PDE-JEPA: Predictive Representation Learning of Latent Dynamics Modeling for Parametric PDEs
PDE-JEPA uses masked-latent predictive pretraining with geometry alignment for parametric PDE dynamics, beating SOTA by 33.4% in-distribution and 51.4% extrapolating.
PDE-JEPA replaces reconstruction-based objectives with predictive representation pretraining for parametric PDE dynamics modeling. It trains an encoder via masked-latent prediction, adds a geometry projector aligning latent trajectory geometry with physical field evolution, and a physics-structured latent predictor that decomposes dynamics into parameter-independent and parameter-dependent components. Across nine widely used PDE benchmarks it outperforms prior state-of-the-art by an average of 33.4% in-distribution and 51.4% when extrapolating to unseen governing parameters.
- Predictive pretraining replaces reconstruction objectives for parametric PDEs
- Geometry projector aligns latent trajectories with physical field evolution
- 33.4% average in-distribution gain across nine PDE benchmarks
- 51.4% improvement extrapolating to unseen governing parameters
From the original article
1 min read at huggingface.co
Physical trajectories contain more than snapshots of a system: they also reveal how its states evolve under governing conditions. However, representation learning for parametric partial differential equations (PDEs) has largely relied on reconstruction-based objectives that emphasize recovering observed physical fields.
By Zhentao Tan, Jianrong Zhang, Ruijie Quan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.