GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training

Neural simulators promise efficient surrogates for physics simulation, but scaling them is bottlenecked by the prohibitive cost of generating high-fidelity training data. Pre-training on abundant off-the-shelf geometries offers a natural alternative, yet faces a fundamental gap: supervision on static geometry alone ignores dynamics and can lead to negative transfer on physics tasks. We present GeoPT, a unified pre-trained model for general physics simulation based on lifted geometric pre-training. The core idea is to augment geometry with synthetic dynamics, enabling dynamics-aware self-supervision without physics labels. Pre-trained on over one million samples, GeoPT consistently improves industrial-fidelity benchmarks spanning fluid mechanics for cars, aircraft, and ships, and solid mechanics in crash simulation, reducing labeled data requirements by 20-60% and accelerating convergence by 2$\times$. These results show that lifting with synthetic dynamics bridges the geometry-physics gap, unlocking a scalable path for neural simulation and potentially beyond. Code is available at https://github.com/Physics-Scaling/GeoPT.

Physics simulation is a foundation of science and engineering, but it is often very expensive to compute. Neural simulators can make predictions much faster, but building foundation models for physics usually requires large amounts of costly simulation data. This paper focuses on this missing piece in constructing a physics foundation model. We introduce GeoPT, a pre-trained model that learns from abundant 3D geometry data instead. Since static shapes alone do not contain physical dynamics, GeoPT adds synthetic dynamics to geometry during pre-training, forming a new lifted pre-training paradigm. This helps the model learn patterns that are useful for downstream physics prediction. On industrial-scale tasks, including fluid simulation and crash simulation, GeoPT improves accuracy, reduces the need for labeled simulation data by up to 60%, and speeds up training. This suggests a scalable path toward a general-purpose physics foundation model.