acceptodds
Under review as a conference paper at ICLR 2027

KMF: Physics-Guided Inference-Time Debiasing for Off-Cadence PDE Foundation Models

Abstract

Scientific workflows require PDE states on temporal grids chosen by downstream analysis, often at a finer resolution than the cadence of available foundation-model outputs or observations. Direct off-cadence model queries can be biased, while standard temporal upsampling ignores the governing dynamics. We introduce Kinematic Manifold Fusion (KMF), a parameter-update-free, physics-guided inference-time correction framework for temporal refinement and autoregressive rollout. When neighboring model snapshots are available, KMF evaluates a cheap spatial PDE operator at those snapshots, constructs a local cubic Hermite physical candidate, and fuses it with a model estimate through a calibration-selected scalar. During rollout, it reuses the same local physics as a causal defect signal before reinjecting the corrected state into the frozen model. It also enforces known constraints through the corresponding exact projection when available. KMF changes no model parameters and uses zero temporal integration substeps, only instantaneous spatial right-hand-side evaluations. Under standard smoothness and operator-error assumptions, we establish a local error bound for the physical bridge and an endpoint-perturbation bound for deployment. Empirical evaluation across PDE systems, temporal cadences, and frozen foundation-model architectures demonstrates accurate temporal refinement and active causal correction in the regimes where local physics is informative, while retaining the efficiency of the underlying model. KMF offers a lightweight complement to learned continuous-time solvers and numerical integration.

Primary area:
learning on time series and dynamical systems
Keywords:
PDE foundation models, inference-time debiasing, temporal refinement, physics-guided inference, off-cadence prediction, scientific machine learning, neural operators, constraint projection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zbnkOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Kinematic Manifold Fusion, a method that uses known PDE dynamics to refine predictions from frozen PDE foundation models without updating their parameters. It combines Hermite interpolation, a correction based on physical consistency, and constraint projection to estimate intermediate states and improve autoregressive rollouts. Experiments on four PDE systems and several model families show improvements at time intervals different from those used in training, with modest inference overhead. However, gains at the native interval often come from projection alone, and some settings show little benefit or worse performance.

Critical strengths
  • KMF works with frozen PDE models without retraining, and the reported inference overhead is modest. This makes it a practical option for improving existing models. The authors measure the cost of the physics evaluations and correction steps, giving readers a concrete picture of the accuracy and runtime trade-off.

  • The authors report failures and sensitivity to calibration and operator mismatch, including cases where calibration disables the correction. These results help establish when the method is useful. They also show where the gains come from projection alone and where the proposed correction adds value, making the positive results easier to assess.

Critical weaknesses
  • Using physics to guide or correct model predictions at inference time has been widely explored [1]. KMF combines established tools such as Hermite interpolation, defect correction, and constraint projection, but the paper does not clearly explain what this combination contributes beyond existing approaches. The limited discussion and comparison with prior work make its novelty difficult to assess.

  • The additional benefit of KMF’s correction depends strongly on the setting. For requested time steps of at least 0.10 seconds, calibration selects a zero correction coefficient in 25 of 32 primary fluid configurations, leaving only constraint projection. Table 16 also reports cases where calibrated KMF performs worse than linear interpolation or the raw model, including Poseidon-L on Wave-Gauss. The authors acknowledge these limitations, but they leave the evidence for benefits beyond projection or retaining the original prediction concentrated in a narrower set of configurations. I would appreciate it if the authors could clarify this.

[1] DiffusionPDE: Generative PDE-Solving Under Partial Observation. Huang et al. NeurIPS 2024.

Questions for authors
  • What is the main methodological advance over existing physics-guided inference and correction methods? Please explain which part of KMF is new and what the experiments show about its benefit beyond combining established numerical techniques.

  • Across the evaluated settings, how often does KMF improve on projection alone or simply retain the raw model prediction? Can you summarize this with confidence intervals over test trajectories?

  • The operator-mismatch experiments recalibrate KMF for each assumed operator. How sensitive is performance to operator errors when the calibration parameters remain fixed?

9 comments
AuthorsOpenReview ↗

On novelty and contribution beyond existing methods

Dear Reviewer zbnk,

Thank you for your constructive feedback. We address your three concerns below.

1. Novelty and contribution beyond existing methods

We agree that physics-guided inference is not new. However, DiffusionPDE addresses a different problem using diffusion-based generation, where the denoising process provides a mechanism for incorporating physical information. This mechanism is not available in general autoregressive PDE foundation models.

KMF addresses two important practical challenges simultaneously: (1) enabling physics-guided correction when the foundation model is available only as a black box, with access limited to its input-output inference interface and the governing PDE operator (unlike diffusion), and (2) enabling accurate predictions at flexible temporal cadences, including off-native intervals where foundation models may otherwise lose accuracy. To the best of our knowledge, KMF is the first training-free framework to jointly address both challenges, without requiring access to model internals, modifying model parameters, or retraining the foundation model.

This setting is important for the practical use of PDE foundation models. When using a pretrained FM, we often have access only to its inference interface, while modifying its parameters or retraining it can be expensive or infeasible. At the same time, scientific applications such as fluid dynamics, climate forecasting, and coupled simulations often require predictions at temporal cadences different from those used during training. Even models supporting variable time intervals can lose accuracy at off-native cadences. KMF addresses these challenges in two stages. First, it enables physics-informed temporal refinement between available FM predictions using a Hermite bridge, without requiring direct off-native queries. Second, it builds on this refinement to perform causal rollout correction, using the Hermite–Simpson physical defect to correct FM predictions before they are fed back into subsequent inference steps. Both stages operate at inference time, without modifying model parameters, retraining, or numerically integrating the PDE forward.

Unlike diffusion-based correction, KMF requires no diffusion scores or denoising trajectory. Unlike Jacobian-based correction approaches such as PhysicsCorrect, it requires no cached Jacobian or pseudoinverse and additionally provides dense intermediate-time predictions.

The experimental evidence supports both the method's effectiveness and transferability. Table 3 isolates the contribution of active physical correction beyond projection, showing improvements of 16.15% on NS-Gauss and 5.73% on FNS-KF. Evaluations across four PDE systems and multiple foundation-model architectures further demonstrate that the same inference-time mechanism can be applied without changing the underlying models.

Thus, our contribution is not the individual numerical components, but a lightweight, architecture-independent approach that jointly addresses off-native temporal refinement and causal physical correction of frozen PDE foundation models.

AuthorsOpenReview ↗

On improvement with correction over projection along & Robustness to operator mismatch without recalibration

2. How often does correction improve over projection alone?

The results demonstrate that KMF's physical correction provides substantial benefits beyond projection alone, rather than merely retaining the original predictions. Across all 126 Task 2 configurations, KMF improves over projection in 81 (64.3%), matches it in 43 (34.1%), and degrades in only 2 (1.6%).

Importantly, correction is actively selected (α∗>0\alpha^*>0) in 83 configurations and improves performance in 81 of them (97.6%). The remaining 43 configurations select α∗=0\alpha^*=0, which is an intentional feature of KMF's calibration: when additional correction is not beneficial on validation data, the method retains the projection-only solution rather than forcing an unnecessary update. Consequently, 124/126 configurations (98.4%) match or improve upon projection, and KMF outperforms the raw FM rollout in 125/126 configurations.

OutcomeAll 126 configurationsActive correction (83)
Improved over projection81 (64.3%)81 (97.6%)
Equal to projection43 (34.1%)0
Degraded2 (1.6%)2 (2.4%)

To establish that these gains are not artifacts of averaging, we additionally report paired results across 20 held-out trajectories for each primary active-correction setting:

DatasetProjection RMSEKMF RMSEGain (95% paired CI)
NS-Gauss0.063320.0530916.15% [15.82%, 16.46%]
FNS-KF0.026230.024725.73% [4.97%, 6.51%]

Both improvements are statistically significant (Holm-adjusted p=7.63×10−6p=7.63\times10^{-6}).

These results establish that physical-defect correction is not merely a fallback to projection: it is actively used in most evaluated configurations and provides consistent, measurable improvements when selected. The only two degradations are 0.10% and 0.22%, demonstrating that failures are rare and small in the evaluated matrix. While no calibration procedure guarantees improvement on every unseen trajectory, KMF combines effective active correction with the ability to avoid unnecessary interventions when validation does not support them.

3. Robustness to operator mismatch without recalibration

To assess sensitivity to operator mismatch, we evaluated KMF using nominal fusion weights calibrated once under the original operator and held fixed across all tested viscosity and forcing perturbations, without recalibrating for each perturbed operator.

DatasetPerturbationsMaximum RMSE difference from recalibrationFixed-weight gain over FM + projection
NS-Gauss102.07×10−62.07\times10^{-6}0–20.16%
FNS-KF506.09×10−56.09\times10^{-5}6.85–13.05%

Even without recalibrating for the perturbed operators, KMF maintains strong performance. On FNS-KF, it improves over FM + projection in all 50 tested settings, with gains of 6.85–13.05%. On NS-Gauss, it achieves improvements of up to 20.16% in five settings, while the remaining five use w∗=1w^*=1 and retain the FM prediction.

Furthermore, recalibrating the fusion weights for each perturbed operator changes RMSE by at most 2.07×10−62.07\times10^{-6} on NS-Gauss and 6.09×10−56.09\times10^{-5} on FNS-KF, demonstrating that the nominal weights remain effective across the evaluated perturbations.

Thus, KMF does not depend on recalibration for every operator variation to maintain its effectiveness in the tested settings. Calibration provides an additional opportunity to adapt the fusion weights, but the results show that repeatedly recalibrating them is not necessary to achieve meaningful improvements across these operator perturbations.

Thank you for your time and thoughtful consideration. We have provided detailed responses and additional experimental evidence to thoroughly address all the points raised in your review.

Please let us know if any substantive concerns remain, as we would be glad to offer further clarification.

If you feel these concerns have been adequately resolved, we kindly ask that you consider raising your overall rating.

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

This paper uses local physical derivatives to construct intermediate states and combines calibrated blending with constraint projection to refine the predictions and future rollout of a frozen neural model.

Critical strengths
  • The framework can be applied to existing predictors without updating their parameters. It clearly separates interpolation using neighboring predictions from causal rollout.

  • The experiments include controls with the same constraint projection and checks of calibration sensitivity and longer rollouts. The authors also report negative results and explain when the correction is not selected.

Critical weaknesses
  • My main concern is about how the blending weight is selected. As written, the weight is fitted to prediction error before applying the physical constraint. However, the final prediction is evaluated after that operation. Errors removed by the constraint can therefore affect the selected weight even though they do not affect the final output. The best weight before projection may not be the best weight afterward. It is not clear whether the calibration objective matches the prediction actually used during evaluation.

  • My other concern is about the evaluation of the continuous trajectory. The dense prediction comparison changes both the endpoint states and the interpolation rule. It compares the physical bridge between corrected endpoints with simple interpolation between uncorrected endpoints. This does not clearly show how much improvement comes from the bridge itself. The main fluid experiment with active correction also lacks intermediate reference states. A comparison using the same corrected endpoints would better support the accuracy claimed between time points.

Questions for authors
  • Are the candidate states projected before fitting the blending weight? If not, could the authors fit the weight using the final constrained prediction and compare the results under the same evaluation setting? This would help show whether calibration chooses the right balance for the output actually used.

  • Could the authors compare the physical bridge and simple interpolation using exactly the same corrected endpoints? Please use a setting where the correction is active and intermediate reference states are available. This would separate the benefit of improving the endpoints from the benefit of the interpolation rule.

6 comments
AuthorsOpenReview ↗

On calibration of the fusion weight and isolating the contribution of the Hermite bridge

Dear Reviewer hRCX,

Thank you for your careful reading and constructive suggestions. We address both concerns below.

1. Calibration of the fusion weight and physical projection

We thank the reviewer for identifying an ambiguity in the mathematical presentation. Our implementation already projects both candidate predictions before fitting the fusion weight, ensuring that calibration is performed on physically projected states.

The manuscript already specifies the projected final estimator in Equation (12):

u∗(t)=PM((1−w∗)uH+w∗uFM).u^*(t)=P_{\mathcal{M}}\left((1-w^*)u_H + w^*u_{\mathrm{FM}}\right).

However, Equations (13)–(14) and Appendix E.4 did not explicitly indicate that the candidates used for calibration were projected. To clarify, the implemented procedure defines

\bar u_{\mathrm{FM}}=P_{\mathcal M}(u_{\mathrm{FM}})$$ and calibrates the weight using $$w^*=\arg\min_{w\in[0,1]}\sum_m \left\|(1-w)\bar u_H^{(m)} +w\bar u_{\mathrm{FM}}^{(m)} -u_{\mathrm{true}}^{(m)}\right\|^2.$$ For the linear Helmholtz–Leray projection, projecting both candidates before fusion is mathematically equivalent to projecting their weighted combination afterward. Thus, **the calibration procedure is consistent with the final estimator already presented in Equation (12)**. **We have clarified this ambiguity in the revised manuscript**, updating Equations (13)–(14), Appendix E.4, and Algorithm 2 to explicitly describe projection-aware calibration. This is a clarification of the mathematical presentation, not a methodological change. **The implementation, calibration procedure, and reported experimental results remain unchanged.** #### 2. Isolating the contribution of the Hermite bridge We agree that comparing interpolation methods using different endpoints does not isolate the benefit of the interpolation rule. Following the reviewer's suggestion, we conducted an additional controlled experiment using **exactly the same KMF-corrected endpoints**, with active physical correction ($\alpha^*=0.20$) and available ground-truth midpoint states. The only difference between the two predictions is the interpolation method: **linear interpolation versus the physics-informed Hermite bridge**. | Dataset | Held-out trajectories | Linear RMSE | Hermite RMSE | Relative improvement (95% paired CI) | | -------- | --------------------: | ----------: | ----------------: | ------------------------------------ | | NS-Gauss | 50 | 0.07838 | **0.05963** | **23.92% [22.37%, 25.36%]** | | FNS-KF | 50 | 0.09317 | **0.09299** | **0.20% [0.16%, 0.24%]** | Both improvements are statistically significant (paired $p<2\times10^{-15}$). **Since both methods use identical corrected endpoints, these results isolate the contribution of the interpolation rule rather than endpoint correction.** On NS-Gauss, the Hermite bridge reduces midpoint RMSE by **23.92%**, demonstrating a substantial independent benefit from incorporating local PDE dynamics. On FNS-KF, the additional improvement is smaller but consistently positive. **This is precisely the motivation behind KMF's design: its components address different sources of prediction error, so a modest improvement from one component does not imply a modest benefit from the complete framework.** When temporal interpolation provides only a small additional gain, physical correction and calibrated fusion can still contribute substantially. For example, on FNS-KF, the complete Task 1 pipeline reduces CNO-FM RMSE from **0.09023 to 0.03059 (66.10%)** relative to the raw FM. In Task 2, KMF reduces projection-only RMSE from **0.02622 to 0.02472 (5.73%)**, with gains of **11.79% and 14.64%** over projection at rollout horizons $H=10$ and $H=20$, respectively. Thus, the controlled experiment establishes the independent contribution of the Hermite bridge, while the broader results support KMF's central design choice: combining temporal refinement, calibrated fusion, and physical correction into a complete pipeline rather than relying on any single component to deliver large improvements in every setting. The components provide complementary benefits, allowing KMF to achieve meaningful end-to-end improvements even when one component's isolated contribution is small. We thank the reviewer for suggesting this experiment, which allows us to distinguish the individual contribution of the interpolation rule from the overall effectiveness of KMF.

Thank you for the clarification. This addresses my first concern. However, the same-endpoint comparison is useful, but it covers CNO-FM on only two systems, with a substantial improvement on NS-Gauss and a much smaller 0.20% improvement on FNS-KF. I maintain my original rating due to limited experiment.

Thank you for the follow-up. We would like to clarify that Appendix Table 5 is a deliberately controlled component ablation, not the extent of the experimental evaluation.

CNO-FM is reported on NS-Gauss and FNS-KF because these are the only benchmark systems compatible with the released CNO checkpoint’s validated velocity-only state contract. ACE is scalar-valued and Wave-Gauss uses a different wave-state representation. Evaluating this checkpoint on those systems would violate its released input contract rather than provide a meaningful generalization test. This restriction is stated explicitly in the caption of Table 1.

The broader evaluation is substantially larger. Section 5.1 evaluates five frozen model families, Poseidon, CNO, Polymathic, DPOT, and MORPH, across four PDE systems and requested cadences from 0.05 to 0.40 seconds, subject to each released interface’s valid state contract. Table 1 reports Task 1 results across all four PDE systems, while Table 2 reports both native and off-native causal Task 2 regimes across all four systems. The complete 126-configuration model–PDE–cadence benchmark is provided in Appendix Tables 19 and 20, including unavailable interfaces and negative transfer cases rather than imputing unsupported evaluations.

The reported 0.20% on FNS-KF should also not be interpreted as the improvement of complete KMF. Table 5 intentionally fixes exactly the same KMF-corrected endpoints for both candidates and changes only the interpolation rule. Therefore, 0.20% measures the additional benefit of Hermite interpolation over linear interpolation after endpoint correction has already occurred. This narrowly controlled experiment was introduced specifically to isolate the bridge contribution requested in the original comment.

End-to-end KMF improvements on FNS-KF are substantially larger:

  • In the CNO-FM Task 1 result in Table 1, complete calibrated KMF reduces RMSE from 0.09023 to 0.03059, improving over raw FM by 66.1% and over the matched linear bridge by 1.7%.
  • In the full causal matrix in Appendix Table 20, CNO-FM KMF improves FNS-KF rollout by 11.96% at 0.05 seconds and 11.48% at 0.10 seconds.
  • In the independent Poseidon-B audit in Table 4, complete calibrated KMF reduces FNS-KF RMSE from 0.012536 to 0.008391, a 33.07% improvement over projected raw FM.

Thus, Table 5 answers the narrow component-isolation question, whereas Tables 1, 2, 4, 19, and 20 establish the breadth and end-to-end performance of the complete framework. The small 0.20% value is a statistically consistent incremental Hermite contribution with endpoint correction held fixed; it is not the overall KMF gain.

We hope this clarification resolves the concern regarding the scope of the evaluation and the interpretation of the controlled ablation. If it does, we would be very grateful if the reviewer would kindly reconsider their assessment in light of the extensive evidence reported in the paper.

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

This paper proposes KMF, an inference-time correction framework for frozen PDE prediction models. For temporal refinement between model-generated endpoints, the method constructs a cubic Hermite interpolant using endpoint states and PDE right-hand-side evaluations, then combines it with a direct model prediction through calibrated scalar fusion. For causal autoregressive prediction, it corrects the predicted endpoint using a Hermite–Simpson defect, applies an appropriate constraint projection, and feeds the corrected state into the next prediction step. The authors provide local interpolation and endpoint-perturbation analyses and evaluate accuracy, calibration stability, and computational cost across multiple PDE systems, model interfaces, and temporal cadences.

Critical strengths

1.The deployment assumptions and sources of improvement are carefully distinguished. The paper separates temporal refinement using future model-generated endpoints from causal prediction without future observations, and keeps calibration trajectories disjoint from test trajectories. It also explicitly states that many native-cadence results reduce to projection-only correction and that the fourth-order interpolation bound does not establish fourth-order accuracy for the deployed system. These qualifications improve the credibility of the conclusions.

2.Targeted experiments demonstrate that active correction can provide meaningful benefits. Table 3 compares KMF against projection, residual-gradient correction, live Gauss–Newton correction, and a cached-Jacobian control. The appendix also reports calibration resampling and longer rollouts. For example, at the off-native cadence, the 20-step gains over projection-only rollout are approximately 14.00% on NS-Gauss and 14.64% on FNS-KF. These results support sustained benefits from the defect correction in some settings.

Critical weaknesses

1.The methodological and theoretical novelty remains limited. Cubic Hermite interpolation, orthogonal projection onto linear constraints, scalar least-squares fusion, and Simpson-defect relaxation all have clear classical origins. The theoretical results primarily explain these established constructions. Relative to existing inference-time physical correction methods, the main contribution is organizing these ingredients into a unified interface for frozen predictors. This integration is useful, but the paper does not yet extract a sufficiently strong new mechanism or generalizable insight to support a high assessment of novelty.

3.The advantage over the closest alternatives is not sufficiently established. In Table 3, KMF achieves an RMSE of 0.053091 on NS-Gauss, compared with 0.053575 for the cached-Jacobian adapter, while their online correction times are 20.03 ms and 5.09 ms, respectively. Paired uncertainty analysis is needed to determine whether this small accuracy difference is consistent. Moreover, the adapter is not an official reproduction of PhysicsCorrect. Appendix Table 15 also reports an ACE error of 0.022755 for one-step RK2, compared with 0.098376 for calibrated KMF fusion, showing that local physical correction does not outperform direct numerical integration in every evaluated setting. These results call for a sharper characterization of where the practical advantage lies.

Questions for authors

1.Contribution clarification: Beyond combining classical interpolation, projection, and scalar calibration, what is the central new insight that transfers to other models or PDEs? Which experiment most directly supports that insight?

2.Comparative evaluation: Is KMF’s small accuracy advantage over the cached-Jacobian adapter in Table 3 consistent under paired statistical analysis on the same trajectories? After accounting for offline caching and online prediction, under what usage volumes or accuracy requirements does KMF provide an advantage?

3.Numerical-evaluation clarification: In Appendix Table 15, increasing RK4 from 1 to 16 substeps on Wave-Gauss increases the error from approximately (1.64\times10^6) to (7.83\times10^{123}). Do these rows use the same total integration interval, and how is the substep size chosen? What property of the discretized operator or its stability explains this behavior? This affects the interpretation of the solver comparison.

4 comments
AuthorsOpenReview ↗

Novelty of KMF

Dear Reviewer eGTp,

Thank you for your valuable insights on the paper.

Concern 1: limited novelty.

The novelty lies in the setting in which we are doing physics-based corrections. Inference-time debiasing methods already exist, but two important challenges remain insufficiently addressed.

  1. Generalizing beyond diffusion-based foundation models. Many existing inference-time correction approaches focus on diffusion models, where diffusion-specific information, such as score estimates, is available to guide physical corrections. However, scientific foundation models can also be based on architectures such as Fourier Neural Operators (FNOs) and Transformers, where such information is not necessarily accessible. Therefore, physics-guided correction methods should be designed for a more general setting, where the foundation model is treated as a black box and only its input-output inference interface and the governing physical operator are assumed to be available. Such methods should correct the predicted states without requiring access to the model's internal representations or modifying its parameters.
  2. Generalizing beyond native temporal cadences. Now, we recently had couple of methods like PhysicsCorrect that addresses 1st concern. But, even when black-box inference is possible, another important limitation remains: PDE foundation models are typically trained at specific temporal cadences, which can restrict their effectiveness at other time intervals. Although some models like Posseidon support user-specified cadences, their predictions may become substantially less accurate outside their native training cadence. This limits their applicability to scientific workflows that require flexible temporal resolution.

These challenges are particularly relevant to applications such as climate forecasting and fluid dynamics, where different physical processes, downstream analyses, and coupled simulations may require predictions at different temporal resolutions. Since foundation models are intended to generalize across scientific settings, their utility should not be restricted to a particular model architecture or training cadence. And we believe, to the best of our knowledge, inference-time debiasing is not done in this setting (considering both), while it remains important to address it.

KMF addresses these two challenges through a unified, inference-time physics-guided correction framework. It operates on frozen, black-box PDE foundation models without accessing their internal parameters or requiring retraining, and enables physically informed predictions at off-native temporal cadences. By improving temporal flexibility while preserving computational efficiency, KMF aims to make existing foundation models more broadly applicable to scientific workflows with varying temporal requirements.

Experimental evidence supporting the central insight. The transferability of KMF is demonstrated through evaluations across four PDE systems (NS-Gauss, FNS-KF, ACE, and Wave-Gauss) and multiple foundation-model architectures, including Poseidon, CNO, and Polymathic models, without modifying their parameters. Most directly, the off-native causal correction experiments (Table 3) show improvements of 16.15% on NS-Gauss and 5.73% on FNS-KF over projection alone, isolating the contribution of physical-defect correction. Furthermore, the complete Task 2 evaluation (Appendix Table 17) shows improvements over projection in 81 of 126 configurations, with 43 ties and only two minor degradations. These results support the central insight that local PDE information can be used to adapt frozen foundation-model predictions across architectures and physical systems, without retraining or accessing model internals.

AuthorsOpenReview ↗

On advantage over the closest alternatives & performance and stability of RK4

Concern 2: Advantage over the closest alternatives

We first address the concerns regarding PhysicsCorrect, and clarify our adaptation, empirical comparison, and KMF's distinct advantages.

1. Why a PC-style adapter instead of the original implementation? Our evaluation targets off-native cadences using Poseidon, which outputs a two-channel 128×128128\times128 velocity field. PC's original Navier–Stokes implementation corrects scalar vorticity. We therefore adapted it by converting velocity to vorticity, applying the correction, reconstructing velocity, and enforcing incompressibility. We explicitly label this a PC-style adapter, not an exact reproduction.

2. Accuracy and computational trade-offs. KMF consistently achieves modest but statistically significant improvements over the adapter:

DatasetKMF / PC RMSEGain (95% CI)Holm pp
NS-Gauss0.05309 / 0.053570.90% [0.22%, 1.72%]0.0215
FNS-KF0.02472 / 0.024940.85% [0.78%, 0.92%]<10⁻⁵

PC has lower online latency (5.09 vs. 20.03 ms on NS-Gauss; 3.39 vs. 4.46 ms on FNS-KF), but requires offline Jacobian-cache construction (488.64 and 91.87 ms). Consequently, KMF has lower cumulative correction cost before approximately 33 and 86 steps, respectively; PC becomes faster after cache amortization.

Moreover, a dense Jacobian for a full two-channel 128×128128\times128 field alone requires approximately 4.3 GB in float32, with additional storage for its pseudoinverse. The evaluated adapter uses a reduced-state cache; this estimate illustrates full-resolution scaling rather than its measured memory usage.

3. What does KMF enable beyond PC?

  • Arbitrary intermediate-time queries: KMF reconstructs states anywhere between available endpoints, even when the FM does not expose an off-native inference interface. PC corrects available predictions but does not itself generate missing intermediate-time states. Thus, PC alone cannot provide off-native predictions when the FM does not support those queries, whereas KMF can refine between available endpoints.
  • Cadence flexibility: KMF directly incorporates the requested time interval without rebuilding a cadence-dependent Jacobian cache.
  • Operator flexibility: Changing PDE parameters requires reevaluating the spatial operator rather than reconstructing a cached Jacobian, although calibration may still need adjustment.
  • Lightweight deployment: KMF requires neither dense Jacobian construction nor offline pseudoinverse caching.

Thus, KMF's contribution is not merely a small accuracy gain over PC, but physics-guided off-cadence temporal refinement and causal correction of frozen PDE foundation models without Jacobian preprocessing or retraining.

Concern 3: Performance and stability of RK4

We acknowledge that RK2 outperforms KMF on ACE. However, this comparison highlights two important advantages of KMF.

1. Numerical stability and computational cost. Appendix E.6 establishes that explicit RK4 is subject to resolution-dependent stability restrictions. Coarse time steps can violate its stability region, while recovering stability requires additional temporal substeps and PDE evaluations. To clarify the Wave-Gauss results, we conducted a state-complete audit using ut=v,  vt=c2Δhuu_t=v,\;v_t=c^2\Delta_hu, rather than the scalar closure used in the original control. The original experiment remains a diagnostic of that closure, but the revised solver comparison uses the complete wave system. All runs integrate over the same 0.05-second midpoint interval, with h=0.05/nh=0.05/n.

RK4 substepsωmax⁡h\omega_{\max}hInterior RMSESuccessive refinement difference
117.3000.02474—
28.6500.024190.01742
44.3250.024180.001598
82.1630.024187.53×10⁻⁵
161.0810.024185.41×10⁻⁶
1280.1350.024181.30×10⁻⁹

RK4 satisfies the predicted stability condition ωmax⁡h≤22\omega_{\max}h\leq2\sqrt2 for the first time in our tested substep sweep at 8 substeps, requiring 32 RHS evaluations, compared with only 2 endpoint RHS evaluations for KMF.

2. Preserving the advantage of neural surrogates. A central motivation for neural PDE surrogates is to avoid the computational expense of repeated numerical integration. Replacing FM inference with conventional time stepping can undermine this advantage, especially when stability requires many substeps. KMF instead exploits already-available FM endpoints, introducing physical information through inexpensive local evaluations without forward temporal integration.

Thus, KMF is not intended to universally outperform stable numerical solvers. Its advantage is enabling physics-guided refinement while preserving the computational efficiency and deployment flexibility of pretrained foundation models.

We thank the reviewer for their time and consideration. We have provided detailed responses and additional experimental evidence addressing all the concerns raised in the review.

We would appreciate knowing if any substantive concerns remain, and we would be happy to provide further clarification.

If the reviewer finds that the concerns have been adequately addressed, we kindly request reconsideration of the overall rating

Dear Reviewer eGTp,

Thank you again for your time. Regarding your concerns, we have some additional input that came up during experimentation in the discussion phase. As stated in our earlier response, the primary novelty lies in the setting KMF addresses, which is very important for practical considerations, and remains less addressed or unaddressed in existing methods. We would also like to clarify that the framework itself is not an arbitrary combination of classical components.

Deliberate, information-matched design

Over one FM interval, we naturally get four reliable quantities in the scenario that KMF is tackling:

F(\tilde u_k),\qquad F(\tilde u_{k+1}).$$ It is important for addressing why we chose Cubic Hermite. Cubic Hermite is the minimum-degree representation that uses all four exactly, serving as a perfect choice for this component. The Hermite midpoint supplies a dynamically informed interior state. Simpson quadrature then gives an endpoint candidate consistent with the integrated PDE dynamics: $$u_{k+1}^{\mathrm{Simpson}} = \hat{u}_k + \frac{\Delta t}{6} \left[ F(\hat{u}_k) + 4F(\tilde{u}_{k+1/2}) + F(\tilde{u}_{k+1}) \right].$$ The difference \$ d_k = u_{k+1}^{\mathrm{Simpson}} - \tilde{u}_{k+1} \$ is a mathematically motivated physical-defect direction. The non-obvious choice is not to solve the collocation equations or force this defect to zero. Doing so would replace the learned prediction with a numerical integration procedure and could move toward a physically consistent but less accurate state. Instead, KMF takes a bounded step from the learned endpoint in the defect direction: $$\tilde{u}_{k+1}^{\mathrm{corr}} = \tilde{u}_{k+1} + \alpha d_k = (1-\alpha)\tilde{u}_{k+1} + \alpha u_{k+1}^{\mathrm{Simpson}}, \qquad \alpha\in[0,1].$$ Calibration determines how much the physical direction should be trusted and may select $\alpha=0$ when the FM is already preferable. The corrected endpoint is projected and reinjected. Thus, the same endpoint-rate representation provides dense off-cadence states and constructs a cautious physical correction direction without turning KMF into an optimizer or PDE integrator. ### Direct ablation of the joint design We compared direct off-native autoregression using only Task 2 with the complete coupled Task 1 + Task 2 pipeline: | PDE | Task 2 only: direct off-native FM rollout | Task 1 + Task 2: full KMF | Gain | |---|---:|---:|---:| | NS-Gauss | 0.05919 | **0.04941** | **16.52%** | | FNS-KF | 0.02675 | **0.02487** | **7.02%** | | ACE | 0.28498 | **0.25460** | **10.66%** | | Wave-Gauss | 0.11131 | **0.10959** | **1.54%** | The coupled design reduces four-step window RMSE on all four systems. Active correction is selected on NS-Gauss, FNS-KF, and ACE, while Wave-Gauss uses the conservative $\alpha=0$ fallback. Thus, decomposing the problem into temporal refinement and causal correction, then coupling them through corrected-state reinjection, provides a measurable benefit beyond direct off-native rollout. ### Why this particular construction? This was not the first arbitrary combination we attempted. Less suitable alternatives revealed the requirements that led to KMF: - Linear interpolation was inexpensive but discarded the available physical rates. - Higher-order splines required additional states or derivatives whose errors would also need control. - Residual minimization required iterative optimization and could favor physically admissible but inaccurate states. - Solving Hermite-Simpson collocation would turn the method into a numerical integrator. - Post-hoc interpolation after rollout could not correct the states reinjected into the FM. These failures led to the present design: use the minimal representation supported by the available information, reuse it to construct a bounded correction direction, and reinject only the corrected state. The novelty is therefore both the black-box, off-native setting and the specific integrated mechanism developed for that setting. Reviewer zbnk also raised essentially similar concerns about novelty and whether the joint design contributes beyond its classical components. After this clarification and ablation, they indicated that the concern was resolved and raised their score. We hope this additional evidence also addresses your concern. If so, we would be very grateful if you would consider revisiting your assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.