MIT CSAIL, Tsinghua Build GeoPT, Cutting Physics Training Data by 60%
Updated
Updated · MIT News · Aug 10
MIT CSAIL, Tsinghua Build GeoPT, Cutting Physics Training Data by 60%
1 articles · Updated · MIT News · Aug 10
Summary
GeoPT, a new pre-training method from MIT CSAIL and Tsinghua University, lets AI simulation models reach peak performance twice as fast while using up to 60% less labeled data than leading systems.
1.3 million synthetic-dynamics samples taught the model physics by simulating particles striking and stopping on 3D objects, giving it reusable intuition before task-specific training.
Benchmarks showed gains across industrial tasks, including wind and pressure on complex shapes, fighter-jet aerodynamics, boat hulls facing air and waves, vehicle crash deformation, and light transport.
100 million mesh-point simulations ran in seconds, suggesting engineers could test cars, planes, ships and robots with fewer physical experiments.
The team says GeoPT is an early step toward a broader physics foundation model that could later extend to materials, weather and realistic video generation.