A single AI model simulates the behavior of many substances, from sliding sand to folding proteins

Researchers at MIT, the University of British Columbia and several other institutions have built WorldParticle, a deep-learning system intended to replace the separate, hand-coded solvers scientists typically use for different physical processes, MIT News reports. The team includes co-corresponding authors Peter Yichen Chen, a former MIT postdoc now an assistant professor at UBC, and MIT postdoc Minghao Guo, along with MIT professor Wojciech Matusik. The work is due to be presented at the ACM SIGGRAPH conference in Asia.

The system breaks any material, whether a sand dune or a protein, into thousands or millions of 3D particles. A neural network then takes their positions and physical properties and forecasts their next state, which Chen compares to the way ChatGPT predicts text one word at a time. External forces are separated out before each step and folded back in afterward, which the researchers say helps the model handle materials and conditions absent from its training. A merging technique compresses particles to keep the computation manageable.

According to the article, the team tested WorldParticle on six phenomena, including deformable objects, granular materials and proteins, and achieved higher accuracy than numerical simulation methods. It modeled protein structural changes about 100 times faster than numerical simulators, and it captured sand collapsing into a pile where some baseline methods left grains floating.

The approach could let engineers study interactions that normally require two simulators, such as oil sloshing against an elastic valve. Chen calls it a first step toward a foundation model for simulation, and the researchers hope to extend it to soft robotics, RNA evolution and galaxy mergers.