betaFeedback

WorksPhysics and Astronomy

Learning noisy phase transition dynamics from stochastic partial differential equations

Abstract

The non-equilibrium dynamics of mesoscale phase transitions are fundamentally shaped by thermal fluctuations, which not only seed instabilities but actively control kinetic pathways, including rare barrier-crossing events such as nucleation that are entirely inaccessible to deterministic models. Machine-learning surrogates for such systems must therefore represent stochasticity explicitly, enforce conservation laws by construction, and expose physically interpretable structure. We develop physics-aware surrogate models for the stochastic Cahn-Hilliard equation in 3D that satisfy all three requirements simultaneously. The key innovation is to parameterize the surrogate at the level of inter-cell fluxes, decomposing each flux into a deterministic mobility-weighted chemical-potential gradient and a learnable noise amplitude. This design guarantees exact mass conservation at every step and adds physical fluctuations to inter-cell mass transport. A learnable free energy functional provides thermodynamic interpretability, validated by independent recovery of the bulk double-well landscape, interfacial excess energy, and curvature-independent interfacial tension. Tests demonstrate accurate reproduction of ensemble statistics and noise-accelerated coarsening, with generalization to spatial domains 64 times larger in volume and temporal horizons 160x longer than those seen during training. Critically, the stochastic surrogate captures thermally activated nucleation in the metastable regime, a qualitative capability that no deterministic surrogate can provide regardless of training, thus establishing flux-level stochasticity as an architectural necessity rather than an optional enhancement.

Introduction

Non-equilibrium phase transitions are central to materials science and chemistry, governing the formation, transformation, and stability of matter across a wide range of conditions and compositions. A decisive, and often underappreciated, factor in these processes is stochasticity: thermal fluctuations and atomistic perturbations that are not mere background noise but are fundamental drivers of kinetic pathways and emergent phenomena. In spinodal decomposition, an initially uniform, thermodynamically unstable mixture spontaneously separates into coexisting phases through continuous amplification of composition fluctuations at all wavelengths [1]; stochastic noise not only seeds this instability but actively accelerates coarsening and sets the characteristic length scale of the resulting microstructure. In nucleation and precipitation from a metastable state, a thermodynamic barrier must be surmounted by a rare thermal fluctuation—a process entirely absent in a deterministic description—and the crossing probability is exponentially sensitive to the noise amplitude, as captured by classical nucleation theory [2]. Understanding and accurately reproducing such noise-driven dynamics is therefore not optional but required for quantitative mesoscale modeling, with implications ranging from the design of advanced alloys and polymer blends to self-organization in soft matter and biological assemblies.

Computational methods for mesoscale phase transitions can be classified by how they represent stochasticity. Particle-based methods, such as molecular dynamics (MD) [3] and kinetic Monte Carlo (KMC) [4], resolve thermal fluctuations explicitly through individual particle trajectories. They are physically complete but prohibitively expensive at the spatiotemporal scales relevant to microstructural evolution. Deterministic phase-field models [1, 5] overcome the mesoscale barrier through a continuum order-parameter description, but entirely suppress stochasticity and consequently miss fluctuation-driven phenomena such as noise-accelerated coarsening, the stochastic spread of nucleation times, and the sensitivity of pattern selection to early-time fluctuations. Stochastic partial differential equations (SPDEs), of which the stochastic Cahn–Hilliard equation (CHE) [6, 7] is the canonical mesoscale example, offer a more faithful description by coupling a deterministic driving force with a noise field whose statistics are grounded in the fluctuation-dissipation theorem. However, SPDEs remain computationally expensive at scales relevant to microstructural evolution and require a priori specification of noise statistics that may not be readily available outside of well-characterized model systems.

The rapid development of machine-learning (ML) surrogate models for physical simulations offers an attractive path toward cost-effective mesoscale dynamics, but existing approaches are largely restricted to the deterministic regime [8, 9, 10, 11, 12, 13]. Physics-informed neural networks [14, 15, 16] and residual-based methods can incorporate stochastic forcing terms in principle, but the optimization process effectively averages over noise realizations, yielding a smooth mean-field solution rather than a genuine stochastic ensemble. Operator-learning architectures [17, 18] and sequence models [19, 20, 21] trained on stochastic trajectories learn state-to-state transition maps that approximate the conditional mean of the next state; they implicitly absorb stochasticity as part of the prediction target rather than modeling it explicitly, and provide no mechanism for enforcing conservation laws during roll-out [21]. Generative approaches—diffusion models [22, 23] and normalizing flows [21]—introduce randomness directly at the level of the predicted state field, enabling diverse trajectory sampling, but they decouple noise from its physical origin in inter-cell transport. As a consequence, even when trained on strictly conserved data, state-level generative models do not guarantee that individual samples conserve mass, because the injected noise is not constrained to arise from a divergence-free flux.

Related works in chemistry and biophysics addresses the inverse problem of learning effective stochastic dynamics, including potential of mean force (PMF) and the diffusion coefficient, from single-molecule or particle-tracking trajectories [24, 25]. The present work deals with an analogous identification problem of learning free energy, mobility, and noise amplitude from stochastic phase-field trajectories, though in a qualitatively different context. The concentration field c(x,t)c(\mathbf{x},t) plays the role of the collective variable, yet it is either a continuous field (function) or discretized at a very high-dimensional mesh rather than atomic coordinates. Consequently, the learned objects are field-level functionals (free energy A[c]A[c], mobility field MijM_{ij}, noise amplitude BijB_{ij}) rather than functions of coordinates.

The above literature survey points to three requirements for a physically faithful surrogate for stochastic conserved microstructure dynamics. First, stochasticity must be represented explicitly and at the correct scale: randomness should be introduced at the level of inter-cell fluxes, where it has a well-defined physical meaning tied to microscopic fluctuations of current density, rather than at the level of the predicted concentration field. Second, mass conservation must be guaranteed by construction during every step of an autoregressive rollout, not merely encouraged through a soft loss term. Third, the surrogate should encode physically interpretable structure: the free energy landscape, mobility, and noise amplitude that govern the dynamics should be learnable as distinct, explainable modules rather than absorbed into an opaque black-box map. No existing ML surrogate framework fully satisfies all three requirements simultaneously.

In this work we address this gap by developing a conservative and stochastic surrogate for the stochastic Cahn–Hilliard equation, chosen as a prototypical and well-characterized model of noise-driven conserved dynamics. The key architectural innovation is to parameterize the surrogate at the level of inter-cell fluxes rather than states. At each autoregressive step, the predicted concentration update is computed as the divergence of a learned flux field comprising a deterministic mobility-weighted chemical-potential gradient and a stochastic noise term, directly mirroring the Bronchart–Finel discretization of the CHE [7]. Mass conservation is guaranteed by the antisymmetry of the learned flux, Fij=−FjiF_{ij}=-F_{ji}, enforced by construction rather than by a penalty term. The chemical potential is obtained by automatic differentiation of a learnable scalar free energy functional, encoding the correct thermodynamic structure while learning the free energy landscape entirely from trajectory data. Mobility and noise amplitude are parameterized as symmetric functions of learned latent cell representations, sharing the chemical-potential backbone. We evaluate the framework across model variants obtained by ablating the free energy constraint and the stochastic flux component, enabling a controlled analysis of each physical ingredient.

The scope of the present work is deliberately a study of model design and architectural feasibility, or the inner loop of a broader surrogate modeling program. Our primary training and test regime is spinodal decomposition, in which the mean concentration lies within the thermodynamically unstable region (∣⟨c⟩∣<0.5|\langle c\rangle| < 0.5 spinodal; ∣⟨c⟩∣<0.85|\langle c\rangle| < 0.85 solubility limit). This choice is intentional: in the spinodal regime, the SPDE generates abundant, information-rich training trajectories, and the physical effects of stochasticity are immediate and direct: noise-accelerated coarsening dynamics and inter-trajectory divergence. Spinodal decomposition therefore provides stringent tests of the architecture. Nucleation and precipitation from metastable initial conditions are qualitatively harder: the rare-event nature of nucleation means that training data near the barrier is intrinsically sparse, and reproducing the full distribution of nucleation waiting times requires sufficient exposure to barrier-crossing configurations. We include a preliminary extrapolation test near the metastable boundary (⟨c⟩=−0.51\langle c\rangle= -0.51, just outside the spinodal), which demonstrates that the learned free-energy based surrogate can capture mean nucleation behavior in near-boundary cases, while also revealing that, as a direct consequence of training data concentrated in the spinodal region, the predicted distribution of nucleation times is narrower than the ground truth. Extending surrogate performance into the deeply metastable precipitation regime requires a qualitatively different outer loop: systematic data curation and active learning strategies that adaptively sample rare-event trajectories and expand the training distribution toward and beyond the solubility limit. Such a data-curation strategy, while critical for a comprehensive and quantitative study across the whole concentration spectrum, is beyond the architectural questions addressed here and is deferred to future work.

Numerical experiments in one and three spatial dimensions demonstrate that the stochastic, free energy constrained variant (FE-MV) accurately reproduces ensemble statistics, noise-accelerated coarsening dynamics, and spatial microstructure on the training grid, and generalizes to spatial domains 4x larger in linear dimension (64x larger volume) than those seen during training. The learned free energy function correctly recovers the double-well bulk energy landscape, interfacial excess energy, and the curvature-independence of the ground-truth interfacial tension—providing stringent, interpretable validation of the thermodynamic consistency of the surrogate. The remainder of this paper is organized as follows. § 2 introduces the stochastic CHE, the flux-preserving surrogate architecture, and the training procedure. § 3 analyzes the noise-driven dynamics of the reference SPDE system in one and three dimensions. § 4 evaluates surrogate model performance, covering in-distribution inference, spatial extrapolation, free energy verification, and concentration-space extrapolation. § 5 summarizes findings and outlines directions for future work.

Methods

Stochastic Cahn–Hilliard equation

The stochastic CHE for the concentration field c(x,t)c(x,t) is

∂c∂t=−∇⋅f=−∇⋅(M∇μ+σdWdt),(1)\frac{\partial c}{\partial t}=-\nabla\cdot\mathbf{f}=-\nabla\cdot\left(M\nabla\mu+\sigma\frac{\mathrm{d}\mathbf{W}}{\mathrm{d}t}\right), \tag*{(1)}

where f\mathbf{f} is the concentration flux, MM is the mobility, σ\sigma is the noise amplitude, and dW/dt\mathrm{d}\mathbf{W}/\mathrm{d}t is a vector-valued white noise field acting on the fluxes. The chemical potential μ\mu is the functional derivative of the total free energy A[c]A[c]:

μ=δA[c]δc,(2)\mu=\frac{\delta A[c]}{\delta c}, \tag*{(2)}
A=Ab[c]+Ai[c]=∫(ab+ai) dV,(3)A=A_{\mathrm{b}}[c]+A_{\mathrm{i}}[c]=\int(a_{\mathrm{b}}+a_{\mathrm{i}})\,\mathrm{d}V, \tag*{(3)}

with bulk and interfacial free energy densities

ab=c44−c22+T(1+c2log⁡1+c2+1−c2log⁡1−c2),(4)a_{\mathrm{b}}=\frac{c^{4}}{4}-\frac{c^{2}}{2}+T\left(\frac{1+c}{2}\log\frac{1+c}{2}+\frac{1-c}{2}\log\frac{1-c}{2}\right), \tag*{(4)}
ai=κ2∣∇c∣2,(5)a_{\mathrm{i}}=\frac{\kappa}{2}\lvert\nabla c\rvert^{2}, \tag*{(5)}

parameterized by temperature TT and gradient-penalty stiffness κ\kappa. As shown in Figure 1, the double-well structure of ab(c)a_b(c) at T<1T<1 drives phase separation, with spinodal boundaries at ∣c∣=0.5\lvert c\rvert=0.5 and equilibrium solubility limits at ∣c∣≈0.85\lvert c\rvert\approx0.85 for T=0.2T=0.2. We solve Eq. (1) numerically following the coarse-grained lattice discretization of Bronchart et al. [7], using Euler–Maruyama time integration on a regular mesh. The flux between neighboring cells ii and jj is

Figure 1. Bulk free energy and chemical potential at T=0.2T = 0.2. The concentrations of solubility (±0.85\pm0.85) and spinodal (±0.5\pm0.5) are highlighted.

fij=DijT(μj−μi)δt+(2Dij)1/2δt ε(6)f_{ij}=\frac{D_{ij}}{T}\left(\mu_j-\mu_i\right)\delta t+\left(2D_{ij}\right)^{1/2}\sqrt{\delta t}\,\varepsilon \tag*{(6)}

where Dij=Dji=(12∣1−ci2∣∣1−cj2∣+0.5)(1+0.1T)D_{ij}=D_{ji}=\left(\frac{1}{2}\lvert1-c_i^2\rvert\lvert1-c_j^2\rvert+0.5\right)(1+0.1T) is a symmetric, concentration-dependent diffusivity derived from the fluctuation-dissipation theorem [7], ξ∼N(0,1)\xi\sim\mathcal{N}(0,1) is drawn independently for each ordered pair (i,j)(i,j), and ε\varepsilon is an adjustable noise amplitude hyperparameter. The factor 2Dij\sqrt{2D_{ij}} in the stochastic term ensures that the variance of the discrete flux is proportional to the local diffusivity, consistent with the fluctuation-dissipation relation at the lattice level. The interfacial contribution to the chemical potential is discretized as

μii=−S∑Δr(ci+Δr+ci−Δr−2ci)(7)\mu_i^{\mathrm{i}}=-S\sum_{\Delta r}\left(c_{i+\Delta r}+c_{i-\Delta r}-2c_i\right) \tag*{(7)}

where the sum runs over the neighbor lattice directions and the constant S=0.002S=0.002 is fixed throughout. Mass conservation follows from the antisymmetry constraint fji=−fijf_{ji}=-f_{ij}: each undirected bond is visited exactly once, so ∑ici\sum_i c_i is conserved at every step. The concentration update is

ci(t+δt)=ci(t)−∑j∈N(i)fij(8)c_i(t+\delta t)=c_i(t)-\sum_{j\in\mathcal{N}(i)}f_{ij} \tag*{(8)}

where N(i)\mathcal{N}(i) denotes all neighbors jj for which fijf_{ij} is computed with ii as the source cell.

Flux-preserving surrogate model

Overall structure. We develop autoregressive neural network surrogates that predict the concentration field at a coarsened time step Δt≫δt\Delta t\gg\delta t, skipping N=Δt/δtN=\Delta t/\delta t solver steps per surrogate evaluation and thereby achieving computational speedup proportional to NN. Mirroring the structure of Eq. (8), the concentration is updated as a flux divergence:

c^i(t+Δt)=ci(t)−∑j∈N(i)Fθ,ij(9)\hat{c}_i(t+\Delta t)=c_i(t)-\sum_{j\in\mathcal{N}(i)}F_{\theta,ij} \tag*{(9)}

where Fθ,ij=−Fθ,jiF_{\theta,ij}=-F_{\theta,ji} is the learned inter-cell flux. This antisymmetric constraint is enforced by construction (see below), guaranteeing exact mass conservation during roll-out.

Flux decomposition. The learned flux mirrors the structure of the SPDE flux in Eq. (6), decomposing into a deterministic mean and a stochastic noise component:

Fθ,ij=M^θ1,ij(μ^θ2,i−μ^θ2,j)Δt+B^θ3,ij2Δt ξ(10)F_{\theta,ij}=\hat{M}_{\theta_1,ij}\left(\hat{\mu}_{\theta_2,i}-\hat{\mu}_{\theta_2,j}\right)\Delta t+\hat{B}_{\theta_3,ij}\sqrt{2\Delta t}\,\xi \tag*{(10)}

where M^θ1,ij\hat{M}_{\theta_1,ij} is the learned mobility, μ^θ2\hat{\mu}_{\theta_2} is the learned chemical potential, B^θ3,ij\hat{B}_{\theta_3,ij} is the learned noise amplitude, and ξ∼N(0,1)\xi\sim\mathcal{N}(0,1). The 2Δt\sqrt{2\Delta t} scaling of the stochastic term reflects the incoherent accumulation of NN independent solver-step noise contributions: the standard deviation of the summed noise scales as N⋅2Dij εδt=2Dij εΔt\sqrt{N}\cdot\sqrt{2D_{ij}\,\varepsilon}\sqrt{\delta t}=\sqrt{2D_{ij}\,\varepsilon}\sqrt{\Delta t}, so B^θ3,ij\widehat{B}_{\theta_3,ij} absorbs the mobility-dependent prefactor and is learned from data without being pre-specified. The network parameters θ={θ1,θ2,θ3}\theta=\{\theta_1,\theta_2,\theta_3\} are trained end-to-end.

Chemical potential and free energy module. We propose two variants for the chemical potential surrogate μ^θ2\widehat{\mu}_{\theta_2}. In the first, a neural network directly predicts the chemical potential field from the concentration and temperature, μ^θ2=F^θ2(c,T)\widehat{\mu}_{\theta_2}=\widehat{F}_{\theta_2}(c,T), without imposing any thermodynamic structure. In the second, which we term the free energy (FE) variant, the network instead learns a scalar free-energy functional:

E^θ2, h^θ2=F^θ2(c,T),(11)\widehat{E}_{\theta_2},\ \widehat{h}_{\theta_2}=\widehat{F}_{\theta_2}(c,T), \tag*{(11)}
μ^θ2=∂E^θ2∂c,(12)\widehat{\mu}_{\theta_2}=\frac{\partial\widehat{E}_{\theta_2}}{\partial c}, \tag*{(12)}

where h^θ2∈Rd\widehat{h}_{\theta_2}\in\mathbb{R}^{d} is a latent encoding of the local thermodynamic state and μ^θ2\widehat{\mu}_{\theta_2} is obtained by automatic differentiation through E^θ2\widehat{E}_{\theta_2}. This formulation embeds the correct thermodynamic structure—the chemical potential is by construction the gradient of a scalar potential—and exposes the learned free energy landscape as an independently verifiable quantity. A ConvNeXt backbone [26] is used in all cases to map the concentration field and temperature to E^θ2\widehat{E}_{\theta_2} and h^θ2\widehat{h}_{\theta_2}.

Mobility and noise amplitude modules. The mobility and noise amplitude are modeled as functions of the latent representations of the two cells sharing a diffusion edge (i,j)(i,j), together with the inter-cell distance dijd_{ij}:

Mθ1,ij=Ψθ1(h^θ2,i,h^θ2,j,dij),(13)M_{\theta_1,ij}=\Psi_{\theta_1}\left(\widehat{h}_{\theta_2,i},\widehat{h}_{\theta_2,j},d_{ij}\right), \tag*{(13)}
M^θ1,ij=12(Mθ1,ij+Mθ1,ji),(14)\widehat{M}_{\theta_1,ij}=\frac{1}{2}\left(M_{\theta_1,ij}+M_{\theta_1,ji}\right), \tag*{(14)}

and analogously for the noise amplitude:

Bθ3,ij=Ψθ3(h^θ2,i,h^θ2,j,dij),(15)B_{\theta_3,ij}=\Psi_{\theta_3}\left(\widehat{h}_{\theta_2,i},\widehat{h}_{\theta_2,j},d_{ij}\right), \tag*{(15)}
B^θ3,ij=12(Bθ3,ij+Bθ3,ji).(16)\widehat{B}_{\theta_3,ij}=\frac{1}{2}\left(B_{\theta_3,ij}+B_{\theta_3,ji}\right). \tag*{(16)}

The explicit symmetrization enforces M^ij=M^ji\widehat{M}_{ij}=\widehat{M}_{ji} and B^ij=B^ji\widehat{B}_{ij}=\widehat{B}_{ji}, which combined with the antisymmetric flux in Eq. (10) ensures that Eq. (9) conserves mass exactly at every step. Both Ψθ1\Psi_{\theta_1} and Ψθ3\Psi_{\theta_3} are multi-layer perceptrons (MLPs). The distance dijd_{ij} is provided as an input because each surrogate step spans many solver steps, so the effective transport range can extend beyond the nearest-neighbor shell; including dijd_{ij} allows the MLP to modulate contributions from bonds at different ranges. We include bonds up to the third-neighbor shell in the present study; the dependence of surrogate accuracy on this cutoff as a function of Δt/δt\Delta t/\delta t is an open question left for future work.

Model variants and baseline. We study four structured variants obtained by switching the free energy formulation and the stochastic flux independently. FE-MV (free energy, mean+variance) uses the FE variant for the chemical potential and the full flux in Eq. (10); this is the primary proposed method. FE-M (free energy, mean only) uses the FE variant but retains only the deterministic flux mean, setting B^θ3,ij=0\hat{B}_{\theta3,ij}=0 during both training and roll-out. nonFE-MV (non-FE, mean+variance) uses the full stochastic flux but replaces the AD-derived chemical potential with a direct network prediction. nonFE-M (non-FE, mean only) combines direct chemical-potential prediction with the deterministic flux mean only. As an additional baseline, we include a black-box surrogate non-flux, in which a single network Φθ\Phi_\theta directly predicts the concentration update:

c^i(t+Δt)=ci(t)+Φθ,i(c(t)).(17)\hat{c}_i(t+\Delta t)=c_i(t)+\Phi_{\theta,i}(c(t)). \tag*{(17)}

Unlike the four structured variants, non-flux does not decompose the update into a flux divergence, and therefore neither enforces mass conservation nor exposes physically interpretable modules. A summary can be found in Table 1. The general framework is illustrated in Figure 2.

Figure 2. Flux-preserved training and sampling framework.

FE-MVFE-MnonFE-MVnonFE-Mnon-flux
modelFree energy, mean+varianceFree energy, mean onlyNon-FE, mean+varianceNon-FE, mean onlyNon-flux black-box
Conserv.YesYesYesYesNo
Value updatec^i(t+Δt)=ci(t)−∑jFij\hat{c}_i(t+\Delta t)=c_i(t)-\sum_j F_{ij}samesamesamec^i(t+Δt)=ci(t)+Φθ,i(c(t))\hat{c}_i(t+\Delta t)=c_i(t)+\Phi_{\theta,i}(c(t))
FijF_{ij} eval.M^ij(μ^i−μ^j)Δt+B^ij2Δt ξ\hat{M}_{ij}(\hat{\mu}_i-\hat{\mu}_j)\Delta t+\hat{B}_{ij}\sqrt{2\Delta t}\,\xisame but B^=0\hat{B}=0same as FE-MVsame but B^=0\hat{B}=0N/A
Chemical potentialautograd μ^=∂E^θ2/∂c\hat{\mu}=\partial\hat{E}_{\theta_2}/\partial csamedirect F^θ2(c,T)\hat{F}_{\theta_2}(c,T)sameN/A
StochasticYesNoYesNoNo
Free en.YesYesNoNoNo
PrimaryYesNoNoNoNo

Table 1. Comparison of surrogate model variants.

Data generation and model training

SPDE simulations. All SPDE simulations are run at fixed temperature T=0.2T=0.2. The primary results use noise level ε=0.2\varepsilon=0.2 unless otherwise noted. Training trajectories are generated from 100 independent SPDE trajectories on a 16316^3 mesh. Each is initialized with c0=⟨c⟩+u0c_0=\langle c\rangle+u_0, including tiny initial perturbations u0∼U(−0.01,0.01)u_0\sim\mathcal{U}(-0.01,0.01) to start the demixing and a mean concentration ⟨c⟩∼U(−0.5,0.5)\langle c\rangle\sim\mathcal{U}(-0.5,0.5) chosen randomly within the spinodal region (∣c0∣<0.5\lvert c_0\rvert< 0.5). Therefore all training trajectories correspond to spinodal decomposition. They are integrated for Nt=24,000N_t = 24{,}000 solver steps at δt=10−3\delta t = 10^{-3}, and frames are saved every nt=100n_t = 100 steps, yielding 240 frames per trajectory, of which frames 25 through 240 are used in training. A validation set, 10 trajectories randomly split from the training set, was used for choosing the best surrogate model but otherwise not presented. Finally, separate test sets were obtained in much longer rollouts to assess the long-term stability and accuracy of the chosen model. These tests will be discussed in detail.

Training procedure. The model is trained on an NVIDIA H100 GPU using the AdamW optimizer [27] with a learning rate of 10−410^{-4} and a batch size of 32. To encourage temporal consistency and suppress roll-out error accumulation, the training loss is evaluated over five consecutive autoregressive steps, from c^t+1\hat{c}_{t+1} through c^t+5\hat{c}_{t+5} conditioned on ctc_t. To ensure numerical stability of the stochastic models, the output of the noise amplitude module is passed through a sigmoid activation function, which bounds B^θ3,ij\hat{B}_{\theta_3,ij} from above at the level corresponding to ε=1\varepsilon= 1 in the reference SPDE. This prevents the learned noise amplitude from drifting to unphysically large values during early training, where the NLL loss is sensitive to the variance prediction.

Loss functions. The deterministic variants non-flux, nonFE-M, and FE-M are trained with the mean squared error (MSE) loss:

LMSE=1Ns∑i=1Ns∥ci−c^i∥22,(18)\mathcal{L}_{\mathrm{MSE}} = \frac{1}{N_s}\sum_{i=1}^{N_s}\lVert c_i-\hat{c}_i\rVert_2^2, \tag*{(18)}

where ii indexes grid cells and cic_i, c^i\hat{c}_i are the ground-truth and predicted concentrations, respectively. The stochastic variants nonFE-MV and FE-MV are trained with the Gaussian negative log-likelihood (NLL):

LNLL=1Ns∑i=1Ns[log⁡v^i+(ci−c^i)22v^i],(19)\mathcal{L}_{\mathrm{NLL}} = \frac{1}{N_s}\sum_{i=1}^{N_s}\left[\log\hat{v}_i+\frac{(c_i-\hat{c}_i)^2}{2\hat{v}_i}\right], \tag*{(19)}

where v^i\hat{v}_i is the predicted variance of the concentration update at cell ii, obtained by summing the independent edge contributions:

v^i=∑⟨ij⟩2Δt B^θ3,ij2.(20)\hat{v}_i = \sum_{\langle ij\rangle} 2\Delta t\,\hat{B}_{\theta_3,ij}^{2}. \tag*{(20)}

This expression assumes that the noise contributions on different edges are statistically independent, which matches the ground-truth SPDE discretization in Eq. (6) where a separate ξ\xi is drawn for each bond. The resulting diagonal covariance structure is the natural leading-order approximation; off-diagonal correlations between edges could in principle arise from multi-step dynamics within a single surrogate step, but are neglected here as a simplifying assumption. The NLL loss simultaneously optimizes the predicted mean c^i\hat{c}_i and the predicted variance v^i\hat{v}_i, penalizing both systematic bias and miscalibrated uncertainty.

Results on simulating stochastic CHE

In this section, we present a brief demonstration of the SPDE results before delving into surrogate models in the next one.

One-dimensional case

We qualitatively analyzed the stochastic dynamics by a one-dimensional simulation result of CHE systems. We run simulations in a periodic one-dimensional system with Nx=256N_x = 256 grid points and δt=10−3\delta t = 10^{-3}. The concentration field cc was initialized with average ⟨c⟩=0\langle c\rangle= 0 and random uniform noise from U(−0.01,0.01)\mathcal{U}(-0.01, 0.01). We compare the results without noise ε=0\varepsilon= 0 and with noise ε=0.2\varepsilon= 0.2. In Figure 3, the subplot (a) shows the spatial value evolution versus time. It has a clear trend that adding noise will make the trajectory much noisy. Subplot (b) shows the standard deviation (std) of all voxel values in a frame versus time. The noisy trajectory has a higher std overall, suggesting that the presence of thermal fluctuations accelerates the demixing dynamics. This is seen more quantitatively from the representative frames at different time stamps (subplot (c) of Figure 3). More results corresponding to one-dimensional analysis for different grid size and grid size c0=0.4+U(−0.01,0.01)c_0 = 0.4 + \mathcal{U}(-0.01, 0.01) is presented in the appendix § 12.§ 1 and B.2.

Figure 3. Results of 1D CHE initialized with c∼0+U(−0.01,0.01)c \sim0 + \mathcal{U}(-0.01, 0.01). (a) Time evolution of cc without and with noise. (b) Standard deviation of cc over time. (c) Representative frames without and with noise.

The rest of the paper will focus on 3D results.

Three-dimensional case

3D simulations were performed in similar settings except a 3D periodic grid. Step-size convergence tests are shown in the appendix, B.3. Here we consider two representative examples, 50:50 mixture (⟨c⟩=0\langle c\rangle= 0) with large, interconnected domains of the two immiscible phases, and 70:30 mixture (⟨c⟩=0.4\langle c\rangle= 0.4) which demixes many disconnected droplets of the minority phase merging into larger droplets, i.e. Ostwald ripening.

In the first case, given the same initial configuration ⟨c⟩=0\langle c\rangle= 0 including slight perturbations drawn from U(−0.01,0.01)\mathcal{U}(-0.01, 0.01), the SPDE with noise level ε=0.2\varepsilon= 0.2 in three independent runs leads to different trajectories due to stochasticity, as shown Figure 4. The clear differences justify our need of building probabilistic machine learning surrogates. The trajectories demonstrate typical demixing behaviors of spinodal decomposition. Additionally, supplementary Figure S6 in Sec. B.4 shows the effect of noise on the evolution of the SPDE system. Finite noise (ε=0.2\varepsilon= 0.2) notably accelerates the evolution of the system compared to deterministic ε=0\varepsilon= 0 in 3D, similar to 1D.

Figure 4. Example 3D SPDE trajectories with 16316^{3} grid of 50:50 mixture (⟨c⟩=0\langle c\rangle= 0) with ε=0.2\varepsilon= 0.2 starting from the same initial condition.

Image not extracted

Figure S6. Top: The evolution of mean ±\pm standard deviation of voxel values of 10 random SPDE samples with respect to different stochastic noise levels, ε=0\varepsilon= 0 and ε=0.2\varepsilon= 0.2. Middle: The corresponding histogram of voxel values. Bottom: The concentration frames from SPDE ε=0\varepsilon= 0 and ε=0.2\varepsilon= 0.2 Left: Initial concentration field c0=U(−0.01,0.01)c_0 = \mathcal{U}(-0.01, 0.01). Right: Initial concentration field c0=0.4+U(−0.01,0.01)c_0 = 0.4 + \mathcal{U}(-0.01, 0.01).

The second case on the more dilute 70:30 mixture (⟨c⟩=0.4\langle c\rangle= 0.4) provides quantitative analysis for the number of clusters and their volume fractions of the precipitated minority species (A). Here the volume is defined as the count of the phase-separated cc minority voxels with c<−0.8c < -0.8, and a corresponding minority cluster is a set of such voxels connected via voxel face, edge or corner while accounting for the periodic boundary condition. For better statistics, the simulations were performed on a larger periodic 64364^3 grid for 4×1064 \times10^6 steps with time interval 1×10−31 \times10^{-3} and total simulation time 4×1034 \times10^3. Figure 5a shows the average over 10 trajectories ε=0\varepsilon= 0 and ε=0.2\varepsilon= 0.2. The number of clusters decreases faster in the presence of noise up to time t≈10t \approx10 with about 10 precipitate clusters remaining, suggesting that

Figure 5. Example 3D SPDE trajectories of 70:30 mixture (⟨c⟩=0.4\langle c\rangle= 0.4) on a 64364^3 grid with ε=0.2\varepsilon= 0.2, starting from the same initial condition with slight perturbation U(−0.01,0.01)U(-0.01, 0.01). (a) Evolution of the number of precipitate clusters (voxels with c≤−0.8c \le-0.8) and the precipitate volume fraction. (b) Precipitate structure evolution for ε=0\varepsilon= 0 and (c) ε=0.2\varepsilon= 0.2.

the system evolves faster with stochastic fluctuations. We refrain from drawing quantitative conclusions beyond t≈10t \approx10 due to insufficient statistics. The volume fraction (sum of cc) of the minority clusters is higher with noise in the very beginning, then becomes lower than ε=0\varepsilon= 0. Figure 5b (ε=0\varepsilon= 0) and c (ε=0.2\varepsilon= 0.2) show example snapshots from t=4t = 4 to t=4000t = 4000.

Results on surrogate models

In this section, we will evaluate the performance of the ML surrogate models outlined in § 2.2. The blackbox, non-flux model is first evaluated and ruled out in § 4.1. The flux-based models were analyzed for in-distribution performance in § 4.2 and for spatiotemporal extrapolation tests in § 4.3. Moreover, we validate the learned free energy in § 4.4. Lastly, we also explore the extrapolation ability for precipitation at concentrations unseen in training, in § 4.5.

Due to the stochastic nature of the ground truth data source and the goal of long-term predictions, voxel-wise metrics alone are insufficient for model assessment. We therefore focus on statistical and physics-based criteria, including rollout stability, phase-separation kinetics, learned mobility, free-energy interpretability, and precipitation behavior. The models were trained on trajectories with random initial ⟨c⟩∼U(−0.5,0.5)\langle c\rangle\sim U(-0.5, 0.5), and tested on two representative cases, ⟨c⟩=0\langle c\rangle= 0 and ⟨c⟩=0.4\langle c\rangle= 0.4. We first show that the black-box non-flux baseline fails in spatial extrapolation and exclude it from further consideration. We then compare the flux-based surrogates, showing that FE-MV best matches the stochastic dynamics, while the nonFE-M and nonFE-MV are less physically consistent as judged by the learned mobility relation.

Blackbox non-conservative baseline vs conservative models

All five models defined in § 2.2 were trained on SPDE trajectories with ε=0.2\varepsilon= 0.2 on a 16316^3 grid. First, the rollout stability was evaluated on test data of unseen trajectories (time duration of 4000) on the same grid that are much longer than the training set (duration of 24). We tested for different initial concentrations ⟨c⟩=0\langle c\rangle= 0 and 0.40.4 with additional random uniform perturbations U(−0.01,0.01)\mathcal{U}(-0.01,0.01). As shown in Sec. B.5, the black-box non-conservative baseline non-flux surrogate matches the overall trend of temporal evolution and marginal concentration histograms on the training grid. However, when evaluated for spatial extrapolation on a 64364^3 grid, non-flux fails at late evolution stages and does not preserve spatial structure, as shown in Figure 6b for a 70:30 mixture (⟨c⟩=0.4\langle c\rangle= 0.4). This shows that without translation equivariance built in through the flux structure, the network cannot generalize to unseen spatial extents or to long time horizon. By contrast, the flux-based conservative surrogates (Figure 6c-Figure 6f) all remain stable and generate physically plausible morphologies starting from the same initial configuration. This comparison clearly indicates that a purely black-box autoregressive model is insufficient for robust very-long and large-scale spatiotemporal generalization in this problem. We therefore exclude non-flux from the remainder of the discussion and focus on conservative flux-based surrogates.

Figure 6. Generalization test of learned surrogate models with ⟨c⟩=0.4\langle c\rangle= 0.4 on a 64364^3 grid for a long rollout. (a) Ground truth SPDE trajectory with ε=0.2\varepsilon= 0.2, (b) non-flux predictions, (c) FE-MV, (d) FE-M, (e) nonFE-MV, and (f) nonFE-M.

In-distribution test of conservative models

Next the conservative surrogate models are tested on the 16316^3 grid, the same as used in training. The standard deviation of voxel values along a trajectory is taken as a statistical measure of demixing, as the mean is constant in these conservative models. Figure 7 shows the standard deviation over 10 trajectories for the same two representative compositions 50:50 (⟨c⟩=0\langle c\rangle= 0) and 70:30 (⟨c⟩=0.4\langle c\rangle= 0.4). We find that the FE-MV and nonFE-MV models, which predict both a mean and a variance and therefore produce stochastic rollouts, matches the ground truth with ε=0.2\varepsilon= 0.2 better, whereas the mean-only models FE-M and nonFE-M show substantial delay in the early stage of demixing dynamics. This is attributed to the physically learned flux fluctuation that accelerates phase separation, in agreement with the observed difference between the SPDE and PDE in Figure 4.

Four plots showing the standard deviation of voxel values over time for FE-M, FE-MV, nonFE-M, and nonFE-MV models

Figure 7. The standard deviation of voxel values of ML models. Free energy based models FE-MV, FE-M for (a) ⟨c⟩=0\langle c\rangle= 0 and (b) ⟨c⟩=0.4\langle c\rangle= 0.4, and non free energy nonFE-MV, nonFE-M for (c) ⟨c⟩=0\langle c\rangle= 0 and (d) ⟨c⟩=0.4\langle c\rangle= 0.4. The mean-only models show substantial delay in demixing.

While the free-energy and non-free-energy based variants show similar behavior in the above tests, a difference can be revealed in the learned mobility MM. Since there is no constraint for the magnitude of the learned MM, its scale is arbitrary (one may freely scale M→αMM \to\alpha M, Fij→1αFijF_{ij} \to\frac{1}{\alpha}F_{ij}). Therefore one just expects a linear relationship between a perfectly learned MM and the ground truth values. To verify if the fitted MM captures the correct physics, we plot it against the analytical mobility of Eq. (6) in Figure 8. Figure 8 indeed shows nearly linear plots for free energy based models FE-MV and FE-M with conservative driving forces (chemical potential) μ=∂E/∂c\mu= \partial E/\partial c, as well as substantial non-linear deviation for non-free energy models, which are under non-conservative driving forces. Similar results were obtained for ⟨c⟩=0\langle c\rangle= 0 and 0.40.4. This suggest that while the non-free energy models can reasonably learn the dynamics, they have cancellation of errors in the non-conservative driving force and corresponding kinetic coefficients, resulting in neither quantity being physically meaningful.

Machine learned mobility of nonFE-M, FE-M, nonFE-MV, and FE-MV models compared with ground truth

Figure 8. Machine learned mobility MM of different models compared to the ground truth from 100 snapshots up to t=10t = 10 with ⟨c⟩=0\langle c\rangle= 0 on the 16316^3 grid.

Generalization tests: long-term agglomeration

Beyond in-distribution test, the extrapolation test is also important to check if the model learns the physics well. We performed a long-term agglomeration test for ⟨c⟩=0.4\langle c\rangle= 0.4 on a 64364^3 grid. As shown previously in Figure 6, all the conservative surrogate models are stable compared to the ground truth up to t=4000t = 4000 and reach similar stationary state of a single (nearly) spherical precipitate. Moreover, we calculated the number of clusters and volume fractions. As shown in Figure 9, both free-energy based variants’ predictions are generally aligned with the ground truth. We also plot the total energy per voxel versus time and show it in Figure 10. This curve is computed as A/ΩA/\Omega where AA is the total free energy defined in Eq. (3) and Ω\Omega is the volume. Due to the arbitrary scaling for multiplicative mobility MM and energy EE, the model predictions are scaled to match the overall scale of the ground truth. The good alignment between machine learning predictions and the true data shows that our free energy model has learned the real physics well.

Extrapolated agglomeration test for clusters and volume fraction evolution

Figure 9. Extrapolated agglomeration test for clusters and volume fraction evolution of ⟨c⟩=0.4\langle c\rangle= 0.4 on a 64364^3 grid

Total energy curves

Figure 10. Total free energy evolution for the same tests.

Machine learned free energy verification

Since the learned energy model E^(c)\hat{E}(c) is a complicated, high-dimensional neural network rather than simple expressions and hard to digest directly, we query its behavior to see if it actually makes physical sense. We plot the learned E^(c)\hat{E}(c) of FE–MV in Figure 11. Subplot (a) shows the bulk free energy landscape of uniform field E(c≡c)E(\mathbf{c} \equiv c) versus cc. With an affine calibration to match the scales (same as in Figure 10), the learned bulk free energy closely follows the double well features of the ground truth in Figure 1.

FE–MV learned free energy curves

Figure 11. FE–MV learned free energy curves. (a) Bulk free energy landscape. (b) Free energy for flat interface along the zz axis. (c) Interfacial free energy per area vs radius (spherical interface).

Figure 11b shows the free energy profile for a manually created flat interface unit-cell of two-phase coexistence. The total height of the cell is 40 and the two sharp peaks at around z=10z = 10 and z=30z = 30 indicate the interfacial regions where cc transitions between A and B phases. The plotted energy was averaged on the x−yx-y plane. The energy peaks represent the excess free energy cost of creating the interface. These results indicate that our machine-learned free energy model correctly grasped the key physics involved in immiscible phases.

Figure 11c shows the interfacial free energy change versus radius. Each data point was evaluated with a sphere of minority of various radius embedded in the majority phase. The model predicts an interfacial tension between 0.56–0.60, which is approximately independent of curvature, in agreement with the ground truth isotropic model. In short, all the three plots confirm the proposed model’s ability to learn the correct thermodynamic behavior for phase separation dynamics. Remarkably, the model learned the thermodynamics purely from observing the trajectories of order parameters with no knowledge of energy or chemical potential.

Generalization to nucleation dynamics: the case for stochastic surrogates

The spinodal decomposition regime used for training provides an ideal setting for learning and validating the surrogate architecture, but it does not expose the most fundamental advantage of flux-level stochasticity: the ability to simulate thermally activated, barrier-crossing events that are entirely inaccessible to deterministic models. We now demonstrate this directly by evaluating both FE–MV and FE–M on initial concentrations in the metastable region, outside the spinodal training region ∣⟨c⟩∣<0.5|\langle c\rangle| < 0.5.

At ⟨c⟩=−0.51\langle c\rangle= -0.51, just outside the spinodal boundary, the deterministic FE–M surrogate produces no nucleation events at all without seeds. This is not a failure of training or ar-

chitecture tuning; it is a fundamental limitation of any deterministic model, which cannot spontaneously generate the rare thermal fluctuation required to cross the nucleation barrier. By contrast, the stochastic FE–MV successfully nucleates without seeds, in qualitative agreement with the ground truth SPDE (Figure 12). To assess this quantitatively, we define the precipitation time tpt_p as the first time at which one or more voxels appear with ∣⟨c⟩∣>0.8\lvert\langle c\rangle\rvert> 0.8, and evaluate an ensemble of 40 independent FE–MV rollouts. The surrogate yields a mean precipitation time ⟨t^p⟩=6.2\langle\hat{t}_p\rangle= 6.2, close to the ground truth value of 7.0. The predicted distribution is narrower than the ground truth (Figure 13), which is expected: the surrogate was trained exclusively on spinodal trajectories, so it has no direct exposure to the statistics of barrier-crossing fluctuations in the metastable region. Extending deeper into the metastable regime, where barrier-crossing nucleation events are rarer in the training data, requires a dedicated data strategy, such as active learning or importance sampling of rare trajectories, which is deferred to future work.

Comparison for extrapolation trajectory

Figure 12. Comparison for extrapolation trajectory ⟨c⟩=−0.51\langle c\rangle= -0.51. (a) Ground truth value ε=0.2\varepsilon= 0.2. (b) FE-MV model.

Distribution of precipitation time $t_p$ over 40 independent runs, FE–MV VS stochastic PDE simulation

Figure 13. Distribution of precipitation time tpt_p over 40 independent runs, FE–MV VS stochastic PDE simulation.

These results establish a clear distinction between the deterministic and stochastic variants: in the spinodal regime both perform comparably on long-time statistics, though the former exhibits delayed early-time dynamics. But only the stochastic surrogate is a viable model class for the metastable regime. This distinction is not a quantitative difference in accuracy, but a qualitative one. A surrogate without flux-level noise cannot model nucleation at all, regardless of how well it is trained. This section therefore serves as the clearest justification for the architectural choices made in this work.

Conclusion

In this work, we developed and evaluated a family of physics-aware surrogate models for the stochastic Cahn–Hilliard equation, a prototypical model of noise-driven conserved phase separation dynamics. The central architectural innovation is to parameterize the surrogate at the level of inter-cell fluxes rather than predicted states, which guarantees exact mass con servation by construction and allows stochasticity to be introduced at its physically correct scale — the fluctuating current between neighboring cells.

Through systematic ablation across five model variants, we find that all three design choices are necessary for a physically consistent surrogate: flux-level conservation, a learnable free energy function, and explicit stochastic flux amplitude. Removing conservation (non-flux baseline) produces a model that fails catastrophically under spatiotemporal extrapolation. Removing the free energy constraint (nonFE-MV, nonFE-M) yields models that reproduce coarsening statistics reasonably well but learn a mobility field that is physically inconsistent, exhibiting error cancellation between a non-conservative driving force and miscalibrated kinetic coefficients. Removing stochasticity (FE-M) produces a deterministic surrogate that matches long-time ensemble statistics in the spinodal regime but introduces systematic delay in early-time demixing kinetics. More critically, it is entirely unable to model nucleation or barrier crossing in the metastable regime. Only FE-MV, the full stochastic free-energy variant, satisfies all three physical requirements simultaneously and generalizes across regimes.

As an early attempt at quantitative surrogate modeling of stochastic dynamical systems, this work made several simplifying assumptions. The covariance matrix for fluxes was assumed to be diagonal. More complicated generative models with nonvanishing covariance between different fluxes or even non-Gaussian noises will be interesting subjects for the future. We also fixed the temperature to a single value. The nucleation and growth mechanism for metastable concentrations beyond the present unstable spinodal decomposition reactions will also be the subject of future studies. The most important generalization of this paper is towards applications in molecular dynamics data. Work is underway to develop surrogate models trained on trajectories of molecular dynamics simulations and will be published separately.

Acknowledgements

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. This work was funded by the Laboratory Directed Research and Development (LDRD) Program at LLNL under project tracking code 25-ERD-002 and 23-ERD-029. Computing support for this work came from the LLNL Institutional Computing Grand Challenge program.

Author Contributions

VHN and LS performed dataset set generation, model refinement and training, and data analysis. FZ initiated the project, wrote the code base, and supervised the research. SL and JK secured the funding. LS, VHN and FZ wrote the paper with inputs from all authors.

Data Availability

The data required to reproduce this work can be requested from FZ.

Code Availability

The source code and a demo for implementing the SD rollout are available at http://www.github.com/llnl/NPS.

Competing interests

The authors state that there is no conflict of interest.

Phase diagram

Phase diagram details are shown in Supplementary Fig. S1.

Bulk free energy and solubility limit at different temperature

Figure S1. Bulk free energy and solubility limit at different temperature.

Additional details on 1D and 3D SPDE solutions

1D SPDE analysis for c0=U(−0.01,0.01)c_0 = \mathcal{U}\left(-0.01, 0.01\right)

Additional 1D statistics on the 32-grid case are shown in Supplementary Fig. S2.

1D SPDE analysis for c0=0.4+U(−0.01,0.01)c_0 = 0.4 + \mathcal{U}\left(-0.01, 0.01\right)

Results for the shifted-initial-condition c0=0.4+U(−0.01,0.01)c_0 = 0.4 + \mathcal{U}\left(-0.01, 0.01\right)case with grid size 32 and 256 are shown in Supplementary Fig. S3 to Figure S4, respectively.

Standard deviation voxel, value frames, Fourier spectrum, and normalized Fourier spectrum for no noise and noisy cases

Figure S4. Supplementary Figure S4: 256 meshgrid with c0=0.4+U(−0.01,0.01)c_0 = 0.4 + \mathcal{U}(-0.01, 0.01) (TOP) The standard deviation voxel (1st MIDDLE) value frames (2nd MIDDLE), Fourier spectrum (BOTTOM) normalized Fourier spectrum.

Convergence of 3D SPDE with different time step size

The SPDE system exhibits almost similar convergence behavior as either using the smaller time step size 10−410^{-4} or a larger one 10−310^{-3}as shown in Supplementary Fig. S5. In our experiments, we opt to generate SPDE data using the 10−310^{-3} time step size, as it is more computationally efficient while still yielding similarly satisfactory results. Note that, the SPDE solution is unstable as solving with the time step size of 10−210^{-2}.

3D SPDE analysis

Supplementary Fig. S6 shows the growth speed of voxel value mean ±\pm standard deviation evolution, averaging over 10 SPDE trajectories, for both ε=0\varepsilon= 0 and ε=0.2\varepsilon= 0.2. The corresponding histograms of voxel value evolution are shown in the second and third rows, respectively. It can be seen that stochastic noise promotes faster phase separation for both initial concentration fields, ⟨c⟩=0\langle c\rangle= 0 and 0.4. Concentration field samples at the time point of 0.4 and 2 for ⟨c⟩=0\langle c\rangle= 0 and 0.4, respectively, are presented in the bottom row of Figure S6. The concentration state simulated with noise reaches the coarsening phase much earlier than the one without noise.

Results for a 32² grid with ⟨c⟩ = 0

Figure S2. Results for a 32232^{2} grid with ⟨c⟩=0\langle c\rangle= 0 (TOP) The standard deviation voxel (1st MIDDLE) value frames (2nd MIDDLE), Fourier spectrum (BOTTOM) normalized Fourier spectrum.

32 meshgrid with no-noise and noisy solutions

Figure S3. 32 meshgrid with c0=0.4+U(−0.01,0.01)c_0 = 0.4 + \mathcal{U}(-0.01, 0.01) (TOP) The standard deviation voxel (1st MIDDLE) value frames (2nd MIDDLE), Fourier spectrum (BOTTOM) normalized Fourier spectrum.

Image not extracted

Figure S5. Top The standard deviation of voxel values from 10 independent 3D simulations at ⟨c⟩=0\langle c \rangle= 0 with different noise levels ε=0\varepsilon= 0, 0.02, 0.2 and integrated with time step size of 10−310^{-3} and 10−410^{-4}. Middle The histogram of voxel values from 10 PDE simulations (ε=0\varepsilon= 0) solved with different time step size. Bottom Same for for SPDE simulations (ε=0.2\varepsilon= 0.2). Left column: time step of 10−310^{-3}, right: time step of 10−410^{-4}.

In-distribution test for non-flux model

Additional in-distribution test results for non-fluxmodel are shown in Supplementary Fig. S7.

Image not extracted

Figure S7. The standard deviation of voxel value of the non-flux model averaged over 100 validation trajectories on a 16316^3 grid and the corresponding cc value histogram at t=10t = 10. Left: c0=U(−0.01,0.01)c_0 = \mathcal{U}(-0.01,0.01) Right: c0=0.4+U(−0.01,0.01)c_0 = 0.4 + \mathcal{U}(-0.01,0.01).

Additional in-distribution result

Additional results for flux-based NN models are shown in Supplementary Fig. S8.

Image not extracted

Figure S8. Prediction and GT (no noise and noise ε=0.2\varepsilon= 0.2) histograms at t=10st = 10s on mesh grid 16×16×1616 \times16 \times16. Top row: c0=U(−0.01,0.01)c_0 = \mathcal{U}(-0.01,0.01) Bottom row: c0=0.4+U(−0.01,0.01)c_0 = 0.4 + \mathcal{U}(-0.01,0.01)

References

  1. [1]Cahn, J. W.; Hilliard, J. E. Free energy of a nonuniform system. I. Interfacial free energy. The Journal of chemical physics 1958, 28, 258–267.
  2. [2]Becker, R.; Döring, W. Kinetische Behandlung der Keimbildung in übersättigten Dämpfen. Annalen der Physik 1935, 416, 719–752.DOI
  3. [3]Binder, K. The Monte Carlo method for the study of phase transitions: A review of some recent progress. Journal of Computational Physics 1985, 59, 1–55.DOI
  4. [4]Voter, A. F. In Radiation effects in solids; Springer: 2007, pp 1–23.
  5. [5]Chen, L.-Q. Phase-field models for microstructure evolution. Annual review of materials research 2002, 32, 113–140.DOI
  6. [6]Cook, H. Brownian motion in spinodal decomposition. Acta metallurgica 1970, 18, 297–306.DOI
  7. [7]Bronchart, Q.; Le Bouar, Y.; Finel, A. New Coarse-Grained Derivation of a Phase Field Model for Precipitation. Physical Review Letters 2008, 100, 015702.pubmed.ncbi.nlm.nih.gov/18232784
  8. [8]Yang, K.; Cao, Y.; Zhang, Y.; Fan, S.; Tang, M.; Aberg, D.; Sadigh, B.; Zhou, F. Self-supervised learning and prediction of microstructure evolution with convolutional recurrent neural networks. Patterns 2021, 2, 100243.
  9. [9]Montes de Oca Zapiain, D.; Stewart, J. A.; Dingreville, R. Accelerating phase-field-based microstructure evolution predictions via surrogate models trained by machine learning methods. npj Computational Materials 2021, 7, 1–11.DOI
  10. [10]Fan, S.; Hitt, A. L.; Tang, M.; Sadigh, B.; Zhou, F. Accelerate microstructure evolution simulation using graph neural networks with adaptive spatiotemporal resolution. Machine Learning: Science and Technology 2024, 5, 025027.arxiv.org/abs/2310.15153
  11. [11]Bertin, N.; Bulatov, V. V.; Zhou, F. Learning dislocation dynamics mobility laws from large-scale MD simulations. npj Computational Materials 2024, 10, 192.arxiv.org/abs/2309.14450
  12. [12]Ji, K.; Sun, L.; Liu, S.; Zhou, F.; Heo, T. W. Scalable Autoregressive Deep Surrogates for Dendritic Microstructure Dynamics. 1–29.arxiv.org/abs/2511.03884
  13. [13]Tian, Z.; Suwandi, E.; Oppelstrup, T.; Bulatov, V. V.; Harley, J. B.; Zhou, F. Scaling Kinetic Monte-Carlo Simulations of Grain Growth with Combined Convolutional and Graph Neural Networks, 2025.
  14. [14]Raissi, M.; Perdikaris, P.; Karniadakis, G. E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics 2019, 378, 686–707.DOI
  15. [15]Sun, L.; Gao, H.; Pan, S.; Wang, J.-X. Surrogate modeling for fluid flows based on physics-constrained deep learning without simulation data. Computer Methods in Applied Mechanics and Engineering 2020, 361, 112732.DOI
  16. [16]Wang, S.; Teng, Y.; Perdikaris, P. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing 2021, 43, A3055–A3081.
  17. [17]Lu, L.; Jin, P.; Karniadakis, G. E. DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193 2019.DOI
  18. [18]Li, Z.; Kovachki, N.; Azizzadenesheli, K.; Liu, B.; Bhattacharya, K.; Stuart, A.; Anandkumar, A. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895 2020.arxiv.org/abs/2010.08895
  19. [19]Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, Ł.; Polosukhin, I. Attention is all you need. Advances in neural information processing systems 2017, 30.arxiv.org/abs/1706.03762
  20. [20]Han, X.; Gao, H.; Pfaff, T.; Wang, J.-X.; Liu, L.-P. Predicting physics in mesh-reduced space with temporal attention. arXiv preprint arXiv:2201.09113 2022.arxiv.org/abs/2201.09113
  21. [21]Sun, L.; Han, X.; Gao, H.; Wang, J.-X.; Liu, L. Unifying predictions of deterministic and stochastic physics in mesh-reduced space with sequential flow generative model. Advances in Neural Information Processing Systems 2023, 36, 60636–60660.DOI
  22. [22]Gao, H.; Han, X.; Fan, X.; Sun, L.; Liu, L.-P.; Duan, L.; Wang, J.-X. Bayesian conditional diffusion models for versatile spatiotemporal turbulence generation. Computer Methods in Applied Mechanics and Engineering 2024, 427, 117023.arxiv.org/abs/2311.07896
  23. [23]Du, P.; Parikh, M. H.; Fan, X.; Liu, X.-Y.; Wang, J.-X. Conditional neural field latent diffusion model for generating spatiotemporal turbulence. Nature Communications 2024, 15, 10416.DOI
  24. [24]Haas, K. R.; Yang, H.; Chu, J.-W. Expectation-Maximization of the Potential of Mean Force and Diffusion Coefficient in Langevin Dynamics from Single-Molecule FRET Data Photon by Photon. J. Phys. Chem. B 2013, 117, 15591–15605.DOI
  25. [25]Frishman, A.; Ronceray, P. Learning Force Fields from Stochastic Trajectories. Phys. Rev. X 2020, 10, 021009.
  26. [26]Liu, Z.; Mao, H.; Wu, C.-Y.; Feichtenhofer, C.; Darrell, T.; Xie, S. A ConvNet for the 2020s. arXiv: 2201.03545 2022.arxiv.org/abs/2201.03545
  27. [27]Loshchilov, I.; Hutter, F. Decoupled weight decay regularization. arxiv: 1711.05101 2017.arxiv.org/abs/1711.05101

Paper details

vlnc:W_axn3fr9m
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents