AeroJEPA: Learning Semantic Latent Representations for Scalable 3D Aerodynamic Field Modeling
Francisco Giral, Abhijeet Vishwasrao, Andrea Arroyo Ramo, Mahmoud Golestanian, Federica Tonti, Adrian Lozano-Duran, Steven L. Brunton, Sergio Hoyas, Hector Gomez, Soledad Le Clainche, Ricardo Vinuesa
cs.LG
2026-05-07
AeroJEPA predicts flow latents from geometry, hitting 0.0048 pressure Rel-L2 on HiLift at 57 TFLOPs, with CL/CD probes at R² 0.930/0.996.
Aerodynamic design is a many-query job: the same family of shapes and operating points has to be evaluated as high-dimensional fields, over and over. High-fidelity CFD remains the reference, and it is too expensive for optimization and large screens. Neural surrogates already predict fields. Two problems remain. Realistic 3D fields are huge. HiLiftAeroML stores about 15 million surface points and 50 million volume points in the boundary layer; pointwise models trained to emit a fixed discretization cannot hold that output in one shot. Even when the field is accurate, the latent is rarely useful for design. It fits the field. It does not organize geometry, operating condition, or lift and drag.
AeroJEPA puts a Joint-Embedding Predictive Architecture on this job. The network does not regress the full field from geometry. It predicts a target latent of the flow from a context latent of geometry plus operating conditions, then optionally decodes through a continuous implicit network at arbitrary query points. Prediction cost is decoupled from field resolution, and the latent is pushed to become something that can be read and searched.
Geometry is stripped to a point cloud and subsampled with farthest-point sampling: about 131,000 points on HiLift, 8,192 on SuperWing. A context encoder turns that cloud into a fixed token set (3,072×64 on HiLift, 512×128 on SuperWing) by learning centroids, running neighborhood message passing, then a point transformer. During training a target encoder does the same to an independently sampled flow cloud. A predictor maps context tokens plus the condition vector (angle of attack, Reynolds number, Mach) to predicted flow tokens. The target encoder is dropped at inference.
The decoder is an implicit neural representation: predicted tokens and a query coordinate (plus signed distance in volumetric settings) produce velocity and pressure at that point. Reconstruction queries are sampled independently of the encoder points, which is a masked-supervision setup. The loss has three terms: latent L2 matching, reconstruction, and SIGReg. SIGReg projects latents onto random low-dimensional directions and pulls those projections toward an isotropic Gaussian, replacing the EMA teacher and stop-gradient tricks common in JEPA. Weights are 1.0 / 1.0 / 0.01. All reported runs train encoder, predictor, and decoder together.
HiLiftAeroML is the stress test: 205 training cases, 50 held-out, ten angles of attack from 4° to 22° per geometry. Baselines train on chunks of 131,000 points and stitch the field at test time. AeroJEPA encodes geometry once and decodes continuously.
| Model | Pressure Rel L2 | Velocity u Rel L2 | Inference TFLOPs |
| AeroJEPA | 0.0048 | 0.1445 | 57.00 |
| FigConvUNet | 0.0206 | 0.4761 | 88.34 |
| Transolver | 0.0288 | 0.6328 | 189.44 |
| GeoTransolver | 0.0289 | 0.6510 | 309.13 |
On unseen geometry LHC013 at 18° angle of attack, velocity-magnitude error is 1.18% in MAE over the max ground-truth mean. Linear probes recover flap and slat deflections from the context latent at R² 0.965 to 0.988, and recover CL and CD from the predicted latent at R² 0.930 and 0.996, even though those integrated coefficients were never training targets.
SuperWing has 4,239 parameterized wings, 28,856 RANS solutions, and about 32,000 surface points per field. In the chunked regime AeroJEPA beats the baselines across skin friction and pressure. When baselines may consume the full field in one pass, GeoTransolver reaches pressure Rel L2 0.0309 against AeroJEPA's 0.0644: at moderate resolution, pointwise models still win on raw field error. On predicted latents, CL R² is 0.984 and CD R² is 0.965. Nine of 54 design parameters are stably readable from the context latent (cross-validated R² ≥ 0.85). A constrained search in the 128-dimensional context latent, using differentiable probes for lift-to-drag, lands on the high-efficiency frontier of the dataset envelope and retrieves a nearby real wing. That is a proof of concept, not a design pipeline.
For 3D aerodynamic surrogates, this puts "the field does not fit in one forward pass" and "the latent is not a design object" in one architecture. On HiLift, error and compute fall together, which says the bottleneck was chunking, not model capacity. A latent that linearly recovers unlabeled geometry and forces can be probed, interpolated, and searched under constraints without remeshing at every inner-loop step.
SuperWing draws the other boundary. When the field still fits in one pass, Transolver-style pointwise models can beat AeroJEPA on field error. The claim worth keeping is resolution-independent decoding plus a searchable latent, not a win in every cell of every table.
The authors restrict the claim: two datasets, mostly steady flow, limited geometries and conditions. HiLift baselines are forced to chunk, so the comparison is not perfectly matched, though chunking is exactly why those models cannot take 15 million points. SuperWing latent optimization uses a trust region and a dataset-envelope ceiling; the retrieved wing belongs to a high-efficiency family already in the corpus. Strong linear probes do not mean the latent stores nonlinear physics. Wrong surrogate predictions outside the calibrated regime will bias design. The paper positions AeroJEPA as decision support, not a CFD replacement.