MIT and Tsinghua introduce GeoPT: physics as AI's third modality

MIT News AI · rss · 2026-08-11

Researchers from MIT CSAIL and Tsinghua University introduced GeoPT, a new pre-training approach that gives AI simulation models a fundamental understanding of physics, enabling them to accurately simulate real-world scenarios like wind tunnel tests and collision deformations.

Core Mechanism: Synthetic Dynamics

Traditional methods rely on time-consuming numerical solvers to calculate physical properties, making large-scale data collection infeasible. GeoPT learns physics by studying 1.3 million samples of synthetic dynamics. It simulates tiny particles moving at various speeds and angles until they hit and "stick" to a 3D shape. This gives the model a tangible feel for physical interactions before training on labeled data.

Performance and Efficiency Breakthroughs

Data Efficiency: In tests simulating boat hulls handling wind and waves, GeoPT required 60% less labeled data than state-of-the-art baselines to reach peak accuracy.
Speed: It reached peak performance up to 4 times faster than top baselines.
Industrial-Grade Fidelity: The system can handle over 100 million mesh points, generating high-fidelity physical simulations in seconds.

The team proposes that physics is the third modality for AI models, after text and pixels. This research is considered a crucial step toward building a general-purpose physics foundation model.

Original post →

MIT and Tsinghua introduce GeoPT: physics as AI's third modality

More from Embodied