US: MIT team develops smart model to simulate sand, fluids and protein movement

Researchers at MIT and the University of British Columbia (UBC) have built one transformer-based simulator design that predicts how particles move in six types of physical systems, including sand and proteins, according to a paper and an MIT release.

Physics simulators are usually built for one job. A solver for flowing water relies on different equations than one for folding proteins, and each is coded by hand, the release said. The new system, called WorldParticle, aims to replace those separate programs with one learned design, though each material type still gets its own trained network, according to the paper.

WorldParticle treats every material as a cloud of particles, each tracked by position and velocity, and predicts their next positions one step at a time. Peter Yichen Chen, an assistant professor at UBC and co-corresponding author, compared the approach to ChatGPT, which predicts text one word at a time.

How the model works

The team’s arXiv preprint splits each step in two. A basic calculation first moves the particles under known outside forces, such as a push. A learned corrector then adjusts positions and velocities to account for how particles affect one another. To keep the computing cost down, the model merges similar particles into a smaller set of summary tokens before passing information back to every particle, the paper said.

The researchers trained a separate network for each of six categories: cloth, elastic solids, two kinds of fluids, sand and protein motion. Training data came from established solvers. Against three other neural simulators on fluid, cloth and sand tasks, the paper reports lower errors and steadier long runs. Those rivals produced floating sand grains and exploding cloth over long rollouts, according to the paper.

Protein runs took 50-femtosecond steps, 100 times longer than the 0.5-femtosecond step of conventional molecular simulation, the paper said. A femtosecond is one quadrillionth of a second.

Uses and open limits

Because the first step handles outside forces directly, the model can respond to pushes it never saw in training, the paper said. The team built a real-time tool where a user applies forces to a soft object. On three real-world manipulation videos, predicted motion matched observed motion qualitatively, and the paper reports no error figures.

Several limits apply. The paper names memory as the main one: the first layer looks at every particle at once, which caps scenes at about 50,000 particles on a single 80-gigabyte graphics processor. The authors also say one set of weights shared across all categories remains future work, and that the model gives up some accuracy compared with specialized solvers.

“We want to have a GPT moment for simulation,” Chen said, describing a hoped-for foundation model that works for any scenario. The team plans to adapt WordParticle to more physical systems, including soft robot tasks, and build toward a general simulation pipeline for scientific discovery, MIT said.

MIT announced the work Nov. 9 and said the team will present it at ACM SIGGRAPH Asia.