acceptodds
Under review as a conference paper at ICLR 2027

Open-Loop, Closed-Form: Solver-Driven Reservoir Computing for Coarse-Grid PDE Forecasting

Abstract

Hybrid numerical–learning models can improve coarse-grid PDE forecasts, but training corrections through coupled rollouts is computationally expensive. We propose an open-loop solver–reservoir framework with closed-form readout training. A numerical solver initialized solely from the coarse initial state generates a base trajectory. A fixed reservoir transforms this trajectory into history-dependent features, and a linear Fourier readout, shared across forecast times, predicts the downsampled high-fidelity states. Training reduces to ridge regression independently across Fourier modes, without differentiation through the solver or reservoir. Learned predictions never enter either state update, so readout errors cannot alter subsequent features. We analyze the guarantees and limitations of this design: feature consistency for a fixed construction under a common train–test trajectory law, a prediction-risk decomposition, and an expected-risk bound using independent trajectories as the sampling units. Across five configurations of the Kuramoto–Sivashinsky, Burgers, and Navier–Stokes equations, our method reduces test nRMSE by 31–88% relative to the uncorrected solver and achieves the lowest error among the compared methods in three configurations. Measured training times, including validation-based model selection, correspond to speedups of 4.0–38.9 over solver-in-the-loop hybrids with FFNO and S4FFNO correctors under the evaluated protocols, although the best such hybrid remains more accurate in the other two configurations. Our method also outperforms the evaluated closed-loop hybrid reservoir baseline in all four shared configurations.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
reservoir computing, partial differential equations, hybrid physics-ML modeling, scientific machine learning, coarse-grained simulation, closed-form learning, spatiotemporal forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MkT5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an open-loop solver–reservoir framework for coarse-grid PDE forecasting. A fixed numerical solver generates a trajectory from the initial state, which is processed by a recurrent reservoir to extract history-dependent features. Predictions are obtained through a Fourier-domain linear readout trained using closed-form ridge regression, without feeding predictions back into the solver. The paper also presents theoretical analyses concerning the properties of this framework. Experiments on the Kuramoto–Sivashinsky, Burgers, and Navier–Stokes equations compare its forecasting accuracy and training efficiency against solver-in-the-loop neural correctors and closed-loop hybrid reservoir models.

Critical strengths

Computationally efficient training without differentiating through the numerical solver. The proposed two-stage approach constructs solver-driven reservoir features and fits only a linear Fourier readout using closed-form ridge regression. This avoids backpropagation through coupled solver–network rollouts and enables reuse of numerical trajectories during hyperparameter selection.

Competitive forecasting accuracy with substantially reduced training time. The method improves upon the uncorrected numerical solver in all five configurations and achieves the lowest nRMSE among the evaluated methods in three. It reports training speedups of 4.0–38.9× over the SITL baselines under the tested protocols.

Effective use of temporal information through fixed reservoir dynamics. The reservoir incorporates information from the numerical solver's history without requiring recurrent-weight optimization. The KS memory ablation demonstrates a substantial improvement over an instantaneous feature representation, supporting the usefulness of temporal processing in this setting.

Careful identification of theoretical and experimental limitations. The paper provides a formal analysis of feature consistency, prediction-risk decomposition, conditional stability, and generalization. Although their theoretical significance is limited, the authors appropriately distinguish the properties established from stronger claims that are not justified, including limitations concerning contraction, generalization bounds, and experimental comparisons.

Critical weaknesses

Insufficiently established methodological novelty and distinctive advantage. The main components, such as numerical solvers, fixed reservoirs, Fourier readouts, and ridge regression, are established techniques. The paper does not convincingly identify what advantage is unique to their combination, particularly relative to other open-loop models using the same numerical trajectories.

Limited theoretical novelty and practical significance. Proposition 1 follows almost immediately from the model construction and elementary probability and linear algebra. Proposition 2 is a standard conditional-expectation decomposition, while Theorem 1 adapts existing regularized-learning stability arguments and yields a numerically vacuous bound in the reported experiments. Apparently, none establishes an architecture-specific superiority or a quantitative advantage over alternative feature representations.

Insufficient baselines and ablations to isolate the proposed method's benefits. The experiments lack important comparisons with open-loop FFNO or S4FFNO correctors, simpler history-based regression, and alternative fixed-feature representations. Consequently, the relative contributions of open-loop coupling remains insufficiently isolated.

Limited fairness and comprehensiveness of the experimental evaluation. The SITL baselines are not systematically tuned for accuracy, and the comparisons do not establish accuracy–computation Pareto frontiers under varying budgets. In addition, the experiments cover only short forecast horizons, fixed physical parameters, and periodic domains. Moreover, longer-term forecasting, parameter changes, and distribution shifts remain untested.

Insufficiently clear mathematical notation and methodological specification. Several important operators and conventions are only informally introduced or incompletely specified, including the downsampling operator Pf\mathcal P_f, the correction-insertion map Icorr\mathcal I_{\mathrm{corr}}, and the teacher-forced and closed-loop solver operations. Intermediate SITL rollout initialization is also unclear. Together with overloaded notation and implicit indexing conventions, these issues unnecessarily complicate understanding and reproduction.

Questions for authors

I will consider responses to the weaknesses as answers to my questions.

Reviewer_HDMgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

A coarse solver is rolled out from u0, a fixed local reservoir turns that trajectory into features, and a Fourier ridge readout predicts the high-fidelity coarse states. There is no feedback and no backpropagation. The method beats the solver on all five KS, Burgers and NS settings, is best in three, and trains 4–39× faster than SITL.

Critical strengths

Simple and cheap: The pipeline has no backpropagation through the solver and needs only one set of solver rollouts, reused across all hyperparameter candidates. That reuse is a practical advantage over both SITL and closed-loop RC.

Honest scoping: The paper admits that Thm. 1 is numerically vacuous, that the contraction certificate fails for 4/5 reservoirs, that the SITL baselines were not tuned, and that speedups are not at matched accuracy.

Reproducibility: Solver, data and hyperparameter details in Appendices B–C are thorough.

Critical weaknesses

Unfair baselines: SITL was not tuned while RC got a grid search. SITL–FFNO even does worse than the plain solver in two settings.

Missing key baseline: There is no comparison against a readout on the solver output alone, with no reservoir. Without it, the reservoir’s contribution is unclear.

Short horizons: All experiments use only 20–25 forecast steps. Because the readout cannot correct the solver trajectory, errors should grow once the coarse solver drifts from the truth. The evaluated horizons may be too short to show this.

Confounded memory ablation: The readout is shared across all forecast times and receives no time index. The memoryless variant therefore has no way to know which step it is predicting. The full reservoir starts from h0 = 0 and has very slow leak rates (down to 10−310^{-3}), so its state implicitly encodes time. The 3x gain in Table 3 may come largely from this time encoding rather than from using the solver’s history.

Weak theoretical contribution: Propositions 1–3 in Section 5 are standard results. Theorem 1 seems meaningless, which the authors acknowledge: they state the coefficient “exceeds 103n10^{3n}. With D = 1000 and M = 2048, it is actually about 2×1052\times 10^{5} even at the most favorable λ\lambda. The uniform stability result of Proposition 3 also fails to hold for 4 of the 5 selected reservoirs. The analysis is presented as a main contribution in the Introduction.

Questions for authors
  • With modest tuning of the learning rate, unroll length and batch size, does SITL–FFNO beat the uncorrected solver on KS f = 64 and Burgers f = 32?

  • How does a ridge readout on the solver output alone, without a reservoir, compare to the full method?

  • Can authors show error versus forecast step, and results on longer horizons for KS and NS? At what point does SITL overtake proposed method?

  • Does the memoryless variant close the gap in Table 3 if the forecast step n is added as an input feature?

  • Most results in Section 5 would hold for any fixed feature map combined with ridge regression. Can the analysis explain why the reservoir or its memory should improve accuracy, for example by showing when it reduces the representation error ErepE_{rep} in Proposition 2?

Reviewer_C4jEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an open-loop solver–reservoir framework for coarse-grid PDE forecasting. A fixed numerical solver generates an approximate trajectory from the observed initial condition, which drives a fixed recurrent reservoir. A shared Fourier-space linear readout is then trained using closed-form ridge regression, avoiding backpropagation through the solver or reservoir. The authors provide conditional feature-consistency, risk-decomposition, stability, and generalization results. Experiments on Kuramoto–Sivashinsky, Burgers, and Navier–Stokes demonstrate improvements over the uncorrected solver in all five configurations, outperforming the compared methods in three, with reported training speedups of 4.0–38.9× over gradient-based solver-in-the-loop baselines.

Critical strengths
  • The proposed architecture is simple, computationally efficient, and avoids expensive solver-in-the-loop backpropagation.

  • The empirical results are promising across three PDE families, particularly the KS 64-grid and Burgers experiments.

  • The reservoir-memory ablation provides evidence that temporal features improve forecasting relative to instantaneous features in the tested KS setting.

Critical weaknesses
  • Limited methodological novelty. Hybrid physics–reservoir forecasting and closed-form reservoir training are well established (e.g., Pathak et al., 2018; Wikner et al., 2020). The main contribution appears to be a relatively straightforward open-loop variant with a Fourier readout.

  • Missing simple baselines. The paper does not compare against direct Fourier ridge regression or nonlinear autoregressive regression on the same solver-generated trajectories. Without these controls, it is difficult to determine how much of the performance gain comes specifically from recurrent reservoir features.

p Potentially unbalanced comparisons. The SITL baselines are not systematically tuned for accuracy, whereas the reservoir methods receive hyperparameter selection. Consequently, the reported accuracy and training-time advantages do not establish a matched-budget or matched-accuracy comparison.

  • Limited theoretical contribution. The feature-consistency and risk-decomposition results largely follow from standard assumptions and conditional-expectation arguments. The expected-risk bound is acknowledged to be numerically vacuous, and the sufficient contraction condition holds for only one of five evaluated configurations.

  • Limited robustness and efficiency analysis. The main results lack multi-seed uncertainty estimates, inference-time measurements, and longer-horizon or parameter-shift evaluations.

Questions for authors
  1. Could the authors compare against a simpler Fourier ridge readout using the same solver states, ideally including a few lagged states or nonlinear autoregressive features? This seems feasible using the existing cached trajectories and would help establish the necessity of the reservoir.

  2. Could the authors clarify the methodological distinction from established hybrid reservoir forecasting, particularly Pathak et al. (2018) and Wikner et al. (2020)?

  3. Could the authors report inference-time costs and discuss the efficiency–accuracy trade-off, given that a full numerical solver and reservoir must still be evaluated at deployment?

  4. How sensitive are the main results to random reservoir initialization? Could the authors provide additional seed results for one or two representative configurations?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.