acceptodds
Under review as a conference paper at ICLR 2027

FlowGEN: A Finite-Horizon Truncation Kernel for Unstable 3D Flow Forecasting Across Meshes

Abstract

Fast surrogate models for unsteady computational fluid dynamics (CFD) need to maintain representation accuracy under complex spatial discretizations and evolution stability in long-term prediction. However, existing methods still have limited joint modeling of the temporal and spatial dimensions. As a result, they cannot maintain accuracy in both dimensions. We propose FlowGEN, a graph evolution network centered on a Finite-Horizon Truncation Kernel (FHTK). FHTK is a separable temporal support kernel over history lag and forecast offset. Along the history axis, a transient-state mask activates only the latest three states, so an older prediction loses its direct feedback edge after leaving the finite-memory window while its dynamically relevant information remains summarized by the evolved recent states. Along the forecast axis, a normalized finite-support kernel determines which recursively generated states contribute loss and gradient. Its support expands from three to six steps during training, and a nested six-/twelve-step validation kernel balances local transition accuracy with delayed instability beyond the optimized horizon. Every supported state is supervised by standardized data fidelity, kinetic energy (KE) matching, and mixed graph divergence (Div) regularization. Across structured and unstructured unstable flows at four Reynolds numbers, FlowGEN obtains the lowest reported mean in all 48 condition–metric comparisons. Relative to the strongest baseline for each metric, it reduces the error-growth ratio by 32.1%, intermediate-horizon MSE by 8.6%, KE error by 31.5%, and Div error by 3.7%. These results show that explicitly controlling state activation and recursive credit provides a practical mechanism for stable flow forecasting across discretizations.

Primary area:
learning on time series and dynamical systems
Keywords:
Neural PDE Solvers, Long-Horizon Autoregressive Forecasting, Graph Neural Networks, Computational Fluid Dynamics, Mesh-Agnostic Learning

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

The paper claims CFD needs to maintain representation accuracy in long-term prediction, and current work has limited modeling in both temporal and spatial dimensions, limiting accuracy. So the author proposes FlowGen with FHTK. It uses a transient-state mask that activates only the latest three states. A normalized finite support kernel is used to determine which states contribute to the loss. The proposed method obtains the lowest reported mean in all 48 condition--metric comparisons.

Critical strengths
  1. Part of the motivation is from physical intuition, which provides a clear background and structure.
  2. The paper provides all 48 condition-metric comparisons for the evaluation of the method.
Critical weaknesses
  1. The use of limited history windows has been explored in prior methods, and similar designs can be found for other proposed components. The main contribution appears to lie in integrating these components into a structured framework, but this requires clearer justification and more direct comparisons with simpler alternatives. Moreover, the proposed methods do not consistently outperform simpler baselines in some settings, raising concerns about their effectiveness and necessity.
  2. Beyond the concerns mentioned above, the experimental evaluation remains insufficient. The ablation studies do not clearly isolate the contributions of individual components or adequately support their claimed benefits. A more detailed analysis of the design choices would be valuable, and the key findings should be included in the main text to better justify the proposed framework.
Questions for authors
  1. What are the key novel contributions beyond existing methods with similar designs?
  2. Why is the proposed framework necessary when simpler methods achieve comparable performance?
  3. Can the authors provide more targeted ablations to isolate and validate the contributions of individual components?
Reviewer_ww4UOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work establishes a neural surrogate (FlowGEN) that predicts the next 3D velocity field of an unsteady flow over a cylinder (claimed from Gerris data, but can't find any citation). It is a standard autoregressive rollout, meaning feeding the prediction back as input for the next step. The authors also want one model to handle both structured and unstructured meshes. FlowGEN treats every mesh as a point cloud (this already adapts naturally to arbitrary mesh types). It (i) builds a k-nearest-neighbour graph with k = 12 from the coordinates and adds Fourier positional features; (ii) applies two EdgeConv layers and six residual MLP blocks; (iii) predicts the velocity increment, so the model is effectively "previous velocity + learned correction"; (iv) is trained with a data loss, a kinetic-energy (KE) matching loss and a graph-based divergence ("Div") penalty. The main idea (Finite-Horizon Truncation Kernel, FHTK) is essentially three training and selection rules, which are (i) using 3 most recent states (History window), (ii) applied to the first H rollout steps (Loss horizon); and (iii) selecting the checkpoint with the best validation score on a 6-step plus 12-step rollout with equal weights.

Critical strengths
  1. The use of physics-aware losses is encouraging. Matching KE and component-wise second moments is cheap and physically meaningful. The ablation suggests KE matters a lot on unstructured data.
Critical weaknesses
  1. The author is repackaging a common practice into a novel method. For example, the history mask K=3 in FHTK is a very common and ordinary practice in every autoregressive model, and calling it a "transient-state mask" does not make it novel. Second, the forecast kernel k_H, which is a uniform average of the step losses over the first H (3 to 6) rollout steps, is essentially the same family as the pushforward, temporal bundling technique, a typical multi-step unrolled loss with a horizon curriculum. No comparison is done with them.
  2. Suspicious formulations and poor figure representation. The joint kernel K_{K,H}(j,r) in Eq. (13) is never used by itself. And, the first-order Jacobian error expansion in Eq. (15) is never measured or checked. Is the"∞" upper limit in Eq. (14) a finite sum? The kernel notation adds an appearance of theory, not a result. In addition, Figure 2 is hard to read, no clylinder body can be observed in the near field of a flow around a cylinder, and the color contours of the bottom image are meaningless without any scale and definition.
  3. I am concerned about how the dataset is split, and its existence. All data are one geometry (a cylinder) and one simulation per condition. The test set is the final 25% of the same trajectory, so the model never meets a new geometry, new initial condition or new Re. Training and test snapshots come from the same single flow condition. "Across meshes" means two sampling types of one flow. It does not mean different geometries or resolutions. There is no test of how the model generalises to a different adaptive refinement pattern. The Gerris dataset is not cited and its details are missing: domain, cylinder size, number of points, number of frames, Δt, and how many shedding periods or flow-through times 35 steps cover. For an "unstable flow" claim, a 35-step horizon may be short in physical time.
  4. There is a missing cost, reproducibility, and clarity section. No training or inference cost is reported. Back-propagating through 6 recursive steps costs more memory and time than one-step training. The motivation is fast surrogates, but there is no speed-up vs CFD and no cost comparison with baselines.
Questions for authors
  1. How is the history mask different from the standard 3-frame input window? What is gained from the "separable kernel" formalism beyond naming the loss horizon and the validation horizon?
  2. Was Eq. (15) (the Jacobian error expansion) ever tested, for example by measuring how rollout error depends on K or H?
  3. Are the baselines trained with the same multi-step rollout loss and the same 3→6 curriculum? If not, please add that control. Also compare FlowGEN trained with (i) one-step loss only, (ii) a plain 6-step or 12-step unrolled loss, and (iii) pushforward or noise injection, so that the specific contribution of FHTK is visible.
  4. What are the Δt, the number of points, the number of frames and the physical time covered by 35 steps, and where does the Gerris dataset come from? Does the 12-step validation window (and the 35-step test) cover a full shedding period at Re = 160?
  5. The paper states that the benefit is largest for the most unstable (Re = 160) cases. Is this because Re = 160 is the most laminar or periodic (and easiest to extrapolate), while at Re = 6400 the benefit vanishes?
Reviewer_jqzHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

FlowGEN predicts 3D Gerris flow with an autoregressive graph model. It uses three past velocity frames as input. FHTK limits feedback to those three states, increases the training horizon from three to six steps, and uses a six-/twelve-step validation kernel for checkpoint selection. The paper tests structured and unstructured data at four Reynolds numbers over 35-step rollouts with three seeds.

Critical strengths
  1. The paper separates which states are fed back, which receive gradients, and which are used for checkpoint selection. This is a useful way to think about autoregressive training. The use of validation rather than test trajectories for extended selection is also sensible.

  2. The rollout results look promising. Table 4 covers eight mesh–Re conditions and six metrics, with the best mean in all 48 entries

Critical weaknesses
  1. The experiments do not isolate FHTK. The w/o FHTK variant only changes the selection kernel from twelve to six steps. It keeps the feedback window and progressive training. The history-length and loss-transfer results announced in the text are missing, so the central attribution is not convincing

  2. The generality of the evidence is limited. The paper tests only Gerris and splits train, validation, and test within temporal sequences. There are no unseen geometries, mesh resolutions, or Reynolds numbers. Three seeds without significance tests are also not enough when several differences are close to the standard deviations

Questions for authors
  1. Can you provide a same-backbone factorial ablation for each part of FHTK and report the missing history-length and loss-transfer experiments?

  2. How many independent trajectories are tested per condition, and do the results hold on unseen geometries, mesh resolutions, or Reynolds numbers?

  3. Can you add per-seed or paired statistical comparisons and validate the graph divergence proxy against a standard divergence measure on structured meshes?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.