One AI model can simulate sand, water and folding proteins, a job that now takes separate programs, MIT researchers say
WorldParticle breaks objects into tiny dots and predicts how each one moves. The team says it could help design soft robots and medicines.
Researchers at MIT, the University of British Columbia and other schools have built one AI system that predicts how many kinds of matter move, from collapsing sand to sloshing liquid to folding proteins, according to an MIT News release published Nov. 9.
What it replaces
Today, a scientist who wants to simulate a physical process on a computer usually picks a program built for that one job, the release says. One handles flowing water. Another handles how a protein folds. Each is hand-coded around the equations for that single process.
How it works
The new system, called WorldParticle, breaks every object into thousands or millions of tiny 3D dots. A neural network, a type of AI that learns from examples, then predicts where each dot moves next, one step at a time. Peter Yichen Chen, a co-author and assistant professor at the University of British Columbia, compares it to ChatGPT, which predicts one word at a time. Because the system learns general rules of motion, the team says, it can handle materials and forces it never saw in training.
In tests on six kinds of physical behavior, the model was more accurate than standard simulation methods, the researchers said. It modeled protein shape changes about 100 times faster. It also showed sand collapsing into a pile, while some older methods left grains floating in midair.
Why it matters
The team says the tool could help engineers design soft robots that touch sand or soil, study how oil sloshing in a tank hits a valve, and develop medicines. The results are the researchers' own. The work will be presented at SIGGRAPH Asia, a computer graphics conference.
Source: MIT News release, Nov. 9, 2026.