acceptodds
Under review as a conference paper at ICLR 2027

Simulating Fluid Dynamics with a Parameter-Efficient Neural Architecture

Abstract

Neural operators provide fast surrogates for fluid forecasting, but accurate prediction with a compact model remains challenging. Fluid evolution combines spatial transport with changes in field values, yet many architectures represent both through the same latent transformations, which can increase the representational burden. To address this, we introduce SIDDY, a parameter-efficient neural architecture that reuses observed content through explicit transport and learns a refinement process for the remaining evolution. Specifically, a shared encoder first summarizes the full observation history to guide both transport and refinement, while the latest field supplies the state that is explicitly evolved. Semi-Lagrangian sampling then transports this field once, after which reaction-diffusion layers refine the transported state through learned nonlinear updates and fixed Laplacian stencils. This physics-inspired decomposition enables accurate forecasting with a compact parameterization. Experiments on both simulated and real-world dynamics show that SIDDY achieves the lowest error across all datasets while using only a fraction of the parameters.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-Informed Machine Learning, Deep Learning Architecture

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

This paper introduces SIDDY, a compact neural architecture for fluid forecasting. Its design is motivated by the potential representational burden of modeling spatial transport and changes in field values through the same latent transformations. SIDDY encodes the observation history into a shared context that guides both transport and refinement, while explicitly transporting only the latest observed field. A single transport stage per prediction step is followed by multiple reaction–diffusion refinement layers that learn corrections to the transported prediction. This organization separates historical conditioning from state evolution, avoiding repeated transport of historical observations. Experiments on simulated and real-world dynamics report lower forecasting errors with fewer parameters than the evaluated baselines.

Critical strengths
  1. SIDDY achieves a favorable trade-off between predictive accuracy and model size, obtaining lower relative L2L_2 errors than the evaluated baselines across Navier–Stokes, ink-video, and atmospheric forecasting tasks while using fewer parameters.
Critical weaknesses
  1. The separation between transport and non-transport evolution has direct precedents in previous works. NowcastNet (2023) uses a shared history encoder to predict motion fields and intensity residuals, combining semi-Lagrangian transport with additive updates. LUPIN (2025) similarly separates learned advection from advection-free intensity evolution, while as early as de Bezenac et al. (2018) already infer motion from history and warp the latest observed field. The authors should clarify what additional methodological insight SIDDY provides beyond these principles and isolate the benefits of its particular organization under comparable parameter and computational budgets.
Questions for authors
  1. The paper motivates SIDDY by suggesting that representing transport and field-value changes through the same latent transformations increases the representational burden. Could the authors define this burden more precisely and provide theoretical reasoning or controlled experimental evidence that specifically supports its reduction through the proposed decomposition?

  2. Could the authors make the motivation behind the main architectural choices in Section 3 more explicit? In particular, what motivates applying transport only once per prediction step, followed by multiple refinement layers, and using a shared history context to condition both stages?

  3. Could the authors clarify the physical interpretation of the transport and refinement modules in Figures 1c and 1d? Beyond their computational definitions, concrete examples explaining the roles of the learned displacements, nonlinear updates, and diffusion stencils would be helpful. Please also distinguish physically grounded properties from architectural analogies, particularly because refinement operates in latent feature space.

  4. Does each iteration of the deduction require recalculating the historical context? How does this affect the inference time?

Reviewer_pqTPOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

SIDDY forecasts 2D fluids with a compact grid model: a history encoder builds a context field, a semi-Lagrangian module predicts backward displacements and samples the latest frame, and gated reaction-diffusion layers with fixed Laplacian stencils refine the result. On Navier-Stokes at three viscosities, ERA5-Z500 and five Spread Ink videos it reports the lowest relative L2L_2 in all nine settings with the fewest parameters.

Critical strengths
  • The ablations are candid and thorough.
  • SIDDY outperforms baselines.
Critical weaknesses
  • The main table has some soundness concerns.
  • The reaction-diffusion framing is insufficiently justified.

See questions for further details.

Questions for authors

Presentation

Clarifications

  1. Is Table 5's latency per step or per rollout? 0.159 s for FNO seems long for one step.

Soundness

Clarifications

  1. Which LRSA configuration and batch size were used, and does SIDDY's margin over LRSA survive LRSA at its published level? On ERA5, the square roots of Table 1's MSE give RMSE 643 for HelmFluid and 1017 for FNO against their respectively published 521 and 597, but HelmFluid's released script evaluates on the 2020 validation year rather than App. B.2's 2021-22 test years; which was evaluated here?
  2. What is the spread over the three runs? The depth sweep at ν=10−3\nu = 10^{-3} zigzags, and widening the no-context model changes error significantly.
  3. Under Eq. 19 a zero forecast scores 1.0, yet HelmFluid scores 1.1615 at ν=10−4\nu = 10^{-4} and 1.03 on Synthetic 2, and ADRNet 1.7-4.6 on every video; Flowers at ν=10−3\nu = 10^{-3} and HelmFluid at ν=10−3\nu = 10^{-3} are unexpected. Were these diagnosed?
  4. What sign do the learned layer scales αℓ\alpha_\ell take?

Errata

  1. Table 1's FNO row matches the FNO-2D row of Li et al. (2021) to four digits in all three settings, yet Sec. 4 says Table 1 reports "the mean values of three independent runs" and App. D describes FNO as trained here. Was it copied?
  2. Is the abstract's credit to "this physics-inspired decomposition" accurate? In Table 3, removing refinement costs 6-15%, removing transport 35-78% and removing the encoder 3.0-9.9x.
  3. Table 1's Δ\Delta rows report 97.9% and 99.4% reductions against failed runs, but the text cites margins against the strongest baseline; does this accurately portray SIDDY's performance?
  4. How are Sec. 4.3's readings, that the zigzag "suggests that the optimal refinement depth depends on the dynamics of the forecasting task" and that S=1S = 1 "gives the lowest error at all three viscosities" distinguished from run variance?

Informativeness

Clarifications

  1. The "w/o context" variants remove the encoder and, presumably, all cross-frame mixing except the softmax, whereas every Navier-Stokes baseline receives ten frames. Without a variant that gives transport and refinement plain access to the history, how is the encoder's design distinguished from merely the presence of history?
  2. The "reaction-diffusion" strengths are positive, but Eq. 9 scales the increment by an unconstrained αℓ\alpha_\ell, so the effective coefficient "can take either sign" and acts in feature space (Apps. A.2, E). The Laplacian branch lowers L2L_2 but raises VGG loss on both videos. Is "diffusion" meant physically?
  3. Transport needs velocity, which depends nonlocally on the vorticity the model sees, yet Sec. 3.1 motivates the global branches only by "flow outside these local neighborhoods", and Table 3 never ablates attention or the spectral branch. Is velocity inference their actual role? What do they actually contribute?
  4. For the interleaved and repeated-transport variants, where do the extra parameters come from, and how much of the gain is capacity rather than organization?
  5. A semi-Lagrangian step has a single departure point. What motivates K>1K > 1 candidates mixed by a softmax, and how does their role differ from the Laplacian branch's?
  6. Despite fewer FLOPs than five of six baselines, SIDDY's latency is 2nd highest. Why?

Addenda

  1. Do the learned displacements match the physical velocity, and when in training does that emerge? Do the candidates specialize?
  2. How is each step's change divided between transport and refinement, and how does that vary with viscosity and spatial scale?
  3. Repeated bilinear sampling damps high wavenumbers over a 10-40 step rollout. Does refinement compensate, and how does that interact with the ν=10−3\nu = 10^{-3} tail in Fig. 4?
  4. What do the refinement layers do spectrally, and does that explain why the Laplacian branch lowers L2L_2 but raises VGG loss on the synthetic videos?
  5. How does the transport prior handle patterns that do not move with the flow (e.g. Rossby waves in Z500, source-dominated changes) and inflow boundaries, where border padding samples clamped values?
  6. Where does single-step sampling break down as the per-step displacement grows?
  7. Sampling is not conservative. How much does total ink drift over the synthetic rollouts?
Reviewer_6b52OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SIDDY addresses parameter-efficient autoregressive forecasting of fluid fields. It encodes the observation history into a shared context, uses semi-Lagrangian-inspired sampling to transport values from the latest observed physical field once per prediction step, and applies reaction–diffusion refinement to predict the remaining change. The authors report lower relative L2 error with fewer parameters than six evaluated baselines across three Navier–Stokes settings, five Spread Ink videos, and ERA5-Z500 forecasting.

Critical strengths

The separation between physical-field transport and learned refinement is clear and useful. The model is compact, the Navier–Stokes L2 improvements are substantial, and the paper includes component ablations and long-horizon rollouts.

Critical weaknesses

The novelty of the broad transport-plus-correction framing appears limited relative to closely related work. In particular, LUPIN and Nowcast3D are absent from Related Work, although both transport an observed physical field and model subsequent changes separately. The paper reports means over three runs but no run-level results, standard deviations, or confidence intervals, making the smaller gains on ERA5-Z500 and some ink videos difficult to assess. Each Spread Ink setting trains and tests on different segments of one video, so the results do not establish generalization to new videos.

Evidence supporting the above. The overlap with prior work is concrete, although the methods are not identical. LUPIN (2025) learns differentiable semi-Lagrangian transport of observed rainfall and adds a learned source–sink correction. Nowcast3D (2025 preprint) encodes radar history, advances the latest observed physical volume through advection and diffusion, and adds a source term. SIDDY's single transport stage, shared context in refinement, and parameter budget may distinguish it, but Section 2 does not explain those differences against these close precedents. I do not regard omission of a preprint alone as grounds for rejection.

The quantitative pattern is mixed. Table 1 reports relative L2 reductions of 25.2–57.0% over the strongest evaluated baseline across the Navier–Stokes settings, but the ERA5-Z500 reduction is 3.2%. In Table 2, the L2 margins are 1.8% on Synthetic 1 and 3.8% on Real 3. Section 4 says the results are means of three independent runs; the tables do not show their spread or a statistical test. Appendix B describes each ink setting as one video with an observed prefix and a later test segment. These facts support the reported point estimates, but leave the reliability of small margins and transfer to unseen videos unresolved.

Overall assessment. SIDDY offers a compact and effective combination of known ideas, but its distinctive contribution needs sharper positioning against nearby methods. I would also need stronger evidence that the modest gains outside Navier–Stokes are reliable and that the ink results extend beyond continuation of the same videos. These concerns appear addressable and could change my assessment.

Questions for authors

Clarification requests:

  1. Which technical and empirical benefits are specific to SIDDY when compared with LUPIN and Nowcast3D, especially the choice to transport only the latest field and to reuse one history context in refinement?

  2. What are the per-run errors or standard deviations for the main comparisons in each of the three task families? Do the smaller ERA5-Z500 and Spread Ink margins persist across runs?

  3. Is there a held-out-video evaluation, or a modest additional test on videos not used for training, that shows the ink forecasting gains transfer beyond each sequence's later frames?

Suggestions (not required for my assessment to change):

Please cite LUPIN and Nowcast3D in Related Work and briefly state the relevant distinctions. Reporting per-run results with mean and standard deviation, and adding a limited held-out-video experiment, would directly address the concerns behind my recommendation. A parameter-matched comparison with a close transport-based baseline would further clarify the benefit of SIDDY's particular architecture, although I do not regard a large new benchmark suite as necessary for the response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.