Researchers use one AI architecture to simulate sand, fluids and proteins

Sliding sand and shifting proteins usually call for different simulation tools. Researchers at MIT, the University of British Columbia and other institutions have developed WorldParticle, an AI architecture designed to handle both. In a November 9 account, MIT described tests across six physical phenomena—a step toward studying different kinds of matter within one simulation framework.

The work addresses a practical divide in scientific computing. A tool built for flowing water is not necessarily the tool used to study a folding protein. Conventional simulators are hand-coded around equations for particular processes. WorldParticle instead learns physical behavior and uses the same architecture to predict how different materials move.

Different materials, the same moving dots

WorldParticle starts by representing an object as thousands or millions of three-dimensional particles. This approach, called a Lagrangian representation, follows those particles as they move through space and time. Their positions, velocities and physical properties provide the information needed to describe the material’s changing state.

A neural network takes the particles’ positions and properties as inputs, then predicts their positions and properties in the next simulation frame. Repeating that prediction produces motion over time. Peter Yichen Chen, a co-corresponding author now at UBC, compares the architecture to ChatGPT: the next prediction concerns moving particles rather than words.

External forces get separate treatment. Before each simulation step, WorldParticle separates out forces acting on the system, such as an impact when a wave strikes a seawall. It then adds their effects back after its initial particle prediction, correcting the predicted behavior. The researchers say this helps it handle unfamiliar materials, environments and forces.

Scale creates another problem: millions of particles can make the calculation too demanding. The team uses a merging technique to combine and compress particles while retaining fine details of their behavior. MIT postdoc and co-lead author Minghao Guo says making the architecture general and scalable required extensive engineering and architectural adjustments.

The tests span collapsing grains and changing proteins

The researchers demonstrated the framework on six physical phenomena, including deformable objects, granular materials and proteins. They report higher accuracy than numerical simulation methods in those demonstrations. In one granular-material example, WorldParticle captured an object collapsing into a pile, while some comparison methods left sand particles suspended in the air.

For proteins, Guo reports that WorldParticle modeled structural changes about 100 times faster than numerical simulators. That speed comparison concerns protein motion; it is not a reported 100-fold gain across every tested phenomenon. The accuracy and speed findings are the research team’s results, described in MIT’s account.

Our vision is to have a foundation model that works out of the box for any scenario you want to simulate. This is a first step in that direction.

Peter Yichen Chen, WorldParticle co-corresponding author, speaking to MIT News

Where different physics meet

One motivation is to study interactions that cross the boundaries between specialized tools. The researchers describe potential uses such as oil sloshing against an elastic valve, or sand and soil moving when touched by a soft robot. A unified framework could let engineers explore the fluid and deformable-solid behavior together.

Those applications remain prospective. The team wants to adapt WorldParticle to additional systems, including soft-robot manipulation, RNA evolution and merging galaxies, and ultimately build a general-purpose simulation pipeline for scientific discovery. The research is scheduled for presentation at ACM SIGGRAPH Asia. Its demonstrated scope is six phenomena, not every physical scenario.