acceptodds
Under review as a conference paper at ICLR 2027

When Do Physical Constraints Correct Neural PDE Rollouts?

Abstract

Autoregressive neural PDE surrogates are routinely regularized and corrected with physical constraints, yet the relationship between enforcing a constraint and reducing prediction error remains poorly understood. We study this relationship for gradient-flow rollouts, where the free energy must decrease along the trajectory. Enforcing this temporal condition on fixed pretrained surrogates yields a discrete Lyapunov certificate, and the same certificate produces sharply different field correction: on Allen–Cahn it removes 99.9% of the structural error and 75.6% of the field error; on Cahn–Hilliard it removes 96.5% of the structural error and 1.1% of the field error. We explain this contrast through the certificate-to-error ratio . The same-step field benefit obeys , and for small corrections , where is the correction–error alignment. The bound holds exactly across 15,412 activated steps, and the alignment sign predicts whether enforcement corrects or degrades a rollout across systems, distribution shifts, backbones, and independent training seeds. Autoregressive feedback can amplify a small local correction into a complete long-horizon rescue, while lower field error alone does not ensure correct dynamics. Constraint satisfaction is therefore an incomplete proxy for prediction quality: certification, field correction, and dynamical fidelity are distinct measurable properties of a learned rollout.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural PDE surrogates, neural operators, autoregressive rollouts, physics-constrained machine learning, physical constraints, scientific machine learning, dynamical systems, Lyapunov stability

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_EA9yOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether enforcing physical constraints improves the accuracy of autoregressive neural PDE surrogates. The authors focus on gradient-flow systems and enforce free-energy monotonicity through inference-time convex-splitting corrections. They establish a discrete energy monotonicity guarantee and introduce a certificate-to-error ratio and correction-error alignment to characterize prediction improvements. Experiments on Allen–Cahn, Cahn–Hilliard, and grain-growth systems show that satisfying the same energy constraint can lead to substantially different improvements in prediction accuracy. The paper also examines the effects of corrections on long-horizon rollouts and physical evolution kinetics.

Critical strengths

Strength

  1. The paper tried to address a meaningful question concerning the relationship between physical constraint satisfaction and prediction accuracy. In particular, the experiments distinguishing energy monotonicity, field accuracy, and coarsening kinetics are useful. The inclusion of unsuccessful corrections, mis-specified energy functionals, and different enforcement strategies provides a more informative picture than simply demonstrating improved prediction accuracy.

  2. The paper evaluates multiple gradient-flow systems, neural architectures, and training configurations. The contrasting results for Allen–Cahn and Cahn–Hilliard provide an interesting empirical motivation for investigating the limitations of energy-based constraints.

Critical weaknesses

Weaknesses

  1. Inference-time physical constraint enforcement for neural PDE surrogates has already been studied. For example, Rochman Sharabi and Louppe (2025) enforce governing-equation constraints through training-free projections. The distinction made in Section 2 between “temporal free-energy sublevels” and “instantaneous residuals” is not sufficiently justified. The prior work explicitly discusses temporal consistency and jointly enforces PDE constraints over entire space-time trajectories. In contrast, the present paper enforces energy monotonicity sequentially between consecutive states. Moreover, the convex-splitting correction relies on established numerical techniques. The theoretical results are also rather elementary.

  2. The emphasis of this paper seems lie in analysis on the difference correction effectiveness. However, the analytical depth is insufficient:

    (1) Insufficient explanation of the Allen–Cahn and Cahn–Hilliard difference. The paper does not establish how the underlying PDE dynamics and prediction error modes affect the ability of the energy certificate to detect and correct errors. This is particularly relevant since the two systems share the same free-energy functional but exhibit substantially different correction outcomes. Moreover, their baseline prediction errors differ considerably, making it difficult to isolate the role of the underlying dynamics.

    (2) Post-hoc characterization rather than explanation. The proposed ratio and correction-error alignment are computed after the correction and are directly related to the resulting local improvement. They characterize whether a correction is effective, but do not explain why a particular physical constraint produces an effective correction in the first place.

    (3) The role of the enforcement mechanism is unclear. Section 5.2 shows that generic backtracking and convex-splitting corrections satisfy the same energy monotonicity condition but produce very different coarsening exponents. This suggests that the enforcement operator itself plays an important role. In particular, convex splitting already incorporates gradient-flow structure, which may contribute to the observed improvements. However, the paper does not sufficiently separate the effects of the certificate from those of the correction method.

    (4) Limited connection to long-horizon behavior. The theoretical analysis focuses on single-step corrections, whereas the main experimental observations concern autoregressive rollouts. The authors themselves show that local and rollout-level improvements can differ substantially. For example, Appendix F, Table 4 reports cases with negative median local benefits but positive rollout benefits. The proposed analysis does not sufficiently explain how local correction properties translate into long-horizon prediction accuracy.

Questions for authors
  1. What additional insight does the correctability analysis provide beyond the geometric identities? Since the local benefit is algebraically determined by RR and cosθcos\theta, can the authors identify a nontrivial implication that is specific to physical constraint enforcement or the underlying PDE dynamics?

  2. Can the authors provide a more direct explanation of the different correction effectiveness for Allen–Cahn and Cahn–Hilliard? In particular, how does this relate to their different dynamics and prediction error modes? Would the conclusions remain the same when controlling for baseline surrogate accuracy?

  3. How much of the observed improvement comes from the energy certificate itself, rather than the convex-splitting correction operator? Given the substantially different kinetics obtained with the same energy constraint, can the authors clarify the role of the enforcement mechanism and its influence on the effective rollout dynamics?

5 comments
AuthorsOpenReview ↗

Response to Reviewer EA9y (1/3)

Thank you for the report, and in particular for the Allen-Cahn / Cahn-Hilliard question; the replies follow your numbering, and the shared answer for each concern sits at the end of the section that raised it first.

W1: inference-time enforcement is prior work

"Inference-time physical constraint enforcement for neural PDE surrogates has already been studied. For example, Rochman Sharabi and Louppe (2025) enforce governing-equation constraints through training-free projections. The distinction made in Section 2 between "temporal free-energy sublevels" and "instantaneous residuals" is not sufficiently justified."

We accept that inference-time projection is prior art and do not claim the projection; Section 3.3 and App. J say the same. What Section 2 now separates is the algebraic type and cost of the constraint: the closest prior work enforces a PDE residual at inference, a space-time equality that requires the operator, while we enforce a scalar temporal inequality that requires only FhF_h and a descent step. The prior work you cite does enforce consistency over space-time trajectories; that is a space-time equality, and it is a different object from a sequential inequality between consecutive accepted states. The elementary-algebra half of your point is answered below. Revision. Section 2 states the distinction as algebraic type and cost, and Section 2, Section 3.3 and App. J state that the projection mechanism is not claimed as novel.

W2(1): the Allen-Cahn / Cahn-Hilliard difference

"Insufficient explanation of the Allen-Cahn and Cahn-Hilliard difference. The paper does not establish how the underlying PDE dynamics and prediction error modes affect the ability of the energy certificate to detect and correct errors."

The two systems differ in how much of the error the energy can see at their own interface widths, and that is what RR measures. At matched visibility the system dummy on RR is −0.0013±0.032-0.0013 \pm 0.032 (n=252n=252; App. O, Table 12), so the system label adds nothing detectable once visibility is fixed. Cahn-Hilliard sits at median R=0.013R=0.013-0.0170.017 with median Blocal=2B_{\text{local}}=2-3×10−53\times10^{-5} across three seeds, and over the Cahn-Hilliard steps the maximum is Blocal=2.9×10−3B_{\text{local}}=2.9\times10^{-3} (Section 4.3, App. F.3). The all-archive maximum, 0.6240.624 at A1 seed 0, is an Allen-Cahn step and is now named as such. The 627-run scan applies three correctors to controlled perturbations of reference states; it is not a re-run of the trained surrogates, so it does not inherit their baseline errors. Revision. Section 4.3 names the Cahn-Hilliard range together with the all-archive maximum, and App. O records the matched-visibility check.

W2(2): post-hoc characterization

"Post-hoc characterization rather than explanation. The proposed ratio and correction-error alignment are computed after the correction and are directly related to the resulting local improvement."

You are right that both quantities are measured after the correction, and we now say so in Section 1 and in the Fig. 3(b) caption. What Section 4.2 adds is a condition the definition does not contain: the certificate can correct only when its budget can close the gap that activated it, and that condition can fail. The identity-versus-capacity split is answered below. The alignment term is presented as a consistency check to O(Rδ)O(R\delta), not as an explanation. Revision. Section 1, Section 4.2 and the Fig. 3(b) caption label the geometry as geometry.

AuthorsOpenReview ↗

Response to Reviewer EA9y (2/3)

W2(3): which part comes from the operator

"The role of the enforcement mechanism is unclear. Section 5.2 shows that generic backtracking and convex-splitting corrections satisfy the same energy monotonicity condition but produce very different coarsening exponents."

The certificate fixes the target τn\tau_n and hence the gap that must be closed; the operator decides how far past it the state travels and in what direction. Section 5.2 now reads Table 3 as an operator ablation on one setting: three interventions reach zero violation and end at relative L2L_2 0.81 (energy descent), 0.72 (monotone) and 0.36 (rate target), against 3.69 unconstrained. The generic-backtracking control is the cleanest split, because it holds the certificate fixed and changes only the operator: it reaches the same monotone target and nearly freezes coarsening, with fitted exponent 0.001 against the reference 0.35. So the certificate supplies feasibility and a bounded budget, and the gradient-flow direction of the corrector supplies the kinetics; Section 3.3 states that convex splitting is established machinery used as a controlled intervention. The rate-target gain is setting-dependent: it over-dissipates a stronger Allen-Cahn rollout and is frequently infeasible for Cahn-Hilliard (Section 5.4).

W2(4): single-step analysis and rollouts

"Limited connection to long-horizon behavior. The theoretical analysis focuses on single-step corrections, whereas the main experimental observations concern autoregressive rollouts."

The single-step question is answered below. The Table 4 rows you cite are the reversals App. F.4 now prints, and the A2 seed 1 row is the escape case where a negative local benefit coexists with a full rollout rescue. Revision. Section 3.3, Section 5.1 and App. F.4.

Q1: insight beyond the identities

"What additional insight does the correctability analysis provide beyond the geometric identities?"

Beyond the identities, the analysis makes one claim the 627 runs could have refuted: an energy inequality can correct only when its correction budget closes the gap that triggered it, which is why the same certificate removes 75.6% of the field error on Allen-Cahn and 1.1% on Cahn-Hilliard (Table 1; abstract). The capacity condition holds in 627/627 controlled runs under the drift-corrected form and in 348/627 under the raw form, with 279 violations (App. O, Table 12). Revision. Section 4.2 and App. O, Table 12.

Q2: controlling for baseline accuracy

"Can the authors provide a more direct explanation of the different correction effectiveness for Allen-Cahn and Cahn-Hilliard? In particular, how does this relate to their different dynamics and prediction error modes? Would the conclusions remain the same when controlling for baseline surrogate accuracy?"

The control you ask for is the matched-visibility regression in App. O: the system indicator has no detectable residual effect on RR once visibility is fixed, and the 627 controlled perturbations do not inherit the trained surrogates' baseline errors. Because the conclusions are stated in terms of RR rather than the system label, a different baseline accuracy moves RR and moves the prediction with it. Revision. App. O, Table 12 and Section 4.3.

AuthorsOpenReview ↗

Response to Reviewer EA9y (3/3)

Q3: certificate versus correction operator

"How much of the observed improvement comes from the energy certificate itself, rather than the convex-splitting correction operator? Given the substantially different kinetics obtained with the same energy constraint, can the authors clarify the role of the enforcement mechanism and its influence on the effective rollout dynamics?"

The operator ablation in Table 3 and the generic-backtracking control separate the two, and the kinetics follow the operator's direction rather than the certificate. Revision. Section 5.2 and Section 5.4.

What the analysis adds beyond the geometric identities

"Since the local benefit is algebraically determined by RR and cosθcos\theta, can the authors identify a nontrivial implication that is specific to physical constraint enforcement or the underlying PDE dynamics?"

Blocal≤RB_{\text{local}}\le R is the reverse triangle inequality, and Section 1 states this. Blocal≈Rcos⁡θB_{\text{local}}\approx R\cos\theta is first order, and the Fig. 3(b) caption now labels it an identity to O(Rδ)O(R\delta) - a consistency check rather than a test. The statement that is not an identity is the capacity condition: the energy gap that activated the certificate must be closed along the segment, so the mean value theorem and Cauchy-Schwarz give a lower bound on the correction the corrector must be able to express. Across 627 controlled perturbations with three correctors (ideal projection, convex splitting, gradient-flow line search), R≥ν~R\ge\tilde\nu holds in all 627 with minimum slack 3.9×10−53.9\times10^{-5}, while R≥νR\ge\nu holds in 348 (Section 4.2, App. O, Table 12). The two forms differ by the reference's own dissipation, 0.44 at the median run, which is why the naive form fails. This concern was also raised by Reviewer MqNi (W2) and Reviewer seKT (W1); their answers point here.

Single-step analysis, long-horizon rollouts

"The authors themselves show that local and rollout-level improvements can differ substantially. For example, Appendix F, Table 4 reports cases with negative median local benefits but positive rollout benefits."(4)

The analysis is a single-step statement, and Section 3.3 now states this: the certificate constrains admissibility, not the size of the correction, and it is not a convergence statement for the learned trajectory. Alignment predicts the sign of the rollout effect on the seed-resolved rows, which we reconcile in Reviewer MqNi's answer. The size is set by autoregressive feedback: on the unstable U-Net rollout, sustained correction rescues the trajectory (Brollout=1.000B_{\text{rollout}}=1.000 at a per-step R=0.034R=0.034), while a single certified intervention at an early step leaves it to re-diverge (Brollout≈0.002B_{\text{rollout}}\approx0.002; Section 5.1, Fig. 4). We do not have a theorem that maps local correction properties to long-horizon accuracy, and the paper does not claim one. This concern was also raised by Reviewer seKT (W3, Q2); that answer points here.

Reviewer_MqNiOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies when enforcing a physical constraint improves autoregressive neural PDE predictions. The authors apply an inference-time corrector to fixed pretrained surrogates for gradient-flow systems (Allen–Cahn, Cahn–Hilliard), enforcing non-increasing free energy. They observe that the same constraint yields very different field-error reductions across systems, and analyze this through the certificate-to-error ratio and correction–error alignment. They also distinguish local correction from its effect over a rollout. Control experiments illustrate that satisfying the energy constraint, improving field accuracy, and preserving the dynamics do not necessarily go together.

Critical strengths
  • The paper distinguishes constraint satisfaction, field accuracy, and dynamical fidelity. The generic-descent and mis-specified-energy controls help illustrate why these properties need separate evaluation.
  • The authors clearly report limitations, including the classical solver's advantage in the runtime comparison and failures of the dissipation-rate target. The code and archived measurements support reproducibility.
Critical weaknesses
  1. The central framing obscures the role of the physical constraint. Free-energy monotonicity is a requirement of physically valid trajectories. Its enforcement is justified by physical consistency, regardless of whether it reduces field error. The empirical question is how much additional error correction the chosen enforcement procedure provides, and what determines that benefit. Yet the paper presents successful certification without substantial field correction as a "puzzle" and draws as a central conclusion that constraint satisfaction is "an incomplete proxy for prediction quality". Although the authors acknowledge the necessary-but-insufficient nature of the condition, the organizing narrative gives insufficient weight to this distinction and presents an expected limitation as a substantive finding. The contribution therefore needs to rest on explaining which errors the constraint and its enforcement procedure can correct, beyond demonstrating that physical admissibility does not ensure accuracy. Reframing alone would not resolve this concern; the missing analysis is discussed in point 3.

  2. The analytical and empirical evidence does not fully support the stated claims. The bound Blocal≤RB_{\mathrm{local}}\le R follows from the triangle inequality for any correction, and Blocal≈Rcos⁡θB_{\mathrm{local}}\approx R\cos\theta is a first-order expansion. Verifying the bound over 15,412 steps therefore provides little additional evidence about physical constraint enforcement. The broader claim that alignment predicts the direction of the rollout effect also needs qualification. Table 4 reports opposite signs for median local benefit and final rollout benefit in U-Net A2 seeds 0–1 and Cahn–Hilliard seed 2. Figure 3(b) reports close agreement between model-level medians of BlocalB_{\mathrm{local}} and Rcos⁡θR\cos\theta. If this agreement also holds for these rows, their sign reversals would contradict the abstract's broad alignment-based claim. Could the authors report the corresponding alignment statistics and reconcile these cases with that claim? Appendix F.1 excludes cases with activation rates below 0.10, yet A1 seeds 0 and 2 and D seed 1 fall below this threshold and appear to remain in the analysis. The two A1 seeds are also included in the five-seed claim. This discrepancy needs clarification.

  3. The analysis does not sufficiently characterize which errors the constraint can detect and correct. The reported contrast does not isolate the system from the surrogate's error structure: within Allen–Cahn, enforcement ranges from degradation to full rescue across models and seeds. The Table 2 caption attributes gains to "energy-visible" errors, but this property is not examined directly. The ratio RR describes the realized correction relative to the error; it does not explain why the correction is small or poorly aligned in particular regimes. Controlled perturbations of reference states, or an analysis of error components relative to local energy sensitivity, could help separate the effects of the system, error structure, and correction procedure. Such evidence would provide a more substantive explanation of the observed contrast. Since RR and alignment require reference trajectories, they do not themselves provide a deployment-time diagnostic. To support the claim in Section 4.4 that the analysis quantifies how far activation can be trusted as a correction signal, the authors would need to establish whether reference-free observables can indicate when enforcement is likely to help.

Questions for authors

The main requests are in the Weaknesses section. The following clarifications would also be helpful; additional experiments are welcome but not expected.

  1. Positioning relative to prior work. Energy dissipation has been incorporated into neural operators for gradient flows (e.g., Zhang et al., "Energy-Dissipative Evolutionary Deep Operator Neural Networks", J. Comput. Phys., 2024), and conservation constraints have been enforced on learned PDE predictions after training (e.g., Hansen et al., "Learning Physical Models that Can Respect Conservation Laws", ICML 2023). Could the authors clarify how their setting and findings relate to these works?

  2. Headline percentages. How are the structural and field reductions in Table 1 computed, including the evaluation time and the order of normalization and averaging? For the same weak-surrogate Allen–Cahn setting, Table 3 reports a decrease from 3.69 to 0.72, corresponding to about 80.5%, whereas Table 1 reports 75.6%. Could the authors explain this difference?

  3. Fallback frequency. At the default correction budget, how often is the bisection fallback used in the main experiments, particularly for Cahn–Hilliard and the seeds where enforcement degrades accuracy? Appendix M.3 reports fallback frequencies only for a single Allen–Cahn diagnostic run.

4 comments
AuthorsOpenReview ↗

Response to Reviewer MqNi (1/2)

We thank the reviewer for their comments.

W1: framing of the constraint's role

"The central framing obscures the role of the physical constraint. Free-energy monotonicity is a requirement of physically valid trajectories. Its enforcement is justified by physical consistency, regardless of whether it reduces field error."

We agree: monotonicity is justified by physical consistency, and field correction is an empirical question about the enforcement procedure, not a reason to enforce. The contribution is not the observation that admissibility does not ensure accuracy; it is which errors the certificate and its corrector can reach, which is the capacity and alignment analysis of Section 4.2-Section 4.3. The abstract states certification, field correction and dynamical fidelity as three distinct measurable properties, and Section 6 turns them into a measurement rule. Those two analysis blocks are answered in our response to Reviewer EA9y and under Local sign versus rollout sign below. Revision. abstract, Section 4.2-Section 4.3 and Section 6.

W2: the bound and the sign reversals

"The bound Blocal≤RB_{\mathrm{local}}\le R follows from the triangle inequality for any correction, and Blocal≈Rcos⁡θB_{\mathrm{local}}\approx R\cos\theta is a first-order expansion. Verifying the bound over 15,412 steps therefore provides little additional evidence about physical constraint enforcement."

We agree on both counts, and the bound half is answered in our response to Reviewer EA9y; Section 4.2 treats the bound as a cap, not as the evidence, and puts the weight on RR, which orders regimes by realized benefit better than the raw energy overshoot does. For the cases you name, A≥0.10A\ge0.10 is a count, not an exclusion. A1 seeds 0 and 2 and D seed 1 sit below it and remain in Table 4 and in the five-seed count, while the 3/3 count applies the cut and App. F.1 excludes only detection-negative rows (A=0A=0). The sign reversals in Table 4 are answered below. Revision. Section 4.4 reports both counts, App. F.3 gives the seed-level table, and App. F.4 names the 6/6 stratum and the below-cut reversals.

W3: which errors, and the reference-free diagnostic

"The analysis does not sufficiently characterize which errors the constraint can detect and correct. The reported contrast does not isolate the system from the surrogate's error structure: within Allen-Cahn, enforcement ranges from degradation to full rescue across models and seeds."

The controlled-perturbation scan is that isolation, and it is the scan you proposed: three correctors on perturbed reference states, with the system label carried as a regression dummy rather than as a difference in baseline error (App. O, Table 12). Across the 627 perturbations the drift-corrected condition holds in 627/627 and the raw form in 348/627 (279 violations), so the scan separates the two forms rather than confirming a bound. On the reference-free half of this point we concede: we have not established an observable that indicates, without the reference, when enforcement is likely to help. Section 4.4 no longer claims that the analysis quantifies how far activation can be trusted; what stands is narrower - activation is a reference-free monitor of the encoded law, and the correction content requires the reference (Section 3.5, Section 6). The reference requirement itself is answered below. Revision. Section 3.5, Section 4.4 and Section 6.

The main requests are in the weaknesses, and we have answered them there; the scan you proposed in W3 is now Section 4.2, and the three clarifications below need no further runs.

Q1: positioning relative to prior work

"Energy dissipation has been incorporated into neural operators for gradient flows (e.g., Zhang et al., "Energy-Dissipative Evolutionary Deep Operator Neural Networks", J. Comput. Phys., 2024), and conservation constraints have been enforced on learned PDE predictions after training (e.g., Hansen et al., "Learning Physical Models that Can Respect Conservation Laws", ICML 2023). Could the authors clarify how their setting and findings relate to these works?"

Your two references build a physical constraint into the model, dissipative in one case and conservative in the other. We keep the surrogate fixed and enforce a scalar temporal inequality at inference, and we ask what that enforcement reveals about the error that remains (Section 2). The findings are complementary rather than competing: making a law hold and measuring what holding it buys are different claims, and the second is regime-dependent here. Revision. Section 2 separates training-time structure from inference-time enforcement in one paragraph and states the setting as complementary.

AuthorsOpenReview ↗

Response to Reviewer MqNi (2/2)

Q2: headline percentages

"How are the structural and field reductions in Table 1 computed, including the evaluation time and the order of normalization and averaging? For the same weak-surrogate Allen-Cahn setting, Table 3 reports a decrease from 3.69 to 0.72, corresponding to about 80.5%, whereas Table 1 reports 75.6%."

Table 1's percentages are the relative drop in the mean final-step error over trajectories, so the average over trajectories is taken first and the ratio second: n=6n=6, T=250T=250 for Allen-Cahn and Cahn-Hilliard, and n=4n=4, T=50T=50 for grain growth. The structural column applies the same relative drop to the energy error. The two Allen-Cahn numbers come from different runs: 75.6% is the Table 1 run at T=250T=250, and 80.49% is the separate Table 3 run with n=4n=4 and T=300T=300. Table 12 records both rows with their trajectory counts and horizons, and the Table 1 caption now states the protocol. Revision. Table 1 caption, Table 12, App. O.

Q3: fallback frequency

"At the default correction budget, how often is the bisection fallback used in the main experiments, particularly for Cahn-Hilliard and the seeds where enforcement degrades accuracy? Appendix M.3 reports fallback frequencies only for a single Allen-Cahn diagnostic run."

The archived frequency comes from the κmax⁡\kappa_{\max} sweep in Table 10, which is one diagnostic Allen-Cahn trajectory; we do not hold a per-seed fallback count for Cahn-Hilliard or for the degrading seeds, and we say so rather than extrapolate. In that sweep the fallback is not reached at the default budget (0% at κmax⁡≥2\kappa_{\max}\ge2), and only at κmax⁡=1\kappa_{\max}=1 is it invoked, on 43.1% of steps, where the certificate still holds at zero violations (Table 10). Revision. App. M.3 states that the sweep is a single diagnostic trajectory, and Section 3.3 points to it for the backstop.

Local sign versus rollout sign

"The broader claim that alignment predicts the direction of the rollout effect also needs qualification. Table 4 reports opposite signs for median local benefit and final rollout benefit in U-Net A2 seeds 0-1 and Cahn-Hilliard seed 2."

The paper now states the denominator. Under the paper's A≥0.10A\ge0.10 rule the sign of Rcos⁡θR\cos\theta matches the sign of BrolloutB_{\text{rollout}} in 3/3 Allen-Cahn A1 seeds (exact one-sided p=0.125p=0.125). Across all five A1 seeds it matches in 5/5 (p=0.031p=0.031), and the two seeds below the cut (A1 seeds 0 and 2) are themselves aligned and improve, which is why the five-seed count is also reported (Section 4.4; App. F.3). The abstract's 6/6 is a count over the six seed-resolved rows with ∣Blocal∣≥0.02|B_{\text{local}}|\ge0.02, and App. F.4 lists the six above-cut rows (A1 seeds 0-4 and D seed 1; the E aggregate row is excluded because it is not per-seed). The Cahn-Hilliard row in your quote is a numerical zero at that resolution, and the A2 seed 1 reversal is the amplification case answered in our response to Reviewer EA9y. App. F.4 prints the reversals below the cut, so the 6/6 is a count over a stated stratum rather than a statement about every row.

Which quantities need the reference

"Since RR and alignment require reference trajectories, they do not themselves provide a deployment-time diagnostic."

Correct. Activation and the violation rate are reference-free; RR, BlocalB_{\text{local}} and BrolloutB_{\text{rollout}} are evaluation quantities that require the reference. Section 3.5 states the division, and Section 6 now states it once: Section 6 no longer calls the alignment term reference-free. This concern was also raised by Reviewer seKT (W2); that answer points here.

Thank you for the response. I agree that enforcing monotonicity is justified on physical grounds. The rollout check verifies the bound Blocal≤RB_{\text{local}} \le R. To tell what enforcement actually corrects, one has to look at the measured field errors. I understand that the authors are withdrawing the Section 4.4 claim because they have yet to establish a reference-free observable that indicates when enforcement helps. The clarification of the A≥0.10A \ge 0.10 stratum, the ∣Blocal∣≥0.02|B_{\text{local}}| \ge 0.02 stratum used for the 6/6 result, and the reversals listed in Appendix F.4 resolves my question about which rows the reported claims cover.

With the Section 4.4 claim withdrawn, the practical value of the remaining analysis is limited. Physical consistency already gives practitioners a reason to enforce monotonicity. The reported measurements quantify how much field error is corrected when a reference trajectory is available. In settings where practitioners lack that reference, the results give them little guidance on when to expect better field accuracy.

The remaining question for me is which kinds of field error the constraint can correct. The Table 2 caption attributes the gains to errors that are visible to the energy functional. Does the perturbation scan in Appendix O report the realised correction separately for the error components visible to the energy functional and those it cannot detect? I found RR and alignment reported across 627 perturbations, but I could not find where that attribution was tested. If the scan reports only RR and alignment, it extends the reference-dependent analysis to more states and leaves open which error components were corrected.

The withdrawal addresses my objection to the Section 4.4 claim. With that claim gone and the error-component attribution still untested in the results I could find, I find the remaining contribution too limited to raise my score.

AuthorsOpenReview ↗

Follow-up Response to Reviewer MqNi

Thank you for the follow-up. You are right that Appendix O prints RR, ν\nu and the alignment and not the split. Based on your feedback, we have added it as Table 13 (App. O in the revised pdf), computed from the full run file of the scan with no row excluded. The split is an identity, Blocal=Bvis+BtanB_{\text{local}} = B_{\text{vis}} + B_{\text{tan}} (Eq. 8), and it holds to 1.0×10−101.0\times10^{-10} on the run file.

Across runs the visible term carries 78% of the realised benefit under the minimum-norm projection (252 runs), 74% under convex splitting (123 Allen-Cahn runs) and 64% under the line-search gradient flow (252 runs). In the Cahn-Hilliard line-search cell it carries 56%. Most of the tangential share of the line search sits in the 157 runs that reach the bisection backstop, where the mean BtanB_{\text{tan}} is 0.38, against 0.05 in the 95 runs that converge. The tangential term is negative in 31% of the projection runs and in 17% of the convex-splitting runs. The correction acts on the energy-visible component, and this component alone does not account for the full realised benefit.

As panel (b) of Table 13 shows, the split follows the injected structure. It ranges from 97-98% visible for amplitude and smoothing perturbations to 20% for the tangent family (median ν=0.37\nu = 0.37), with i.i.d. noise at 78%. Following your comment, we have also updated the caption of Table 2. It now reports the controlled Cahn-Hilliard split and points to Table 13.

At the step level the archive agrees. Among the activated Allen-Cahn steps, none of the 5,552 with ν<0\nu < 0 reduces the field error, while 2,374 of the 2,660 with ν>0\nu > 0 do.

We hope this addresses your remaining question. If this is indeed the case, we kindly ask you to consider increasing your score. Please let us know if you still have any concerns.

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

In this paper, the authors investigate an interesting problem: whether enforcing physical constraints on neural PDE surrogates can improve prediction accuracy. The authors find that applying the same physical constraint to different systems can lead to significantly different outcomes. To analyze this problem, they propose a temporal certification procedure that defines a time-dependent feasible set based on the free-energy condition and, when applicable, a mass conservation constraint. As long as the correction operator returns a state within this feasible set at each step, the discrete free energy is guaranteed to be non-increasing. Furthermore, the authors introduce the certificate-to-error ratio, which measures the magnitude of the physical correction relative to the prediction error. They theoretically show that the same-step field improvement is bounded above by this ratio. They further analyze the alignment between the correction and the prediction error, showing that the same-step improvement can be approximated by the certificate-to-error ratio multiplied by the cosine of the alignment angle when the correction is small relative to the prediction error. Finally, the authors conduct experiments across different PDE systems and neural network architectures to validate their analysis and demonstrate how physical constraint enforcement can improve, have little effect on, or even degrade prediction accuracy.

Critical strengths

1:The paper investigates an interesting and important problem in physics-informed machine learning. Understanding whether enforcing physical constraints can actually improve prediction accuracy is valuable for neural PDE surrogates.

2: The authors provide an intuitive analysis of the relationship between physical constraint enforcement and prediction error. The certificate-to-error ratio and correction alignment help explain why the same physical constraint can lead to different prediction outcomes across systems.

3: The experiments are comprehensive, covering different PDE systems, neural network architectures, distribution shifts, and independent training seeds. The authors also include additional control experiments to analyze the effectiveness and limitations of physical constraint enforcement.

Critical weaknesses

1: The main theoretical contributions are based on standard vector geometry, such as the triangle inequality and correction-error alignment. While the analysis provides an interesting perspective, the mathematical novelty is relatively limited for a theory-oriented paper.

2: The proposed certificate-to-error ratio and alignment analysis require access to the ground-truth solution, which can be expensive to obtain for PDE problems. This limits the practical applicability of the proposed metrics and makes it difficult to extend the conclusions to general PDE settings where reference solutions are unavailable.

3: The theoretical analysis mainly focuses on single-step correction. Although the authors provide experiments demonstrating long-horizon rollout improvements, a more comprehensive theoretical analysis of how local corrections affect long-term prediction accuracy is still missing.

Questions for authors

1: Can the authors further clarify how the current theoretical results can be generalized to broader PDE systems, especially those without an explicit energy-dissipation structure?

2: Can the authors provide further theoretical insights into long-term prediction accuracy? In particular, how do single-step corrections affect error accumulation and the stability of long-horizon rollouts?

3 comments
AuthorsOpenReview ↗

Response to Reviewer seKT

We thank the reviewer for their comments.

W1: mathematical novelty

"The main theoretical contributions are based on standard vector geometry, such as the triangle inequality and correction-error alignment. While the analysis provides an interesting perspective, the mathematical novelty is relatively limited for a theory-oriented paper."

We agree that the geometry is standard, and the Fig. 3(b) caption now labels Blocal≈Rcos⁡θB_{\text{local}}\approx R\cos\theta a consistency check rather than a test. The claim we defend is the capacity condition, which holds in 627/627 runs under the drift-corrected form and fails in 279 of 627 under the raw form (App. O, Table 12). Revision. Section 1, Section 4.2 and the Fig. 3(b) caption.

W2: reference requirement

"The proposed certificate-to-error ratio and alignment analysis require access to the ground-truth solution, which can be expensive to obtain for PDE problems."

RR and the alignment require the reference, and activation and the violation rate do not. The reference half is answered in our response to Reviewer MqNi, and the applicability half below. The paper claims an evaluation protocol rather than a deployment controller, and Section 3.5 now says that in one place. Revision. Section 3.5 and Section 6.

W3: long-horizon theory

"The theoretical analysis mainly focuses on single-step correction. Although the authors provide experiments demonstrating long-horizon rollout improvements, a more comprehensive theoretical analysis of how local corrections affect long-term prediction accuracy is still missing."

The single-step scope is stated in Section 3.3, and we do not have a theorem that maps local correction to long-horizon accuracy. The paper adds the measured separation between local correctability and rollout outcome, with the escape case reported rather than averaged away (App. F.4). Revision. Section 3.3, Section 5.1 and App. F.4.

Q1: broader PDE systems

"Can the authors further clarify how the current theoretical results can be generalized to broader PDE systems, especially those without an explicit energy-dissipation structure?"

Section 6 now states the formal scope: the certificate covers gradient-flow systems with an evaluable free energy, and extensions to other temporal structures need their own guarantees. What transfers without that structure is the analysis rather than the certificate, because any correction-returning constraint admits the same ratio, bound and alignment decomposition (Section 6). Revision. Section 6.

Q2: long-term error accumulation

"Can the authors provide further theoretical insights into long-term prediction accuracy? In particular, how do single-step corrections affect error accumulation and the stability of long-horizon rollouts?"

The one structural statement is Proposition 2, which bounds the energy sequence along the rollout but not the field error (Section 3.3). Error accumulation itself is measured rather than bounded, and the single-intervention control is what isolates accumulation as the mechanism: sustained correction rescues the rollout, and one early certified step does not (Section 5.1). Revision. Section 3.3 and Section 5.1.

Deployment boundary

"This limits the practical applicability of the proposed metrics and makes it difficult to extend the conclusions to general PDE settings where reference solutions are unavailable."

The limit is real, and it is measured. On the inexpensive 2D benchmark, a coarse-step energy-stable solver reaches relative L2=0.01L_2=0.01 in 0.34 s with exponent 0.47, against 2.18 in 0.61 s for the corrected surrogate (Table 5). That is the appropriate tool when a classical forward model is available; the construction is relevant when a learned surrogate is already the forward model and a structural rollout property is required (Section 6).

I really appreciate the detailed explanation. I will keep my score. I encourage the author to further improve their manuscript based on the review from all reviewers here.

AuthorsOpenReview ↗

Thank You to Reviewer seKT

Thank you very much again for taking the time in reviewing our paper and in the rebuttal. Following the discussion with the other reviewers, we have added a split of the realised correction into an energy-visible and a tangential term (Eq. 8 and Table 13 in the revised pdf). Please let us know if any further information would be helpful. Thank you again for your time and effort.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for their constructive feedback and valuable insights. The reviews converge on the same concern: the bound Blocal≤RB_{\text{local}}\le R is elementary geometry, and the quantities that carry the analysis need the reference. The paper treats the bound as an identity and puts the weight on the one statement the geometry leaves open, the capacity test in Section 4.2: across 627 controlled perturbations of reference states with three correctors (ideal projection, convex splitting, gradient-flow line search), the drift-corrected condition R≥ν~R\ge\tilde\nu holds in 627/627 with minimum slack 3.9×10−53.9\times10^{-5}, while the raw condition R≥νR\ge\nu holds in only 348/627, leaving 279 violations (App. O, Table 12). That gap between the two forms is the substance the identities leave open, and it is explained by the reference's own dissipation, 0.44 at the median run. Section 3.5 and Section 6 state the reference division once: activation and the violation rate are reference-free, while RR, BlocalB_{\text{local}} and BrolloutB_{\text{rollout}} require the reference. Revisions are marked in blue in the revised PDF.

AuthorsOpenReview ↗

Summary of the Rebuttal for the Area Chair and Reviewers

We thank the Area Chair and the reviewers for their time and valuable feedback. As the discussion period is ending, we thought it could be useful to summarize the current state of the rebuttal.

We would first like to highlight the main result of the paper. Physical constraints are already routinely used to regularize and correct neural PDE surrogates, and a satisfied constraint is often read as a sign of a better prediction. Before this work there was, to the best of our knowledge, no controlled test of what enforcing a constraint does to the prediction error. Our paper provides this test on gradient-flow rollouts. The same certificate removes 99.9% of the structural error and 75.6% of the field error on Allen-Cahn, and 96.5% and 1.1% on Cahn-Hilliard. The certificate-to-error ratio RR explains this contrast. The alignment term Rcos⁡θR\cos\theta follows the measured benefit across 15,412 steps (Pearson 0.991), and the capacity condition holds in 627/627 controlled runs. All three reviews describe this question as meaningful or important, and two of them list the separation of constraint satisfaction, field accuracy and dynamical fidelity as a strength of the paper.

We initially received scores 1, 1, 3. Based on reviewer feedback, we added the following to our manuscript:

  • A new exact split of the realised correction into an energy-visible and a tangential term (Eq. 8), reported for all 627 controlled runs. The visible term carries 78% of the realised benefit under the minimum-norm projection, 74% under convex splitting and 64% under the line search, and the split follows the injected error structure (App. O, Table 13).
  • A new controlled scan of 627 perturbations of reference states with three correctors. The capacity condition holds in 627/627 runs in the drift-corrected form and in 348/627 in the raw form (Section 4.2, App. O, Table 12).
  • A new matched-visibility check. Once visibility is matched, the system indicator has no detectable effect on RR (−0.0013±0.032-0.0013 \pm 0.032, n=252n = 252; App. O, Table 12).
  • A seed-level report with stated denominators. The sign of Rcos⁡θR\cos\theta matches the sign of the rollout effect in 5/5 Allen-Cahn A1 seeds and in 6/6 seed-resolved rows with ∣Blocal∣≥0.02|B_{\text{local}}| \ge 0.02 (Section 4.4, App. F.3, App. F.4).
  • A statement of which quantities need the reference (Section 3.5, Section 6).

During the discussion phase, we provided detailed responses to the concerns raised by all three reviewers and revised the paper accordingly. The rating changes and current status of the three reviewers are as follows:

ReviewerOriginal RatingAfter Discussion
EA9y11 (Replied on Nov. 19)
MqNi11 (Replied on Nov. 17)
seKT33 (Replied on Nov. 19)

All three reviewers replied. Reviewer MqNi confirmed that our clarifications resolve the questions about the reported strata and asked for the split of the correction by error component, which is now Table 13. Reviewer EA9y acknowledged the clearer distinction between the certificate and the correction operator and asked one technical question about the capacity condition, which we have answered. Reviewer seKT appreciated the detailed explanation. We have answered every follow-up question, and we hope the Area Chair can take these additions into account. We also kindly ask the reviewers to read our latest responses and to update their scores if their concerns are resolved.

We are extremely grateful for the feedback given by the reviewers, and sincerely believe their recommendations improved the paper. We are happy to answer any question about the paper or the rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.