- 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
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.
| Dataset | Baseline | +MP | +RC | +MP+RC | PDE-Refiner |
|---|---|---|---|---|---|
| RB | 1.00× | 1.01× | 1.70× (2) | 1.83× (6) | 1.20× |
| SF | 1.00× | 1.01× | 1.88× (20) | 1.91× (26) | 1.64× |
| AM | 1.00× | 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).
| Dataset | Baseline | +MP | +RC | +MP+RC |
|---|---|---|---|---|
| RB | 1.00× | 1.01× | 1.79× (13) | 1.66× (7) |
| SF | 1.00× | 0.99× | N/A | 1.92× (6) |
| AM | 1.00× | 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.
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).