acceptodds
Under review as a conference paper at ICLR 2027

AeroJEPA: Learning Semantic Latent Representations for Scalable 3D Aerodynamic Field Modeling

Abstract

High-fidelity CFD is essential for aerodynamic design, but repeated simulations are computationally expensive, motivating surrogate models for rapid evaluation across geometries and operating conditions. Most existing surrogates are designed for direct field regression, requiring the evaluation of millions of field points even when only an aerodynamic quantity or a localized region is needed, while their internal representations are not intended for direct use in downstream tasks. We introduce AeroJEPA, a framework inspired by joint-embedding predictive architectures that represents the problem in two distinct latent spaces: context tokens encode geometry, while predicted tokens encode the aerodynamic state. Both representations remain directly accessible for downstream tasks, such as linear readouts of design variables and aerodynamic quantities without decoding and integrating the full field. When spatial detail is needed, a continuous implicit decoder evaluates the field only at the requested coordinates while reusing the encoded geometry. We evaluate AeroJEPA on HiLiftAeroML, with multi-million-point fields, and SuperWing, which spans a broad family of transonic wings. Compared with state-of-the-art direct-regression surrogates, AeroJEPA trades peak full-field accuracy for compact, reusable representations. In our selective-decoding experiment, however, AeroJEPA substantially outperforms the evaluated direct-regression surrogates while avoiding predictions over the remainder of the aircraft. The learned representations further support controlled interpolation, concept-vector arithmetic, and preliminary constrained latent-space optimization. These results show how predictive representations can support aerodynamic analysis with or without full-field reconstruction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Aerodynamic surrogate modeling, Predictive representation learning, Continuous field decoding

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_eE3oOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AeroJEPA, a machine‑learning substitute for 3‑D aerodynamic simulations built on the JEPA joint‑embedding prediction framework.

Standard CFD aerodynamic simulations demand huge computing resources. Most existing neural surrogate models output complete flow fields. Even if engineers only need local flow information or basic aerodynamic coefficients, these models still calculate data for millions of grid points. Besides, their internal features are hard to repurpose for downstream engineering work.

AeroJEPA builds two distinct latent token representations. One captures aircraft geometry; the other stores flow information for given flight conditions including angle of attack, Reynolds number and Mach number. During inference, the geometry embedding is calculated just once. Design parameters and aerodynamic coefficients can be retrieved straight from these latent outputs, with no need to reconstruct the whole flow field. If spatial flow data is required, a continuous implicit decoder computes values only at requested locations.

The model is evaluated using two datasets: HiLiftAeroML for high‑lift aircraft and SuperWing for transonic wings. It sacrifices some peak full‑flow‑field accuracy to gain flexible, reusable latent embeddings. In tests of selective local‑region decoding, AeroJEPA substantially beats baseline alternatives. Its latent space enables linear readout of values, smooth state interpolation, concept‑vector operations and initial constrained design optimization inside latent space.

One drawback is reduced full‑field accuracy on the SuperWing dataset versus top direct‑regression transformer models. This work handles only steady flow. Future development will target unsteady flows, multi‑fidelity datasets and full CFD checks for shapes optimized in latent space.

Critical strengths
  1. This paper borrows the JEPA embedding idea for 3D aerodynamic surrogate models. It splits internal representations for aircraft shape and flight‑driven airflow, plus a selective implicit decoder. Traditional models always generate the complete flow field. This model skips that heavy work for common engineering jobs like getting aerodynamic coefficients, and only computes local flow values when asked. It fixes a real‑world computing burden and brings useful new methods for aerodynamic machine learning.

  2. The authors test more than just full‑flow‑field prediction error on two difficult aerodynamic datasets (HiLiftAeroML and SuperWing). They check practical functions: local selective decoding, linear value extraction, concept‑vector operations and latent‑space constrained optimization. It notably outperforms transformer baselines on local wing prediction. Instead of focusing purely on reconstruction quality, these experiments prove the real‑world value of its learned representations.

Critical weaknesses
  1. The paper uses AeroJEPA’s latent space for gradient‑driven wing design. Optimized geometries are only evaluated via the frozen model’s internal predictions, with no high‑fidelity CFD checks. Model hallucinations may yield unrealistic aerodynamic results. Without real CFD validation, we cannot prove latent‑space optima deliver genuine aerodynamic gains, undermining this key use case of the framework.

  2. On the SuperWing test, existing transformer baselines outperform AeroJEPA on full‑field prediction. Authors treat lower accuracy as a deliberate trade‑off for good latent representations. Yet they barely explore architecture or loss tweaks to reduce this gap. AeroJEPA works best for latent‑focused pipelines, and falls short for scenarios requiring top‑level full‑field fidelity, limiting its general‑purpose surrogate value.

  3. The paper should compare their methods with Transolver++.

Questions for authors
  1. For latent‑space design optimization: you generate wing candidates only using latent‑space linear probes and a fixed predictor. Do these linear probes have built‑in biases or known failure cases that output non‑physical aerodynamic values for out‑of‑distribution latent samples? Without CFD simulation, what built‑in checks inside the AeroJEPA pipeline can tell you when an optimized latent candidate gives bad performance predictions?

  2. On the SuperWing test, AeroJEPA has higher full‑field reconstruction error than GeoTransolver / Transolver. The paper says this is a trade‑off between reconstruction fidelity and latent representation quality. If you tune loss weights or model size to shrink most of this full‑field error gap: how much will representation quality (linear readout for design and aerodynamic variables) get worse? Is there a practical setting that delivers good full‑field accuracy and still keeps strong latent representations? Or is this trade‑off unavoidable for this model architecture?

  3. In selective‑decoding experiments, AeroJEPA is much better than baselines when testing only a local spatial area. Why do direct‑regression transformer baselines do so badly on these local‑area queries? Does this weak performance come from the baseline models’ own design, or is it side‑effect from your local‑region evaluation and batching setup?

  4. SIGReg regularizer is important for stable latent representations. Your ablation shows larger λs\lambda_s hurts both reconstruction and probe results. Can this chosen SIGReg hyperparameter work on new unseen aerodynamic datasets? Or do you need heavy per‑dataset tuning of the SIGReg weight to get usable latent spaces?

Reviewer_ArF7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes AeroJEPA, which introduces the JEPA paradigm into three-dimensional aerodynamic surrogate modeling. The core idea is to avoid directly making dense predictions on high-resolution 3D flow fields. Instead, it uses a context encoder to represent geometric information and operating-condition variables as a fixed number of latent tokens. A predictor predicts the target latent representation of the flow field from the context latent. The flow field is reconstructed at any query coordinate through an implicit neural representation. This design decouples latent prediction from the spatial resolution of the final field, thus providing a new modeling approach for the computation and storage of large-scale 3D CFD data.

Critical strengths
  1. AeroJEPA uses a straightforward design. It represents the flow field with a fixed-size latent vector and uses an implicit decoder to predict values at any spatial location as needed. This approach reduces the computational and memory demands compared to predicting full high-resolution fields directly.
  2. The paper uses experiments like linear probing, latent interpolation, and latent-space design optimization to show that the latent representation holds useful geometric and aerodynamic information. The results go beyond just reconstructing the flow field and explore whether JEPA-style latent learning can create structured, meaningful representations.
Critical weaknesses
  1. The paper shows that AeroJEPA’s latent representation contains structured geometric and aerodynamic information, as demonstrated by latent probing, interpolation, concept-vector analysis, and comparison with VAE. However, these experiments mainly show that the latent space is meaningful, but do not clearly prove that this benefit comes from JEPA’s latent prediction or matching objective. The current experiments still do not fully separate the effect of latent predictive loss from other factors like the latent bottleneck, the INR decoder, and model capacity. The paper also discusses both coupled and decoupled training, but does not provide enough systematic evidence to explain why coupled training is necessary or to clarify the roles of the two-stage training approach in latent quality, field prediction, and physical consistency. As a result, the main claim that JEPA-style predictive latent learning is the key reason for AeroJEPA’s advantage is not yet fully supported.
  2. AeroJEPA avoids the baseline’s repeated chunk-wise prediction by using one-time geometry encoding and continuous implicit decoding in large-scale cases like HiLiftAeroML. This gives it clear advantages in TFLOPs, inference time, and model size. However, this benefit depends on the chunking or inference protocol. When SuperWing can do one-pass prediction, the computational cost of Transolver and GeoTransolver is similar to AeroJEPA, and some baselines even achieve higher field prediction accuracy. So, the current results mainly show AeroJEPA’s advantage in very large-scale scenarios where chunked prediction is needed, but do not fully prove its general efficiency or scalability. Also, the paper does not compare AeroJEPA with newer scalable methods made for industrial-scale use, like AB-UPT[1] and Transolver-3[2], so it is still unclear if AeroJEPA’s scalability advantage applies to broader high-resolution CFD situations.

References:

[1] Alkin, B., Bleeker, M., Kurle, R., Kronlachner, T., Sonnleitner, R., Dorfer, M., & Brandstetter, J. (2025). "AB-UPT: Scaling neural CFD surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers." TMLR 2025.

[2] Zhou, Hang, et al. "Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries." ICML2026.

Questions for authors
  1. The paper shows that AeroJEPA’s latent representation has a strong semantic structure. However, it is hard to tell whether this comes from JEPA-style latent prediction, a latent bottleneck, the INR decoder, or the model’s capacity. Could you directly compare the full AeroJEPA with a model that uses only reconstruction or regularization objectives, keeping the encoder, latent dimension, decoder, and model capacity the same, and report on their field prediction and latent representation performance?
  2. The paper talks about both coupled and decoupled training, but it is not clear what causes the differences between these two strategies. Does the claim that coupled training better preserves the physical validity of predicted latent variables have support from quantitative experiments? If decoupled training gives similar results for field accuracy, latent semantic structure, or physical consistency, why is coupled training needed? 3.The main efficiency advantage of AeroJEPA comes from separating geometry encoding from output resolution, while the baseline needs chunk-wise prediction for large-scale cases. But in the one-pass SuperWing setting, the baseline’s computational cost is already similar to AeroJEPA, and some baselines are more accurate. Can you clarify whether the scalability claim applies only to large-scale cases where the entire field cannot be predicted at once? If other methods can do full-field or memory-efficient inference, would AeroJEPA still have efficiency and accuracy advantages?
  3. The paper’s scalability results are mainly based on FigConvUNet, GeoTransolver, and Transolver. Since newer methods now use memory-efficient or full-mesh inference and physical-state caching for large-scale tasks, how can we show that AeroJEPA’s latent-field decoupling offers real advantages over these scalable alternatives? If we compare actual inference time and peak memory for different numbers of output points, will AeroJEPA’s advantage keep growing as field resolution increases?
Reviewer_Ky2tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes AeroJEPA, which use joint embeddings to predict fluid field values. One branch encodes the input geometry of the airfoil surface to obtain context latent embeddings, while the other branch encodes the fluid states on the surface to obtain target latent embeddings.

The context latent embeddings are combined with operating variables related to the environment to predict latent embeddings that match those obtained from the fluid field. At the same time, the predicted latent embeddings are then passed to a decoder to predict the fluid field values with the per-point coordinates as the inputs.

Critical strengths

The paper presents clear computational experiments demonstrating the effectiveness of the sampling-based method. The proposed framework achieves state-of-the-art inference efficiency, requiring only 43 seconds for inference. Further improvements in training efficiency would make the method even more attractive.

The proposed idea is reasonable and interesting. It introduces a two-branch framework for learning latent representations, providing a promising pipeline for fluid field modeling. The method also demonstrates better predictive performance than FIGConvUNet, which has comparable inference efficiency.

Critical weaknesses

Based on Table 2, there is still a performance gap between the proposed method and GeoTransolver. However, GeoTransolver has approximately three times as many parameters as the proposed method. Would the authors mind discussing whether increasing the number of parameters could further improve the performance and make it comparable to GeoTransolver? What will the computational comparsion at this time?

Questions for authors

N/A.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.