acceptodds
Under review as a conference paper at ICLR 2027

Segmented Physics-Consistent Diffusion Anchors with Deterministic Rollouts for Allen–Cahn Dynamics

Abstract

We propose a segmented diffusion–Euler method for conditional coarse-to-fine reconstruction of Allen–Cahn dynamics. Residual-centered diffusion reconstructs sparse high-resolution anchors from coarse observations, and deterministic fine-grid Euler steps update each segment interior. An Allen–Cahn grid residual provides physical guidance, with the ReMD multigrid correction kept as an optional refinement. For any fixed finite spatial dimension, we derive a rollout bound that separates discretization bias, anchor risk, and segment-local error amplification. Controlled experiments show that periodic re-anchoring substantially improves rollout accuracy over unsegmented original ReMD. The Allen–Cahn-aware variant achieves comparable field accuracy with lower mean energy mismatch, and independent training runs show low variability. These results identify anchor reconstruction as the main accuracy bottleneck and quantify the tradeoffs between learned reconstruction and deterministic dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Conditional diffusion models, physics-informed machine learning, coarse-to-fine reconstruction, partial differential equations, error analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_s57qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses a conditional generative model with data and physics corrections to reconstruct high-resolution states from coarse observations and combines these states with numerical integration to recover phase-field dynamics.

Critical strengths
  • The framework separates learned reconstruction from numerical propagation. This makes the role of each part clear and helps explain how reconstruction errors affect the predicted trajectory.

  • The paper evaluates both field accuracy and energy behavior. It also reports mixed results from the component comparisons and discusses the limitations.

Critical weaknesses
  • My main concern is about how the learned correction is trained and applied. During training, it receives the reference state and a residual calculated from the coarse input. During sampling, it receives the predicted state and a residual calculated from that prediction. The training objective learns the correction through the perturbed input. It does not directly check whether the correction improves a generated state. The paper acknowledges these different inputs but does not clearly show why the correction should remain helpful when the prediction contains errors.

  • My other concern is about the experimental support for the full design. Some simpler configurations give comparable or better field accuracy. The existing comparisons are useful but do not directly show whether the learned correction improves the same predicted states before and after the update. This makes it difficult to separate the effect of the correction itself from differences caused by training.

Questions for authors
  • Why are different state inputs used to train and apply the correction? Please explain how the training objective prepares it to handle imperfect predictions during sampling. Could the authors test a version that uses more consistent inputs during training and sampling?

  • Could the authors apply the correction to the same predicted states and compare reconstruction error and physical consistency before and after the update? Please also separate the effect of the raw residual update from the learned correction. This would help explain when each component is useful and why the full design should be preferred.

If the authors answer these questions well and provide convincing evidence that the correction improves the actual sampling process, I am willing to change my recommendation to acceptance.

5 comments
AuthorsOpenReview ↗

Response to Reviewer s57q :Regarding whether consistent training and sampling inputs improve the correction

Dear Reviewer s57q,

Thank you for your constructive suggestions. We conducted both requested tests: a same-state, four-branch correction experiment and a controlled comparison using more consistent training and sampling inputs. These experiments identify the distinct contributions of the raw and learned corrections and provide evidence that input alignment improves their deployment behavior.

1. Regarding whether consistent training and sampling inputs improve the correction

The difference originates from the inherited training construction: the correction receives the reference HR state and a residual computed from the interpolated coarse input during training, but the predicted clean state and its corresponding residual during sampling.

Although the reconstruction objective uses perturbed inputs, this does not directly train the correction on deployment-time prediction errors. We agree that these two forms of robustness should be distinguished.

To test your suggestion, we fine-tuned the same checkpoint under two conditions with matched optimizer settings, training-sample and timestep sequences, and a budget of 1,000 updates. The original-input condition retains the original construction. The prediction-input condition uses frozen pilot predictions generated from training-set coarse observations as both the correction’s state input and the state used to compute its residual.

Prediction-input fine-tuning improves both rollout RMSE and energy agreement relative to original-input fine-tuning. RMSE decreases by approximately 0.050%, with a paired difference of −3.15 × 10⁻⁵ and a reported 95% interval of [−5.69 × 10⁻⁵, −7.12 × 10⁻⁶]. Mean energy error decreases by approximately 1.39%. The adverse RMSE increment introduced by M2NO decreases by approximately 57.8%.

Training conditionRollout RMSE ↓Mean energy error ↓M2NO-induced RMSE increment ↓
Unfine-tuned reference0.0626039940.001395805+0.0000136210
Original-input fine-tuning0.0629333560.001280002+0.0000139236
Prediction-input fine-tuning0.0629018880.001262153+0.0000058700

The last column reports FULL minus raw-only within each model; positive values indicate increased rollout RMSE. This indicates a practical benefit from input alignment, although the evidence comes from a single training seed, the frozen pilot inputs only approximate deployment inputs, and neither fine-tuned model exceeds the unfine-tuned checkpoint in overall field RMSE.

The training and inference inputs are clarified in Section 2.4 (Eqs. 21–22); the matched-budget fine-tuning protocol and results are reported in Appendix I.3 and Table 21.

For the remaining questions, please refer to the other responses. Thank you very much.

AuthorsOpenReview ↗

Response to Reviewer s57q:Regarding whether the correction improves the same predicted states

2. Regarding whether the correction improves the same predicted states

Following your suggestion, we tested no correction, raw residual only, learned M2NO correction only, and both corrections. For the local diagnostic, all four branches act on exactly the same pre-correction prediction from the FULL reference sampling chain. We also evaluated complete sampling chains starting from common noise, distinguishing local update effects from accumulated trajectory effects.

The same-state results reveal complementary but stage-dependent behavior:

  • The raw residual update reduces mean coarse-data inconsistency at all 15 sampling steps. It increases mean HR error at the first three steps but decreases it thereafter.
  • Adding M2NO improves local energy error during the later sampling steps, orders 6–14. It does not improve mean HR error, coarse-data inconsistency, or vector-field inconsistency relative to raw-only. Complete-chain results using a single fixed checkpoint | Correction branch | Anchor RMSE ↓ | Rollout RMSE ↓ | Mean energy error ↓ | |---|---:|---:|---:| | None | 0.065819701 | 0.062595221 | 0.001400185 | | Raw only | 0.065815809 | 0.062590373 | 0.001398266 | | Learned only | 0.065837212 | 0.062608999 | 0.001397717 | | Full | 0.065833173 | 0.062603994 | 0.001395805 |

The complete-chain evaluation is consistent with this distinction. Raw-only achieves the lowest rollout RMSE, approximately 0.0077% lower than without correction; 17/20 trajectories improve. FULL achieves the lowest mean energy error among the four branches, approximately 0.176% lower than raw-only; adding M2NO improves energy error on 17/20 trajectories. Its rollout RMSE is slightly higher, by approximately 0.0218% relative to raw-only.

An additional three-training-seed evaluation supports the same interpretation: adding M2NO reduces mean energy error by 2.35%, with a 0.186% increase in rollout RMSE. The energy benefit has the same direction across all three seeds.

Three-training-seed aggregate results

ConfigurationRollout RMSE ↓Mean energy error ↓
Raw only0.062945190.00144299
Full0.063062490.00140905

The same-state diagnostics and complete-chain intervention are reported in Appendix I.3 and Table 20. The three-training-seed aggregate and per-seed correction effects appear in Tables 12 and 13, respectively.

For the remaining questions, please refer to the other responses. Thank you very much.

AuthorsOpenReview ↗

Response to Reviewer s57q :Regarding when the full configuration should be preferred

The requested experiments clarify the components’ roles: the raw residual provides a small field-accuracy benefit, while M2NO provides an energy-matching refinement. Thus, FULL is useful when reference-energy agreement is an important objective; raw-only is preferable when field RMSE alone determines the choice. Energy agreement here does not imply monotone energy dissipation.

We consider M2NO when agreement with the reference energy is an important objective alongside field accuracy. In our three-seed evaluation, adding M2NO reduces mean energy error by 2.35% with a 0.186% increase in rollout RMSE. Therefore, M2NO is a useful refinement.

The resulting energy–accuracy tradeoff is discussed in Section 4.4 and Appendix H.3, with the corresponding aggregate results in Table 12.

Thank you for prompting these targeted tests. They provide direct evidence about actual correction behavior and a concrete improvement through prediction-input training. We hope this evidence supports a more favorable assessment.

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

This paper proposes a coarse-to-fine conditional reconstruction method for the two-dimensional Allen–Cahn equation. At sparse coarse time points, corresponding to segment heads, a residual-centered conditional diffusion model generates high-resolution anchors. Deterministic fine-grid explicit Euler evolution is then performed between adjacent anchors, and a new anchor is generated at the next observation time. The method inherits the generative and multiscale correction framework of the prior ReMD method and introduces a coarse–fine grid consistency residual based on the Allen–Cahn vector field. The paper distinguishes inherited components from newly introduced designs with reasonable clarity and explicitly limits its claims regarding acceleration, energy stability, and generalization.

Critical strengths
  1. The paper focuses on conditional reconstruction from coarse observations when complete high-resolution states are unavailable. It periodically uses a diffusion model to recover high-resolution states, performs explicit Euler evolution on the fine grid, and then reconstructs new anchors. This approach improves trajectory reconstruction within the existing ReMD framework and has practical feasibility and relevance.
  2. The authors clearly state the contributions and limitations of the proposed method. The experiments include GT-anchor+Euler, which replaces generated anchors with ground-truth high-resolution anchors. Its substantially lower rollout errors in both experimental suites support the conclusion that anchor reconstruction is the dominant source of error in the current pipeline. In addition, the segmented and unsegmented ReMD variants reuse the same model checkpoint, and the paper reports that multigrid correction provides only limited error reduction and that performance deteriorates at a lower absolute resolution. These results help clarify the limitations of the proposed method.
Critical weaknesses
  1. Table 2 reports a rollout RMSE of 0.062606 for FULL, which is slightly higher than the value of 0.061956 for segmented original ReMD, while FULL reduces the mean energy mismatch by 9.7%. Table 1 further shows that removing M2NO produces slightly lower anchor and rollout errors than FULL. These findings suggest that the performance gains of the complete model may be metric-dependent, and the current experiments do not establish the necessity of the newly introduced modules. I recommend conducting multi-seed experiments for FULL, segmented original ReMD, NO-AC, and NO-M2NO, and reporting between-method differences with confidence intervals. The authors should explain why the modules have opposing effects across different metrics, identify the conditions under which each module provides a significant benefit, and clarify why the multigrid module is retained despite its limited error reduction.
  2. Section 4.3 compares segmented and unsegmented ReMD using the same checkpoint. However, the unsegmented method generates a high-resolution anchor only once from the initial coarse-grid state and then performs continuous Euler evolution, whereas the segmented method regenerates anchors from the corresponding coarse-grid states at ten segment heads. Reusing the same checkpoint controls for differences in model training, but it does not remove the effects of unequal information and computational budgets during inference. Consequently, it is unclear whether the accuracy improvement arises primarily from segmentation itself or from the continued correction of the trajectory using additional observations. I recommend that Section 4.3 explain the rationale for the different inference procedures used by segmented and unsegmented ReMD. The paper should also include controlled experiments using the same coarse-grid snapshots, update frequency, and number of fine-grid integration steps, so that any significant accuracy gain can be attributed to the proposed method rather than to additional operations.
  3. The paper lacks some necessary explanatory figures, provides relatively few visualizations of the experimental results, and is not always written fluently. As a result, the central research problem and the core method are difficult to understand clearly and intuitively. I recommend adding an intuitive illustration of the research problem, a diagram demonstrating the proposed method, and other necessary visual explanations.
Questions for authors
  1. The paper shows that, with the existing ReMD reconstructor, periodically generating anchors from new coarse observations and then applying Euler evolution is more accurate than generating only one initial anchor and applying Euler evolution thereafter. However, the paper does not address whether a diffusion model is necessary to obtain this improvement or whether independently generating a new anchor is required. Could the authors explain these points in detail?
  2. In Section 2.4, the residual is computed from the interpolated coarse-grid state during training, and the ground-truth high-resolution target is used as the state input to M2NO. During inference, however, the residual is computed from the clean state predicted by the network, and this predicted state is also provided to M2NO. Why are different inputs used during training and inference? Have the authors evaluated the effect of this train-inference input mismatch, and if so, what is its impact?
  3. Equation (77) defines the base diffusion error as the root-mean-square deviation of the generated output from the Bayes conditional mean and states that this quantity can be recorded directly on validation anchors. However, Appendix E.3 explicitly notes that the Bayes conditional mean cannot be directly obtained from a finite validation set and that the corresponding theoretical quantity therefore cannot be measured directly. Please clarify whether the experiment records reconstruction error relative to the ground-truth high-resolution sample or error relative to an estimate of the conditional mean. If the former is used as a proxy, how are the irreducible conditional variance and coarse-grid conditioning bias included in that error handled?
  4. The paper builds on ReMD by introducing segmented anchor reconstruction and an Allen-Cahn residual. Could the authors further explain why ReMD was selected as the foundation? Given the paper's observation that the additional benefit of multigrid correction is limited, how can the generality of the proposed framework be established? To what extent does the method's effectiveness depend on design choices specific to ReMD, and how much of the benefit comes from a segmentation strategy that could be applied to other reconstruction models? Please provide further clarification.
5 comments
AuthorsOpenReview ↗

Response to Reviewer FDkz: Regarding whether diffusion and independent re-anchoring are necessary and the impact of training–inference input mismatch

Dear Reviewer FDkz,

Thank you for your constructive assessment. Your comments motivated matched-observation comparisons, targeted correction diagnostics, and clearer theoretical definitions. We first address your four questions, then summarize the revisions corresponding to the weaknesses.

1. Regarding whether diffusion and independent re-anchoring are necessary

The framework separates repeated observation-conditioned reconstruction from the choice of reconstruction model. To evaluate this distinction directly, we added a deterministic-SR control using the same UNetModelSwin backbone. FULL and SR each use three training seeds and receive the same ten coarse observations, ten anchor reconstructions, and 200 fine-grid Euler updates. SR predicts HR anchors directly, without diffusion sampling or residual corrections.

Under the same ten-observation and 200-step Euler protocol, FULL reduces rollout RMSE by approximately 29.6% relative to repeated bicubic reconstruction, demonstrating the added value of learned reconstruction beyond access to new observations alone.

This matched-information comparison identifies a clear field–energy tradeoff. FULL achieves approximately 14.0% lower mean energy error than SR. The paired FULL-minus-SR difference has a 95% interval of [−4.20 × 10⁻⁴, −5.08 × 10⁻⁵]. SR has lower rollout RMSE, approximately 7.61% lower than FULL. Thus, diffusion is not required for the framework’s field-accuracy benefit, while the reconstructor determines the balance between field accuracy and energy agreement.

Matched-observation comparison: three training seeds per learned method

ReconstructorCoarse observationsEuler updatesRollout RMSE ↓Mean energy error ↓
Deterministic SR102000.0582647110.001637783
FULL102000.0630615520.001408908

Independent anchors offer a clean interface between current observations and segment-local propagation, and permit batched execution. We also tested a simple coarse-innovation update of the previous terminal state; it did not outperform repeated bicubic reconstruction under the tested protocol. Interface diagnostics show that FULL reduces field error at approximately 94.5% of evaluated resets, with mean benefits persisting after one and five Euler steps. These results support independent re-anchoring as a practical design choice: it incorporates fresh observations to correct accumulated trajectory error, enables segment-local error analysis.

The SR and simple-update controls are reported in Appendix I.2, Tables 18–19; the same appendix provides the reset and short-term continuation diagnostics. Section 4.4 summarizes the findings.

2. Regarding the impact of training–inference input mismatch

The difference follows the inherited training construction: training uses the HR reference as the correction’s state input and evaluates the residual from the interpolated coarse state; inference uses the predicted clean state for both. We clarify that reconstruction training with perturbed inputs does not directly guarantee robustness of the correction to generated-state errors.

We evaluated input alignment through matched-budget fine-tuning from the same checkpoint, using 1,000 updates with matched optimizer settings, sample order, and timesteps. The prediction-input condition uses frozen pilot predictions generated exclusively from training-set coarse observations.

Compared with original-input fine-tuning, prediction-input fine-tuning improves rollout RMSE by approximately 0.050% and mean energy error by approximately 1.39%. The paired RMSE difference has a 95% interval of [−5.69 × 10⁻⁵, −7.12 × 10⁻⁶]. The correction-induced RMSE increment also decreases by approximately 57.8%. Controlled input-alignment experiment: one training seed

Training conditionRollout RMSE ↓Mean energy error ↓M2NO-induced RMSE increment ↓
Unfine-tuned reference0.0626039940.001395805+0.0000136210
Original-input fine-tuning0.0629333560.001280002+0.0000139236
Prediction-input fine-tuning0.0629018880.001262153+0.0000058700

The last column reports FULL minus raw-only within each model; positive values indicate increased rollout RMSE.

These results provide direct evidence that input alignment partially mitigates the mismatch. We retain the unfine-tuned reference to distinguish this improvement from an overall improvement over the original checkpoint. The experiment and its scope are documented in “Same-state correction diagnostics and controlled input alignment.”

Section 2.4 specifies the training and inference inputs, while Appendix I.3 and Table 21 report the controlled prediction-input fine-tuning experiment.

AuthorsOpenReview ↗

Response to Reviewer FDkz:Regarding how reconstruction error is measured and why ReMD was chosen and how the framework gene

3. Regarding how reconstruction error is measured

The experiments measure error against paired HR reference samples, not an estimated Bayes conditional mean. Thank you for identifying the inconsistent wording; we have corrected the direct-measurability statement.We replaced the statement that conditional-mean error can be measured directly on validation anchors with an explicit distinction between empirical HR-reference error and theoretical conditional-mean error (Appendix E.2, Eqs. 99–101).

Let Y denote the HR target, C the conditioning information, m(C) = E[Y | C], and G(C,Z) the generated anchor. When sampling randomness Z is independent of Y conditional on C, E[||G − Y||²] = E[||G − m(C)||²] + E[||Y − m(C)||²].

Accordingly, sample-reference risk includes irreducible conditional variance. We retain this contribution rather than subtracting an unobservable quantity. The identity concerns expected squared norms, not a sum of averaged RMSE values.

Discrepancies between ideal and supplied coarse conditions are treated separately through conditioning-error terms under the stated stability assumptions. Numerical-reference discrepancy is handled through the reference-risk comparison. The revised analysis establishes a clear link between measurable reconstruction risk and end-to-end rollout error, while explicitly distinguishing conditional uncertainty, coarse-conditioning discrepancy, and numerical-reference error.

The distinction between observable HR-reference errors and conditional-mean errors is explicit in Appendix E.2, Eqs. (99)–(101). Proposition D.10 in Appendix D.7 relates numerical-reference risk to the rollout analysis.

4. Regarding why ReMD was chosen and how the framework gene

ReMD provides an established residual-centered conditional reconstructor with a mechanism for incorporating physical residuals. It is a natural foundation for studying reconstruction coupled with numerical propagation, and its inherited architecture and sampling machinery are explicitly attributed.

The analytical interface is anchor risk, not a particular denoiser or multigrid architecture. The framework therefore accommodates reconstruction operators satisfying the stated assumptions, including configurations without M2NO. The same-backbone SR comparison demonstrates this modularity empirically: another reconstructor fits the same propagation framework and produces a different accuracy–energy balance.

We consequently distinguish the core contribution—the analyzed reconstruction–propagation coupling—from optional reconstruction components. The analysis provides a modular foundation for extending the framework to other reconstructors satisfying its assumptions, complemented by controlled validation on Allen–Cahn dynamics.

Sections 2.3–2.5 describe the reconstruction–propagation interface; Theorem 3.6 in Section 3.2 provides the modular rollout bound. Appendix I.2 and Table 18 present the deterministic-SR instantiation.

AuthorsOpenReview ↗

Response to Reviewer FDkz: Revisions addressing the weaknesses

W1. Regarding whether the added modules provide consistent benefits

We added same-state four-branch interventions and complete-chain comparisons, separating correction effects from training differences. The three-training-seed FULL/raw-only evaluation shows 2.35% lower mean energy error with a 0.186% increase in rollout RMSE, identifying M2NO as an optional energy-oriented refinement. Energy mismatch also reflects spatial gradients and nonlinear potential terms, so improved energy agreement need not coincide with lower field RMSE.Paired confidence intervals quantify the evaluated differences, while the three-seed FULL/raw-only comparison characterizes the correction’s energy–accuracy tradeoff across training runs.

Tables 12–13 report aggregate and per-seed correction effects; Appendix I.3 and Table 20 provide the complementary fixed-checkpoint interventions.

W2. Regarding whether the comparisons use matched information and computational budgets

ates single-observation framework comparisons from controls with identical coarse snapshots, reconstruction frequency, and Euler-step counts. We explicitly report both matched information budgets and execution costs, enabling transparent comparisons of reconstruction quality and computational efficiency.

Section 4.4 and Appendix I.2 distinguish observation budgets and matched-update controls; Tables 18–19 give the corresponding equal-observation comparisons.

W3. Regarding how the method and results are presented

We added a compact method schematic showing coarse observation → HR reconstruction → Euler propagation → re-anchoring, with residual guidance and optional M2NO located within reconstruction. We also streamlined the contribution statement, experimental interpretation, and conclusion, and clarified the distinction between inherited components and the proposed coupling.

Figure 1 illustrates the workflow alongside Section 2.5; the contribution statement in Section 1 and the synthesis in Section 5 have also been revised.

We appreciate that your questions helped strengthen the paper’s experimental fairness, analytical precision, and accessibility. We hope these additions support a more favorable assessment, and would welcome clarification of any remaining issue decisive for your recommendation.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Reviewer zZTe: Contributions, Theory, and Experimental Evidence

Dear Reviewer zZTe,

Thank you for your detailed review. Some passages are affected by mathematical rendering and text overflow. We respond below to the clearly readable comments and would appreciate a plain-text or corrected version of the remaining passages.

1. Contribution: connecting learned reconstruction with numerical propagation

Our contribution is a modular framework that couples observation-conditioned HR reconstruction with deterministic segment-wise propagation and connects anchor quality to end-to-end trajectory accuracy.

Building on ReMD's residual-centered reconstruction machinery, we incorporate an Allen–Cahn grid residual and formulate an explicit reconstruction–propagation interface. This interface makes the roles of anchor estimation, numerical discretization, and segment length individually analyzable and experimentally testable. The revised manuscript clearly attributes the inherited components and identifies the contribution of their analyzed coupling.

New controlled experiments demonstrate the value of this design. With the same ten observations and 200 Euler updates, FULL reduces rollout RMSE by approximately 29.6% relative to repeated bicubic reconstruction. A same-backbone deterministic-SR control further identifies the reconstruction tradeoff: SR achieves lower field RMSE, while FULL provides approximately 14.0% lower mean energy mismatch. These results establish the value of learned reconstruction over simple interpolation and clarify how different reconstructors serve different accuracy objectives.

2. Theoretical value: translating anchor risk into rollout accuracy

The analysis provides a common interface between learned reconstruction risk and numerical error propagation. Its modular bound separates discretization bias, anchor error, and segment-local amplification.

This separation makes the result actionable: improvements in anchor reconstruction enter through the anchor-risk term, while segment length and numerical stability determine subsequent amplification. Re-anchoring replaces the propagation history with a new observation-conditioned estimate, allowing error to be analyzed over individual segments. The same analytical structure accommodates different reconstructors satisfying the stated assumptions.

We have strengthened the connection to empirical evaluation by distinguishing paired-HR-reference error from Bayes conditional-mean error and explicitly accounting for conditional uncertainty, coarse-conditioning discrepancy, and numerical-reference error. The resulting analysis provides a precise foundation for interpreting measured reconstruction quality in terms of trajectory accuracy.

3. Efficiency enabled by independent segments

The independently reconstructable segments offer an execution advantage: their anchor generation and within-segment propagation can be batched.

Under the same sampling formulas and precision, batched execution reduces mean time from 17.163 to 6.016 seconds per trajectory—a 2.853-fold speedup, or 64.95% lower wall-clock time. The paired-bootstrap 95% interval for the speedup is [2.789, 2.912].

This is a same-method execution improvement with all fine-grid Euler updates retained. We report the timing protocol and baseline costs separately, allowing readers to assess reconstruction quality, energy agreement, and computational cost together. Simpler reconstructors remain faster, while the new measurements demonstrate an effective way to reduce the cost of the proposed framework.

4. Ablations identify the roles of the residual and learned correction

The controlled comparisons identify a consistent energy-oriented contribution. Relative to segmented original ReMD, FULL reduces mean energy mismatch by approximately 9.7%, with slightly higher field RMSE.

The same-state four-branch experiment further separates the explicit residual update from the learned M2NO correction. The raw update provides a small complete-chain field-accuracy improvement over no correction, while the full configuration achieves the lowest mean energy error among the four tested branches.

Across three training seeds, adding M2NO reduces mean energy error by 2.35%, with a 0.186% increase in rollout RMSE. This gives a concrete criterion for including M2NO: it is useful when closer agreement with reference energy is important alongside field accuracy and the measured tradeoff is acceptable. The framework accommodates either configuration, enabling component selection according to the reconstruction objective.

We have incorporated these findings into the paper. Together, the revisions make the paper's contribution more explicit: an analyzed reconstruction–propagation framework with demonstrated accuracy benefits over simple reconstruction, quantified energy–accuracy tradeoffs, and improved execution efficiency.

We hope these clarifications and targeted results support a more favorable assessment.

AuthorsOpenReview ↗

Title: Summary of Revisions and Additional Evidence

Dear Reviewers and Area Chair,

Thank you for your thoughtful feedback. We have revised the manuscript and conducted targeted experiments addressing correction effectiveness, training–inference consistency, and the contribution of repeated reconstruction.

  1. Repeated reconstruction and controlled comparisons. Matched-observation experiments compare reconstruction operators using the same ten coarse observations, ten reconstruction opportunities, and 200 Euler updates. Interface diagnostics show that re-anchoring reduces field error at most evaluated interfaces, with mean improvements persisting through subsequent numerical propagation. These results quantify the value of repeated observation-conditioned reconstruction.

  2. Correction diagnostics and input alignment. Four-branch interventions on identical predicted states separate the effects of raw residual updates and learned M2NO corrections. Across three training seeds, M2NO reduces mean energy mismatch by 2.35%, with a 0.186% increase in rollout RMSE, supporting its use when energy agreement is an important objective. Prediction-input fine-tuning also improves field and energy accuracy relative to original-input fine-tuning.

  3. Execution efficiency. Batched execution achieves a 2.853-fold wall-clock speedup over serial execution of the same method, with a paired-bootstrap 95% interval of [2.789, 2.912], demonstrating the computational benefit of independently reconstructable segments.

  4. Theory and presentation. We clarified the distinction between measurable HR-reference reconstruction risk and Bayes conditional-mean error, and revised the supporting bounds and proofs. The analysis connects reconstruction quality to end-to-end rollout accuracy. We also added a method-flow schematic and clarified experimental protocols.

The revised manuscript incorporates these results, and our reviewer-specific replies provide detailed evidence and manuscript locations. Thank you for considering these revisions.

15:51

AuthorsOpenReview ↗

Response to Reviewer zZTe: A new experiment highlighting the advantages of the method

Dear Reviewer zzte

Thank you very much for your questions about our paper and for your valuable comments. After conducting numerous experiments without obvious results, we finally calmed down and reflected on why the experimental results fell short of expectations. We examined our theoretical assumptions and requirements, and found that our method should have a significant advantage under stable initial value conditions. Therefore, we conducted experiments with changed initial value conditions. We found that the results of this experiment can well address your questions.

We considered the smooth initial fields u0(x,y)=Acos⁡(2πx+ϕx)cos⁡(2πy+ϕy)u_0(x,y)=A\cos(2\pi x+\phi_x)\cos(2\pi y+\phi_y), with A∈[0.7,1.3]A\in[0.7,1.3] and independently sampled phases. We trained all three learned reconstructors from scratch using the same 160 training fields, 20 validation fields, and 2,000 updates per model. Checkpoints were selected by validation RMSE. The table reports mean first-frame errors for 64×64→256×25664\times64\to256\times256 reconstruction on 20 held-out initial fields. Lower is better in every column; relative L2L^2 is reported as a ratio.

MethodField RMSERelative L2L^2Coarse-node RMSEAbs. energy errorLow-freq. RMSEHigh-freq. RMSE
Bicubic2.303×10−32.303\times10^{-3}4.639×10−34.639\times10^{-3}004.461×10−54.461\times10^{-5}3.159×10−43.159\times10^{-4}2.281×10−32.281\times10^{-3}
Direct SR5.185×10−35.185\times10^{-3}1.034×10−21.034\times10^{-2}5.324×10−35.324\times10^{-3}7.838×10−47.838\times10^{-4}5.093×10−35.093\times10^{-3}8.916×10−48.916\times10^{-4}
ReMD (local legacy implementation)2.214×10−32.214\times10^{-3}4.482×10−34.482\times10^{-3}2.466×10−32.466\times10^{-3}1.874×10−41.874\times10^{-4}1.967×10−31.967\times10^{-3}9.714×10−49.714\times10^{-4}
FULL (legacy protocol)2.078×10−32.078\times10^{-3}4.144×10−34.144\times10^{-3}2.176×10−32.176\times10^{-3}3.984×10−43.984\times10^{-4}1.953×10−31.953\times10^{-3}6.476×10−46.476\times10^{-4}

Relative to direct SR, FULL reduced mean field RMSE by 59.9% and mean absolute energy error by 49.2%. Thus, on this family, its energy benefit over SR did not come at the expense of field accuracy. FULL also had the lowest mean high-frequency error of the four reported methods. Bicubic remained strongest in low-frequency error and energy agreement and reproduced the coarse observations exactly. In addition, our method also has a clear advantage in computational efficiency.

This result identifies a concrete setting in which the full reconstructor improves on direct SR at the anchor.

We appreciate your suggestion to clarify when the added reconstruction components are useful. We hope this comparison helps inform your final assessment.

AuthorsOpenReview ↗

A new experiment highlighting the advantages of the method

Dear Area Chair,

We have submitted a very important new experiment that demonstrates the advantages of our method. We kindly ask you to give special consideration to this latest experiment. Thank you very much. We considered the smooth initial fields u0(x,y)=Acos⁡(2πx+ϕx)cos⁡(2πy+ϕy)u_0(x,y)=A\cos(2\pi x+\phi_x)\cos(2\pi y+\phi_y), with A∈[0.7,1.3]A\in[0.7,1.3] and independently sampled phases. We trained all three learned reconstructors from scratch using the same 160 training fields, 20 validation fields, and 2,000 updates per model. Checkpoints were selected by validation RMSE. The table reports mean first-frame errors for 64×64→256×25664\times64\to256\times256 reconstruction on 20 held-out initial fields. Lower is better in every column; relative L2L^2 is reported as a ratio.

MethodField RMSERelative L2L^2Coarse-node RMSEAbs. energy errorLow-freq. RMSEHigh-freq. RMSE
Bicubic2.303×10−32.303\times10^{-3}4.639×10−34.639\times10^{-3}004.461×10−54.461\times10^{-5}3.159×10−43.159\times10^{-4}2.281×10−32.281\times10^{-3}
Direct SR5.185×10−35.185\times10^{-3}1.034×10−21.034\times10^{-2}5.324×10−35.324\times10^{-3}7.838×10−47.838\times10^{-4}5.093×10−35.093\times10^{-3}8.916×10−48.916\times10^{-4}
ReMD (local legacy implementation)2.214×10−32.214\times10^{-3}4.482×10−34.482\times10^{-3}2.466×10−32.466\times10^{-3}1.874×10−41.874\times10^{-4}1.967×10−31.967\times10^{-3}9.714×10−49.714\times10^{-4}
FULL (legacy protocol)2.078×10−32.078\times10^{-3}4.144×10−34.144\times10^{-3}2.176×10−32.176\times10^{-3}3.984×10−43.984\times10^{-4}1.953×10−31.953\times10^{-3}6.476×10−46.476\times10^{-4}

Relative to direct SR, FULL reduced mean field RMSE by 59.9% and mean absolute energy error by 49.2%. Thus, on this family, its energy benefit over SR did not come at the expense of field accuracy. FULL also had the lowest mean high-frequency error of the four reported methods. Bicubic remained strongest in low-frequency error and energy agreement and reproduced the coarse observations exactly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.