MIT and UBC unveil WorldParticle: A single AI model simulates sand flow and protein folding

Researchers from MIT, UBC, and other institutions have introduced WorldParticle, a unified deep learning system that uses the same architecture to simulate diverse physical phenomena ranging from flowing sand grains to protein folding. The system represents materials as large numbers of particles, with neural networks progressively predicting particle positions and properties, eliminating the need to manually write equations or build solvers for each phenomenon. The team demonstrated six categories of phenomena including deformable objects, granular materials, and proteins, achieving accuracy superior to numerical simulation methods.