acceptodds
Under review as a conference paper at ICLR 2027

Projection, Convection, and Response Correction for Long-Horizon Neural Fluid Simulation

Abstract

Neural network models accelerate fluid simulation through autoregressive rollouts that feed each forecast into the next prediction, but accumulated errors limit long-horizon accuracy. Complementing unrolling and pushforward training, we introduce Mass Projection (MP) and Scalar Consistency (SC) to couple incompressible velocity prediction with scalar transport. MP enforces exact discrete incompressibility, including on walled domains. SC evolves persistent subcell states with the current and projected velocities, retaining scalar structure within grid cells. Its closed-domain transport conserves scalar amount and preserves bounds independently of learned weights, while diffusion, sources and boundary exchange are included explicitly. Both components are differentiable, allowing scalar errors to train the velocity model. We further propose Response Correction (RC), a post-training method that learns how the trained model would continue from the ground-truth next state and uses the predicted response to correct its forecasts. Across three flows from The Well database, MP+RC lowers full-rollout error in 15 of 18 architecture–training settings. In matched one-step comparisons, MP+RC reduces full-rollout error by 89% on Rayleigh–Bénard convection, and MP+SC by 82% on shear flow.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Differentiable Projection, Conservative Transport, Physics-Constrained Learning, Long-Horizon Prediction

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

To reduce accumulated errors and improve long-horizon accuracy in fluid simulation, the paper introduces MP and SC to jointly improve incompressible velocity prediction. The two components are integrated into an end-to-end training framework. Across 18 training settings, the proposed method achieves lower full-rollout errors.

Critical strengths

The theoretical analysis is sufficient and provides clear support for the proposed method.

Experiments across diverse settings validate the effectiveness of the method, with additional analyses providing further insights.

Critical weaknesses

The number of evaluated rollouts is limited, raising concerns about the statistical significance of the reported improvements. Moreover, the gains on some metrics are relatively modest, particularly compared with strong baselines. More extensive evaluations would help establish the consistency and robustness of these improvements and validate the effectiveness of the method.

The paper lacks sufficient efficiency analysis, including computational complexity and actual running time. These comparisons are important for assessing the practical applicability of the proposed method.

Questions for authors

Are the reported improvements statistically significant across different rollout settings, particularly when compared with strong baselines?

How does the proposed method compare with existing approaches in terms of computational complexity, training efficiency, and inference time?

3 comments
AuthorsOpenReview ↗

Response to reviewer Mamm: Part 1

W1 & Q1: Thank you for the careful review. Each of our test trajectories is rolled out over its full length (194 steps for Rayleigh–Bénard and shear flow and 76 steps for active matter). The full-rollout error averages over all predicted steps, and the windows up to step 30 are meant to show how the error grows early in the rollout. A VRMSE above 1 means the prediction is worse than the spatial mean of the true field (Section 4). The mean error of every Baseline passes this level before the end of its rollout, after 0–17 steps on Rayleigh–Bénard, 14–151 steps on shear flow and 8–16 steps on active matter (Table 3(a)). On shear flow, all three FNO Baselines exceed 10 over the full rollout (Table 1).

Across the 18 settings, which combine three datasets, two models and three training strategies, MP+RC lowers the steps 13–30 error in 17 and the full-rollout error in 15. In the remaining three, every variant exceeds 10 over the full rollout (Section 4.1). RC lowers the steps 13–30 error in 36 of its 43 deployments in the main configurations and the full-rollout error in 32 (Section 4.2). The components also carry over to Transolver++, where +MP+SC improves on +MP with every rollout method (Table 3(b)).

+MP+RC has significantly lower error than the Baseline in all 12 Rayleigh–Bénard comparisons, including the unrolling and pushforward Baselines designed for long rollouts, and in all six active-matter full-rollout comparisons (our reply to Reviewer kMiu, W4). RC also makes better use of its extra compute than further training does. On average over bounded pairs, one more hour of ordinary training lowers the steps 13–30 error by 2 %, whereas RC lowers it by 20 % (Section 4.2, Appendix E.4). Spending that hour on a two-step loss instead leaves a higher full-rollout error than RC in every setting compared (our reply to Reviewer YHSM). Our additional experiments show that PDE-Refiner, a method designed for stable long rollouts, is also less accurate than +MP+RC over long horizons (our reply to Reviewer kMiu).

The gains of our framework are largest over long horizons. RC leaves the first prediction unchanged by design, while MP lowers the step-1 error in 23 of 24 models, by 21 % on average (Section 4.2). Over the full rollout, MP+RC reduces the error by 89 % for one-step FNO on Rayleigh–Bénard (8.620 to 0.977), and MP+SC by 82 % for one-step CNextU-Net on shear flow (5.047 to 0.933). MP+RC also extends the forecast horizon in 14 of 18 settings (Section 4.1).

AuthorsOpenReview ↗

Response to reviewer Mamm: Part 2

W2 & Q2: We sincerely thank the reviewer for suggesting a more detailed efficiency analysis, which matters for practical use. MP applies one exact projection to each prediction and adds no network evaluation. The projection uses Fourier transforms on periodic domains and a Fourier–Chebyshev solve on walled domains. SC advances the transported fields with a conservative update on the same grid. Its circulation head has only 257 parameters. RC acts only at inference and only from its starting step. Each corrected step evaluates the companion once and the frozen model once with one backward pass. The model is then evaluated on the candidate frame, and this evaluation is reused as the prediction of the next step. For comparison, PDE-Refiner with K = 3 evaluates its network four times for every prediction. Unrolling and pushforward add cost only during training.

On training, MP models use the same 12 GPU-hour budget as the Baselines on an RTX 5090 and complete 89–95 % of the Baseline's one-step epochs (Table 4). SC adaptation and RC companion training each take one additional GPU-hour on the same GPU, against 12 GPU-hours for the model (Appendix A).

We measured the training times of PDE-Refiner, the Baseline, and the running times in Tables R1 and R2 on one RTX PRO 6000. PDE-Refiner is trained on the same CNextU-Net model and data, and for the same number of updates its training takes 1.04×–1.38× the Baseline's time.

Table R1. CNextU-Net, running time of a full test run relative to the Baseline on Rayleigh–Bénard (RB), shear flow (SF), and active matter (AM). Parentheses give the starting steps of RC.

DatasetBaseline+MP+RC+MP+RCPDE-Refiner
RB1.00×1.01×1.70× (2)1.83× (6)1.20×
SF1.00×1.01×1.88× (20)1.91× (26)1.64×
AM1.00×1.01×1.67× (15)1.71× (15)1.54×

Table R2. FNO, same as Table R1. N/A: no starting step qualifies, so RC stays off and the time equals the Baseline's (Appendix C.3).

DatasetBaseline+MP+RC+MP+RC
RB1.00×1.01×1.79× (13)1.66× (7)
SF1.00×0.99×N/A1.92× (6)
AM1.00×1.00×1.42× (16)1.45× (14)

In every setting, a full test run with RC takes less than twice the Baseline's time. +MP changes it by at most 1 %, which is within run-to-run variation. MP+SC takes 1.31× the Baseline's time for one-step CNextU-Net on shear flow, the main setting for SC. RC applies to an already trained model, including models trained with unrolling and pushforward. PDE-Refiner instead requires training the model from scratch with its own objective.

Thanks for the additional experiments. I prefer to maintain my score.

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

This paper proposes three separate mechanisms for improving ML emulators of fluid solvers: a mass projection that enforces incompressibility and boundary conditions, a scalar consistency formulation that enforces scalar mass conservation, and a response correction that learns a post-processing step involving a two-step predictor to correct the one-step prediction. These improve forecasts on several standard fluid prediction datasets and architectures.

Critical strengths

The paper successfully improves upon architectures trained on standard fluid datasets.

Critical weaknesses
  1. The presentation is very poor. Based on the main text, it is not clear at all what each proposed method is actually doing. Notation is not defined, and the vague descriptions don't provide enough information to judge the novelty of the methods.
  2. This paper, in general, seems very problem-specific and may be better suited for a more field-specific venue rather than a general ML audience.
Questions for authors
  1. Mass projection seems like a very common and easily performed correction that exists broadly across the fluid literature. What is the novelty here?
  2. Is scalar consistency just using the current velocity field to solve a scalar advection problem? Doesn't this just mean running a standard PDE solver for the scalar? It seems obvious that running a PDE solver will be more accurate but slower than an ML surrogate model.
  3. I don't understand why response correction should be helpful at all. If the goal is to predict two steps ahead more accurately, why not just do that during training rather than apply a correction term afterward?
5 comments
AuthorsOpenReview ↗

Response to reviewer YHSM: Part 1

W1 & Q1: We have revised Sections 1–3 to make these distinctions explicit, define the notation, and explain the mechanisms alongside their equations.

MP: We agree that projection is established in fluid simulation. Incompressibility is a physical constraint within the incompressible Navier–Stokes system, and projection is one numerical means of enforcing it. Our neural network predicts the next velocity directly, without solving the coupled momentum and pressure equations, so its predictions do not automatically satisfy this constraint. Our specific contribution is a differentiable MP implementation for autoregressive prediction on both periodic and walled grids. In particular, the Fourier–Chebyshev formulation jointly enforces discrete incompressibility and zero normal velocity at the physical walls, including when stored samples exclude the wall endpoints. The constraint is incorporated during training and every rollout step. We position this as a specific extension and evaluation of established projection principles, rather than claiming that projection or its use in ML is new.

Appendix B.1 establishes that MP satisfies the discrete constraints and cannot increase the distance to a target satisfying those constraints in the specified metric. Empirically, on Rayleigh–Bénard with FNO trained for one step, MP reduces full rollout VRMSE over 194 steps from 8.620 to 1.535, an 82% reduction. A separate CNextU-Net control shows that training through MP reduces steps 13–30 error from 7.279 with projection applied only during rollout to 4.731, improving 28 of 35 trajectories. These MP results support enforcing the physical constraint within the learned transition, while we distinguish these measured benefits from a general guarantee of rollout stability.

W2: We appreciate the concern about relevance to a general ML audience. Fluid dynamics is an established setting for methodological research at major ML conferences. Recent examples include neural surface fluid simulation by Wang et al. at ICLR 2025 [1], physically constrained neural surrogates by Huang and Greenberg at ICML 2025 [2], and turbulence simulation with EddyFormer at NeurIPS 2025 [3]. These papers illustrate how a focused physical domain can support contributions relevant to the broader ML community.

Our work similarly addresses methodological questions in learned simulation, including enforcing constraints during recurrent prediction, retaining spatial information between predictions, and correcting forecasts through a frozen predictor’s response. We evaluate these methods across three flow systems, multiple architectures, and different training strategies. Comparisons against additional training and direct state correction further assess their methodological value. We have revised the manuscript to explain these contributions more clearly for ML readers. We therefore believe the work is appropriate for ICLR as a contribution to the design, training, and inference of neural simulators.

References

[1] Haoxiang Wang, Tao Yu, Hui Qiao, and Qionghai Dai. Neural Fluid Simulation on Geometric Surfaces. International Conference on Learning Representations (ICLR), 2025.

[2] Yunfei Huang and David S. Greenberg. Geometric and Physical Constraints Synergistically Enhance Neural PDE Surrogates. International Conference on Machine Learning (ICML), PMLR 267, pp. 25265–25310, 2025.

[3] Yiheng Du and Aditi Krishnapriyan. EddyFormer: Accelerated Neural Simulations of Three-Dimensional Turbulence at Scale. Advances in Neural Information Processing Systems (NeurIPS), volume 38, 2025.

AuthorsOpenReview ↗

Response to reviewer YHSM: Part 2

Q2: SC includes a conservative numerical transport update, but it does more than advect a scalar using the current velocity. It uses both the current velocity and the projected next velocity predicted by the neural model. It retains four scalar averages within each coarse cell across predictions, preserving spatial information that a single cell average would lose. A shared network with 257 parameters learns internal fluid motion while preserving the volume transferred across cell boundaries. Scalar prediction errors train this network and the velocity predictor jointly. Conservation and scalar bounds follow from the transport construction under the stated conditions, independently of the learned weights.

The architecture still predicts fluid motion, so SC does not replace the neural simulator with a solver for the full coupled fluid equations. Moreover, numerically transporting a scalar with an inaccurate predicted velocity does not automatically produce an accurate trajectory. Our experiments therefore evaluate the complete learned system. SC lowers error over steps 13–30 in seven of eight shear flow models compared with both MP and MP given an additional hour of training. SC adds transport computation, and these results establish predictive benefits rather than a speed advantage over conventional solvers.

More broadly, conventional numerical solvers remain effective alternatives, while ML offers additional ways to represent solutions, incorporate observations and learn dynamics. For example, Raissi et al. developed PINNs for forward and inverse problems involving established PDEs, including equations already accessible to classical numerical methods [1]. Their methodological contribution illustrates the value of exploring alternative formulations. Our corresponding question is how persistent scalar information, learned transport and physical constraints can improve autoregressive neural prediction.

[1] Maziar Raissi, Paris Perdikaris, and George E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707, 2019.

AuthorsOpenReview ↗

Response to reviewer YHSM: Part 3

Q3: Thank you for the question. RC is not meant to predict two steps ahead more accurately. It adjusts the one-step prediction using the model's next prediction made from it, so that the error caused by feeding back the model's own prediction does not accumulate. Let x^n+1=F(hn)\hat x_{n+1}=F(h_n) be the prediction that is fed back and xn+1gtx_{n+1}^{gt} the ground truth. The error of the next prediction can be written as

x^n+2−xn+2gt=[F(hn+1(x^n+1))−F(hn+1(xn+1gt))]+[F(hn+1(xn+1gt))−xn+2gt].\hat x_{n+2}-x_{n+2}^{gt}=\big[F(h_{n+1}(\hat x_{n+1}))-F(h_{n+1}(x_{n+1}^{gt}))\big]+\big[F(h_{n+1}(x_{n+1}^{gt}))-x_{n+2}^{gt}\big].

The second term is the model's one-step error from the true frame and does not depend on x^n+1\hat x_{n+1}. The first term is the error introduced by feeding back x^n+1\hat x_{n+1} instead of xn+1gtx_{n+1}^{gt}, and this term is what accumulates during a rollout. RC is designed to reduce the first term. The companion predicts F(hn+1(xn+1gt))F(h_{n+1}(x_{n+1}^{gt})), so the gap rr in Eq. (5) estimates the first term, with opposite sign, without containing the second. The correction δx=PcJ⊤r\delta x=P_cJ^\top r follows the derivative of the response with respect to the fed-back frame, and the acceptance test in Eq. (6) keeps a bounded step only when the gap decreases sufficiently (Section 3.3; Appendix C.4 gives the condition for a strict decrease of the field error). Matching xn+2gtx_{n+2}^{gt} instead would also try to cancel the second term through the fed-back frame, that is, distort x^n+1\hat x_{n+1} to compensate for an error it did not cause.

A two-step loss acts on both terms through the weights of FF, on average over the training data. RC leaves FF unchanged and corrects the frame of the current rollout at inference. The two are therefore complementary, and the paper already applies RC on top of models trained with five-step unrolling and pushforward: over the 43 RC deployments in the main configurations, RC lowers the steps 13–30 error in 36 cases and the full-rollout error in 32 (Section 4.2, Appendix D).

In order to test the suggestion directly, we trained with a two-step loss using the paper's unrolling method with a two-step window and gradients through both predictions. We trained in two ways. First, each 12-h one-step Baseline was trained for one extra hour with this two-step loss on the RTX 5090 (32 GB), the same budget as RC's companion (Table R1). Second, on Rayleigh–Bénard, both models were trained from scratch with the two-step loss for the full 12 h on the RTX 5090, and we also included experiments that applied RC on top of the one-hour two-step models (Table R2).

Table R1. One extra hour of training with a two-step loss or for RC. Full-rollout test VRMSE (lower is better; bold marks the best value in each row). RB: Rayleigh–Bénard, SF: shear flow, AM: active matter; Baseline: 12 h one-step training.

SettingBaseline+1 h one-step+1 h two-stepBaseline + RC
RB, FNO8.6208.3817.6914.164
RB, CNextU-Net3.8503.8703.8951.117
SF, CNextU-Net5.0475.7303.8882.450
AM, FNO1.5291.5641.5351.377
AM, CNextU-Net1.3701.3701.3421.292

Shear-flow FNO is omitted because its full-rollout VRMSE exceeds 10 in every column.

Over the full rollout, Baseline + RC has the lowest error in all five settings. The extra hour of two-step training changes the Baseline's error by between a 1 % increase and a 23 % decrease, whereas RC reduces it by 6–71 %. The difference between RC and the two-step hour is significant in four of the five settings. The two-step loss mainly improves the short horizon, reducing the steps 6–12 error by 3–13 % in four of the five settings.

Table R2. Two-step training on Rayleigh–Bénard. Full-rollout test VRMSE (lower is better; bold marks the best value per model; Baseline: 12 h one-step training).

TrainingFNOCNextU-Net
Baseline8.6203.850
Two-step, 12 h>10>10
Baseline +1 h two-step7.6913.895
Baseline +1 h two-step + RC3.5720.963

Trained with the two-step loss for the full 12 h, both models reach a full-rollout VRMSE above 10, well above the one-step Baseline (8.620 and 3.850). Starting from the trained Baseline, one more hour of two-step training changes the error moderately, and RC applied on top reduces it further (from 7.691 to 3.572 for FNO and from 3.895 to 0.963 for CNextU-Net). A two-step loss does not replace RC, and RC still helps on top of it.

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

This paper proposes three components to stabilize the fluid dynamics forecasting model's rollout, including components that enforce the divergence and boundary. The method is evaluated on three datasets in the Well dataset.

Critical strengths
  • The experimental setup seems to be consistent and includes different good baseline models.
Critical weaknesses
  • The paper is heavily AI-written, including the entire related work section, and major parts of the method section, making the method hard to understand. For example, the divergence remove operation and the boundary description in sec. 3.1 are mixed together.

  • MP and SC do not appear to result in consistent improvement in experiment results.

  • RC does seem to result in improvement, but it would require significant additional inference time since it uses an additional trained model. The additional parameters and training time may also make the comparison unfair.

  • The test dataset size seems to be small, which may result in high variance and sensitivity to outliers.

Questions for authors
  • How does the proposed method compare to other stabilization methods, such as PDE-refiner?

  • How is boundary handled in MP? Can you describe SC and RC in more natural languages?

3 comments
AuthorsOpenReview ↗

Response to reviewer kMiu: Part 1

W1: Thank you for highlighting the presentation issues. We have revised the abstract, introduction, related work, and methods to improve clarity for readers without a CFD background. Section 3.1 now distinguishes discrete incompressibility from wall boundary conditions and explains how the projection satisfies both. We also define the notation before use and explain SC’s retained scalar information and RC’s response target and correction procedure in natural language alongside the key equations. These revisions are highlighted in the updated manuscript.

W2: The benefits of individual components depend on the physical system, training strategy and prediction horizon. Our experiments demonstrate improvements across multiple settings, with broader gains from the combined framework.

MP lowers the first forecast error in 23 of 24 models, by 21% on average (Section 4.2). It can provide substantial improvements over longer rollouts. For Rayleigh–Bénard FNO trained with a single step loss, MP alone reduces full rollout error from 8.620 to 1.535, an 82% reduction. Training through MP further improves on applying projection only during rollout, reducing steps 13–30 error from 7.279 to 4.731 for CNextU-Net (Section 4.2, Appendix E.1).

SC lowers steps 13–30 error in seven of eight shear flow models relative to MP, averaging 38% over pairs with both errors at most 10 (Section 4.2; Tables 10 and 11, Appendix E.2). It also outperforms an additional hour of ordinary MP training in seven of eight models in this window. On Rayleigh–Bénard, SC lowers full rollout error in all eight MP models. These results support the value of coupling predicted velocity with persistent conservative scalar transport.

With RC and PA for pushforward training, MP+RC improves steps 13–30 in 17 of 18 settings and full rollouts in 15 of 18 (Section 4.1, Table 1). MP and SC provide physical constraints within each transition, while RC addresses remaining prediction errors through the trained model’s response.

W3: RC adds work only at inference. At each corrected step it evaluates the companion once, the frozen model FF once together with one backward pass for the derivative JJ, and FF once more on the candidate frame (Section 3.3 of the rebuttal revised PDF). This last evaluation is reused as the prediction of the next step, so it replaces the forward pass of FF there. RC corrects from a starting step selected on training or validation data (Appendix C.3). We measured the time costs in Tables R1 and R2 and the training time of PDE-Refiner on one RTX PRO 6000, running the Baseline and all compared variants on the same card.

Table R1. CNextU-Net, time cost of a full test run relative to the Baseline on Rayleigh–Bénard (RB), shear flow (SF), and active matter (AM). Parentheses give the starting steps of RC.

DatasetBaseline+MP+RC+MP+RCPDE-Refiner
RB1.00×1.01×1.70× (2)1.83× (6)1.20×
SF1.00×1.01×1.88× (20)1.91× (26)1.64×
AM1.00×1.01×1.67× (15)1.71× (15)1.54×

Table R2. FNO, same as Table 1. N/A: no starting step qualifies for RC, so RC stays off and the time equals the Baseline's (Appendix C.3).

DatasetBaseline+MP+RC+MP+RC
RB1.00×1.01×1.79× (13)1.66× (7)
SF1.00×0.99×N/A1.92× (6)
AM1.00×1.00×1.42× (16)1.45× (14)

RC takes less than twice the Baseline's time, and +MP changes it by at most 1 %, within run-to-run variation.

RC does not retrain FF. Its companion has the same architecture and is trained once for 1 GPU-hour, against 12 GPU-hours for FF (Appendix A). It only supplies the response target and never produces a forecast. Against the same model trained for one more hour, RC lowers steps 13–30 error in 32 of 40 comparisons, by 20 % on average against 2 % for the extra training (Section 4.2, Appendix E.4), and one more hour of two-step training also leaves a higher full-rollout error than RC (our reply to Reviewer YHSM). We further included PDE-Refiner, which takes 1.20×–1.64× the test time (Table 1) and 1.04×–1.38× the training time of the Baseline, while +MP+RC is more accurate over long rollouts (our reply to Q1 below).

AuthorsOpenReview ↗

Response to reviewer kMiu: Part 2

W4: Thank you for raising this concern. We now assess uncertainty and sensitivity to individual trajectories directly. We compare MP+RC with the matching Baseline across both architectures and three training strategies, using PA for pushforward. Confidence intervals are computed using 100,000 paired bootstrap resamples of complete trajectories, preserving the pairing between methods. Frames are not treated as independent samples.

The following results summarize the comparisons. A positive confidence interval indicates lower mean error with MP+RC.

DatasetEvaluation windowTest trajectoriesComparisons with a positive 95% confidence intervalComparisons retaining improvement after removing any three trajectories
Rayleigh–BénardSteps 13–30356/66/6
Rayleigh–BénardFull rollout356/66/6
Active matterSteps 13–30263/64/6
Active matterFull rollout266/66/6
Shear flowSteps 13–30132/62/6
Shear flowFull rollout131/3 finite comparisons1/3 finite comparisons

For Rayleigh–Bénard, improvements occur on 28–35 of the 35 trajectories in each comparison. All 12 paired sign tests remain significant after Bonferroni correction across 36 planned comparisons, with a maximum adjusted p=0.0183p=0.0183.

To assess outlier sensitivity, we removed the three trajectories showing the largest paired gains in each comparison. Mean improvement remains positive in all 12 Rayleigh–Bénard comparisons and all six active matter full rollout comparisons. Consequently, removing any three trajectories cannot reverse these observed improvements.

These checks support robustness in those settings. Shear flow provides weaker statistical evidence, which we distinguish explicitly. Three shear flow full rollout comparisons contain nonfinite scores and therefore receive no bootstrap interval. All finite scores, including those above 10, are retained. Bootstrap intervals are pointwise and describe uncertainty across test trajectories for the evaluated trained models.

AuthorsOpenReview ↗

Response to reviewer kMiu: Part 3

Q1: Thank you for the suggestion. We trained PDE-Refiner (Lippe et al., 2023) on the same CNextU-Net model as our Baseline, with the same data, input history, optimizer and learning-rate schedule. It uses the residual (state-difference) target, K = 3 refinement steps, a minimum noise standard deviation of 10⁻³ and EMA weights with decay 0.995, and it evaluates its network K + 1 = 4 times per prediction. We measured the training time of PDE-Refiner and the Baseline on one RTX PRO 6000, running both on the same card for the same number of updates. PDE-Refiner takes 1.07×, 1.04× and 1.38× the Baseline's time on Rayleigh–Bénard, shear flow and active matter, respectively.

Tables R3–R5 give the test VRMSE of CNextU-Net (lower is better, and bold marks the best value in each column).

Table R3. Rayleigh–Bénard.

MethodStep 1Steps 6–12Steps 13–30Full Rollout
PDE-Refiner0.0450.5244.5182.914
Baseline0.0780.9538.7493.850
+MP0.0580.4534.7313.036
+RC0.0780.7132.0241.117
+MP+RC0.0580.3800.7390.871

Table R4. Shear flow.

MethodStep 1Steps 6–12Steps 13–30Full Rollout
PDE-Refiner0.0100.1900.6223.647
Baseline0.0040.0500.1625.047
+MP0.0040.0420.1444.283
+RC0.0040.0500.2032.450
+MP+RC0.0040.0420.1421.369

Table R5. Active matter. PDE-Refiner failed to learn refinement here (its refinement loss stayed at the noise level), so its errors mainly reflect this failure.

MethodStep 1Steps 6–12Steps 13–30Full Rollout
PDE-Refiner0.6372.5982.5441.945
Baseline0.0900.5681.6491.370
+MP0.0910.6651.7581.464
+RC0.0900.5681.4991.292
+MP+RC0.0910.6651.5781.204

On Rayleigh–Bénard, PDE-Refiner improves the plain Baseline, and it has the lowest step-1 error which +MP's is close to, but over steps 13–30 and the full rollout it stays well above both RC models. Compared with PDE-Refiner, +MP+RC lowers the error by 84 % over steps 13–30 (0.739 against 4.518) and by 70 % over the full rollout (0.871 against 2.914). On shear flow, PDE-Refiner is above all our models in every window up to step 30, and +MP+RC lowers the full-rollout error from 3.647 to 1.369. On active matter (Table R5), all our models are below PDE-Refiner in every window.

The methods also differ in what they require. PDE-Refiner changes the training objective and the inference procedure of the model itself, evaluating the network four times per prediction. MP and SC constrain each prediction of the model, and RC keeps the trained model unchanged and corrects the frame that is fed back during the rollout, at the cost reported in our reply to W3.

Q2: MP handles periodic directions through its Fourier representation. On walled domains, the Fourier–Chebyshev solve jointly enforces zero divergence and zero velocity normal to the physical walls, even when the stored grid excludes boundary points. This prevents fluid from crossing the walls. MP does not prescribe tangential wall velocity.

SC stores four scalar averages, such as dye concentrations, within each grid cell and retains them across predictions. The current and projected velocities transport these values, while a small network adjusts fluid movement inside each cell without changing the fluid volume crossing its boundaries. Each transfer removes an amount from one region and adds it to another. This conserves scalar amount on closed domains, while step size restrictions preserve bounds. Diffusion, sources, and wall exchange are treated explicitly.

RC trains a companion network initialized from the frozen predictor to predict the predictor's response to the ground-truth next field (how it would continue if that ground-truth field were appended to its input). During inference, this response target is compared with the predictor's response to its own forecast. Their difference guides a bounded correction to the forecast through input gradients. The corrected candidate is accepted only if it sufficiently reduces this difference, else the forecast is kept unchanged. Correction starts at a step selected on training or validation data, and the first prediction is never changed. At inference, RC needs the same input as the predictor alone, and the predictor's weights stay unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.