Advanced and frontier / Advanced methods

JEPA

Predict the latent of the flow from geometry and condition; decode a field only when you need one.

Joint-embedding predictive architectures (JEPA) learn by predicting an abstract latent representation of the answer, instead of the answer itself. AeroJEPA (Vinuesa and colleagues, 2026) brings this to aerodynamics: from the geometry and flight condition, predict the latent of the flow field; decode a field only when you need one.

• Context encoder: the geometry point cloud becomes a fixed set of tokens.
• Target encoder: during training only, the solver’s flow field is encoded into target tokens.
• Predictor: conditioned on angle of attack, Reynolds or Mach number, it maps geometry tokens to predicted flow tokens; the loss compares tokens, not cells.
• Anti-collapse regularisation (SIGReg) keeps the latent space informative, so it cannot cheat by mapping everything to the same point.
• An optional implicit neural decoder reads the field at any query point, so prediction cost is independent of mesh size: tested on 15M surface and 50M volume points (HiLiftAeroML).
• The latent space organises itself by physics: simple linear probes recover lift, drag and control-surface deflections from it, and moving through it is a way to explore designs.

Industrial example: The experienced aerodynamicist’s intuition

A senior engineer looks at a new flap setting and says “this will behave like the configuration we tested last spring, with a bit more lift”. They are not computing every pressure tap; they predict at the level of meaning, in a mental map where similar flows sit close together. AeroJEPA learns that map: designs with similar physics land near each other, and lift or drag can be read off it with a ruler.

References

• AeroJEPA: Semantic Latent Representations for Scalable 3D Aerodynamic Field Modeling (Giral, …, Vinuesa et al., 2026)
• Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA) (Assran et al., CVPR 2023)

Related

• Latent space: Compress millions of points into a few hundred learned slots that hold the physics.
• Scaling to millions of points: Attention grows with the square of the points; latent and linear schemes keep it tractable.
• Recent papers: Recent field-surrogate architectures from NeurIPS, ICML and ICLR, 2023 to 2026, in order.