Researchers at MIT and the University of British Columbia have built WorldParticle, a single deep-learning model that simulates many kinds of physical behavior, MIT News reported on November 9. The system handles fluids, deformable solids, granular materials such as sand, and proteins, all without a separate hand-coded solver for each.
WorldParticle represents every material as small 3D particles and predicts how those particles move one time step at a time, learning the underlying dynamics from data rather than from hand-written equations. Because everything is a particle, it can also simulate interactions between materials, such as a fluid sloshing against a deformable solid.
According to MIT, the model handled six types of physical phenomena more accurately than conventional numerical simulation methods, and modeled how protein structures change over time about 100 times faster than numerical simulators. In a granular collapse test, it formed the sand pile correctly, while some baseline methods left particles floating in the air. MIT's article does not give per-task accuracy figures.
The work is led by Peter Yichen Chen, a former MIT postdoc who is now an assistant professor at UBC, with MIT postdoc Minghao Guo, UBC graduate students Caoliwen Wang and Siyuan Chen, and MIT CSAIL professor Wojciech Matusik, plus collaborators from Georgia Tech, Inria, Meta, the University of Pennsylvania, the University of Utah, UCLA and the University of Hong Kong. The paper, "WorldParticle: Unified World Simulation of Lagrangian Particle Dynamics via Transformer," was presented at SIGGRAPH Asia and is open access on arXiv.
"We want to have a GPT moment for simulation," Chen said. A general-purpose learned simulator could replace stacks of specialized solvers in graphics, robotics training and molecular modeling. MIT's announcement does not mention a public code or model release, so for now the paper is the main resource for anyone who wants to build on the approach.