1. Introduction
Turbulent mixing in multi-component materials is a fundamental process observed in astrophysical (Porth, Komissarov & Keppens Reference Porth, Komissarov and Keppens2014; Hillier Reference Hillier2018) and geophysical flows (Fernando Reference Fernando1991; Gregg et al. Reference Gregg, D’Asaro, Riley and Kunze2018), and also in engineering applications such as inertial confinement fusion (ICF) (Lindl Reference Lindl1995; Nakai & Takabe Reference Nakai and Takabe1996; Hurricane et al. Reference Hurricane2024; Abu-Shawareb et al. Reference Abu-Shawareb2024). Controlling the perturbations at the fuel/ablator interface of an ICF capsule is crucial to reduce the mixing, which has depleting effects on the ignition. In fluids or plasmas, turbulent mixing often occurs where density gradients are present along with strong accelerations (denoted by
$g$
). This configuration is at the origin of many phenomena such as gravity waves, the Richtmyer–Meshkov instability (Thornber et al. Reference Thornber, Drikakis, Youngs and Williams2010) and the Faraday instability (Briard, Gostiaux & Gréa Reference Briard, Gostiaux and Gréa2020; Cavelier et al. Reference Cavelier, Gréa, Briard and Gostiaux2022). Complex acceleration histories have been shown to have an important impact on the mixing layer dynamics (Dimonte, Ramaprabhu & Andrews Reference Dimonte, Ramaprabhu and Andrews2007; Ramaprabhu, Karkhanis & Lawrie Reference Ramaprabhu, Karkhanis and Lawrie2013; Livescu, Wei & Brady Reference Livescu, Wei and Brady2021; Aslangil, Lawrie & Banerjee Reference Aslangil, Lawrie and Banerjee2022; Morgan Reference Morgan2022). The Rayleigh–Taylor instability (RTI) (Rayleigh Reference Rayleigh1882; Taylor Reference Taylor1950) is particularly important among buoyancy-driven flows due to its ability to efficiently convert potential energy into mixing (Davies Wykes & Dalziel Reference Davies Wykes and Dalziel2014).
The RTI thus occurs when a light fluid pushes a heavy one or equivalently when a heavy fluid is placed on top of a lighter one in a gravity field (Boffetta & Mazzino Reference Boffetta and Mazzino2017; Zhou Reference Zhou2017; Zhou et al. Reference Zhou2021). Classical RTI dynamics involve the growth of a mixing zone from an initially perturbed interface, successively punctuated by different regimes (Sharp Reference Sharp1984). In the context of miscible fluids, the early linear phase (Chandrasekhar Reference Chandrasekhar1961; Duff, Harlow & Hirt Reference Duff, Harlow and Hirt1962) occurs immediately, or after a diffusive process, depending on the characteristic length of the main interface defects. When all initial modes are stable, with their wavelengths being smaller than the cut-off scale determined by the dispersion relationship, unstable larger wavelength can still be produced by a slow backscatter process. The mixing zone later enters a potential nonlinear regime where the dominant spikes or bubbles evolve towards a constant terminal velocity (Layzer Reference Layzer1955; Cook & Dimotakis Reference Cook and Dimotakis2001; Goncharov Reference Goncharov2002; Ramaprabhu et al. Reference Ramaprabhu, Dimonte, Young, Calder and Fryxell2006; Bian et al. Reference Bian, Aluie, Zhao, Zhang and Livescu2020; Gréa & Briard Reference Gréa and Briard2023). Secondary shear instabilities, produced between the rising and falling structures, trigger the transition to turbulence, leading to the formation of a turbulent mixing zone (Cook, Cabot & Miller Reference Cook, Cabot and Miller2004).
The special case where the density contrast, expressed by the Atwood number (
$\mathcal A$
), is small has been extensively studied (Andrews & Dalziel Reference Andrews and Dalziel2010; Schilling Reference Schilling2020). In such configurations, variable density effects are primarily evident through the buoyancy force and the mixing layer maintains a top/bottom statistical symmetry (Ramaprabhu & Andrews Reference Ramaprabhu and Andrews2004; Schilling & Mueschke Reference Schilling and Mueschke2010; Livescu et al. Reference Livescu, Ristorcelli, Petersen and Gore2010). There is a strong correspondence between experimental results (Mueschke, Andrews & Schilling Reference Mueschke, Andrews and Schilling2006) and simulations (Mueschke & Schilling Reference Mueschke and Schilling2009a
,Reference Mueschke and Schilling
b
) in the transitional regime. Furthermore, mixing quantities and scaling laws have been thoroughly investigated by Boffetta et al. (Reference Boffetta, Mazzino, Musacchio and Vozella2010). In the later stages, dimensional analysis confirms the emergence of a final regime where the mixing zone width,
$L$
, grows according to the formula
$ 2 \alpha \mathcal A g t^2$
with
$t$
representing time and
$\alpha$
denoting the classical self-similar growth rate of the bubbles (Dalziel, Linden & Youngs Reference Dalziel, Linden and Youngs1999; Ristorcelli & Clark Reference Ristorcelli and Clark2004; Vladimirova & Chertkov Reference Vladimirova and Chertkov2009; Morgan et al. Reference Morgan, Olson, White and McFarland2017).
Whether the growth parameter is universal or not has been an ongoing debate in the past decades. A wide range of
$\alpha$
values, from
$0.02$
to
$0.10$
, have been reported in both experiments or simulations, which, from a modelling perspective, has significant implications (Young et al. Reference Young, Tufo, Dubey and Rosner2001; Dimonte et al. Reference Dimonte2004; Ramaprabhu, Dimonte & Andrews Reference Ramaprabhu, Dimonte and Andrews2005). Beyond the difficulty of reaching the RTI’s final regime, particularly due to the confinement of the growing large-scale structures, the most resolved simulations (Cook & Cabot Reference Cook and Cabot2006; Livescu Reference Livescu2013; Morgan & Black Reference Morgan and Black2020; Briard, Gréa & Nguyen Reference Briard, Gréa and Nguyen2022) or well-controlled experiments (Roberts & Jacobs Reference Roberts and Jacobs2016) agree on the lower values
$\alpha \sim 0.02{-}0.03$
. Furthermore, the perturbations at the initial interface seem to have an impact (Youngs Reference Youngs2013). Depending on whether the initial modes are at large or small scales, it has been observed that the final value of
$\alpha$
either slightly evolves or remains constant (Dimonte Reference Dimonte2004). This phenomenon has been linked to mode competition or mode coupling. Therefore, it appears that the RTI, even in its final self-similar turbulent regime, retains partial memory of its initial conditions.
In addition, it comes as no surprise that the mixing layer growth is also related to the large scales of the flow. From the rapid acceleration model, a formula for
$\alpha$
has been derived, showing its dependence on global mixing and the dimensionality parameter that expresses the elongation of large-scale structures along the vertical direction (Gréa Reference Gréa2013). The imprint of the coherent structures on the dynamics of RTI has also been evidenced through the analysis of the infrared slope of turbulent spectra (Poujade & Peybernes Reference Poujade and Peybernes2010; Soulard, Griffond & Gréa Reference Soulard, Griffond and Gréa2015). However, except for homogeneous configurations that satisfy the generalised principle of permanence of large eddies, these studies cannot explicitly demonstrate the relationship between the final regime and the initial conditions of buoyancy-driven flows such as RTI.
More recently, Kord & Capecelatro (Reference Kord and Capecelatro2019) demonstrate the ability of adjoint-based methods to evaluate the sensitivity of Rayleigh–Taylor (RT) mixing to its initial conditions. The study also successfully derives the optimal interface perturbations to dampen or enhance the RTI growth in its late stages. Beyond the challenges of applying an adjoint method when the initial perturbation involves a very large number of modes, this study clearly suggests that the distribution of the initial perturbation is crucial when modelling the RTI.
In this work, we aim to examine the extent to which the RTI retains or loses information about its initial conditions, thereby exploring the ‘memory’ of the flow. We address this problem using a Bayesian approach to infer initial conditions. Rather than evaluating how complete knowledge of the mixing zone state would allow for a full reconstruction of the initial conditions, we focus on how zero-dimensional (0-D) observations can recover a reduced set of parameters that express the statistical properties of the initial conditions. This approach paves the way for a statistical Bayesian modelling of the Rayleigh–Taylor instability. After inferring the most probable initial conditions, we use this information to predict the evolution of turbulent quantities.
This work is organised as follows. In § 2, we present the RTI database obtained from direct numerical simulation (DNS) and the surrogate model developed to reproduce the 0-D turbulent quantities. In § 3, we perform a sensitivity analysis to identify the influence of each initial parameter across the different RTI regimes. Finally, in § 4, we describe the methodology for inferring the initial conditions of RTI from 0-D observations and for modelling the RTI transition to turbulence based on the maximum a posteriori (MAP) initial conditions.
2. Direct numerical simulations and surrogate model
2.1. Basic equations and RTI initial set-up
The binary mixture of two miscible fluids with densities
$\rho _1$
and
$\rho _2$
(
$\rho _1\gt \rho _2$
) can be characterised by the velocity vector
${\textbf {u}} (\textbf {x},t)=(u_x,u_y,u_z)^T$
and the concentration of the heavier fluid
$C (\textbf {x},t)$
, where
$\textbf {x}= (x,y,z )^T$
defines the position in a Cartesian frame and
$t$
denotes time. In the limit of small Atwood number (
$\mathcal A = (\rho _1-\rho _2)/(\rho _1+\rho _2)\rightarrow 0$
) and neglecting compressibility effects, the density can be expressed by
$\rho =\rho _0 (1-\mathcal A +2 \mathcal A C)$
(with
$\rho _0=(\rho _1+\rho _2)/2$
). The problem is thus governed by the incompressible Navier–Stokes equations under the Boussinesq approximation (Boffetta & Mazzino Reference Boffetta and Mazzino2017):
Equation (2.1a
) stands as the incompressibility condition, while in the momentum equation, (2.1b
),
$\Pi$
is the reduced pressure (pressure divided by mean density) and
$\nu$
the kinematic viscosity. For this configuration, the acceleration (
$g\gt 0$
), appearing in the buoyancy term, is oriented downward along the
$z$
vertical direction indicated by the unit basis vector
$\boldsymbol {e}_z$
. In the transport equation for the concentration
$C$
, (2.1c
), the diffusion coefficient
$\kappa$
is here fixed such that the Schmidt number is unity (
$Sc=\nu /\kappa =1$
). The system (2.1a
)–(2.1c
) can be efficiently solved with numerical pseudo-spectral methods. By contrast, the variable density equations, as introduced by Joseph (Reference Joseph2013) and Cook & Dimotakis (Reference Cook and Dimotakis2001), provide a more rigorous framework for binary mixtures for RTI studies, but are also more expensive to solve. That is why they have not been considered for the current study.
At the initial time (
$t=0$
), we assume that the fluid is at rest (
${\textbf {u}}=0$
) and a heavy fluid (concentration
$C=1$
) is located above a lighter fluid (concentration
$C=0$
) separated by a diffuse interface at
$z=0$
and of width
$\delta$
. The initial concentration field
$C$
is thus defined by
with
$\eta (x,y)$
being a zero mean
$2 \pi$
-periodic initial perturbation at the interface. The use of periodic initial conditions in our study is necessitated by the pseudo-spectral method employed to solve the equations, as detailed in the following section. The perturbation is further characterised by the horizontal discrete Fourier modes
$\hat \eta$
of wavevector
$\textbf {k}=(k_x,k_y)^T \in \mathbb {Z}^2$
and modulus
$k=\|\textbf {k} \|$
such that
The realisability condition,
$\eta \in \mathbb {R}$
, imposes that for the complex Fourier modes,
$\hat \eta (-\textbf {k})=\hat \eta ^*(\textbf {k})$
. In this work, we further consider an annular spectrum for the interface perturbation, (2.3), as in Dimonte et al. (Reference Dimonte2004) of the form
Here,
$k_0$
is the mean wavenumber,
$\eta _0$
is the root-mean-square (r.m.s.) amplitude,
$\Delta k$
is the bandwidth of the perturbation,
$\phi$
the phase of modes and
$P(k)$
the perturbation spectrum of the interface. Figure 1 provides an illustration of the initial perturbation in which each phase
$\phi$
is randomly sampled.
Initial perturbation of the interface with: (a) the Fourier power spectrum of the perturbation amplitude; (b) the phase of the Fourier modes; and (c) the perturbation amplitude in the physical space.

Therefore, given the phase
$\phi$
and the wavenumber
$k_0$
, the initial conditions are fully determined by four non-dimensional numbers, namely
In (2.5),
$\mathsf {R}$
stands as an initial perturbation Reynolds number based on the classical linear growth rate
$\sqrt {\mathcal {A}gk_0}$
, and expresses the balance between the inertial and the dissipative effects. In addition,
$\mathsf {B}$
indicates a narrow or large band multi-mode perturbation. The renormalised diffusion thickness is provided by
$\mathsf {D}$
, and
$\mathsf {S}$
measures the steepness of the interface perturbation.
These four numbers, while not accounting for the phases of the modes or the local structures of the disturbance, were shown by Thévenin (Reference Thévenin2024) to provide a robust determination of the non-dimensional trajectories of the volume-averaged 0-D quantities described in § 2.3.
2.2. A large database of RTI simulations
To explore the impact of initial conditions on the RTI, we solve the Navier–Stokes equations under the Boussinesq approximation, (2.1a
)–(2.1c
), using our pseudo-spectral code
$\mathsf{Stratospec}$
(Viciconte et al. Reference Viciconte, Gréa, Godeferd, Arnault and Clérouin2019; Gréa & Briard Reference Gréa and Briard2019; Briard et al. Reference Briard, Gostiaux and Gréa2020Reference Briard, Gréa and Nguyen2022). Our study encompasses a series of three-dimensional (3-D) simulations within a triply periodic domain sized
$(2\pi )^2\times (4\pi )$
, elongated in the vertical direction
$z$
. These simulations are conducted on a
$1024^2\times 2048$
grid, starting from a state where the flow velocity is at rest, and the initial perturbed interface is aligned with the set-up described earlier in § 2.1. Further details regarding the implementation of initial conditions can be found from Briard, Gréa & Nguyen (Reference Briard, Gréa and Nguyen2024).
We have compiled a substantial database of DNS by varying the quartet of the initial conditions, denoted as
$ \mathsf {I}=(\mathsf {R},\mathsf {B},\mathsf {S},\mathsf {D})^T$
, while maintaining a consistent random seed for the phase
$\phi$
. Throughout these simulations, the Atwood number, the Schmidt number and the viscosity are held constant at
$\mathcal {A}=0.05$
,
$Sc=1$
and
$\nu =7.5\times 10^{-4}$
(in arbitrary unit), respectively. The chosen viscosity ensures that, for a gravity
$g\leqslant 150$
, the smallest turbulent scales are wellresolved, with the ratio between the Kolmogorov scale
$\ell _K$
and the grid spacing
$\Delta x=(2\pi )/1024$
consistently exceeding
$ {1}/{2}$
. The maximum initial non-zero wavenumber in these simulations is
$k=341$
.
To mitigate the effects of large-scale confinement, which could adversely affect the dynamics of the mixing layer at later stages, we exclude wavenumbers below
$20$
that have non-zero amplitude in the initial spectral band. This restriction, alongside ensuring that the initial interface thickness (
$\delta$
) and the perturbation amplitude (
$\eta _0$
) are sufficiently resolved (
${\delta } \geqslant 5 \Delta x$
and
$\eta _0 \geqslant 3 \Delta x$
, respectively), outlines the numerical constraints challenging the simultaneous achievement of a high Reynolds number (
$\mathsf R$
) and a broad initial mode bandwidth (
$\mathsf B$
). The distribution of our DNS across the space of initial parameters is illustrated in figure 2, highlighting areas where simulations become numerically infeasible due to these constraints. In total, the database comprises 493 DNS, executed on massively parallel architectures at an approximate computational cost of 30 million CPU hours. All data, including the initialisation parameters and the volume-averaged 0-D dynamical quantities described in § 2.3, are freely accessible online via the French Fluid Dynamics Database (Thévenin et al. Reference Thévenin, Gréa, Kluth and Briard2025).
Distribution of the DNS in the four-dimensional initial conditions space
$\{\mathsf {R},\mathsf {B},\mathsf {S},\mathsf {D}\}$
. Every marker corresponds to a simulation and it is coloured according to its value of
$\mathsf {R}$
(red corresponds to a high value and purple to a low one). The grey areas correspond to infeasible DNS due to constraints and precautions taken to ensure a good numerical resolution. The blue and yellow star symbols indicate the parameters for the diffusive and inertial simulations detailed in § 2.3. The rectangles delimit the domain where the global sensitivity analysis is performed.

Using the linear stability analysis of a diffuse interface driven by the RTI, it is possible to express
$\gamma _0^{\star}$
, the non-dimensionalized exponential growth rate corresponding to the mean wavenumber
$k_0$
, as a function of the parameters
$\mathsf {R}$
and
$\mathsf {D}$
\begin{equation} \gamma ^{\star} _0(\mathsf {R},\mathsf {D})=\frac {\nu ^{1/3}}{(\mathcal A g )^{2/3}} \gamma _0= \frac {1}{\mathsf {R}^{4/3}}\left (\left ( \frac {\mathsf {R}^2}{ \psi (\mathsf {D})} +1 \right )^{1/2}-2 \right ) \ \text {with} \ \psi (\mathsf {D})= 1+\frac {2}{3} \frac {\sqrt {2}}{\pi } \mathsf {D}. \end{equation}
Equation (2.6) is obtained using the results of Duff et al. (Reference Duff, Harlow and Hirt1962) in the low-Atwood-number limit and taking
$1/\| \mathrm {d} C/\mathrm {d} z \|_{ {max}}=2 \delta /3$
for the diffusion thickness. Also from (2.6), one recovers the inertial limit for
$\mathsf {R} \gg 1$
and
$\mathsf {D} \ll 1$
leading to the classical inviscid growth rate,
$\gamma _0=\sqrt {\mathcal A g k_0}$
. In contrast, the RTI is stabilised for
$\mathsf {R} \ll 1$
, with a damping rate of
$\gamma _0=-\nu k_0^2$
. While perturbations decay in this regime, the layer still evolves through diffusion, with growth governed by
$\gamma _D=\kappa \delta ^2=\nu /\delta ^2$
, or in non-dimensional form,
$\gamma _D ^{\star} =\mathsf {D}^{-2} \mathsf { R}^{-4/3}$
.
In figure 3, we illustrate the distribution of initial conditions in the DNS database across inertial and diffusive regimes, depending on whether
$\gamma _0^{\star}$
is larger or smaller than
$\gamma _D ^{\star}$
, respectively.
(a) Linear growth rate
$\gamma ^{\star} _0$
and (b) stability diagram of the mean wavenumber
$k_0$
for all simulations. Circular markers (
$\circ$
) correspond to simulations in which there is at least one unstable wavenumber, whereas downward triangles (
$\triangledown$
) correspond to simulations where all wavenumbers in the initial spectral band are stable. The symbol colour is related to the diffusive thickness
$\mathsf {D}$
(in panel a), or to the parameter
$\mathsf {B}$
(in panel b) expressing if the simulation is narrow or large bandwidth. The grey region (in panel b) indicates the diffusive region where
$\gamma ^{\star} _D \gt \gamma ^{\star} _0$
. The two blue and yellow
${\star}$
symbols indicate the parameters for the diffusive and inertial simulations discussed in § 2.3. The rectangles define the hyper-cubic domain
$\mathcal I$
where is performed the global sensitivity analysis presented in § 3.

At large Reynolds numbers (
$\mathsf {R} \gt 10$
), the linear growth rate scaling follows the inertial limit, with
$\gamma ^{\star} _0 = \mathsf {R}^{-1/3}$
. However, diffusion can still play a role in certain limits: when
$\mathsf {D} \ll 1$
, it thickens very sharp interfaces and when
$\mathsf {D} \gg 1$
, it corresponds to a large diffuse layer where the Rayleigh–Taylor instability is stabilised. Similarly, at low
$\mathsf {R}$
, a significant number of simulations fall within the diffusive region, where the mean perturbation is damped. Depending on
$\mathsf {B}$
, some or all modes may be linearly stable. The remaining simulations are distributed in the intermediate range, between the neutral stability curve and the maximum growth rate, as shown in figure 3.
It is also important to note that this analysis does not consider the steepness parameter
$\mathsf {S}$
, which, when large, indicates significant nonlinear effects. Additionally, the perturbation, characterised by
$\mathsf {S}$
, artificially increases the effective diffusion thickness, leading to further damping of the linear growth rate.
2.3. The dynamics of the 0-D quantities
To examine the dynamics of the RTI, we focus on the 0-D turbulent quantities derived from averaging across the entire width of the mixing layer, which are commonly used in mix models.
Given that the RTI is statistically homogeneous in the horizontal
$(x,y)$
plane, it is practical to introduce the horizontal average
$\overline {Q}$
and the fluctuation
$q$
of a quantity
$Q$
, as determined by the Reynolds decomposition
$Q=\overline {Q}+q$
. Within the context of the Boussinesq approximation, the RTI exhibits a zero mean velocity,
$\overline {\textbf {u}}=0$
. Moreover, the concentration field can be expressed as
$C(\textbf {x},t)=\overline {C}(z,t)+c(\textbf {x},t)$
, allowing the definition of the mixing zone size as (Andrews & Spalding Reference Andrews and Spalding1990)
This definition for
$L$
, (2.7), derives from a piecewise mean concentration profile that accurately represents a turbulent Rayleigh–Taylor mixing zone. Notably, for a diffuse interface without perturbation,
$L$
is equivalent to
$\delta$
from (2.2).
To more accurately characterise the velocity field
$\textbf {u}$
, this study considers the 0-D turbulent kinetic energy
$\mathcal K$
and its dissipation rate
$\varepsilon$
Similarly, concentration fluctuations
$c$
are captured by the scalar variance
$\mathcal K_{cc}$
and the scalar dissipation rate
$\varepsilon _{cc}$
,
Additionally, the global mixing parameter
$\Theta$
is considered as an alternative representation to
$\mathcal K_{cc}$
(Young et al. Reference Young, Tufo, Dubey and Rosner2001)
\begin{equation} \Theta =\frac {\int _{-\infty }^{+\infty } \overline {C (1-C)}\,\mathrm {d} z}{\int _{-\infty }^{+\infty } \overline {C} (1-\overline {C})\,\mathrm {d} z}=1-6 \mathcal K_{cc}. \end{equation}
The system is further characterised by the vertical concentration flux
$\mathcal F$
, which relates to the injection of potential energy. The growth rate of the mixing layer and the elongation of density structures along the vertical direction are quantified by the self-similar coefficient
$\alpha$
and the dimensionality parameter
$\sin ^2(\Gamma )$
, respectively
where
$\mathcal {E}_{cc}(\theta ,t)$
is the concentration covariance spectra in spherical coordinates integrated along the radial and azimuthal coordinates. The dimensionality parameter,
$\sin ^2(\Gamma )$
, introduced by Gréa (Reference Gréa2013), measures the anisotropy of the density structures, well elongated along the vertical direction in RT flows. From its definition and for an axisymmetric flows, it equals 1 for pure vertical density structures and
$2/3$
when it is isotropic. Together with the global mixing parameter
$\Theta$
, they indicate the intensity of the buoyancy forces, driving the dynamics of the mixing layer.
The 0-D quantities are interconnected, as illustrated by the budget equations for kinetic energy and scalar variance (derived, for instance, by Briard et al. Reference Briard, Gréa and Nguyen2024),
Noticeably, while the budget equation for kinetic energy, (2.9a ), is exact, the equation for the scalar, (2.9b ), relies on a uniform mean concentration gradient inside the mixing layer.
The mixing zone size
$L$
, all the other 0-D quantities
$\dot {L}, \mathcal K$
,
$\varepsilon$
,
$\Theta$
,
$\varepsilon _{cc}$
,
$\mathcal F, \sin ^2(\Gamma )$
and
$\alpha$
are computed in the various DNS simulations at all times. To investigate the RTI phenomenology and to shed light on the transition to turbulence, it is important to work with non-dimensional state variables. This is done by using the viscosity
$\nu$
and the buoyancy acceleration
$\mathcal {A}g$
to rescale the time and the length. The non-dimensional quantities are thus marked with a star
${\star}$
symbol such that, for instance,
$t^{\star} =t (\mathcal A g)^{2/3}/\nu ^{1/3}$
and
$L^{\star} =L (\mathcal A g)^{1/3}/\nu ^{2/3}$
.
The Rayleigh–Taylor simulations are stopped either when the mixing zone width reaches half of the vertical computational domain (i.e.
$L \geqslant 2 \pi$
) or when the integral scale of turbulence becomes too large. This latter condition is defined from the 0-D quantities as
$\ell _I=\mathcal {K}^{3/2}/\varepsilon \gt 1.15$
or, equivalently, in term of wavenumber
$k_I=2 \pi / \ell _I \lt 5.4$
. In practice, it is mainly the criterion on the integral scale that determines the end of the simulations. This procedure aims at reducing the risk of the lateral confinement of the large eddies which would alter the growth of the mixing layer.
Time evolution of the 0-D quantities
$L$
,
$\dot L$
,
$\alpha ,$
$\mathcal K$
,
$\varepsilon$
, the turbulent Reynolds number
$Re_T$
,
$\Theta$
,
$\varepsilon _{cc}$
and
$\sin ^2 \Gamma$
evaluated for all the RT DNS. The trajectories are coloured as a function of the initial Reynolds number
$\mathsf {R}$
. The dashed curves evaluate the reconstruction of
$\dot L ^{\star}$
using
$-12 \mathcal F^{\star}$
.

By differentiating (2.7) and assuming a linear mean concentration profile, we derive that
$\dot L^{\star} =-12 \mathcal F ^{\star}$
. This relationship indicates that the vertical flux represents the injection rate of potential energy into the system. This relationship is consistently confirmed by our simulations, leading us to preferentially present
$\dot L ^{\star}$
in our results to avoid redundancy.
We now discuss the dynamics of the 0-D quantities computed from the DNS that are presented in figure 4. The fact that the simulations end at different values of
$t^{\star}$
is not only the consequence of the stop criterion detailed before, but also to the non-dimensionalisation procedure as inertial trajectories with small initial Reynolds number, for instance, are run with a small buoyancy acceleration
$\mathcal A g$
.
A large dispersion of the mixing zone width is observed among the DNS due to the initial conditions. Figure 4 highlights the importance of the Reynolds number which, as expected, mainly selects between inertial or diffusive behaviours. The diffusion
$\mathsf {D}$
, the steepness
$\mathsf {S}$
and the spectral bandwidth
$\mathsf {B}$
also play a role, which will be investigated later in § 3.
Before examining how the trajectories differ, it is worth noting that the classical Reynolds number, based on the mixing width, can be calculated using
$L^{\star} \dot L^{\star}$
with values exceeding
$10^4$
by the end of the simulations. Similarly, the turbulent Reynolds number
$Re_T$
can be derived from the formula
$(\mathcal K ^{\star} )^2/\varepsilon ^{\star}$
, with values rising above
$10^3$
. This shows that the mixing zone has already transitioned to turbulence and developed an inertial range. We even remark that most of the DNS reach the self-similar regime. This is characterised by a quadratic growth in time for
$L^{\star}$
and
$\mathcal K ^{\star}$
, a linear growth for
$\dot L^{\star}$
,
$-\mathcal F^{\star}$
and
$\varepsilon ^{\star}$
, and a constant for the global mixing and dimensionality parameters
$\Theta$
and
$\sin ^2(\Gamma )$
, the coefficient
$\alpha$
and the scalar dissipation rate
$\varepsilon ^{\star} _{cc}$
. In particular, the values for the self-similar coefficient and for the global mixing parameter collected at the end of the simulations are
$\alpha =0.019 \pm 0.002$
and
$\Theta =0.81 \pm 0.01$
, which are consistent with the usual DNS results. This is also well consistent with a self-similar regime driven by a mode-coupling phenomenology as explained by Dimonte (Reference Dimonte2004). However, depending on how large is the initial bandwidth
$\mathsf B$
, mode competition effects can also take place during the transients.
We now compare more specifically an inertial and a diffusive DNS in figure 5.
(a) Two-dimensional slices at different times of the concentration field
$C$
extracted from a DNS corresponding to an inertial configuration with
$\mathsf {R}=12.5$
,
$\mathsf {B}=0.14$
,
$\mathsf {S}=3.0$
and
$\mathsf {D}=3.0$
. (c) Same as in panel (a) but for a diffusive case with
$\mathsf {R}=2.35$
,
$\mathsf {B}=0.14$
,
$\mathsf {S}=3.0$
and
$\mathsf {D}=3.0$
. (b) Time evolution of the 0-D quantities for the inertial and diffusive trajectories. The star symbols indicate the times of the 2-D concentration slices presented in the figure. The end of the simulations corresponds to the stop criterion detailed in § 2.3. The dashed curves evaluate the reconstruction of
$\dot L ^{\star}$
using
$-12 \mathcal F^{\star}$
.

Renormalised velocity and scalar spectra for the inertial and diffusive trajectories corresponding to figure 5 and for times
$t^{\star} \in \{0;12;23;36;75;109\}$
(same colours). Note that at
$t^{\star} =0$
, the velocity spectra is zero as we initially consider a pure interface perturbation.

The diffusive trajectory having a smaller linear growth, even negative for some modes, the convergence towards the self-similar regime is therefore delayed with respect to the inertial configuration. The transition to turbulence is very rapid in the inertial case, occurring around the peak of
$\varepsilon ^{\star}$
at
$t^{\star} =20$
with a turbulent Reynolds number
$Re_T \approx 100$
. This transition, associated with the secondary instabilities between bubbles and spikes, also corresponds to an increase of the global mixing parameter
$\Theta$
and to a diminution of the mixing growth exponent
$\alpha$
. By contrast, it seems more progressive and delayed in the diffusive configuration where it becomes less convenient to define a clear transition time.
The convergence towards the self-similar regime can also be clearly observed for both inertial and diffusive trajectories in the renormalised energy and scalar spectra shown in figure 6. The initial imprints of the annular condition and interface thickness gradually give way to the classical inertial spectra, characterised by a
$k^{-5/3}$
scaling law in the inertial range. During this process, larger perturbation modes are generated through backscattering effects, allowing unstable Rayleigh–Taylor modes to develop even in the diffusive configuration. The collapse of the renormalised spectra at large scales confirms that the self-similar regime is well established.
Note that a relationship between the growth of the mixing layer and the parameters for mixing in the nonlinear regime has been proposed by Gréa (Reference Gréa2013):
where
$\sin ^2(\Gamma ) =1$
for purely vertical structures. This result has been confirmed with an accuracy of up to
$10 \,\%$
in our DNS database, particularly for
$t^{\star} \geqslant 80$
, as illustrated in figure 7. This relation also yields
$\alpha =0$
for
$\Theta =1$
, consistent with the absence of buoyancy forces when there are no density fluctuations across the layer. Although this theory does not directly compute the growth rate coefficient
$\alpha$
as defined by (2.8d
), it provides valuable insight into how zero-dimensional quantities are interconnected in Rayleigh–Taylor flows.
Additionally, figure 5 demonstrates that the relationship
$\dot L^{\star} =-12 \mathcal F ^{\star}$
is consistently verified in our simulations, as evidenced by the near-perfect superposition of the curves.
2.4. A physics-informed surrogate model for the RT simulations
Although the DNS contain ample information for exploring the memory of the RTI, their significant computational demands render them impractical for analyses requiring extensive evaluations. To make possible a global sensitivity analysis and the resolution of Bayesian inverse problems, it becomes advantageous to develop a surrogate model of our DNS database.
This surrogate model must be sufficiently precise to ensure the validity of our conclusions. Given the voluminous DNS data and the problem’s high degree of nonlinearity, neural networks (NNs) appear particularly wellsuited for this purpose. In recent years, NNs have been employed in numerous turbulent flow studies (Ling, Kurzawski & Templeton Reference Ling, Kurzawski and Templeton2016; Duraisamy, Iaccarino & Xiao Reference Duraisamy, Iaccarino and Xiao2019; Brunton, Noack & Koumoutsakos Reference Brunton, Noack and Koumoutsakos2020; Guastoni et al. Reference Guastoni, Güemes, Ianiro, Discetti, Schlatter, Azizpour and Vinuesa2021; Buaria & Sreenivasan Reference Buaria and Sreenivasan2023; Solera-Rico et al. Reference Solera-Rico, Sanmiguel Vila, Gómez-López, Wang, Almashjary, Dawson and Vinuesa2024; Zhu et al. Reference Zhu, Jiang, Lefauve, Kerswell and Linden2024), demonstrating their capacity to tackle complex fluid mechanics challenges. In this study, the NN surrogate model aids in elucidating the impact of initial conditions, pinpointing critical moments of the transition to turbulence and identifying key quantities for monitoring.
Accordingly, the NN surrogate model maps the time
$t^{\star}$
and the initial conditions
$\mathsf {I}=(\mathsf {R},\mathsf {B},\mathsf {S},\mathsf {D})^T$
to the 0-D quantities
$\mathsf {Q}=(L^{\star} ,\dot {L}^{\star} ,\mathcal {K}^{\star} ,\varepsilon ^{\star} ,\Theta ,\varepsilon _{cc}^{\star} ,\mathcal {F}^{\star} ,\sin ^2(\Gamma ))^T$
as outputs (see figure 8). This scenario is relatively straightforward since the DNS trajectories are distinctly defined by four dimensionless initial parameters. A simple multi-layer perceptron architecture suffices for creating an initial model that accurately links the five inputs to the seven outputs (with
$\dot L^{\star}$
computed through auto-differentiation). This architecture is crafted to reinforce the self-similar scaling laws of various 0-D quantities for large
$t^{\star}$
values (see Appendix A for more details).
Depiction of the physics-informed neural network as a surrogate for modelling the DNS’s 0-D quantities. The total loss consists of a data-driven component,
$\mathcal J _{{Data}}$
, aiming to minimise the discrepancy between the neural network and DNS outputs, and a physics-informed component,
$\mathcal J_{{PI}}$
, formulated via automatic differentiation and assessed at new collocation points. This approach enhances the conservation of kinetic energy and scalar variance balances, enforces a constraint on
$\alpha$
, ensures
$\dot L^{\star}$
remains positive and upholds the consistency of
$L_0 ^{\star}$
.

Relative error of the surrogate model’s outputs, calculated at various times across the 73 test DNS in the dataset, is displayed in a boxplot format. Due to many simulations not reaching longer durations, the number of simulations available to evaluate the error diminishes over time, as indicated in the figure.

To improve the quality and the robustness of the surrogate model, we try to enforce several physical properties through unsupervised terms in the loss function (see figure 8 and Appendix A). The conservation of the kinetic and potential energy budgets, (2.9a
), (2.9b
), are imposed by the loss terms
$\mathcal J_{uu}$
and
$\mathcal J_{cc}$
. From (2.10) and knowing that
$2/3 \leqslant \sin ^2 (\Gamma ) \leqslant 1$
, we consider
$\mathcal J_{ss}$
for
$t^{\star} \geqslant 150$
to express a self-similar constraint between
$\alpha$
and
$\Theta$
:
Note that this loose inequality, corresponding to the grey region in figure 7, is well verified in all simulations as early as
$t^{\star} \geqslant 80$
. It primarily influences the early stages of training, helping to accelerate the convergence of the neural network, as shown in Appendix A. It is important to emphasise that this constraint does not determine the self-similar value of
$\alpha$
(taking
$\Theta =0.8$
, for instance, leads to
$ 0.15\leqslant \alpha \leqslant 0.33$
). The condition
$\mathcal J_{\dot L^{\star} }$
corresponds to
$\dot {L}^{\star} \geqslant 0$
as the acceleration is always destabilising. Similarly to the constraint
$\mathcal J_{ss}$
, it mainly influences the early stages of neural network training. The last relation
$\mathcal J_{L_0^{\star} }$
aims at recovering
$L_0^{\star} =3/2\mathsf {R}\,\partial L_0^{\star} /\partial \mathsf {R}$
. This comes from dimensional analysis as the initial mixing zone width
$L_0$
does not depend on the acceleration (or
$\mathsf {R}$
). Consequently, the NN emerges from a physics-informed machine-learning methodology (Karniadakis et al. Reference Karniadakis, Kevrekidis, Lu, Perdikaris, Wang and Yang2021).
In figure 9, we provide the relative error of the surrogate model’s 0-D quantities. It shows that the surrogate model is accurate up to
$5\,\%$
across the 73 test DNS which were not used for the training. Incorporating these physics-informed properties not only bolsters the model’s resilience – especially in data-sparse regions – but also boosts confidence in the model as it aligns more closely with fundamental physics principles, offering credible solutions.
3. Sensitivity analysis
In this section, we aim to assess how the initial parameters
$\mathsf {I}=(\mathsf {R},\mathsf {B},\mathsf {S},\mathsf {D})^T$
influence the 0-D quantities. To achieve this, we use the surrogate model introduced in § 2.4 to rapidly evaluate the dynamics and quantify the sensitivity to the initial conditions. This global approach involves sampling the initial conditions across an extended region of the initial parameter space
$\mathcal I$
, as defined below (also see figures 2 and 3):
The domain thus encompasses the initial conditions for the DNS databases. It is crucial to limit the extent of extrapolation to maintain confidence in the surrogate model. Within this domain,
$\mathsf {I}$
is sampled using a joint prior distribution,
$p(\mathsf {I})$
, which is uniform for
$\mathsf {R}$
and
$\mathsf {B}$
, and log-uniform for
$\mathsf {S}$
and
$\mathsf {D}$
. This ensures coverage of both inertial and diffusive trajectories. Initially, we present a comprehensive overview of the dependency on
$\mathsf {I}$
, subsequently focusing more specifically on the late time self-similar dynamics. The final part is dedicated to the conditional sensitivity analysis at a given initial mixing zone width
$L^{\star} _0$
. Additional sensitivity analyses focusing on the diffusive and inertial regimes have also been conducted and are detailed by Thévenin (Reference Thévenin2024). These analyses provide insight into the influence of initial conditions across different instability regimes.
3.1. Main effect of the initial conditions
Evolution and dispersion of the 0-D trajectories
$q (t^{\star} ;\mathsf I \in \mathcal I)$
with initial conditions following the joint prior distribution in the extended domain
$\mathcal I$
. The mean trajectories
$E[q]$
are represented by the plain line and the dashed line is the trajectory corresponding to the mean of the initial conditions
$q (t^{\star} ;E[\mathsf I])$
. The shaded areas indicate the zones delimited by
$\pm \sqrt {V[q]}$
or the
$95 \,\%$
confidence interval.

Evolution of the variances for the 0-D trajectories,
$V[q(t^{\star} ;\mathsf I \in \mathcal I)]$
, with initial conditions sampled with the prior distribution in the extended domain
$\mathcal I$
. The first-order Sobol indices
$s_i^q$
are computed to disentangle the initial condition effects. Their respective contributions are indicated by the coloured areas of width
$s_i^q V[q(t^{\star} ;\mathsf I)]$
in the figure.

Figures 10 and 11 detail how the various 0-D quantities
$q$
, parametrised by the initial conditions
$\mathsf {I}$
, evolve and vary. A considerable dispersion of trajectories is observed, especially for the mixing width
$L^{\star}$
. This highlights the challenge in modelling Rayleigh–Taylor turbulence, which remains highly sensitive to the unknown initial conditions. The variance of the mixing zone width
$V[L^{\star} ]$
– represented by the shaded area width accompanying the trajectories – is primarily driven by the diversity in initial mixing zone width values,
$L^{\star} _0$
. However, the pronounced variance is chiefly attributed to the linear growth rates, which are highly sensitive to
$\mathsf {I}$
, as indicated by (2.6). Therefore, the variance exponentially grows until the transition to turbulence at approximately
$t^{\star} \in \ [25, \ 50 ]$
. When entering the self-similar regime at late time, the variance is still growing accompanying the mixing zone evolution. By contrast, the variance of
$\dot L^{\star}$
remains very small in the self-similar regime, echoing a nearly constant value for the self-similar parameter
$\alpha$
at late time shown in figure 10. This pattern aligns with (2.10), as the mixing and dimensionality parameters,
$\Theta$
and
$\sin ^2(\Gamma )$
, also tend towards constant values. The variance of the vertical concentration flux
$V[\mathcal {F}^{\star} ]$
(not shown) mirrors the trend of
$V[\dot {L}^{\star} ]$
since both quantities are nearly proportional within the 0-D equations. The kinetic energy variance
$V[\mathcal {K}^{\star} ]$
is anticipated to follow a scaling law similar to
$V[(\dot {L}^{\star} )^2]$
. Nonetheless, a significant variation occurs at approximately
$t^{\star} =25$
, aligning with the onset of turbulence. This assertion is corroborated by the peak in the dissipation rate variance
$V[\varepsilon ^{\star} ]$
, coinciding with the emergence of secondary Kelvin–Helmholtz instabilities. With the scalar dissipation rate
$\varepsilon ^{\star} _{cc}$
diminishing as
$1/t^{\star}$
in the self-similar regime, its variance also rapidly decreases. Due to the nonlinear dependency on
$\mathsf {I}$
, generally,
$E[q(t^{\star} ;\mathsf {I})] \neq q(t^{\star} ;E[\mathsf {I}])$
.
To pinpoint the initial conditions’ influence, we calculate the temporal evolution of the first-order Sobol indices
$s^q_i$
, also known as the main effects, with
$i \in \{\mathsf {R},\mathsf {B},\mathsf {S},\mathsf {D}\}$
which, for a quantity
$q$
, are defined as (see Appendix B)
\begin{equation} s^q_i=\frac {V [ E \left [q|i \right ] ]}{V \left [q \right ]}. \end{equation}
The Sobol indices thus depend on the joint prior distribution
$p(\mathsf {I})$
taken in the extended domain
$\mathcal I$
, (3.1). A large Sobol index
$s^q_i$
indicates an important sensitivity to the parameter
$i$
. It can be verified that
$\Sigma _i s_i^q \leqslant 1$
, with equality when initial conditions effects are decoupled. In figure 11, we see that the variance of each 0-D quantity is correctly captured by the first-order Sobol indices. The coupling between the parameters, expressed by the values of higher-order Sobol indices, becomes more significant around the transition to turbulence,
$t^{\star} \approx 25$
. It comes as no surprise that the transition regime is the most complex phase to model the RT dynamics. Noticeably, the 0-D quantities are very sensitive to the Reynolds number
$\mathsf {R}$
mainly controlling the growth of the layer. The steepness
$\mathsf {S}$
has also its importance, in particular, to explain the variability of
$L^{\star}$
. The diffusive parameter
$\mathsf {D}$
also plays a role in the global mixing parameter
$\Theta$
variability, in particular, during the first stage of the instability. This is of course expected as a very diffused layer leads to higher
$\Theta$
. Consequently, it mainly appears in the initial
$\alpha$
variance from (2.10). The parameter
$\mathsf {B}$
expressing the bandwidth seems to have a limited effect on the variability of the 0-D quantities in the domain we considered. However and according to linear stability analysis, for an initial condition having a mean wavenumber growth rate around zero, a larger bandwidth has more chances to have at least one unstable mode, which leads to a very different dynamics. So the parameter
$\mathsf {B}$
has a strong influence in this case, in a reduced domain at low
$\mathsf {R}$
(see also Thévenin Reference Thévenin2024).
3.2. Late time sensitivity to initial conditions
Probability density functions (p.d.f.s) for
$\alpha _\infty$
and
$t_\infty ^{\star}$
using the surrogate model, with initial conditions
$\mathcal {I}$
sampled in the extended domain (3.1). The mean and the variance are indicated in the legend. The first-order Sobol indices
$s_i^{t_\infty ^{\star} }$
are also plotted for
$t_\infty ^{\star}$
. Both the standard deviation
$\mathrm{Std}$
of the p.d.f.s and the surrogate error
$\epsilon _{NN}$
are indicated in the figure.

At late times, trajectories are expected to follow self-similarity, and hence are solutions to a buoyancy-drag equation, as demonstrated in previous studies (Fermi & von Neumann 1951; Ramshaw Reference Ramshaw1998; Ristorcelli & Clark Reference Ristorcelli and Clark2004; Cook & Cabot Reference Cook and Cabot2006; Boffetta, Mazzino & Musacchio Reference Boffetta, Mazzino and Musacchio2011):
where
$\alpha _\infty$
denotes an asymptotic growth rate and
$t^{\star} _\infty$
a virtual origin that aligns self-similar trajectories. We may compute the values of
$\alpha _\infty$
and
$t^{\star} _\infty$
by fitting (3.3) over a late time interval. We consider, for instance,
$t^{\star} \in [160, \ 170]$
, such that the DNS trajectories are in the self-similar regime while the surrogate model is not too much extrapolating (the results shown hereafter are not sensitive to this interval choice). By sampling again the initial conditions on the extended domain
$\mathcal I$
(3.1), we obtain the p.d.f. for
$\alpha _\infty$
and
$t^{\star} _\infty$
in figure 12.
The
$\alpha _\infty$
values cluster tightly around their mean
$E[\alpha _\infty ] = 0.020$
, with minimal variance within
$\alpha _\infty \in [0.019, 0.021]$
, aligning with findings from DNS analyses (see § 2.3). Although the surrogate model constrains the mixing zone size
$L^{\star}$
to achieve asymptotic
$(t^{\star} )^2$
growth, the final value of
$\alpha _\infty$
is not dictated by these imposed bounds, which are relatively broad in practice (
$[0.016, \ 0.033]$
for
$\Theta =0.8$
). Hence, the consistent value of
$\alpha _\infty$
emerges not from an artificial constraint on the surrogate, but from the intrinsic dynamics of the RTI trajectories. Furthermore, the dispersion observed in the
$\alpha$
-p.d.f. of figure 12 corresponds roughly to the approximation error of the surrogate model, as depicted in figure 9. The sensitivity analysis clearly supports the idea of a universal growth rate
$\alpha _\infty$
in unconfined self-similar RTI mixing zones. This finding aligns with the early theoretical work of Soulard et al. (Reference Soulard, Griffond and Gréa2015), which suggests that
$\alpha$
should be lower than
$0.025$
.
Conversely,
$t^{\star} _\infty$
exhibits a broader dispersion, ranging from
$t^{\star} _\infty \in [10, 90]$
with a mean
$E[t_\infty ^{\star} ] = 48$
. Thus, in the self-similar regime,
$t^{\star} _\infty$
variations predominantly govern
$V[L^{\star} ]$
variance. This hypothesis is corroborated by using (3.3) and the p.d.f. for
$\alpha _\infty$
and
$t^{\star} _\infty$
to reconstruct the variance for
$L^{\star}$
. The results are presented in figure 13.
Time evolution of
$V[L^{\star} ]$
, evaluated using the neural network surrogate model with initial conditions sampled within the extended domain
$\mathcal {I}$
(solid line). The variance, reconstructed from the buoyancy-drag (BD) solution (3.3), appears as dashed lines, further broken down into contributions from variations in
$\alpha _\infty$
and
$t_\infty$
.

At late times, the reconstructed variance aligns with the true variance of
$L^{\star}$
. Decomposing (3.3) reveals that
$V[L^{\star} ] \approx (t^{\star} )^2 V[\alpha _\infty t^{\star} _\infty ]$
. Variations in
$\alpha _\infty$
, potentially causing a variance growth proportional to
$(t^{\star} )^4$
, exert minimal influence on trajectory dispersion within this time frame. Therefore, the mixing zone width variance,
$V[L^{\star} ]$
, is primarily determined by the variations in
$t^{\star} _\infty$
during the early transition regimes.
Values of
$\alpha _\infty$
and
$t^{\star} _\infty$
sampled using a log uniform distribution of the initial parameters
$\mathcal I$
. The symbols associated with the trajectories are coloured by the steepness
$\mathsf {S}$
.

In figure 12, we also analyse
$t^{\star} _\infty$
’s dependency on
$\mathsf {I}$
through Sobol indices. Given the negligible variance in
$\alpha _\infty$
, possibly attributable to the surrogate model’s precision, we abstain from inferring physical significance from these fluctuations. In addition, at late time, the turbulent mixing zone growth is determined by the very anisotropic structures at large scales. Therefore, the viscosity
$\nu$
, or equivalently
$\mathsf {R}$
, should not play an important role in the self-similar dynamics. This explains why a constant
$\alpha _\infty$
is expected and that
$t^{\star} _\infty \propto \mathsf {R} ^{1/3}$
from dimensional analysis, as evidenced by the large value of
$s_{\mathsf {R}}^{t^{\star} _\infty }$
. In figure 14, we present the sampled characteristics of the asymptotic regimes. The distributions of
$\alpha _\infty$
and
$t^{\star} _\infty \mathsf {R} ^{-1/3}$
have a very small dependence on
$\mathsf {R}$
confirming that the viscosity does not play any role in the late-time regime. Still we observe a small dependence on
$\mathsf R$
, which can be attributed to the diffusive trajectories. Moreover, the large Sobol indices
$s_{\mathsf {S}}^{t^{\star} _\infty }$
shed light on the importance of the steepness
$\mathsf {S}$
on the late-time regime. This trend is more explicitly observed in figure 14. Therefore,
$t^{\star} _\infty$
sensitively increases with
$\mathsf {S}$
. In addition, the results show some dependence on the bandwidth
$\mathsf {B}$
and to a lesser degree on the diffusion thickness number
$\mathsf {D}$
. We thus propose a linear fit for
$\alpha _\infty$
and
$t^{\star} _\infty$
, which allows to evaluate the impact of the initial default of an interface on the late-time dynamics. This leads to
\begin{equation} \left \{ \begin{array}{c} \alpha _\infty \simeq 0.021+\left (0.3 \,\mathsf {R} -4.6\, \mathsf {B} -3.26 \, \ln (\mathsf {S}) -3.73\, \ln ( \mathsf {D})\right )\times 10^{-4}, \\[5pt] t^{\star} _\infty \simeq \mathsf {R}^{1/3} (14.09 -0.05 \, \mathsf {R} + 2.52 \,\mathsf {B}+3.1 \, \ln (\mathsf {S}) -0.08 \,\ln (\mathsf {D})). \end{array} \right . \end{equation}
Equation (3.4) confirms the importance of
$\mathsf {R}$
,
$\mathsf {S}$
and to a lesser degree
$\mathsf {D}$
. The strong coefficient associated with
$\mathsf {B}$
has to be contrasted with its domain of definition,
$\mathsf {B}\in [0, \ 2]$
, so it has a weak effect on the self-similar dynamics. The virtual origin derives from the early RT dynamics, including the transition to turbulence. It is thus larger when the initial mixing layer growth is important. The effect of
$\mathsf {S}$
cannot be provided by the linear theory, as it is only valid in the limit
$\mathsf {S} \rightarrow 0$
. Still, we can consider the rapid acceleration model exhibiting a buoyancy-drag equation parametrised by the turbulence quantities (Gréa Reference Gréa2013). Initially, the growth is mainly controlled by the buoyancy coefficient, which is provided by the global mixing parameter and the dimensionality coefficient as
$\sin ^2 (\Gamma ) (1-\Theta )$
. A large steepness
$\mathsf {S}$
will generate more elongated structure (or large
$\sin ^2 (\Gamma )$
) and less mixing (smaller
$\Theta$
). It results in strong buoyancy effects and therefore larger
$t^{\star} _\infty$
in the self-similar regime. Conversely, the dependence of the late time
$t^{\star} _\infty$
to
$\mathsf S$
and
$\mathsf B$
can also be related to the presence of mode competition phenomenology during the transient (Dimonte Reference Dimonte2004). In addition, a large steepness parameter
$\mathsf S$
reduces the transition time, i.e. the time at which a mode saturates and reaches its terminal velocity, leading to larger
$t_\infty$
.
3.3.
Sensitivity analysis conditioned at a given
$L^{\star} _0$
Thus far, our investigation has covered the global effects of the initial parameters
$\mathsf I$
on RT dynamics. While the initial conditions are often unknown, the initial mixing zone width
$L^{\star} _0$
is usually measurable. Therefore, conducting a sensitivity analysis at a specific
$L_0^{\star}$
is worthwhile.
First, it is beneficial to determine how
$L^{\star} _0$
relates to the initial conditions. Through various approximations, we derive
with
$F =k_0 L_0 =\mathsf {D}$
when
$\mathsf {S} \ll \mathsf {D}$
and
$F\simeq 3 \,\mathsf {S}$
when
$\mathsf {S} \gg \mathsf {D}$
. We also verified that
$L^{\star} _0$
was weakly dependent on
$\mathsf {B}$
.
Conditional and unconditional probability density functions for (a)
$\alpha _\infty$
and (b)
$t^{\star} _\infty$
given
$L^{\star} _0$
. The initial conditions
$\mathcal I$
are sampled using the joint prior distribution. The plain lines show the evolutions of
$E[\alpha _\infty |L_0^{\star} ]$
and
$E[t^{\star} _\infty |L_0^{\star} ]$
.

We present in figure 15 the conditional p.d.f. of
$\alpha _\infty$
and
$t^{\star} _\infty$
characterising the late time dynamics at given
$L^{\star} _0$
. Compared with the unconditional p.d.f. presented in figure 12, the knowledge of
$L^{\star} _0$
reduces the uncertainties while not completely eliminating them. In addition, we observe a very small dependence of
$E[\alpha _\infty |L_0^{\star} ]$
to
$L^{\star} _0$
, while the dependence of
$E[t^{\star} _\infty |L_0^{\star} ]$
to
$L^{\star} _0$
is more pronounced, which can be interpreted using (3.5). For small
$L^{\star} _0$
, the dispersion on
$\mathsf {R}$
and
$\mathsf {S}$
is reduced such that the dispersion on
$\alpha _\infty$
and
$t^{\star} _\infty$
are limited compared with larger
$L^{\star} _0$
. As
$L^{\star} _0$
grows, so do
$\mathsf {R}$
and
$\mathsf {S}$
such that we observe a slight diminution of
$\alpha _\infty$
and growth of
$t^{\star} _\infty$
, as detailed by (3.4). Therefore, assuming
$\mathsf S$
,
$\mathsf D$
and
$\mathsf B$
are fixed, then we can expect
$t^{\star} _\infty \sim (L_0^{\star} )^{1/2}$
(valid for inertial trajectories).
4. Inferring the initial conditions of Rayleigh–Taylor turbulence
We have explored how 0-D quantities respond to the initial conditions
$\mathsf I$
, particularly highlighting the self-similar regime. The enduring impact of the initial conditions on late-time dynamics serves as a memory of Rayleigh–Taylor turbulence. We continue this analysis by identifying the initial conditions
$\mathsf {I}$
and the time
$t^{\star}$
at which 0-D quantities are measured. Our goal is to assess how well these 0-D quantities retain the memory of Rayleigh–Taylor turbulence and their predictive capability regarding the mixing zone’s future dynamics. We frame the issue within the context of classical Bayesian inference, subsequently examining how the memory of initial conditions diminishes across successive RTI stages for two sets of 0-D variables.
4.1. Problem formulation
Here, we assume we have knowledge of a 0-D vector
$\mathsf {Q}$
(for instance, with two variables
$\mathsf {Q}_2=(L^{\star} ,\dot L^{\star} )^T$
) and seek to find the initial conditions
$\mathsf {I}$
and time
$t^{\star}$
corresponding to this measurement. In a Bayesian formulation, this involves computing the joint posterior distribution
$p(\mathsf {I},t^{\star} |\mathsf {Q})$
which, according to Bayes rule, can be expressed as
In (4.1),
$p(\mathsf {Q}|\mathsf {I},t^{\star} )$
classically defines the likelihood,
$p(\mathsf {I},t^{\star} )$
the joint prior and
$p(\mathsf {Q})$
the marginal distribution.
To determine the posterior distribution from (4.1), we first need to specify the joint prior
$p(\mathsf {I},t^{\star} )$
, which we would like as uninformative as possible. The analysis relies on the ability of the surrogate model to correctly predict the RT trajectories, so we do not want to use it in extrapolation. The joint prior distribution is therefore non-zero only for
$ (\mathsf {I},t^{\star} ) \in \mathcal I \times [0, \ 170]$
, where the surrogate model is well representative of the DNS. This domain corresponds to the white area in figure 2, bounded by the black rectangle for the minimum and maximum values. In this five-dimensional (5-D) domain, the prior is taken to be uniform in
$\mathsf {R}$
,
$\mathsf {B}$
,
$t^{\star}$
and log-uniform in
$\mathsf {S}$
,
$\mathsf {D}$
, consistent with the sensitivity analysis proposed in § 3.
The joint likelihood is obtained from a multivariate normal distribution,
$ {\mathcal N} (\hat {\mathsf {Q}}(t^{{\star}} ;{\mathsf I}), {\pmb \Sigma })$
, where the mean is obtained from the neural network surrogate model
$\hat {\mathsf {Q}}(t^{\star} ;\mathsf {I})$
as done by Nadiga, Jiang & Livescu (Reference Nadiga, Jiang and Livescu2019), since in the absence of the surrogate, Bayesian inference as presented here would become infeasible due to the large computational cost of DNS. The covariance matrix is chosen diagonal,
${\pmb\Sigma }=\boldsymbol {\sigma }\boldsymbol {\mathbb {I}}$
, with the variances
$\boldsymbol {\sigma }$
evaluated from the error between the surrogate model and the DNS database. So the likelihood is defined as
where
$N_{\mathsf {Q}}$
is the dimension of the 0-D vector
$\mathsf {Q}$
considered.
Due to the high computational cost of evaluating the marginal distribution in a high-dimensional space, as
$p(\mathsf {Q})=\int p(\mathsf {Q}|\mathsf {I},t^{\star} ) \,p(\mathsf {I},t^{\star} )\,\mathrm {d} \mathsf {I} \mathrm {d} t^{\star}$
, the rapid evaluation of the likelihood with the surrogate model allows us to employ a Markov Chain Monte Carlo (MCMC) method to sample the posterior distribution (Metropolis et al. Reference Metropolis, Rosenbluth, Rosenbluth, Teller and Teller1953; Hastings Reference Hastings1970; Gelman et al. Reference Gelman, Carlin, Stern and Rubin1995). Details of the algorithm are provided in Appendix C. Since we anticipate multi-modal posterior distributions in some cases, owing to the inertial or diffusive nature of the trajectories, we use the parallel tempering algorithm to more efficiently explore the parameter space (Geyer Reference Geyer1992; Falcioni & Deem Reference Falcioni and Deem1999; Sambridge Reference Sambridge2014).
4.2. Results
In this part, we apply the Bayesian methodology introduced in § 4.1 to recover the initial conditions
$\mathsf I$
and time
$t^{\star}$
from both inertial and diffusive trajectories, as depicted in figure 5. For this purpose, we consider two sets of 0-D measurements:
$\mathsf {Q}_2=(L^{\star} ,\dot L^{\star} )^T$
or
$\mathsf Q_7=(L^{\star} ,\dot L^{\star} , \mathcal K^{\star} , \varepsilon ^{\star} , \varepsilon _{cc} ^{\star} , \Theta ,\sin ^2 \Gamma )^T$
. The vector
$\mathsf {Q}_2$
, corresponding to the state variables of a buoyancy-drag model, provides less information compared with
$\mathsf {Q}_7$
, which is roughly associated with more advanced mix models (Grégoire et al. Reference Grégoire, Souffland and Gauthier2005; Banerjee, Gore & Andrews Reference Banerjee, Gore and Andrews2010; Schwarzkopf et al. Reference Schwarzkopf, Livescu, Gore, Rauenzahn and Ristorcelli2011; Morgan & Wickett Reference Morgan and Wickett2015; Schilling & Mueschke Reference Schilling and Mueschke2017). Consequently, better inference is expected with
$\mathsf {Q}_7$
.
Posterior distributions
$p(i,t^{\star} | \mathsf Q_2)$
projected on the 2-D planes
$ (t^{\star} , i)$
(with
$i \in \{\mathsf R, \mathsf B ,\mathsf S,\mathsf D \}$
) and sampled using the MCMC algorithm described in § 4.1. The three observations
$\mathsf Q_2$
corresponding to the buoyancy-drag state vector are taken along (a) the inertial and (b) the diffusive RT trajectories introduced in figure 5 (marked by the star symbols). The plain curves indicate the DNS trajectories and the dashed curve the corresponding neural network surrogate model.

Posterior distributions
$p(\mathsf i,t^{\star} | \mathsf Q_7)$
in the 2-D planes
$ (t^{\star} , i)$
(with
$i \in \{\mathsf R, \mathsf B ,\mathsf S,\mathsf D \}$
) and sampled using the MCMC algorithm described in § 4.1. The eight observations
$\mathsf Q_7$
corresponding to the full 0-D state vector are taken along (a) the inertial and (b) the diffusive RT trajectories introduced in figure 5 (marked by the star symbols). The plain curves indicate the DNS trajectories and the dashed curve the corresponding neural network surrogate model. The black dashed lines exhibited in the
$(t^{\star} ,\mathsf R)$
planes come from the sensitivity analysis, (3.4).

In figures 16 and 17, we show the inference results for
$\mathsf Q _{2,7}$
taken at various times along the inertial and diffusive trajectories of figure 5. The 5-D posterior distributions
$p(\mathsf I,t^{\star} | \mathsf Q_{2,7})$
are here represented in 2-D planes by
$p( i,t^{\star} | \mathsf Q_{2,7})$
(with
$i \in \{\mathsf R, \mathsf B ,\mathsf S,\mathsf D \}$
). Globally, the posteriors derived from the full 0-D variables
$\mathsf Q_7$
performs well to recover the initial conditions
$\mathsf I$
and the measurement time
$t^{\star}$
. The loss of memory that occurs at the transition to turbulence can be seen through the spread of the posterior distributions. By contrast, the inference is degraded when only using the reduced buoyancy-drag variables
$\mathsf Q_2$
.
There is, of course, a correlation between the sensitivity to initial conditions, detailed in § 3, and the ability to infer these same initial conditions. For instance, the bandwidth
$\mathsf B$
is not perfectly inferred by both
$\mathsf Q_{2,7}$
as the trajectories are not very sensitive to this parameter. By contrast, the Reynolds number
$\mathsf R$
can be well recovered by Bayesian inference, in particular, with full knowledge of
$\mathsf Q_7$
. At late time, one can notice the correlation between
$\mathsf R$
and
$t^{\star}$
on the posterior distributions. In the self-similar regime, knowing the buoyancy-drag variables
$\mathsf Q_2$
gives access to
$t^{\star} +t^{\star} _\infty$
, as
$\alpha _\infty$
is almost constant. Therefore, from (3.4) averaged on the initial domain
$\mathcal I$
, it is not surprising that the posterior is aligned along
$t^{\star} +21\, \mathsf R ^{1/3} = 2 L^{\star} /\dot L^{\star}$
, as shown in figure 17.
Inferring the initial conditions from 0-D quantities at
$t^{\star} =0$
presents an interesting area of study. With
$\mathsf {Q}_2$
, accurately determining
$\mathsf {I}$
is not possible. Moreover,
$t^{\star}$
is not properly inferred either, the uncertainty resulting from the fact that the state vector
$\mathsf Q_2$
is frozen at the beginning of the diffusive trajectories. Conversely, why does
$\mathsf Q_7$
perform much better? It can be shown that knowing
$\Theta$
(or
$\mathcal K_{cc}$
),
$\varepsilon _{cc}^{\star}$
and the diffusion coefficient enables the reconstruction of the initial wavenumber
$k_0$
and thus the Reynolds number
$\mathsf R$
.
(a,b) Trajectories represented in the
$(L^{\star} ,\dot L^{\star} )$
plane and sampled from the joint posterior distributions
$p(\mathsf I, t^{\star} | \mathsf Q_{2,7})$
projected (c,d) in the
$(t^{\star} , \mathsf R)$
plane. The figure shows the inference for a diffusive trajectory, where the star symbols in the
$(L^{\star} ,\dot L^{\star} )$
and
$(t^{\star} , \mathsf R)$
planes correspond to the observation of (a,c)
$\mathsf Q_2$
or (b,d)
$\mathsf Q_7$
.

Noticeably, figure 16 exhibits multi-modal posterior distributions
$p(\mathsf {I}, t^{\star} | \mathsf {Q}_{2})$
when considering the state vector
$\mathsf {Q}_2$
around the transition to the turbulence regime. This phenomenon, appearing in both inertial and diffusive trajectories, makes the evaluation of the posterior challenging and justifies the use of an MCMC parallel tempering algorithm. Figure 18 displays the trajectories in the phase plane
$(L^{\star} ,\dot {L}^{\star} )$
, sampled from the posterior distributions. Different trajectories, corresponding here to inertial and diffusive dynamics, converge at the same observation point and cannot be distinguished by knowing
$\mathsf {Q}_2$
alone. Conversely,
$\mathsf {Q}_7$
enables the differentiation between the various trajectory families, thus leading to more accurate inference. This raises the question: is it possible to optimise a better set of variables to model the RT dynamics, from the linear to the self-similar regime?
4.3. Finding an optimal state vector for reconstructing the initial conditions
So far, we have observed that measuring the full 0-D vector
$\mathsf Q_7$
performs much better for inferring the initial conditions than only the buoyancy-drag variables
$\mathsf Q_2$
. If a state vector can correctly infer the initial conditions of a trajectory, then it is an excellent candidate for a model. Therefore, it is important to quantify the ability of a state variable
$\mathsf Q$
to infer the initial conditions to find an optimal set of variables for modelling the full RT dynamics.
Our expectation is to find a posterior distribution close to the initial parameters corresponding to the trajectory we want to reconstruct, i.e.
$\mathsf R_{ {obj}}, \mathsf B_{ {obj}}, \mathsf S_{ {obj}},\mathsf D_{{obj}}$
and the measurement time
$t^{\star} _{ {obj}}$
. The ability of the posterior to recover the true initial conditions can be evaluated using the mean Euclidean distance
with
and where every parameter
$t^{\star}$
,
$i\in \{\mathsf {R},\mathsf {B},\ln \mathsf {S},\ln \mathsf {D}\}$
is normalised between 0 and 1 to have the same weight in the measure. In (4.3), the distance is zero if the posterior corresponds to the initial conditions. This measure can be quickly and directly computed from the samples obtained with the MCMC algorithm.
Ability of the various state vectors
$\mathsf Q$
to infer the initial conditions from the measure
$\mathcal D(\mathsf Q)$
. The plot illustrates the mean and standard deviations of
$\mathcal D(\mathsf Q)$
at various times, evaluated from 400 points uniformly sampled across the initial condition and time domain.

In figure 19, we present the time evolution of the distance
$\mathcal D(\mathsf Q)$
for various 0-D quantities evaluated from 400 points uniformly sampled across the initial conditions and time domain. As expected, the use of all the 0-D quantities,
$\mathsf Q_7$
, allows to well recover the initial conditions. The loss of memory of the initial conditions appears progressively, reaching an asymptotic plateau in the self-similar regime. When less information is present, the inference deteriorates and leads to higher values of
$\mathcal D(\mathsf Q)$
. Still, the measure allows the optimisation of the state variables for a model. Using
$\mathsf Q_3=(L^{\star} ,\Theta ,\varepsilon _{cc}^{\star} )^T$
, for instance, is a good trade-off between inference and number of 0-D variables.
5. The maximum a posteriori (MAP) model
In this section, we propose a complete modelling approach for the Rayleigh–Taylor instability. The model’s state variables consist of 0-D quantities selected based on their ability to infer the initial conditions. Therefore, the goal is to forward-propagate this state vector from
$\mathsf Q^n$
to
$\mathsf Q^{n+1}$
, corresponding to a
$\Delta t^{\star}$
time interval later.
We start by applying the Bayesian inference detailed in § 4.1 to derive the posterior distribution
$p(t^{\star} ,\mathsf I|\mathsf Q^n)$
. Then, we determine the maximum a posteriori (MAP) of this distribution,
$(t^{{\star} n}_{ {map}},\mathsf I ^n_{ {map}})=\text {argmax}\, p(t^{\star} ,\mathsf I|\mathsf Q^n)$
, and use the neural network surrogate model to propagate it forward. This leads to
The procedure, (5.1), can be repeated to derive the complete evolution of the state vector. This MAP model propagates the state variables along the most probable trajectory passing through
$\mathsf {Q}^n$
. In that respect, it is the best model that can be achieved for a given choice of
$\mathsf Q$
.
One of the difficulties of the MAP model is to obtain
$(t^{{\star} n}_{ {map}},\mathsf I ^n_{ {map}})$
from
$\mathsf Q^n$
, which is computationally intensive. Here, to illustrate the procedure and to have access to the uncertainties at every inference, we chose to infer the whole posterior distribution and to take the MAP out of it. However, a more efficient evaluation of the MAP would be to perform a maximum likelihood estimation. It is an optimisation procedure that only seeks the maximum of the distribution, avoiding the costly sampling.
Results of MAP models constructed from (5.1) with
$\Delta t^{\star} =30$
and using various state vectors indicated in the columns. (a–d) Comparison of the MAP model to the true trajectory, which corresponds to the inertial trajectory in figure 5. The grey areas indicate the level of uncertainities of the MAP trajectories. (e–h) Inferred samples of Reynolds
$\mathsf R$
and time
$t^{\star}$
and the posterior distributions
$p(\mathsf R, t^{\star} |\mathsf Q)$
at different instants of the MAP procedure.

In figures 20 and 21, we present the results derived with various MAP models starting from
$t^{\star} =0$
. While it is not surprising that the
$\mathsf Q_7$
model (all the 0-D variables without the vertical flux
$\mathcal F^{\star}$
) is able to reproduce nearly perfectly the target inertial trajectory, a less informed MAP model can also perform very well, like
$(L^{\star} , \Theta , \varepsilon _{cc}^{\star} )$
. Indeed, it is not necessary to infer the initial conditions during the phases with weak dependence on the initial parameters. In the late time self-similar regime, for instance, all the trajectories have the same growth rate so the prediction is not too difficult even with a low dimension state vector.
Notably, the trajectories obtained from the MAP model are almost continuous in their state variables (following the variance
$\boldsymbol {\sigma }$
of the likelihood), but not necessarily differentiable as the inferred MAP initial conditions can jump from one extrema to another. This is particularly true for
$\mathsf Q_2$
, which exhibits multi-modal posterior distributions. This explains the deviation of the MAP
$\mathsf Q_2$
model from the DNS trajectory.
Finally, by providing the most probable trajectories, the MAP model serves as a valuable tool to enhance current mix models used in fields like astrophysics or high-energy density applications. From a data modelling perspective, calibrating or refining the closures of mix models based on MAP trajectories – rather than on less probable, random trajectories – presents a more targeted approach.
6. Conclusion
This work investigates the memory of initial conditions in Rayleigh–Taylor (RT) turbulence using a large dataset of approximately 500 highly resolved low-Atwood direct numerical simulations generated with the code
$\mathsf{Stratospec}$
. These simulations are parametrised by four non-dimensional numbers that represent the statistical properties of the initial interface. These parameters do not capture the phase information of each initial perturbation mode – assumed to be random in this study – and therefore cannot uniquely determine the turbulent fluctuations of a single realisation. Rather, the intent of our study is to gain insight into and characterise robust statistical dependencies of the dynamics of the mixing zone – issues that have far wider applicability. In particular, we focus on how the mixing layer dynamics retains partial memory of its initial conditions via these four non-dimensional parameters.
To achieve this, we develop a surrogate model of the DNS database using a physics-informed neural network strategy. By incorporating soft and hard physical constraints, the neural network can reproduce the time evolution of 0-D turbulent quantities with high fidelity and cost-effectiveness, even in data-scarce regions. This model provides outputs with an estimated accuracy of
$5\,\%$
for given inputs, including time and the four non-dimensional parameters that characterise the statistical properties of the initial interface.
We then use this surrogate model to perform a sensitivity analysis across a broad initial domain, covering both inertial and diffusive trajectories, to disentangle the effects of the initial conditions. It is first demonstrated that, while varying initial conditions leads to a significant dispersion in the 0-D trajectories, the self-similar growth parameter remains nearly constant at
$\alpha \approx 0.020$
in the late-time regime – within the accuracy of the surrogate model and consistent with a mode-coupling phenomenology. This finding suggests that the memory of the Rayleigh–Taylor instability in the self-similar regime is not encoded in its growth rate, but rather in the virtual time origin
$t^{\star} _\infty$
, which varies sensitively with the initial Reynolds number, interface steepness, bandwidth and, to a lesser degree, diffusion thickness.
This finding quantifies the sustained growth of the mixing zone width variance at late times, driven significantly by the early stages of Rayleigh–Taylor dynamics and peaking around the transition to turbulence. While knowledge of the initial mixing zone width can help reduce uncertainties, it cannot fully eliminate them. This limitation arises from the roles of molecular mixing and the anisotropy of RT turbulence, which are not completely captured by measurements of the mixing zone alone. In this context, additional parameters such as the Schmidt number – although not considered in this study – are expected to impact the variability of the mixing zone at late times due to their influence on both linear growth and the transition to turbulence.
Furthermore, the initial perturbation bandwidth affects the mode-coupling or mode-competition dynamics, particularly near the cut-off wavenumber. When the bandwidth is large, the probability of having at least one unstable mode increases, leading to distinctly different dynamics.
To identify key quantities that conserve the memory of Rayleigh–Taylor turbulence, we perform a classical Bayesian inference to recover the initial conditions and the measurement time from a set of 0-D turbulent variables. We employ an MCMC parallel tempering algorithm to compute the posterior distributions, which, in some cases, can be multi-modal, accommodating both inertial and diffusive trajectories. The quality of the reconstruction is monitored throughout the different phases of the RT instability, revealing a rapid expansion of the posterior at the transition to turbulence. As expected, incorporating more 0-D variables into the observation set improves the inference quality. By quantifying the discrepancy between the inferred and true parameters of trajectories, we identify which sets of 0-D quantities best conserve the memory of initial conditions through the different phases of RT turbulence. These sets provide promising candidates for modelling.
Finally, we introduce the maximum a posteriori (MAP) model exemplifying a combined strategy of Bayesian inference and machine learning in turbulence modelling. By identifying the peak of the posterior distribution and advancing the state variables forward with the surrogate model – parametrised by the inferred initial conditions – we achieve precise predictions of the RT dynamics from the initial stages through the late-time self-similar regime. As this model adheres to the trajectories of the surrogate model, it inherits all its beneficial properties stemming from the physics-informed approach. Thus, this model represents the optimal solution achievable with the state vector
$\mathsf Q$
, adhering to the most probable trajectories indicated by the MAP estimate. Furthermore, having complete knowledge of the posterior distribution allows quantification of the uncertainties associated with these predictions (Xiao & Cinnella Reference Xiao and Cinnella2019).
This methodology including sensitivity analysis and Bayesian inference would not have been possible without modern machine learning tools, which facilitate the derivation of an accurate surrogate model from the DNS database. While currently applied to the Rayleigh–Taylor problem, this approach can be adapted to other types of flows, representing a significant advancement in predicting turbulent mixing layers. However, it is important to note that the effectiveness of this method depends heavily on the quality of the database used and is confined to its domain of validity. Despite incorporating physics-informed terms, making predictions far outside the established data range remains challenging.
Acknowledgements
The simulations were performed at the French TGCC centre.
Declaration of interests
The authors report no conflict of interest.
Data availability statement
The data that support the findings of this study are openly available in the French Fluid Dynamics Database (F2D2), Thévenin et al. (2025), at https://doi.org/10.57745/VTM1PN.
Appendix A. Physics-informed neural network as a surrogate model for the RT database
To emulate the DNS dynamics, we consider a multi-layer perceptron (MLP), which is classically used for building deep neural networks. It takes as inputs the initial conditions
$\mathsf {I}$
and the time
$t^{\star}$
, and returns the associated prediction of the dynamical quantities
$\mathsf {Q}_8(t^{\star} ;\mathsf {I})=(L^{\star} ,\dot {L}^{\star} , \mathcal {K}^{\star} , \varepsilon ^{\star} , \Theta , \varepsilon _{cc}^{\star} , \mathcal {F}^{\star} , \sin ^2(\Gamma ))^T$
. The time derivative
$\dot {L}^{\star}$
is directly computed by auto-differentiation of
$L^{\star}$
with respect to
$t^{\star}$
. The MLP consists of seven hidden layers, each containing 32 neurons, with Softplus activation functions, totalling
$6792$
parameters. The architecture has been meticulously designed to ensure physical consistency. We are particularly interested in performing (reasonable) time extrapolations. A simple approach is to square the output of the last neuron, corresponding to
$L^{\star}$
, to capture its quadratic time growth at later stages. This is because a fully activated Softplus MLP will produce a linear response to its inputs, which include time. This approach is referred to as ‘hard constraints’ in the literature, as these properties are enforced through the architecture, and it is represented by the neuron T in figure 8.
The MLP parameters are optimised during the training stage to minimise the loss function
$\mathcal {J}_{{tot}}$
which is composed of two parts:
\begin{equation} \mathcal {J}_{{Data}}= \sum _{t^{\star} _n,I_m \in \mathcal D_{\text {DNS}}} \beta ^{{Data}}_{nm}\left |\left | \text {norm}_{01} \left ( \mathsf {Q}_8(t^{\star} _n;\mathsf I_m) - \widehat {\mathsf {Q}}_8 (t^{\star} _n;\mathsf I_m) \right ) \right |\right |^2 \ \text {and} \end{equation}
The supervised part
$\mathcal {J}_{{Data}}$
, (A1b
), measures the discrepancy between the normalised surrogate predictions and their DNS counterpart. It is evaluated on a time and initial conditions domain
$\mathcal D_{\text {DNS}}$
. The diagonal matrix
$\text {norm}_{01}$
normalises the 0-D quantities in a
$[0, \ 1]$
interval to facilitate training. The weight coefficients
$\beta ^{Data}_{nm}$
take the same values except at
$t^{\star} _n=0$
, where it takes higher values to enforce the right initial conditions.
The unsupervised (physics-informed) part
$\mathcal {J}_{{PI}}$
, (A1c
), penalises predictions which do not satisfy a list
$\mathcal U$
of physical constraints
$u$
and evaluated on a set of sampled collocation times and initial conditions
$\mathcal D _{coll}$
. Therefore,
$\mathsf J _u$
is the residual corresponding to the physical constraints at a given point
$(t_n^{\star} ,\mathsf I_m)$
. The total sum over all collocation points is then provided by
$ \mathcal J _u$
. The physical constraints in
$\mathcal U$
are respectively: the kinetic energy (
$uu$
) and concentration variance (
$cc$
) conservations (2.9a
), (2.9b
), the self-similar relation (2.10) (
$ss$
), the positivity of
$\dot {L}$
(
$\dot L^{\star}$
) and the dimensional relation
$0=L_0^{\star} -3/2\mathsf {R}\partial L_0^{\star} /\partial \mathsf {R}$
(
$L^{\star} _0$
) expressing the fact that the initial mixing zone width does not depend on the acceleration (or
$\mathsf R$
). The weight coefficients,
$\beta ^{PI}_{nmu}$
, allow balancing of the constraints with each other and the use of different sets of collocation points for the constraints. These physics-informed terms are referred to as soft constraints, as the properties are encouraged rather than enforced. To get the most out of it, it is evaluated both at the DNS data points and at randomly drawn collocation points in an extended domain to guide the training towards plausible solutions, even in extrapolation. In practice, several combinations of the weight coefficients are tried and in the end, the best performing surrogate is chosen.
Evolution of the (ai) supervised and (b) unsupervised losses during the training of the physics-informed neural network.

The surrogate model is implemented with the PyTorch deep learning framework (Paszke et al. Reference Paszke2019). The MLP parameters are initialised with the Xavier normal procedure and are optimised using the Adam algorithm (Kingma Reference Kingma2014) with a learning rate of
$5\times 10^{-4}$
on an NVIDIA A100 GPU, for
$2\times 10^4$
epochs done in approximatively
$80$
minutes. A training dataset containing
$294$
trajectories (of
$300$
time points each) is used for optimisation and a validation dataset containing
$126$
is used to monitor overfitting, while the remaining
$73$
are kept for the a posteriori test. At every epoch, the complete loss is evaluated using batches of
$25\,\%$
the size of the training dataset. In addition, the unsupervised part is also evaluated at random points
$5\,\%$
the size of the training dataset. The evolution of the different losses is illustrated in figure 22, showing the good balance between the supervised and unsupervised part. Some terms of the unsupervised loss, such as
$\mathcal J_{ss}$
or
$\mathcal J_{\dot L ^{\star} }$
, diminish rapidly during training, explaining why their primary role is to accelerate the process. This also indicates that the loose inequality corresponding to
$\mathcal J_{ss}$
does not constrain the self-similar evolution of the surrogate model.
Appendix B. Computing the Sobol indices
In the Rayleigh–Taylor configuration, knowing the initial conditions
$\mathsf {I}$
and the time
$t^{\star}$
allows for the complete determination of a zero-dimensional dynamics
$q(t^{\star} ; \mathsf {I})$
. However, if these input parameters are unknown, the dynamics may vary substantially. Sensitivity analysis (SA) describes how uncertainty in the inputs influences this variance and, therefore, identifies the parameters (or group of parameters) that contribute most significantly to it.
A simple SA can be conducted by varying one input at a time while keeping the others constant. This approach is referred to as a local SA. It has the advantage of being straightforward to perform and incurs a very low computational cost. However, it does not account for interaction effects and its effectiveness depends on the point around which the local SA is conducted (Saltelli et al. Reference Saltelli, Aleksankina, Becker, Fennell, Ferretti, Holst, Li and Wu2019). A global SA overcomes these limitations but requires a significantly higher computational cost, which is mitigated by using a surrogate model.
In this paper, we employed a variance-based global SA, see Saltelli (Reference Saltelli2008) for a comprehensive overview. It consists in decomposing the output in a sum of increasing dimension orthogonal functions such that
Here, we omit the time
$t^{\star}$
dependence for simplification. The functions up to second-order are thus defined as
This leads to the Hoeffding–Sobol decomposition of the variance,
\begin{equation} \begin{aligned} V[q] &= \sum _{i\in \mathsf {I}} V[q_i(i)] + \sum _{i\in \mathsf {I}}\sum _{j\in \mathsf {I},j\neq i} V[q_{ij}(i,j)] + \cdots + V[q_{\mathsf {R}\mathsf {B}\mathsf {S}\mathsf {D}}(\mathsf {I})] \\ &= \sum _{i\in \mathsf {I}} V[q_i(i)] + \text {interactions}, \end{aligned} \end{equation}
where
$V[q_i(i)]=V[E[q|i]]$
is known as the first-order partial variance. It represents the part of the variance that is due to a single input parameter
$i\in \mathsf {I}$
. It expresses the main effect of
$i$
. By dividing this partial variance by the total variance
$V[q]$
, one obtains the so-called first-order Sobol index. Its explicit formulation involves several integrals that are analytically intractable and in relatively high dimension, so to perform SA, one must use efficient approximating techniques. In this work, we used an efficient Monte Carlo based procedure first introduced by Soboĺ (Reference Soboĺ1993). It relies on the observation that
$V[E[q|i]]$
equals the covariance
$\textit {Cov}[q,q']$
, where
$q$
and
$q'$
are computed from two sets of inputs that share the same value for input
$i$
only. Therefore, if we construct two independent and identically distributed samples
$q_n$
and
$q'_n$
of size
$N$
, it leads to the following estimate (Saltelli et al. Reference Saltelli, Annoni, Azzini, Campolongo, Ratto and Tarantola2010):
\begin{equation} V[E[q|i]] \simeq \dfrac {1}{N} \sum _{n=1}^{N} q_n q'_n - \left (\dfrac {1}{N}\sum _{n=1}^{N} q_n\right )^2, \end{equation}
where it is assumed that
$ (({1}/{N})\sum _{n=1}^{N} q_n )^2= ( ({1}/{N})\sum _{n=1}^{N} q'_n )^2$
. In our study, we used
$N=3\times 10^4$
samples, where the inputs were randomly drawn within a hypercubic domain because the variance-based global SA requires that the inputs not be correlated. This procedure was executed for various time steps
$t^{\star}$
to reconstruct the time evolution of the initial conditions’ main effects. Additionally, this procedure was repeated 20 times and the results were averaged to obtain a robust estimation through bootstrapping. This was particularly important for assessing the second-order interaction effects, although we ultimately decided to exclude these from the paper due to their minimal contribution.
Appendix C. Bayesian inference by MCMC
As introduced in § 4.1, directly evaluating the joint posterior from Bayes’ rule is computationally heavy due to the evaluation of the marginal law
$p(\mathsf {Q})$
. Markov Chain Monte Carlo (MCMC) is a general class of sampling algorithms that overcomes this issue by avoiding the need to compute this term. It involves constructing a Markov chain whose stationary distribution is the target distribution we seek, in our case, the joint posterior. This means that, after a convergence stage, all samples generated by the Markov chain will follow the target distribution. See Gelman et al. (Reference Gelman, Carlin, Stern and Rubin1995) for a comprehensive introduction.
In this work, as we expect our joint posterior to be multimodal in some cases, we used the parallel tempering algorithm with a Metropolis–Hastings random walk (Metropolis et al. Reference Metropolis, Rosenbluth, Rosenbluth, Teller and Teller1953; Hastings Reference Hastings1970; Geyer Reference Geyer1992; Falcioni & Deem Reference Falcioni and Deem1999; Sambridge Reference Sambridge2014). This algorithm relies on two key elements. First, an artificial dimension known as the temperature
$T$
is added to the problem, necessitating that we now sample the distribution
This temperature has discrete values and one can remark that for
$T=1$
, we recover our target distribution. In addition, for
$T\gt 1$
, the distribution
$p(\mathsf {I},t^{\star} ,T|\mathsf {Q})$
becomes a flattened version of
$p(\mathsf {I},t^{\star} |\mathsf {Q})$
. The second key idea is to run several Markov chains in parallel that have different temperatures
$T\geqslant 1$
. These chains independently advance following a random walk and every once in a while, temperature swaps between those chains are proposed and either accepted or rejected according to an acceptance rule very similar to that of Metropolis and Hastings. This allows an efficient exploration of the whole space, no matter how sharp and multi-modal the original joint posterior is. Indeed, a classical Metropolis–Hastings algorithm would get stuck in a high-probability region and would have trouble exploring another if the two peaks are separated by a low-probability region. In parallel tempering, a chain with
$T=1$
would get stuck, but a higher-temperature chain would not. So, the latter could reach the second peak and then swap temperature with the former, allowing an efficient sampling of our joint posterior (by collecting samples drawn at
$T=1$
).
For each inference, we use
$M=10$
parallel chains with temperatures logarithmically ranging between
$1$
and
$10^4$
. We perform
$N=5\times 10^5$
iterations and discard the first
$10^5$
iterations that correspond to a (conservative) convergence stage. At every iteration, we either propose a temperature swap between pairs of chains, with probability
$p_{ {swap}}=0.25$
, or make a random walk step for every parallel chain, with probability
$1-p_{ {swap}}=0.75$
. The random walk involves making a random step starting from the last state
$(\mathsf {I}_n,t^{\star} _n)$
of the chain. We define the proposal distribution
$\mathcal {P}(\mathsf {I}_j,t^{\star} _j|\mathsf {I}_n,t^{\star} _n)$
as a Gaussian with mean
$(\mathsf {I}_n,t^{\star} _n)$
and covariance matrix
$\boldsymbol {C}$
chosen diagonal with standard deviations set to be equal to
$1\,\,\%$
of the overall domain range. In practice, we conduct two independent parallel tempering runs to assess the statistical convergence of the results, as suggested by Roy (Reference Roy2020); Vehtari et al. (Reference Vehtari, Gelman, Simpson, Carpenter and Bürkner2021). In addition, we have verified that the effective sample size (ESS), which indicates the number of uncorrelated samples in the MCMC chain, is greater than
$400$
, and the potential scale reduction factor
$\hat R$
, which assesses how well the parallel chains are mixed, is less that
$1.01$
, as suggested by Vehtari et al. (Reference Vehtari, Gelman, Simpson, Carpenter and Bürkner2021). Additional details can be found from Thévenin (Reference Thévenin2024).






































































































