acceptodds
Withdrawn from ICLR 2027

FluidJEPA: Joint-Embedding Predictive World Models for Controlled Fluid Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modeling controlled fluid dynamics requires predicting how a flow evolves under a prescribed action sequence. Unlike forecasting under fixed forcing, the same initial state can lead to different futures as the control changes. Accurate next-field prediction, however, does not ensure that a model's internal state can support repeated action-conditioned prediction. We introduce \method, a joint-embedding world model that learns physical-field prediction and latent dynamics together. Action-conditioned latent predictions are matched to encoded future observations. A jointly trained decoder retains physical-field supervision, and a distributional regularizer discourages representation collapse. We construct a benchmark of four two-dimensional flow environments spanning internal forcing and boundary actuation, with 8,000 trajectories organized into shared-initial-state control branches. Evaluation separates field-space feedback from latent rollout and tests control-induced changes and generalization to an action family withheld from training and validation. \method achieves competitive field accuracy. In same-architecture ablations across the four environments, latent alignment reduces 20-step mean physical-field error under latent rollout by 73.6–96.6% relative to training without this alignment. Action-family holdout produces modest degradation in periodic Navier–Stokes latent prediction relative to training on all families. The larger degradation in Pinball highlights a remaining challenge when unseen actions also change actuator coordination.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
fluid dynamics modeling

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 3 reviews, above 27% of the others.
Reviewer_XVRKOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FluidJEPA as a latent model to evaluate the field prediction, latent recurrence, and action-family transfer for the fluid dynamics.

Critical strengths

The strength is that the author introduces sufficient benchmark datasets to test and compare. Ablation study is also done to support the main claim.

Critical weaknesses

My biggest concern is that the proposed framework simply uses previous methods (architecture, loss design, action-conditioned modeling) and technical innovations are limited.

Questions for authors
  1. Some notations in Fig. 1 are used but not explained. E.g., A, B, q_t, F_i.
  2. The technical contributions are quite limited as the model architectures, loss designs, and action-conditioned modeling are all already there.
  3. The gap between accurate decoded predictions and reliable latent recurrence is a known issue. The reported results only partially address it, as latent rollout remains less accurate than decoding and re-encoding in the main experiments.
  4. The baseline comparison is too limited, with only FNO and U-NO as external baselines. Please include methods supporting direct latent recurrence, such as Learning to Accelerate Partial Differential Equations via Latent Global Evolution (Wu et al., 2022) and Toward Goal-Agnostic Joint-Embedding Predictive Control of Partial Differential Equations (Gallagher & Guglielmi, 2026), under matched training and evaluation settings.
  5. There is no drawback discussion and the future work is limited.
Reviewer_Ht5kOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a controlled world model approach combining a latent rollout with fluid field prediction. Going beyond a pure operator learning approach, the paper also considers active fluid control with various control mechanisms. That is, the flow is modulated dynamically conditioned on the actions taken.

To test their paper, the authors simulate 8000 trajectories of controllable 2D fluid flow and train their models theron. Evaluation is performed by a 20 step ahead rollout on trajectories with an initial state unseen in training. The authors compare their models to FNO, UNO and a latent rollout where the objective is is only on the rollout side, finding mostly an advantage of their own method which trains action conditioning and fluid reconstruction simultaneously. The paper furthermore contains detailed studies of the residual error and an extensive appendix with further numerical studies.

Critical strengths
  • The combination of fluid flow prediction and active fluid control is of high novelty. Here the contribution is twofold, on the one hand a data set and on the other hand their JEPA world model and training strategy.
  • The numerical work is very extensive and of high quality bringing convincing evidence for the authors claims.
Critical weaknesses
  • While the FluidJepa model is mostly superior to the competition (FNO, U-NO), the argins are not spectacular on pure fluid roll-out.
  • Some well-known baselines like DeepOnet and Transsolver are not considered in the tests
Questions for authors
  • Would you provide the comparison with missing baselines?
  • How long is you rollout in terms of the characteristic times associated with the largest eddies?
Reviewer_EaK5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the prediction of fluid evolution from an initial observation and a sequence of control inputs. The authors generate 8,000 trajectories across four fluid environments using existing numerical simulators and train a separate model for each environment.

The paper compares two ways of making successive predictions with the same model. Field rollout decodes each predicted representation into a flow field and encodes that field again for the next step. Latent rollout directly passes the predicted representation to the next step, using decoding only to produce the output. Both are evaluated using flow-field prediction errors. Training combines field prediction, matching predicted representations to those of the actual next observations, and SIGReg regularization. Under the reported training setup, representation matching substantially improves latent rollout. Performance on unseen control types varies across environments. All results use the authors' generated data and evaluation settings.

Critical strengths

The comparison reveals a meaningful difference between two rollout procedures using the same model. Both use the same initial observation, controls, and model parameters, and both are evaluated in terms of flow-field error. In NS, removing representation matching changes field error only from 0.0363 to 0.0383, while latent error increases from 0.1039 to 0.9340. Cavity shows a similar pattern (Eq. 7 and Table 18). This directly demonstrates that accurate decoded predictions do not guarantee that the predicted representations can be reused successfully for subsequent steps.

The data design supports fair comparisons and helps isolate the effects of control. Different control sequences start from the same complete simulation state, and trajectories sharing an initial state stay in the same data split. The held-out control type is excluded from training, normalization, and model selection; models trained with and without that type are evaluated on the same test trajectories (Sections A.4–A.5 and B.3; Table 17). FNO and U-NO receive the same control information and have roughly comparable parameter counts (Section B.4; Table 10). These choices strengthen the comparisons within the study.

Critical weaknesses

Limited novelty and an incomplete training comparison. AeroJEPA already combines representation matching, field reconstruction, and SIGReg, while prior JEPA work studies controlled fluid dynamics and latent planning. The contribution therefore rests mainly on the empirical findings. However, two-step training is reported only for the full model (Table 15), leaving unclear how much matching adds beyond addressing the mismatch between one-step training and repeated prediction.

Limited validation of latent control prediction and broader applicability. The direct control-response evaluations in Table 16 and Figure 11 cover only field rollout; full-field errors alone do not establish whether latent rollout accurately captures smaller control-induced changes. Evaluation is also limited to four 2D environments and 20 prediction steps, with fixed geometry and physical parameters in the held-out control tests. Mesh and time-step convergence are not established (Section A.6), further limiting confidence that the findings extend beyond the chosen simulation settings.

Questions for authors

New findings and training setup: What is the most important new finding beyond existing fluid JEPA work, and which results best support it? Can the current evidence distinguish the role of representation matching from the effects of the mismatch between training and inference? If results without matching under the same two-step training budget are available, please include them.

Predicting control effects: Table 16 and Figure 11 evaluate only field rollout. Which results show that latent rollout accurately predicts changes caused by control inputs? Could the existing predictions be used to report the control-response error from Table 16 for latent rollout, allowing a direct comparison with field rollout?

Numerical reliability of the reference data: For the larger time step used in Cavity, are reference results at a smaller time step or finer mesh available to show that the flow fields and paired control responses remain reasonably stable? If not, please explain the rationale for this setting and how it limits interpretation of the findings.

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

Related papers

Discussion (0)

Sign in to comment.