Researchers at MIT CSAIL and Tsinghua University have won the Best Paper Award at ICLR 2026 for GeoPT, a pre-trained model that learns physics fundamentals from synthetic geometric interactions then simulates wind currents, hull stress, car crashes, and light transport with state-of-the-art accuracy using up to 60 percent less labeled data.
GeoPT is trained on 1.3 million samples of "synthetic dynamics": virtual spheres that move at varying speeds and angles until contacting a 3D object, where they stick. These particle-on-mesh interactions give the model a grasp of physical forces—pressure distribution, momentum transfer, structural deformation—before it ever trains on labeled simulation data. The approach sidesteps the prohibitive cost of generating high-fidelity physics labels from numerical solvers, joining a wider effort to make AI models follow fundamental physics laws rather than brute-forcing predictions from data alone.
In benchmarks spanning fluid mechanics for vehicles and solid mechanics in crash tests, GeoPT reached peak performance twice as fast as leading models. When simulating how a boat hull handled simultaneous air and wave forces, it needed 60 percent fewer labeled data points and converged to peak accuracy four times faster than top baselines.The system processed meshes exceeding 100 million points in seconds, which MIT postdoc Haixu Wu said could let engineers test vehicle blueprints without costly physical experiments. GeoPT's reliance on synthetic pre-training, learning physics from geometry before ever seeing a labeled task, parallels recent work in which AI learns to theorize about the world from raw observation, a paradigm shift away from supervised-only training. The result, its authors argue, is a concrete step toward a "physics foundation model."