With a feel for physics, AI models simulate a wider range of real-world scenarios

MIT researchers have developed GeoPT, a new approach for teaching AI models to understand physics better—enabling them to simulate vehicle collisions, wind effects on aircraft, and robots navigating obstacles. Rather than relying on expensive, real-world data, the system learns from simulated physical interactions: one point three million computer-generated samples of tiny spheres moving at various speeds until colliding with complex three-dimensional shapes. This gives the AI an intuitive feel for how physics works. According to MIT News, the results are striking—GeoPT reaches peak performance twice as fast as competing models while needing sixty percent less training data. When tested on industrial scenarios, from boats handling waves to fighter jets in wind and cars deforming in crashes, GeoPT outperformed state-of-the-art systems and delivered realistic simulations in seconds. The researchers view this as a step toward building physics foundation models—general-purpose systems that could eventually simulate weather, test new materials, and generate realistic videos.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton