Stochastic partial differential equations (SPDEs) are PDEs pushed around by random noise at every point in space and time. They model phase separation, noisy waves and turbulent flow. Their hard cases have a "cubic" reaction term, $f(u) = -u^3$. The exact solution is well behaved, but a naive explicit time-stepper can blow up: once noise kicks the state into a region where $\Delta t\,u^2 > 2$, each step overshoots further. Numerical analysts have fixed this with tamed and truncated schemes, which scale the force down when the solution gets large, and with implicit schemes. They have proven sharp convergence rates for these methods. Separately, machine learning papers train neural networks to map noise to solution and report relative errors as low as 0.0009, along with 20–300x speedups. These two literatures never check each other. The proofs rarely meet a real benchmark. The neural networks are scored against solver outputs whose own error is never measured, using per-path metrics that ignore the statistics a stochastic model exists to produce.
This note proposes to make the proven rates do empirical work. First, we measure whether tamed, truncated and implicit schemes actually reach their proven convergence rates at the grid sizes used to make training data. That gives every reference dataset a measured error floor. Second, we score neural surrogates relative to that floor, using ensemble statistics rather than single-path error. The central claim is testable: on the stochastic Allen-Cahn / dynamic $\Phi^4_1$ benchmark, the best reported surrogate errors sit at or below the discretization error of the data they were trained on. If so, those numbers measure agreement with a particular solver, not with the SPDE. When compared on expected observables at matched wall-clock cost, a coarse tamed scheme will be competitive with, or better than, the surrogates.
What is already known, and what is missing
Three analysis papers supply the reference methods and the rates we test. Becker et al. prove that nonlinearity-truncated explicit schemes for stochastic Allen-Cahn with space-time white noise converge at strong rate $1/2-\epsilon$ in space and $1/4-\epsilon$ in time, and show these rates are sharp. Optimal Error Estimates of Galerkin FEM for Stochastic Allen-Cahn proves $O(h^\gamma+\tau^{\gamma/2})$ for backward-Euler finite elements (implicit time steps on a mesh of size $h$), with $\gamma\in[d/3,2]$ set by how spatially correlated the noise is. Tamed-truncated stochastic convolutions gives exponential moment bounds for tamed variants. Weak rates for explicit Allen-Cahn establishes that errors in expected observables ("weak" error) converge roughly twice as fast as errors along individual paths ("strong" error). None of these papers characterizes practical cost or tests rates beyond small 1D slope plots.
Three neural lines motivate the question of what their reported numbers mean. Neural SPDEs learn a mild-solution-style operator. DLR-Net adds regularity-structure features that smooth rough noise before a Fourier neural operator (FNO), a network that works on Fourier modes. Expanding the Chaos feeds Wiener-chaos (Hermite-polynomial) features of the noise into neural operators. All three report per-path relative $L^2$ error against solver-generated data. None compares against tamed or truncated schemes, and none checks moments, observables or behavior in the regime where explosion happens. A symptom is that on the "same" $\Phi^4_1$ benchmark, plain FNO is reported at $l_2 = 0.032$ in DLR-Net and at relative $L^2 = 0.029$ in Expanding the Chaos. The setups differ enough that neither number can be read against the other. PDEBench constrains the design: it shows that RMSE hides physical failures and adds spectral and conservation errors. Those are still per-path, deterministic metrics, and PDEBench has no stochastic, superlinear cases. The validation-gap literature explains why proven rates cannot simply be assumed to hold at practical sizes. Examples include the conservative constants behind Nibble and the hidden $O(\kappa^2)$ cost behind EMGD. Rates are asymptotic, and whether the constants are benign at $N=64$–$512$ grid points is an empirical question.
Core Claim
A proven convergence rate becomes an error bar if it is confirmed empirically. On a hierarchy of grids driven by one shared Brownian noise sample, a scheme that attains its proven slope lets us extrapolate its distance to the true solution at each level. We call this distance the reference floor $\delta_{\text{ref}}(h,\Delta t)$. A surrogate trained on level-$(h,\Delta t)$ data and reporting error $e \lesssim \delta_{\text{ref}}$ has not been shown to approximate the SPDE better than the solver it copied. The meaningful comparison happens on expected observables, which is where the analysis predicts classical schemes gain the most (the doubled weak rate). It also happens under noise intensities and time horizons outside the training range, where the explosion mechanism operates.
Research Program
The first component is an audit of the reference schemes, run as an empirical study of the analysis literature. We implement explicit Euler, tamed Euler (denominator $1+\Delta t\|u_n\|^2$), the truncated spectral-Galerkin scheme of Becker et al., and backward-Euler finite elements following Wang. We run them on 1D stochastic Allen-Cahn with space-time white noise, which is the dynamic $\Phi^4_1$ testbed. We also run 2D Allen-Cahn with trace-class noise across several correlation exponents $\gamma$. All levels of a grid hierarchy share Brownian increments, so pathwise differences isolate discretization error. For each scheme we fit strong and weak slopes. We measure the empirical explosion frequency (the fraction of paths that overflow) as a function of $\Delta t$ and noise amplitude $\sigma$. We also locate the smallest grid at which the proven slope is visible. That last quantity is the practical onset of the asymptotic regime, and it is what the validation-gap literature suggests may be far from where theory is quoted.
The second component defines the stochastic scoring protocol. For each $(h,\Delta t)$ level we report $\delta_{\text{ref}}$ from the fitted rates. On a held-out ensemble of $10^4$ noise draws we evaluate a set of observables. These are mean and second moment of $\|u\|$, the fourth moment (which blows up first when moment bounds fail), the fraction of the domain in each phase well, the 99th-percentile of $\sup_x|u|$, and the probability of a full well-to-well transition within horizon $T$. That last event is rare and governed by metastability, and it is exactly the tail behavior that pathwise $L^2$ averages away.
The third component retrains NSPDE, DLR-Net, F-SPDENO and FNO on the same data under matched budgets. We score them on these observables in-distribution, under a $\sigma$ sweep beyond the training range, and under autoregressive rollouts up to ten times the training horizon.
Baselines
The naive reference point is untamed explicit Euler at the same grid. The analysis predicts divergent moments for it at any fixed $\Delta t$. Measuring its explosion frequency calibrates how severe the regime is, and any method, neural or classical, must at least never exhibit that failure. The strongest classical competitor on this question is the truncated explicit scheme of Becker et al., together with tamed Euler. These are cheap per step, carry proven sharp strong rates and doubled weak rates, and have never been timed. For fairness we select their coarsest grid by one rule: the cheapest level whose observable error matches the surrogate's. We then report the wall-clock cost of generating the full test ensemble on the same GPU, with FFT-based spectral steps as the surrogates use. Backward-Euler finite elements are the conservative, implicit baseline. The comparison records Newton iteration counts as the mesh is refined, to test the stiffness cost raised in the file.
The strongest neural competitors are DLR-Net and F-SPDENO. On $\Phi^4_1$, DLR-Net reports $l_2 = 0.0009$ against FNO's 0.032 with roughly a 20x speedup. Expanding the Chaos reports relative $L^2 = 0.012$ for F-SPDENO, 0.029 for FNO and 0.134 for Neural CDE. On Ginzburg-Landau, Neural SPDEs reports about 0.6% error against 2.7% for FNO, with a 59x speedup. Every neural model gets the same training data, the same tuning budget (a fixed number of hyperparameter trials) and the same GPU-hour cap.
A surrogate "beats" the classical baseline only if it satisfies three conditions together. Its observable errors must be below those of the tamed scheme at equal wall-clock time. Its $L^2$ error must be meaningfully below $\delta_{\text{ref}}$ when scored against a finer reference. It must stay bounded under the $\sigma$ sweep. The key ablation separates the certified reference floor from the ensemble metrics. We rescore all models using ensemble metrics but against the uncertified training-level solver instead of the rate-extrapolated truth. That shows how much of any ranking change comes from certification itself.
Why not just…?
Why not just generate ground truth with a very fine implicit solver and skip the rate audit? This is the most serious alternative, and we partly adopt it: the finest level of our hierarchy is exactly such a solve. The weakness is that "very fine" is certified only by assuming convergence. Without fitted slopes nobody knows how fine is fine enough for white-noise forcing, where the time rate is only $1/4$. The audit is what establishes that the fine solve is trustworthy. If slopes turn out clean at modest sizes, the audit collapses to a cheap sanity check and this alternative becomes the recommended practice. That would still be a useful result to report.
Why not just add PDEBench-style metrics to existing surrogate evaluations? The spectral, boundary and conservation errors in PDEBench are computed path by path against one solver output. They cannot detect a surrogate that is accurate on typical paths but misses rare well-to-well transitions or underestimates fourth moments. We will report fRMSE alongside our observables and check directly whether it ranks models the same way. If it does, the simpler metric suffices.
Why not just shrink $\Delta t$ in explicit Euler? The analysis literature (tamed-truncated bounds) shows that Gaussian forcing makes entering the explosion zone possible for every fixed $\Delta t$, so moments diverge regardless of step size. Our explosion-frequency sweep measures how often this happens in practice at realistic horizons. If the frequency is negligible below some $\Delta t$ at acceptable cost, we will report that plainly as a practical threshold.
Proposed Experiments
The first decisive check is whether each reference scheme shows its proven strong slopes ($1/2$ space, $1/4$ time) and roughly doubled weak slopes, and at what grid size those slopes appear. The second check is whether the reported surrogate errors on $\Phi^4_1$ fall below $\delta_{\text{ref}}$ at the resolutions those benchmarks use. If they do, per-path accuracy claims at that level cannot distinguish the models. If they do not, the reported numbers are validated. The third check is whether surrogates reproduce the fourth moment, the phase-transition probability and the tail quantiles within the reference floor, and whether they stay bounded when $\sigma$ exceeds the training range by 1.5–3x over ten-fold rollouts. The fourth check is whether, at matched observable accuracy, tamed coarse-grid simulation of the full ensemble is slower or faster than surrogate inference. This is the comparison that none of the reported speedups (20x, 59x, 300x) made, because those speedups were measured against fine solvers rather than against the cheapest accurate classical scheme. Each outcome is informative either way, and the second and fourth checks can overturn how the existing neural SPDE results should be read.
References
- Strong convergence rates for explicit space-time discrete numerical approximations of stochastic Allen-Cahn equations
- Optimal Error Estimates of Galerkin Finite Element Methods for Stochastic Allen-Cahn Equation with Additive Noise
- Exponential moment bounds and strong convergence rates for tamed-truncated numerical approximations of stochastic convolutions
- Weak Convergence Rates for an Explicit Full-Discretization of Stochastic Allen-Cahn Equation with Additive Noise
- Neural Stochastic Partial Differential Equations
- Deep Latent Regularity Network for Modeling Stochastic Partial Differential Equations
- Expanding the Chaos: Neural Operator for Stochastic (Partial) Differential Equations
- PDEBench: An Extensive Benchmark for Scientific Machine Learning
- A Local Clustering Algorithm for Massive Graphs and Its Application to Nearly Linear Time Graph Partitioning
- Linear Convergence with Condition Number Independent Access of Full Gradients