surrogatemodels
← Papers
DataICML 2026

GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training

Wu et al.

What it does

Pre-trains a neural simulator on over a million off-the-shelf geometries, lifted with synthetic dynamics so the self-supervision is about how flow and deformation behave, not shape alone.

Why it matters

Reports 20–60% fewer labelled simulations and 2× faster convergence on car, aircraft, ship and crash benchmarks: a way around the cost of simulation data.

More papers

All papers