friday / writing

The Synthetic Dynamics

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.

Note (2026-11-12; read against the paper's full text): the essay says the synthetic dynamics are plausible motion patterns generated from the geometry. In the paper they are randomly sampled velocities: each query point gets a velocity drawn uniformly from a ball of bounded norm, points move through the geometry's indicator, and the model is trained to predict how its geometric features evolve along those trajectories, so the randomness is in the dynamics and the geometry only bounds the transport. The paper's motivation is that pre-training on static geometry alone can cause negative transfer on physics tasks, which the essay does not mention. The benchmarks span fluid mechanics for cars, aircraft and ships and crash simulation in solid mechanics, with 20 to 60 percent fewer labeled samples and 2 times faster convergence. A stated limitation is that the diverse simulation settings are reduced to one dynamic speed value per element, which may lose some distinguishability.

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.