, 2 and 3, respectively. The used numerical schemes are consistent with those in § 2.2: the initial TKE
$\tilde {K}_0=0.0001A_tg\tilde {L}_0$
, turbulent mass flux velocity
$a_0=\sqrt {\tilde {K}_0}/4$
, density-specific volume covariance
$b=1\times 10^{-8}$
. Particularly, the initial turbulent length scale
$\tilde {L}_0$
is adjusted to obtain optimal results. All the cases use the same RANS initialisations.

While BHR-type models advance beyond the
$K$
–
$L$
model through introductions of transport equations for turbulent mass flux velocity
$a_i$
and density-specific volume covariance
$b$
, their predictive accuracy for the density-stratified compressible RT mixing remains inadequate. Here, we evaluate the recently improved
$K$
–
$L$
–
$a$
–
$b$
model (Zhao et al. Reference Zhao, Xie, Xiao, Wang and Zhang2025) across cases of
$S_r$
= 0.5, 1, 2 and 3. Figure 14 reveals two fundamental limitations. First, all predictions exhibit quadratic growth in time and early-stage overprediction, indicating a failure to capture mixing transition dynamics due to the inherent fully developed-turbulence modelling assumption. Second, this
$K$
–
$L$
–
$a$
–
$b$
model cannot capture the mixing width evolutions involving the compressibility effect, substantially under-predicting widths for strong compressible cases of
$S_r=$
2 and 3, despite acceptable accuracy for weak compressible cases of
$S_r=$
0.5 and 1. While adjusting
$\tilde {L}_0$
can improve late-time predictions for specific
$S_r$
, it simultaneously exacerbates early-time errors. This trade-off confirms that parametric tuning cannot resolve the model’s inherent limitations in capturing transition and compressibility physics.
To improve these limitations, possible strategies are proposed as follows. For effective prediction of mixing transition dynamics, the introduction of either an intermittency factor
$\gamma$
(as implemented in the
$K$
–
$L$
–
$\gamma$
model here) or density-specific volume covariance
$b$
(Xiao et al. Reference Xiao, Qi and Zhang2025) to constrain turbulent eddy viscosity shows considerable promise. Regarding compressibility effects, incorporating turbulent heat flux closure proves essential, as discussed previously. A rough analysis of the
$K$
–
$L$
–
$a$
–
$b$
model reveals that while its transport equations for
$a_i$
and
$b$
achieve theoretically sound modelling – explaining its comparatively superior performance than the
$K$
–
$L$
model – its critical deficiency lies in the absence of turbulent heat flux modelling. This fundamental limitation causes systematic underestimation of late-time mixing width and failure to capture mixing growth induced by temperature fluctuations. Therefore, incorporating turbulent heat flux closures in the
$K$
–
$L$
–
$a$
–
$b$
model (even other BHR-type models) to physically account for thermal fluctuation effects is a potential improvement direction.






































































