Newton's Laws vs. Pixels: AI Taught to 'Feel' Physics

Engineers from MIT's Computer Science and Artificial Intelligence Laboratory and Tsinghua University have presented GeoPT, an architecture for a universal physics simulator. Unlike generative models that handle text and images well, GeoPT tackles mechanical and aerodynamic processes. The system was pre-trained on a dataset of 1.3 million scenarios featuring millions of virtual microparticles moving at different angles and colliding with 3D shapes. This helps the algorithm grasp physics on an intuitive level before working with specialized industrial blueprints. Users can upload a 3D model and specify force direction to get a thermal map of deformations, airflows, or light distribution in seconds. In industrial tests on complex geometries with 100 million node points, GeoPT reached peak accuracy four times faster than counterparts when calculating ship stability and fighter jet airflow, requiring 60% less human-annotated data. The developers describe physics as a third key modality for AI and plan to scale the system for weather anomaly prediction and realistic physical video generation.