Neural physics simulators need training data from expensive numerical simulations. A single crash simulation can take hours. The data bottleneck is not compute for training the neural network but compute for generating the physics labels it trains on.
Wu and colleagues (arXiv:2602.20399) observe that geometric models are abundant and cheap — CAD files, meshes, 3D scans. They lack dynamics: a car body mesh doesn't encode how the metal crumples on impact. But geometry constrains dynamics. The shape of the body determines where stress concentrates, how forces propagate, which regions deform first.
GeoPT augments static geometries with synthetic dynamics — not real physics, but plausible motion patterns generated from the geometry alone. The pre-training task is to learn geometric features that are relevant to dynamics without ever seeing real dynamics. This creates dynamics-aware representations from geometry-only supervision.
The approach reduces labeled physics data requirements by 20-60% and achieves 2× faster convergence. Trained on over one million geometry samples across automotive, aerospace, and crash simulation benchmarks, the pre-trained representations transfer to real physics tasks.
The bridge between geometry and physics is the constraint relationship: geometry determines the space of physically possible dynamics. Pre-training doesn't teach the network what happens — it teaches the network what could happen, narrowing the hypothesis space that the physics labels must resolve.
The general observation: when expensive labels are constrained by cheap structure, pre-training on the structure reduces the label requirement. The geometry doesn't replace the physics — it reduces the number of physics examples needed to learn the physics. Structure is a form of prior knowledge that can be extracted from data that doesn't contain the target phenomenon.