MIT and Tsinghua researchers develop GeoPT AI for physics simulations

Researchers from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and Tsinghua University have introduced GeoPT, a new pre-training method designed to provide AI models with physical intuition. By utilizing a ‘Synthetic Dynamics’ training mechanism with 1.3 million virtual samples, the model learns basic physical laws—such as how particles interact with 3D structures—before being exposed to labeled data.

This approach significantly improves the efficiency and accuracy of 3D physics simulations. Key performance metrics include doubling training efficiency and reducing the required labeled data by up to 60%. In complex environmental tests, such as simulating a ship facing simultaneous wind resistance and wave impact, the model achieved peak accuracy four times faster than existing methods while using 60% less data.

GeoPT can complete high-fidelity simulations of over 100 million grid points in seconds. This capability allows engineers to quickly assess force distributions on 3D models of aircraft, ships, or vehicles, facilitating design evaluations without expensive physical experiments. The research team views physics as a ‘third modality’ for AI, following text and pixels, and aims to expand the technology to weather forecasting and new material testing.