betaFeedback

WorksPhysics and Astronomy

Learning Cole-Hopf Coordinates for Stochastic Growth Dynamics: the KPZ class

Abstract

Inferring effective equations from observations of nonlinear stochastic processes is challenging because individual trajectories are noisy realizations, while the relevant physical information is encoded in ensemble statistics. We introduce a data-driven framework that learns a change of variables in which nonlinear stochastic dynamics become linear. The model combines a neural change of representation with a linear stochastic latent evolution and is trained through a short-time moment-matching objective, thereby avoiding trajectory \rev{simulation} during optimization. This construction yields an interpretable generative model: both the learned transformation and the latent evolution operators can be directly inspected. We validate the method on the one-dimensional Kardar–Parisi–Zhang (KPZ) equation, a paradigmatic model of stochastic growth for which the Cole–Hopf transformation maps the nonlinear dynamics to the stochastic heat equation. From simulated height trajectories alone, the model recovers the local exponential Cole–Hopf map and the associated linear stochastic operator. Generated trajectories reproduce the main statistical signatures of the KPZ universality class, including the growth law Var(ht)∼t2β\mathrm{Var}(h^{t})\sim t^{2β} with β=1/3β=1/3 and the GOE Tracy–Widom height distribution for flat initial conditions. These results show that interpretable stochastic equations can be inferred from data by learning representations in which the dynamics closes in a simple form. Furthermore the development of the physical informed learning permits to get insights on the physics of the dynamical process.

Introduction

Many dynamical processes in nature are intrinsically stochastic, from microscopic thermal fluctuations to macroscopic growth phenomena. In such systems, the observed trajectories are noisy realizations of an underlying probabilistic dynamics, and the relevant physical information is encoded not in a single path but in statistical observables such as correlations and probability distributions. Inferring an effective stochastic equation from data therefore remains a central problem across statistical physics and related fields [1–3].

Traditionally, this task relies on physical insight. One proposes an effective Langevin equation or stochastic partial differential equation, guided by symmetries, conservation laws, locality, and dimensional arguments, and then validates the model a posteriori by comparing its predictions with experiments or simulations. A paradigmatic example is the Kardar–Parisi–Zhang (KPZ) equation [4, 5], which describes the stochastic evolution of the height h(x,t)h(x,t) of a growing one-dimensional interface,

∂th(x,t)=ν∂x2h(x,t)+λ2(∂xh(x,t))2+2D η(x,t),(1)\partial_t h(x,t) = \nu\partial_x^2 h(x,t) + \frac{\lambda}{2}\left(\partial_x h(x,t)\right)^2 + \sqrt{2D}\,\eta(x,t), \tag*{(1)}

where the first term smooths the interface, the second accounts for lateral growth, and η(x,t)\eta(x,t) is a Gaussian space-time white noise representing random fluctuations. Despite its deceptively simple form, the KPZ equation captures the universal large-scale behavior of a broad class of non-equilibrium growth processes. KPZ scaling and fluctuations have been observed or proposed in systems ranging from turbulent liquid crystals and polariton condensates to bacterial colonies and other kinetic roughening phenomena [6, 7, 8, 9, 10, 11]. While this phenomenological approach has been remarkably successful, it relies on proposing the appropriate evolution equation from physical intuition. Determining the correct form of the dynamics is often the most challenging step, especially for complex systems where the underlying mechanisms are only partially understood.

This difficulty is amplified by nonlinearity. Most complex phenomena are described by nonlinear evolution equations [12], for which exact analytical treatments are rare. In some special cases, however, a suitable change of variables can drastically simplify the dynamics. A classical example is the Cole–Hopf transformation, which maps the viscous Burgers equation to the linear heat equation [13, 14]. In the stochastic setting, the same idea applies to the KPZ equation: after the transformation

z(x,t)=exp⁡[λ2νh(x,t)],(2)z(x,t) = \exp\left[\frac{\lambda}{2\nu}h(x,t)\right], \tag*{(2)}

the nonlinear KPZ equation is mapped onto the linear stochastic heat equation with multiplicative noise [15, 16],

∂tz(x,t)=ν∂x2z(x,t)+λ2ν2D z(x,t)η(x,t),(3)\partial_t z(x,t) = \nu\partial_x^2 z(x,t) + \frac{\lambda}{2\nu}\sqrt{2D}\,z(x,t)\eta(x,t), \tag*{(3)}

where the noise is interpreted using the Stratonovich convention. KPZ therefore provides an example of a nonlinear stochastic field theory that admits an exact linearizing transformation. This naturally raises the following data-driven question: can such simplifying transformations be learned directly from stochastic trajectories?

Machine learning provides a possible route to this problem. Neural differential-equation models have made it possible to learn continuous-time dynamics from data, and neural stochastic differential equations extend this idea to stochastic trajectories [17, 18, 19, 20]. Recent neural operator and neural SPDE approaches further address stochastic spatiotemporal dynamics at the level of solution operators [21], although they currently remain restricted to specific classes of systems. In parallel, sparse-regression methods such as SINDy and PDE-FIND seek to infer governing equations directly from data [22, 23] [24]. These were originally formulated for deterministic dynamics, but the same sparse-identification idea has been extended to stochastic systems by combining it with Kramers–Moyal estimates of the drift and diffusion coefficients [25]. Together, these developments show that increasingly complex dynamical systems can now be learned directly from observations.

For deterministic dynamics, several methods have been developed to learn interpretable representations in which the evolution becomes simpler. These include Koopman-inspired linear embeddings [26–29] and autoencoder-based approaches combined with sparse regression, such as SINDy-AE [30]. Such methods can identify latent coordinates in which nonlinear deterministic systems evolve approximately linearly or admit a sparse description, i.e. governing equations built from only a handful of terms. These results demonstrate that machine learning can go beyond black-box prediction, and it can be used to recover compact mathematical representations of complex dynamics.

Extending these ideas to stochastic systems is considerably more challenging. In a deterministic system, a model can be trained by requiring its predicted trajectory to remain close to the observed one. In a stochastic system, this criterion is no longer appropriate: even the correct model would generate a different trajectory if driven by a different realization of the noise. Instead, the relevant information is encoded in the statistics of short-time transitions, such as their conditional mean and covariance, rather than in individual trajectories.

Several recent works have addressed this challenge from different perspectives. Stochastic Force Inference [31] estimates drift and diffusion fields from short-time increments by projection onto a known basis, yielding directly interpretable equations, and has been extended to spatially extended systems through a dissipation-based dimensional reduction [32]. Heck et al. [33] introduce moment-matching constraints within a variational latent-SDE framework, where both drift and diffusion are represented by neural networks. Hasan et al. [34] focus on learning latent representations within a variational autoencoder, while simulation-free approaches such as SDE Matching [35, 36] avoid trajectory integration during training. Operator-theoretic methods instead learn the infinitesimal generator directly from observations without introducing an explicit latent representation [37, 38].

These approaches have demonstrated impressive generative performance. However, their primary objective is to reproduce the observed stochastic dynamics rather than to recover an interpretable mathematical description of it. The latent dynamics is typically represented by highly expressive neural networks, making it difficult to extract an explicit stochastic equation that can be inspected and analyzed. Our objective is different: rather than merely learning to generate trajectories, we seek to learn a change of variables in which the observed dynamics becomes a linear stochastic differential equation whose coefficients can be read directly from the trained model.

In this work, we introduce a neural framework designed to recover interpretable stochastic equations from data. The central idea is to learn a local change of variables in which the observed dynamics becomes a linear stochastic evolution. The model combines an autoencoder mapping the observed field to a latent representation with a linear latent SDE, and is trained by matching short-time increment statistics rather than by rolling out full trajectories during optimization. This yields a model that is both generative and interpretable: it can produce new trajectories with the statistics of the training data, while providing explicit objects—the learned transformation, the latent linear operator, and the latent noise structure—from which an effective stochastic equation can be reconstructed.

We validate the framework on the one-dimensional KPZ equation, for which the Cole–Hopf transformation maps the nonlinear dynamics to the linear stochastic heat equation [15, 16]. This benchmark allows us to test simultaneously whether the method recovers the correct linearizing transformation and whether the learned generative model reproduces the universal statistical properties of the KPZ class. We show that the model recovers both the Cole–Hopf map and the associated linear stochastic dynamics directly from height trajectories, while reproducing highly non-trivial KPZ signatures, including the t2βt^{2\beta} roughening law and the GOE Tracy–Widom height distribution for flat initial conditions [39–44].

The remainder of the manuscript is organized as follows. In Sec. II we describe the numerical generation of the KPZ dataset used throughout the paper. In Sec. III we introduce the interpretable latent-SDE framework, including the local change of variables, the linear latent dynamics, the moment-matching loss, and the training procedure. We then assess the generative performance of the learned model and analyze the learned representation, showing that it recovers both the Cole–Hopf transformation and the associated linear stochastic operator. We also discuss in Sec. IV an extension of the same statistical-learning principle to a more flexible autoencoder architecture. Finally, Sec. V contains our conclusions and perspectives.

Numerical generation of KPZ trajectories

As discussed in the Introduction, the training data are generated from the Cole–Hopf representation of the one-dimensional KPZ equation. Rather than discretizing the nonlinear KPZ equation directly, we simulate the corresponding linear stochastic dynamics for the transformed variable zz on a periodic one-dimensional lattice. This guarantees that an exactly linear latent evolution underlies the training data, since on the discretized lattice the Cole–Hopf mapping holds only to leading order in Δx\Delta x. The resulting lattice stochastic equation, using Itô’s formalism, is

dzit=[νzi+1t+zi−1t−2zitΔx2+Dα2Δxzit]dt+α2DΔxzitdtξit(4)d z_i^t = \left[\nu\frac{z_{i+1}^t+z_{i-1}^t-2z_i^t}{\Delta x^2}+\frac{D\alpha^2}{\Delta x}z_i^t\right]dt+\alpha\sqrt{\frac{2D}{\Delta x}}z_i^t\sqrt{dt}\xi_i^t \tag*{(4)}

where α=λ/(2ν)\alpha=\lambda/(2\nu) and ξit∼N(0,1)\xi_i^t\sim\mathcal{N}(0,1) are independent Gaussian random variables with zero mean and unit variance. Equivalently, the quantity dtξit\sqrt{dt}\xi_i^t represents a Brownian increment.

The height field is recovered through the inverse Cole–Hopf transformation

hit=1αlog⁡zit.(5)h_i^t=\frac{1}{\alpha}\log z_i^t. \tag*{(5)}

The deterministic contribution proportional to ziz_i in Eq. (4) is the standard Itô correction associated with multiplicative noise. Because the noise amplitude itself depends on ziz_i, fluctuations contribute not only to the variance of the process but also to its average evolution, generating an additional drift term proportional to ziz_i. This correction is necessary for the reconstructed field Eq. (5) to evolve according to the KPZ equation. Without it, the interface would acquire an additional deterministic growth velocity, while its fluctuation properties would remain unchanged [15, 16]. All trajectories are initialized from a flat interface,

hi0=0,zi0=1.(6)h_i^0=0,\qquad z_i^0=1. \tag*{(6)}

Unless stated otherwise, we use

ν=0.5,λ=1,D=0.5,Δx=1,dt=10−2,N=256.(7)\nu=0.5,\qquad\lambda=1,\qquad D=0.5,\qquad\Delta x=1,\qquad dt=10^{-2},\qquad N=256. \tag*{(7)}

The dataset consists of 10510^5 independent trajectories generated from different realizations of the noise (Figure 1). For the parameters used here, the microscopic lattice time tmicro∼Δx2/ν≃2t_{\mathrm{micro}}\sim\Delta x^2/\nu\simeq2 is not well separated from the estimated EW–KPZ crossover time tc≃7.9t_c\simeq7.9 [45]; consequently, no extended EW scaling plateau is resolved.

Figure 1. Numerical KPZ trajectories generated from a flat initial condition. Left: height field h(x,t)h(x,t) as a function of position xx and time tt for a single noise realization. Right: corresponding interface profiles at selected times. The full dataset consists of independent stochastic trajectories generated from the same flat initial condition and driven by different noise realizations. The ensemble displays the characteristic roughening and fluctuation growth of one-dimensional KPZ dynamics.

Physical interpretable model

Architecture

The structure of the model is illustrated in Figure 2. We consider a dataset of stochastic interface evolutions, given by height fields ht=(h1t,…,hNt)\boldsymbol{h}^t=(h_1^t,\ldots,h_N^t) sampled at discrete times tt, and denote by dht=ht+dt−htd\boldsymbol{h}^t=\boldsymbol{h}^{t+dt}-\boldsymbol{h}^t the observed increment over a short time interval dtdt. Our goal is to infer, directly from these observations, a dynamical model that can reproduce the stochastic evolution of the system.

Figure 2. Schematic of the model. The observed field h(x,t)h(x,t) is mapped site by site to a latent variable z(x,t)=φθ(h(x,t))z(x,t)=\varphi_\theta(h(x,t)), using the same scalar function at every spatial location. In latent space, the dynamics is modeled as a linear stochastic evolution with parameters KθK_\theta and MθM_\theta. The decoder Ψθ\Psi_\theta maps the latent variable back to the physical field. The locality of the transformation makes it directly interpretable as a change of variables, while the linear latent dynamics provides a simple description of the evolution.

Motivated by the idea that complex stochastic dynamics may become simple in a suitable representation, we seek a change of variables in which the evolution takes a linear form. Specifically, we look for a transformation of the form

zit=φθ(hit),i=1,…,N,z_i^t = \varphi_\theta(h_i^t), \qquad i=1,\ldots,N,

where the same scalar function φθ\varphi_\theta is applied independently at each spatial site. This locality assumption is deliberately restrictive: the transformed variable at site ii depends only on the height at the same site. The motivation is twofold. First, such a transformation is directly interpretable, as it reduces to a one-dimensional function that can be inspected a posteriori. Second, several stochastic growth equations of physical interest admit local changes of variables that simplify their dynamics, the Cole–Hopf transform being a notable example [43].

To ensure that the transformation is invertible, we simultaneously learn a map

hit=Ψθ(zit),h_i^t = \Psi_\theta(z_i^t),

and penalize reconstruction errors. The pair (φθ,Ψθ)(\varphi_\theta,\Psi_\theta) therefore defines a learned change of variables between the observed field hh and a new representation zz.

In the transformed variables, we assume that the dynamics takes a simple stochastic form. Specifically, we model the increments of zz as

dzit=∑jKθ,ijzjt dt+∑jMθ,ijzjt dWjt,(8)\mathrm{d}z_i^t = \sum_j K_{\theta,ij} z_j^t \,\mathrm{d}t + \sum_j M_{\theta,ij} z_j^t \,\mathrm{d}W_j^t, \tag*{(8)}

where KθK_\theta and MθM_\theta are N×NN \times N matrices, and dWjt\mathrm{d}W_j^t are independent Gaussian increments with zero mean and variance dt\mathrm{d}t. Equivalently, one can write

dWjt=dt ξjt,\mathrm{d}W_j^t = \sqrt{\mathrm{d}t}\,\xi_j^t,

where ξjt∼N(0,1)\xi_j^t \sim\mathcal{N}(0,1) are independent standard normal variables.

The first term represents a linear deterministic evolution, while the second term introduces multiplicative noise: each component zjtz_j^t sets the amplitude of an independent noise source, which is then linearly mixed across space by the matrix MθM_\theta. The learnt process noise is therefore zero-mean Gaussian with covariance Mθdiag⁡((zt)2)MθTM_\theta\operatorname{diag}((z^t)^2) M_\theta^{\mathsf{T}}. This matrix is symmetric and positive semi-definite by construction, for any MθM_\theta.

The goal of the model is therefore to discover, directly from data, a representation in which the dynamics closes in this linear form. We chose this parametrization for consistency with the linear Cole-Hopf dynamics (Eq. (4)). More general forms generate the data just as well, but make the analytical map harder to recover unambiguously. We therefore report only this specific case.

Moment-matching loss

The model is trained by matching the short-time statistics of the observed increments. The reconstruction loss enforces consistency of the learned change of variables:

Lrec=∑i(Ψθ(φθ(hit))−hit)2.(9)\mathcal{L}_{\mathrm{rec}} = \sum_i \left(\Psi_\theta(\varphi_\theta(h_i^t)) - h_i^t\right)^2. \tag*{(9)}

The dynamical loss should enforce the agreement between dynamical variables, that is increments for a SDE. Since it is meaningless to ask for a full consistency between true and modeled trajectories, the loss is of a statistical kind. It is based on the conditional drift and covariance of the increment predicted by the latent stochastic model. Since hit=Ψθ(zit)h_i^t = \Psi_\theta(z_i^t),

Itô’s formula gives, to leading order in dtdt,

dhit,model=Ψθ′(zit) dzit+12Ψθ′′(zit)(dzit)2.(10)dh_i^{t,\mathrm{model}}=\Psi_\theta'(z_i^t)\,dz_i^t+\frac{1}{2}\Psi_\theta''(z_i^t)(dz_i^t)^2. \tag*{(10)}

We denote by dhit,modeldh_i^{t,\mathrm{model}} the increment predicted by the model, and by dhit,datadh_i^{t,\mathrm{data}} the increment measured in the training data. Substituting Eq. (8) yields the conditional drift

bθ,i(ht):=E[dhit,model∣ht]dt=Ψθ′(zit)∑jKθ,ijzjt+12Ψθ′′(zit)∑jMθ,ij2(zjt)2.(11)\begin{aligned} b_{\theta,i}(h^t):=\frac{\mathbb{E}[dh_i^{t,\mathrm{model}}\mid h^t]}{dt} \\ &=\Psi_\theta'(z_i^t)\sum_j K_{\theta,ij}z_j^t+\frac{1}{2}\Psi_\theta''(z_i^t)\sum_j M_{\theta,ij}^2(z_j^t)^2. \tag*{(11)} \end{aligned}

The second term is the Itô correction induced by the multiplicative latent noise and the nonlinear decoder. The corresponding conditional covariance per unit time is

Cθ,ik(ht):=Cov⁡(dhit,model,dhkt,model∣ht)dt=Ψθ′(zit)Ψθ′(zkt)∑jMθ,ijMθ,kj(zjt)2.(12)\begin{aligned} C_{\theta,ik}(h^t):=\frac{\operatorname{Cov}(dh_i^{t,\mathrm{model}},dh_k^{t,\mathrm{model}}\mid h^t)}{dt} \\ &=\Psi_\theta'(z_i^t)\Psi_\theta'(z_k^t)\sum_j M_{\theta,ij}M_{\theta,kj}(z_j^t)^2. \tag*{(12)} \end{aligned}

The empirical increment dhitdh_i^t is a noisy realization of the local transition law. Therefore, the objective is not to match each observed increment exactly, but to match the drift and covariance in expectation over many short-time transitions. We use the empirical losses

Ldrift=∑i(bθ,i(ht)−dhit,datadt)2,(13)\mathcal{L}_{\mathrm{drift}}=\sum_i\left(b_{\theta,i}(h^t)-\frac{dh_i^{t,\mathrm{data}}}{dt}\right)^2, \tag*{(13)}

and

Lcorr=∑i,k(Cθ,ik(ht)−dhit,datadhkt,datadt)2.(14)\mathcal{L}_{\mathrm{corr}}=\sum_{i,k}\left(C_{\theta,ik}(h^t)-\frac{dh_i^{t,\mathrm{data}}dh_k^{t,\mathrm{data}}}{dt}\right)^2. \tag*{(14)}

In Eq. (14) we use the raw product dhit,datadhkt,datadh_i^{t,\mathrm{data}}dh_k^{t,\mathrm{data}} rather than the centered product. This is equivalent at leading order in dtdt. This correction is therefore beyond the accuracy of the short-time expansion.

The total loss is

L=wrecLrec+wdriftLdrift+wcorrLcorr,(15)\mathcal{L}=w_{\mathrm{rec}}\mathcal{L}_{\mathrm{rec}}+w_{\mathrm{drift}}\mathcal{L}_{\mathrm{drift}}+w_{\mathrm{corr}}\mathcal{L}_{\mathrm{corr}}, \tag*{(15)}

with wrec=wdrift=wcorr=1w_{\mathrm{rec}}=w_{\mathrm{drift}}=w_{\mathrm{corr}}=1. This moment-matching perspective is close in spirit to Kramers–Moyal approaches, which estimate drift and diffusion from short-time increments [46, 47] [25]. It differs from many latent-SDE methods, which typically rely on likelihood-based objectives over ensembles of trajectories and require rollouts (i.e. simulating the process) during training [33]. Here, instead, we use a local-in-time moment-matching principle in a learned latent representation, while imposing a physically interpretable structure: a local scalar change of variables and a linear stochastic latent dynamics.

Implementation and training

The model consists of four components: the encoder φθ\varphi_\theta, the decoder Ψθ\Psi_\theta, and the matrices KθK_\theta and MθM_\theta defining the latent dynamics. All components are trained jointly. The encoder and decoder are scalar feed-forward neural networks with three hidden layers of 128 units and smooth (SiLU) activations. Because they act independently at each spatial site, their evaluation and differentiation remain computationally cheap. The matrices KθK_\theta and MθM_\theta are treated as free N×NN \times N parameters, with no spatial structure imposed a priori. In particular, no Laplacian or locality constraint is enforced. The loss function depends on the first and second derivatives of Ψθ\Psi_\theta, which are computed by automatic differentiation, meaning that the derivatives are obtained analytically by propagating gradients through the layers of the network, rather than by numerical finite differences. This remains efficient since Ψθ\Psi_\theta is one-dimensional.

Training is performed using the Adam optimizer with learning rate 10−510^{-5} and L2L_2 regularization on the encoder and decoder parameters. Each training step uses a batch of independently sampled transitions (hth^t, dhtd h^t). Importantly, the loss is evaluated on single-step transitions and does not require simulating trajectories during training. This makes the procedure significantly cheaper than likelihood-based latent-SDE approaches.

As a validation check, we periodically generate trajectories by encoding an initial condition, evolving the latent dynamics forward in time, and decoding back to the observation space. We monitor the agreement between generated data and empirical distributions of the training data using the symmetric Kullback–Leibler divergence. Training is stopped when this metric saturates.

The model is trained on this dataset using the loss described in section III B. Figure 3 shows the evolution of the loss during training. A notable feature is that the loss does not vanish, and should not. This reflects the stochastic nature of the data: the observed increments dhtd h^t contain an irreducible stochastic contribution, so an exact match between model predictions and observations is not expected.

Figure 3. Training dynamics. The difference between the total loss L\mathcal{L} and the optimal loss Lbest\mathcal{L}_{\mathrm{best}} is shown as a function of the training epoch. Lbest\mathcal{L}_{\mathrm{best}} is computed by evaluating the loss at the exact Cole–Hopf solution. Due to noise fluctuations, Lbest\mathcal{L}_{\mathrm{best}} is not zero.

Generative accuracy

We evaluate the learned model in two complementary ways. First, we test its generative accuracy by comparing trajectories sampled from the model with the empirical KPZ dataset. Second, we inspect the learned transformation and latent operators to determine whether the model has recovered the underlying physical structure.

The first diagnostic is the ensemble-averaged mean height, which grows linearly in time,

⟨ht⟩∼v∞t.(16)\langle h^t \rangle\sim v_{\infty}t. \tag*{(16)}

The second is the height variance, which in the one-dimensional KPZ class grows as

Var⁡(ht)∼t2β,β=13.(17)\operatorname{Var}(h^t) \sim t^{2\beta}, \qquad\beta= \frac{1}{3}. \tag*{(17)}

Finally, we compare the full one-point height distribution. For flat initial conditions, the rescaled height converges in the asymptotic KPZ regime to the GOE Tracy–Widom distribution. We use

h~=σTWh−⟨h⟩Var⁡(h)+μTW,(18)\tilde{h} = \sigma_{\mathrm{TW}} \frac{h-\langle h\rangle}{\sqrt{\operatorname{Var}(h)}} + \mu_{\mathrm{TW}}, \tag*{(18)}

where μTW=−1.2065\mu_{\mathrm{TW}}=-1.2065 and σTW=1.2680\sigma_{\mathrm{TW}}=1.2680 are the mean and standard deviation of the standard GOE Tracy–Widom distribution. For flat KPZ growth one commonly uses the convention σflat=2−2/3σTW\sigma_{\mathrm{flat}}=2^{-2/3}\sigma_{\mathrm{TW}} [48]. This constant factor does not affect the shape of the distribution and here we use the standard GOE convention only to fix the axes of the plots.

Figure 4 shows that the learned model reproduces these KPZ statistics. The trend lines displayed in Figure 4a and Figure 4b are obtained as least squared regression on the Ground Truth points. The generated trajectories exhibit the correct linear growth of the mean, the expected roughening exponent, and a late-time rescaled height distribution close to the GOE Tracy–Widom law. This indicates that the model does not merely fit short-time increments, but also generates trajectories whose long-time statistics remain in the KPZ universality class.

Figure 4. KPZ statistics: generated model trajectories versus ground truth. (a) Ensemble-averaged mean height as a function of time. (b) Height variance, displaying the expected KPZ scaling Var⁡(h)∼t2/3\operatorname{Var}(h) \sim t^{2/3}. (c,d) Distribution of the rescaled height h~\tilde{h} at late times, shown on linear and semi-logarithmic scales. The theoretical reference is the GOE Tracy–Widom law associated with flat initial conditions, which exhibit an asymmetric non-Gaussian decay. We attribute the tail discrepancies to small deviations of the learned components from the true solution.

Recovery of the Cole–Hopf structure

We finally ask whether the model has identified the physical representation underlying the KPZ dynamics. For KPZ, this amounts to recovering the Cole–Hopf transformation and the associated stochastic heat equation.

Figure 5 shows the learned scalar map φθ\varphi_\theta. In semi-logarithmic scale, the learned function is approximately linear, indicating an exponential dependence on hh. A fit gives

Figure 5. Learned change of variables. The learned mapping ∣φθ(h)∣\lvert\varphi_\theta(h)\rvert follows an exponential dependence on hh, in agreement with the analytical Cole–Hopf transform z=eαhz=e^{\alpha h}. The vertical offset corresponds to an arbitrary multiplicative constant in the latent variable, which does not affect the linear latent dynamics.

φθ(h)≃Ceα^h,α^≃0.96,(19)\varphi_\theta(h) \simeq C e^{\widehat{\alpha}h}, \qquad\widehat{\alpha} \simeq0.96, \tag*{(19)}

close to the theoretical value α=λ/(2ν)=1\alpha=\lambda/(2\nu)=1 used in the simulations. The multiplicative constant CC is not identifiable, since rescaling zz by a constant leaves the linear latent dynamics unchanged after a corresponding rescaling of the decoder. Up to this trivial degree of freedom, the model therefore recovers the Cole–Hopf map.

Figure 6 compares the learned drift matrix KθK_\theta with the analytical operator K⋆K^\star corresponding to the discrete stochastic heat equation. The dominant tridiagonal structure of the discrete Laplacian is recovered without imposing any stencil or locality constraint. The learned matrix also captures the periodic boundary conditions. Small deviations from the ideal operator remain and are visible in the difference plot. These errors are consistent with the small discrepancies observed in the far tails of the height distribution, where rare fluctuations are particularly sensitive to the precise latent noise structure. We also show the learnt matrix MθM_\theta in Appendix § 0.3.2. The model autonomously recovers the diagonal structure of the ground truth.

Figure 6. Learned drift matrix KθK_\theta compared with the analytical target K⋆K^\star. Left: learned matrix. Center: analytical discrete diffusion operator. Right: difference Kθ−K⋆K_\theta-K^\star. The first 32×3232 \times32 block is shown. The learned operator recovers the tridiagonal structure of the discrete Laplacian and the periodic boundary conditions, although small residual deviations remain.

Taken together, these results show that the model does not only reproduce KPZ statistics. It also reconstructs explicit, interpretable objects: a local exponential change of variables and a linear latent stochastic operator. These objects are sufficient to infer the underlying stochastic heat equation and, by inversion through the learned decoder, the corresponding KPZ dynamics.

Proof-of-concept extension to a flexible autoencoder

The interpretable latent-SDE model introduced above is deliberately constrained: the change of variables is local and scalar, and the latent dynamics is linear. These assumptions make the learned representation directly readable, and allow us to match the short-time drift and covariance of the observed increments. In this last section, we provide some insights on the question whether the same statistical-learning philosophy could also be useful for a less constrained architecture.

As a proof of concept, we consider a convolutional autoencoder architecture similar in spirit to those used in deterministic latent-linearization approaches [28]. In contrast with the local model of Sec. III A, the encoder and decoder are now general convolutional maps. The latent representation is therefore no longer restricted to be a pointwise change of variables, and the learned coordinates cannot be directly compared with the scalar Cole–Hopf transform. This makes the model less interpretable, but potentially more flexible for systems where the simplifying representation is nonlocal or not known in advance. One could extend the loss formulation of section III B to the multivariate case by using the Itô chain rule for multivariate transformations. This, however, would require the computation of the Hessian matrix of the transform. For discretized spatially extended systems (in our case, 256 sites), this can become computationally expensive. We therefore propose another learning strategy.

In the interpretable model, the latent dynamics has a simple explicit form, which allows us to match the local short-time drift and covariance of the observed increments. For the flexible autoencoder, our goal is different: we do not attempt to extract a readable stochastic equation, but only to test whether statistical constraints improve the quality of generated trajectories. We therefore implement the same learning philosophy at the trajectory level: individual stochastic trajectories are not matched pointwise, while ensemble observables of the generated trajectories are constrained to agree with those of the KPZ dataset.

The model is trained by generating trajectories from the latent dynamics and matching their statistics to those of the KPZ dataset. We keep the notation of Sec. III B: the superscripts model and data distinguish quantities evaluated on the generated trajectories from those measured on the training set, while ⟨⋅⟩\langle\cdot\rangle denotes the average over the NN lattice sites and over the independent noise realizations of the batch. The index t=1,…,Tt=1,\ldots,T runs over the time steps of the rollout.

In addition to the reconstruction loss,

Lrec=1T+1∑t=0T⟨(Ψθ(φθ(ht,data))−ht,data)2⟩,(20)\mathcal{L}_{\mathrm{rec}} = \frac{1}{T+1}\sum_{t=0}^{T}\left\langle\left(\Psi_{\theta}\left(\varphi_{\theta}\left(h^{t,\mathrm{data}}\right)\right)-h^{t,\mathrm{data}}\right)^2\right\rangle, \tag*{(20)}

which has the same meaning as Eq. (9), we impose agreement between generated and empirical observables. The first two terms match the mean height and the height variance already used in Figure 4,

Lmean=1T∑t=1T(⟨ht,model⟩−⟨ht,data⟩)2,(21)\mathcal{L}_{\mathrm{mean}} = \frac{1}{T}\sum_{t=1}^{T}\left(\left\langle h^{t,\mathrm{model}}\right\rangle-\left\langle h^{t,\mathrm{data}}\right\rangle\right)^2, \tag*{(21)}

and

Lvar=1T∑t=1T(Var⁡(ht,model)−Var⁡(ht,data))2,(22)\mathcal{L}_{\mathrm{var}} = \frac{1}{T}\sum_{t=1}^{T}\left(\operatorname{Var}\left(h^{t,\mathrm{model}}\right)-\operatorname{Var}\left(h^{t,\mathrm{data}}\right)\right)^2, \tag*{(22)}

where Var⁡(ht)\operatorname{Var}(h^t) is the height variance, defined in Eq. (17).

We also include a spatial-correlation loss. For a trajectory ensemble, define

Gt(ξ)=⟨(hit−⟨ht⟩)(hi+ξt−⟨ht⟩)⟩,(23)G^t(\xi)=\left\langle\left(h_i^t-\left\langle h^t\right\rangle\right)\left(h_{i+\xi}^t-\left\langle h^t\right\rangle\right)\right\rangle, \tag*{(23)}

with periodic indexing, so that Gt(0)=Var⁡(ht)G^t(0)=\operatorname{Var}(h^t). The correlation loss is

LG=1T∣Ξ∣∑t=1T∑ξ∈Ξ(Gt,model(ξ)−Gt,data(ξ))2,(24)\mathcal{L}_G=\frac{1}{T|\Xi|}\sum_{t=1}^{T}\sum_{\xi\in\Xi}\left(G^{t,\mathrm{model}}(\xi)-G^{t,\mathrm{data}}(\xi)\right)^2, \tag*{(24)}

where Ξ\Xi is the set of spatial lags used during training. The total loss is therefore

Lflex=wrecLrec+wmeanLmean+wvarLvar+wGLG,(25)\mathcal{L}_{\mathrm{flex}}=w_{\mathrm{rec}}\mathcal{L}_{\mathrm{rec}}+w_{\mathrm{mean}}\mathcal{L}_{\mathrm{mean}}+w_{\mathrm{var}}\mathcal{L}_{\mathrm{var}}+w_G\mathcal{L}_G, \tag*{(25)}

with

wrec=0.1,wmean=1,wvar=1,wG=1.(26)w_{\mathrm{rec}}=0.1,\qquad w_{\mathrm{mean}}=1,\qquad w_{\mathrm{var}}=1,\qquad w_G=1. \tag*{(26)}

The training dataset for this setup was generated using λ=2\lambda=2 and dt=0.05dt=0.05.

Figure 7 compares the mean height ⟨ht⟩\langle h^t\rangle and the height variance Var⁡(ht)\operatorname{Var}(h^t) for the ground-truth KPZ trajectories and for trajectories generated by the trained autoencoder. The model captures the linear growth of the mean height and the roughening of the interface, with residual discrepancies at short times and near the end of the rollout.

Figure 7. KPZ statistics: model trajectories versus ground truth. The flexible autoencoder was used for these plots. (a) Mean height and (b) height variance. Dashed lines show the KPZ scalings ⟨h⟩∼t\langle h\rangle\sim t and Var⁡(h)∼t2β\operatorname{Var}(h)\sim t^{2\beta}.

The spatial correlations provide a more stringent test than the first two global moments. As shown in Figure 8, the generated trajectories reproduce the shape and time evolution of the empirical correlation function GtT(ξ)G_{t}^{T}(\xi) with reasonable accuracy. This indicates that the model has captured part of the spatial organization of the KPZ interface, and not only its one-point statistics.

Figure 8. Spatial correlation function GT(ξ)G^{T}(\xi) (Eq. (23)) at final time. The flexible autoencoder model reproduces the spatial correlation length of KPZ.

While this loss formulation is not directly comparable with the one defined in Sec. III B, we think it could be used as a ”proof of concept” for more complex settings. Unlike the physically constrained model, the flexible autoencoder does not provide a directly readable Cole–Hopf map or an explicit stochastic heat operator. However, it shows that statistical losses based on ensemble observables can improve the generative behavior of more expressive architectures. At this early stage, we show that one can encode in the loss relevant statistics of a nonlinear process and produce, in a linear way, new trajectories that follow these key statistics. This suggests that the moment-matching philosophy may be useful beyond the local interpretable setting, especially for stochastic systems where the appropriate simplifying variables are unknown or nonlocal.

Conclusion

We have introduced an interpretable latent-SDE framework for learning changes of variables that transform nonlinear stochastic dynamics into a simple linear stochastic evolution. The model combines a learned local transformation with a linear latent dynamics and is trained through a moment-matching objective based on short-time increments, avoiding trajectory rollouts during optimization.

We validated the approach on the one-dimensional KPZ equation, where the Cole–Hopf transformation provides an exact reference solution. Starting from raw trajectories alone, the model successfully recovers both the local exponential change of variables and the associated linear stochastic operator. The learned representation is therefore not only generative but also interpretable, providing explicit mathematical objects from which the underlying stochastic evolution can be reconstructed.

The generated trajectories reproduce the main statistical signatures of the KPZ universality class, including the linear growth of the mean height, the roughening law Var⁡(ht)∼t2β\operatorname{Var}(h^t) \sim t^{2\beta} with β=1/3\beta= 1/3, and the convergence toward the GOE Tracy–Widom distribution for flat initial conditions. These properties are not imposed during training, but emerge naturally from matching the local statistics of the stochastic dynamics.

The present framework nevertheless has several limitations. It relies on short-time increments and therefore requires data with sufficient temporal resolution. Moreover, the loss constrains only one-step statistics, so small errors may accumulate during long rollouts. Finally, the interpretable model assumes a local scalar change of variables, an assumption well suited to KPZ but potentially too restrictive for more general stochastic field theories. Relatedly, the optimization always returns some transformation, so the loss alone does not certify that a linearizing representation exists. In practice this is assessed a posteriori: if the statistics of the generated trajectories reproduce those of the dataset the learned representation can be considered adequate. A sharper test is to monitor the generative quality while increasing the model expressivity and the size of the training set: a saturation of the performance at a finite error suggests that the mapping is only approximate rather than exact. Even in that case the result remains physically informative, since a simple effective model capturing part of the dynamics is a natural starting point for analytical work.

As a proof of concept, we also investigated a more flexible convolutional autoencoder trained according to the same statistical-learning philosophy, although using trajectory-level observables instead of local increment statistics. While this architecture is no longer interpretable and does not recover the Cole–Hopf transformation explicitly, it reproduces several key KPZ observables with reasonable accuracy. This suggests that statistical moment-matching strategies may also prove useful for more expressive neural architectures.

Schematic representation of the convolutional block used in the flexible autoencoder

Figure 9. Schematic representation of the convolutional block used in the flexible autoencoder.

A one-dimensional convolutional filter is applied locally to the input height profile, followed by a nonlinear activation and pooling. The number of feature maps is increased across successive convolutional layers. The final feature maps are flattened and passed through fully connected layers. A residual connection is added to the output, producing a dimension-preserving transformation from the height field h\mathbf{h} to the latent field z\mathbf{z}.

Future work will explore the application of this autoencoder framework to other equations without known linearizing transformations.

Architecture details for the flexible autoencoder

This appendix describes the implementation details of the proof-of-concept experiment presented in Sec. IV. Its purpose is to investigate whether the statistical-learning philosophy developed for the interpretable latent-SDE model can also improve a more expressive, but non-interpretable, neural architecture.

The encoder and decoder are implemented using one-dimensional convolutional neural networks, a natural choice for spatially extended systems because they exploit locality and weight sharing [49, 50]: each output depends only on a small neighborhood of the input, and the same filter is applied at every spatial position. Convolutions are also a natural fit. for physical systems, since a small kernel can represent local differential operators directly. The architecture is inspired by the deterministic latent-linearization framework of Ref. [51].

Let

ht=(h1t,…,hNt)(27)h^t = (h^t_1,\ldots,h^t_N) \tag*{(27)}

denote the discretized KPZ height profile at time tt, as in Sec. III A. The encoder

φθ:RN→RN(28)\varphi_\theta: \mathbb{R}^{N} \to\mathbb{R}^{N} \tag*{(28)}

maps the height field to a latent field,

zt=φθ(ht),(29)z^t = \varphi_\theta(h^t), \tag*{(29)}

while the decoder

Ψθ:RN→RN(30)\Psi_\theta: \mathbb{R}^{N} \to\mathbb{R}^{N} \tag*{(30)}

maps the latent field back to the height representation,

Ψθ(zt)=Ψθ(φθ(ht)).(31)\Psi_\theta(z^t) = \Psi_\theta(\varphi_\theta(h^t)). \tag*{(31)}

Unlike in Sec. III A, φθ\varphi_\theta and Ψθ\Psi_\theta now act on the whole profile rather than site by site, so the change of variables is no longer local. As in the interpretable model, the latent dimension is chosen equal to the number of spatial lattice sites. This choice, also tested in [51], avoids the information loss imposed by a dimensional reduction and is compatible with an invertible change of variables. To generate trajectories, we initialize the latent field by encoding the flat initial condition,

z0,model=φθ(h0),h0=0,(32)z^{0,\mathrm{model}} = \varphi_\theta(h^0), \qquad h^0 = 0, \tag*{(32)}

and evolve it with the same stochastic linear update as used in the other model [Eq. (8)]. The resulting architecture is more flexible, but no longer directly interpretable.

The generated height trajectories are finally obtained by decoding the latent rollout,

ht,model=Ψθ(zt,model).(33)h^{t,\mathrm{model}} = \Psi_\theta(z^{t,\mathrm{model}}). \tag*{(33)}

Unlike the interpretable architecture, the purpose of this model is not to recover the Cole–Hopf transformation itself, but to assess whether the statistical-learning strategy developed in the main text remains beneficial with an architecture capable of representing more complex transforms.

The learned coupling matrix M

In the main text we reported only the drift matrix KθK_\theta. For completeness, Figure 10 shows the noise coupling matrix MθM_\theta. The model autonomously recovers the diagonal structure of the noise covariance. Because MθM_\theta multiplies a noise term that is symmetric around zero, Mθ,iiM_{\theta,ii} and −Mθ,ii-M_{\theta,ii} are indistinguishable, and both are valid solutions of the optimization problem.

Figure 10. Learned diffusion matrix MθM_\theta compared with the analytical target M⋆M^\star. Left: learned matrix. Center: analytical discrete diffusion operator. Right: difference M⋆−∣Mθ∣M^\star- \lvert M_\theta\rvert. The first 32×3232 \times32 block is shown. The learned operator recovers the diagonal structure of the Identity matrix, although small residual deviations remain.

Acknowledgments

Acknowledgments This project has received funding from the European Union’s Horizon Europe research and innovation programme under the Marie Skłodowska-Curie COFUND grant agreement No 101127936 (DeMythif.AI) and by the French ANR grant Scalp (ANR-24-CE23-1320).

References

  1. [4]Kardar M, Parisi G and Zhang Y C 1986 Phys. Rev. Lett. 56 889–892
  2. [5]Kardar M 2007 Statistical physics of fields (Cambridge University Press)
  3. [6]Takeuchi K A and Sano M 2010 Phys. Rev. Lett. 104 230601
  4. [7]Takeuchi K A, Sano M, Sasamoto T and Spohn H 2011 Scientific Reports 1 34
  5. [8]Fontaine Q, Squizzato D, Baboux F, Amelio I, Lemaître A, Morassi M, Sagnes I, Le Gratiet L, Harouri A, Wouters M, Ravets S, Bloch J, Minguzzi A, Bramati A, Carusotto I, Galopin E and Amo A 2022 Nature 608 687–691
  6. [9]Halpin-Healy T and Takeuchi K A 2015 J. Stat. Phys. 160 794–814
  7. [10]Takeuchi K A 2018 Physica A 504 77–105
  8. [11]Vicsek T, Cserző M and Horváth V K 1990 Physica A: Statistical Mechanics and its Applications 167 315–321
  9. [12]Bohr T, Jensen M H, Paladin G and Vulpiani A 1998 Dynamical systems approach to turbulence
  10. [13]Hopf E 1950 Communications on Pure and Applied Mathematics 3 201–230
  11. [14]Cole J D 1951 Quarterly of Applied Mathematics 9 225–236
  12. [15]Bertini L and Giacomin G 1997 Communications in Mathematical Physics 183 571–607
  13. [16]Hairer M 2013 Annals of Mathematics 178 559–664
  14. [17]Chen R T, Rubanova Y, Bettencourt J and Duvenaud D K 2018 Neural ordinary differential equations Advances in Neural Information Processing Systems vol 31arxiv.org/abs/1806.07366
  15. [18]Tzen B and Raginsky M 2019 arXiv:1905.09883arxiv.org/abs/1905.09883
  16. [19]Li X, Wong T K L, Chen R T Q and Duvenaud D 2020 Scalable gradients for stochastic differential equations Proc. AISTATS
  17. [20]Kidger P, Foster J, Li X and Lyons T J 2021 Neural SDEs as infinite-dimensional GANs Proceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research vol 139) (PMLR) pp 5453–5463arxiv.org/abs/2102.03657
  18. [21]Salvi C, Lเมercier M and Gerasimovics A 2022 Neural stochastic PDEs: Resolution-invariant learning of continuous spatiotemporal dynamics Advances in Neural Information Processing Systems vol 35 pp 1333–1344arxiv.org/abs/2110.10249
  19. [22]Brunton S L, Proctor J L and Kutz J N 2016 Proceedings of the National Academy of Sciences 113 3932–3937
  20. [23]Rudy S H, Brunton S L, Proctor J L and Kutz J N 2017 Science Advances 3 e1602614
  21. [46]Friedrich R and Peinke J 1997 Phys. Rev. Lett. 78 863–866
  22. [47]Siegert S, Friedrich R and Peinke J 1998 Physics Letters A 243 275–280
  23. [48]Prähofer M and Spohn H 2000 Phys. Rev. Lett. 84 4882–4885
  24. [49]LeCun Y, Bottou L, Bengio Y and Haffner P 1998 Proceedings of the IEEE 86 2278–2324
  25. [50]Cacciari I and Ranfagni A 2024 Applied Sciences (2076-3417) 14
  26. [51]Gin C, Lusch B, Brunton S L and Kutz J N 2021 European Journal of Applied Mathematics 32 515–539

Paper details

vlnc:W_447zjuxe
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents