Physics Simulation Gets AI Upgrade

⊲ MIT News ⊳ Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory and Tsinghua University have developed GeoPT, a pre-training method that teaches simulation models the basics of physics by reenacting everyday mechanical interactions in 3D. Trained on 1.3 million samples of particles moving across complex shapes, the system reaches peak accuracy up to four times faster and needs up to 60 percent less labelled data than leading models. It has already outperformed rivals on tasks including wind loading, fighter jet aerodynamics, boat hulls and car crash deformation.
⊲ Meta researcher Fei Sha’s endorsement suggests GeoPT could accelerate an industry shift toward physics foundation models built on synthetic dynamics. ⊳
⊲ Image – arxiv ⊳
www.simulation.news