acceptodds
Under review as a conference paper at ICLR 2027

D-JEPA: Design-Recoverable JEPA Representation with Swappable Physics Decoders

Abstract

Joint-Embedding Predictive Architectures (JEPAs) provide a framework for learning compact representations without directly reconstructing high-dimensional observations. However, in parameterized physical systems, learned representations can entangle geometry with operating conditions and task-specific physical responses, limiting their reuse across prediction tasks. We introduce D-JEPA (Design-recoverable JEPA), a geometry-centric JEPA that computes a compact representation from geometry alone and reuses it across operating conditions and physical response spaces through lightweight physics-specific decoders. An explicit design-recoverability objective encourages the geometry latent to preserve information about the underlying design variables, enabling the representation to support design analysis and optimization. We further identify a case-level collapse failure mode in which target representations become nearly invariant across distinct geometries despite low reconstruction error, and mitigate it using case-level variation constraints and auxiliary target reconstruction. Across four 3D aerodynamic, hydrodynamic, and structural benchmarks, D-JEPA maintains or improves full-field prediction accuracy while achieving near-perfect linear recoverability of design parameters. The frozen geometry representation can be reused at held-out operating conditions and transferred to a structural response task with fewer trainable parameters. Finally, the representation supports differentiable design optimization, with designs validated using high-fidelity CFD, preserving the predicted ranking of candidate designs. These results demonstrate that separating a reusable geometry representation from physics-specific prediction provides apractical representation for scientific surrogate modeling and design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific Machine Learning, Physics AI, Surrogate Modeling, Joint Embedding Predictive Architecture, Representation Learning, Disentangled Representation, Representation Collapse

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

This paper proposes D-JEPA that learns a geometry representation independent of operating conditions and physical responses, enabling physics-specific decoders to predict physical properties. The framework utilizes a JEPA-based latent prediction objective to encourage the geometry encoder to capture information useful for predicting physical responses. Experiments on three aerodynamic benchmarks demonstrate promising performance accuracy compared with existing methods. The authors also show that the learned geometry representation can be reused across different scenarios.

Critical strengths
  • The paper introduces a new method that can learn reusable geometry representations for scientific surrogate modeling. By using efficient physical-field prediction without repeatedly performing computationally expensive CFD simulations, the proposed approach shows promising potential for accelerating industrial design and optimization.
  • The method adapts JEPA to surrogate modeling, enabling physics-aware geometric representation learning. It also identifies a case-level representation collapse issue that conventional token-level regularization may fail to address.
  • D-JEPA achieves promising results across three aerodynamic benchmarks, outperforming several baselines. Ablation studies also demonstrate the effectiveness of the proposed case-level constraints and auxiliary reconstruction objectives.
Critical weaknesses
  1. My major concern is about the motivation behind JEPA, which is the main contribution of this paper. It is unclear why JEPA is critical in the framework, since there are some applicable alternatives. For instance, Physgen[1] has a similar design that encodes 3D shapes into latent codes and then jointly decodes physical values. A VAE-like structure seems promising, but the authors do not have a comparison against it.
  2. In addition to this, as shown in Table 3, the model still achieves strong performance without LlatL_{lat}. Given that this loss plays a crucial role in the JEPA architecture, this observation appears to contradict the authors’ claims regarding the benefits of adopting JEPA.
  3. The method explicitly supervises the latent using ground-truth design parameters, while the other baselines do not receive such information. Therefore, the comparisons may not be entirely fair. It would be better if the authors could provide some comparisons under the same or similar supervision settings.

[1] Yingxuan You, Chen Zhao, Hantao Zhang, Ming Xu, and Pascal Fua. "Physgen: Physically grounded 3d shape generation for industrial design." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 27209-27218. 2026.

Questions for authors

Please refer to my comments in Weaknesses

2 comments
Area_Chair_N5cGOpenReview ↗

Author-reviewer discussion period is ending soon

Dear Reviewer sZWT,

If you haven't done so, please read through authors' response and let authors know if you need further clarification or have more questions. Beginning 11/19, authors are not allowed to respond any more, so this is good time to review the response and ask more.

Thank you.

-AC

I thank the authors for the response and for the additional experiments. However, my concerns about the motivation for using JEPA and LlatL_{lat} still exist. The experiments show that JEPA leads to better performance, but it is still unclear why JEPA is critical. A thorough analysis of this would be valuable, as this is the major contribution of this paper. The variants without LlatL_{lat} still show strong performance, making it a bit unclear whether LlatL_{lat} (JEPA-like design) is important. Therefore, I have to keep my original negative rating.

Reviewer_N15yOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes D-JEPA, a geometry-centric JEPA framework for engineering surrogate modeling. The method encodes geometry into a compact latent, keeps operating conditions and physics-specific outputs in lightweight decoders, adds a design-recoverability loss, and evaluates on aerodynamic, hydrodynamic, structural-transfer, and design-optimization settings.

Critical strengths

D-JEPA compares against strong surrogate baselines such as DoMINO and GeoTransolver, report full-field metrics and integrated force coefficients, and include useful ablations on representation collapse, design recovery, operating-condition interpolation, decoder transfer, and CFD validation. The empirical presentation is much more complete than many surrogate-modeling papers.

Critical weaknesses

My main concern is that the novelty is weaker than the framing suggests. Prior work already studies geometry-aware PDE/CAE surrogates on complex geometries, and AeroJEPA is especially close in using predictive latent representations for 3D aerodynamics and design-oriented operations. D-JEPA mainly adds supervised design recovery and a geometry/decoder factorization, but it does not clearly show that this is a major conceptual step beyond nearby work.

The design-recoverability result is also somewhat circular. The model is trained directly with true design variables through LdisL_{dis}, and the encoder uses hand-crafted global/regional geometric descriptors. Therefore near-perfect design recovery is not surprising. Table 3 shows that removing LdisL_{dis} substantially hurts design R2R^2, which weakens the claim that the representation naturally preserves reusable design information.

The transfer and optimization evidence is limited. Condition reuse is only interpolation over two held-out Mach values. Cross-physics transfer is only CFD-to-structural transfer on AutoHood3D, and by 100–300 epochs the frozen-encoder model is close to training from scratch. CFD validation uses only four candidates on two benchmarks, without comparison to AeroJEPA or other surrogate-guided optimization baselines. Overall, the evidence supports a useful prototype, but not the broad claim of a general reusable representation for engineering design.

Questions for authors

Please refer to weaknesses

1 comment
Area_Chair_N5cGOpenReview ↗

Author-reviewer discussion period is ending soon

Dear Reviewer N15y,

If you haven't done so, please read through authors' response and let authors know if you need further clarification or have more questions. Beginning 11/19, authors are not allowed to respond any more, so this is good time to review the response and ask more.

Thank you.

-AC

Reviewer_D3icOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents D-JEPA, a representation learning framework attempting to unlock downstream generalization/fast adaptation of unseen operating condition (c) and physical quantity of interest (QoI). Within the D-JEPA framework, several components are added: 1. handcrafted geometry descriptors s_glob and s_reg 2. taking feature fusion of geometry and c out of part of the JEPA embedding (compared to AeroJEPA) and delegating the fusion directly into the physics decoder 3. FiLM-based fusion of (coordinate + c + surface normal) and D-JEPA geometry embedding g. The paper includes these notable evaluations: 1. cross geometry, same c (Tab 1). 2. same geometry, cross c (tab 4) and 3. same geometry, same c, new QoI adapatation (tab 5).

Critical strengths

strengths:

  1. the proposed geometry descriptors seem to be highly predictive and effective to aerodynamics simulation like the SHIFT family, as strong results of D-JEPA in Tab 1 and Tab 3 concretely validates.

  2. The integral D-JEPA framework achieves state-of-the-art performance for cross-geometry generalization with fixed c and QoI compared to recent baselines, as demonstrated in Tab 1.

Critical weaknesses

weaknesses:

Several major claims of the paper are not supported by the inherently confounded experiments.

  1. claim of cross c generalization: for Tab 4, c generalization/interpolation is tested, but there isn't any baseline such that a meaningful notion of relative performance of D-JEPA can be established on this simulation parameter interpolation scheme compared to alternative methods?

  2. claim of cross physics generalization ("cross-physics transfer" line 87): is missing from the main text, so either genuine cross-governing-equation simulation tranfer experiments need to be conducted, or the claim should be droped.

  3. claim of cross QoI fast adaptation: Tab 5 showed D-JEPA pretrained E_theta helps the model achieves higher performance ceiling and faster, but this is not surprising considering that the geometry itself can be highly predictive to the unseen quantity, so I'm expecting fast adaption.

Therefore, the meaningful question is, would D-JEPA presents better result compared to compatible pretraining frameworks like AeroJEPA in Tab 5's setup - same geometry, same c, different QoI - evaluated under fair settings?

  1. claim of strong representation quality with design param recovery: the param recovery is already part of the pretraining objective, it's more like a train/test split on the param prediction task, rather than a objective A (pretraining) -> objective B (param recovery) transfer. On top of this fundamental confound of experiment design for representation quality measurement, it's further unclear whether the param being probed / test-time 5-fold data have already appeared during training phase, which would further reduce the scope of the finding from D-JEPA learns good prediction of geometry -> design param to trivial finding that E_theta memorized the seen mapping to geometry -> param.

  2. Confounded effectiveness of each D-JEPA contribution to the success: as a JEPA-derived representation learning framework with many components as presented in paper summary, besides the missing concrete evidence of claims in above bullet, the superior performance in Tab 1 for D-JEPA seems to come from much narrower contribution than the whole D-JEPA design.

Referencing Tab 3, the learned-only variant has the worst result, which would make (D-JEPA minus handcrafted geometry feature) inferior than baselines in Tab 1. This validates a much narrow insight: s_glob and s_reg is very predictive for SHIFT simulations, rather than D-JEPA is a superior representation learning framework, such that it makes superior neural surrogate for SHIFT simulations.

  1. Unclarity on experiment fairness in terms of iso-parameter-budget: can authors clarify the parameter sizes for each baselines across experiments where different models are compared?

Overall, I found the paper to have weakly-supported arguments. In other words, the evaluation supports arguments that are much narrower than claimed scope of superior performance across different settings.

Questions for authors

Please refer to the weaknesses for under-supported claims, and additional experiments with careful design will help clarify the concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.