D-JEPA separates geometry from physics for better design modeling
D-JEPA: Design-Recoverable JEPA Representation with Swappable Physics Decoders
Machine LearningArtificial Intelligence
Summary
Predicting how complex physical systems behave can be tricky because geometry and physical conditions often get mixed up in learned models. The authors introduce D-JEPA, an approach that focuses on representing geometry separately, so it can be reused across different physical situations. This lets the model better understand designs and even helps with optimizing them by keeping track of underlying design details. Their method works well on 3D problems like airflow and structural responses, improving accuracy and making design exploration easier.
What this means in practice
- •For aerospace engineers: Use a compact geometry representation to predict aerodynamic responses across different flight conditions with fewer computations.
- •For mechanical design teams: Perform design optimization of physical parts by reusing geometry information and swapping physics models without retraining the entire system.
- •For cfd software developers: Integrate design-recoverable representations to accelerate simulation-based design processes with validated surrogate models.$Commercial implications: Supports building software products for faster aerodynamic and structural design evaluations validated against high-fidelity simulations.
Authors
Nitin Nagesh Kulkarni, Aashwin Anand Mishra, Yin Yu, Peter Lyu
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 a practical representation for scientific surrogate modeling and design.