MIT and UBC's WorldParticle uses one transformer to simulate sand, fluids, cloth and proteins

Today a scientist who wants to simulate water picks a fluid solver, and one who wants to model a protein picks a molecular one. Each is hand-coded around its own equations[1]. Coupling two regimes, such as oil sloshing against an elastic valve in a tank, usually means stitching separate simulators together.

WorldParticle replaces that patchwork with one transformer. It breaks every object into thousands or millions of 3D particles, then predicts where they will be in the next step, much as a language model predicts the next word[1]. "We want to have a GPT moment for simulation," said Peter Yichen Chen, a former MIT postdoc now an assistant professor at UBC and co-corresponding author.

The paper describes a prediction and correction design[2]. An explicit predictor first moves particles under the known external forces. A learned corrector then fixes the interactions between particles in three stages: a tokenizer that encodes local particle, boundary and topology interactions, an encoder that repeatedly merges tokens into a compact set of "super tokens", and a decoder that lifts those back to full resolution to output per-particle position and velocity corrections.

Halving the token count at each encoder level is what keeps millions of particles tractable. "We needed to do a lot of detailed engineering and employ many tweaks to the architecture for it to be truly generalizable and scalable," said co-lead author Minghao Guo, an MIT postdoc[1].

The team tested the same architecture on six categories of dynamics: cloth, elastic solids, Newtonian and non-Newtonian fluids, granular materials and molecular dynamics[3]. It generalised to unseen materials, boundaries, starting conditions and forces, and beat numerical simulation methods on accuracy[1]. In one test a sand sculpture collapsed into a realistic pile, where some baseline methods left grains floating in the air. Protein structure changes were modelled about 100 times faster than numerical simulators.

The paper also shows interactive control, inverse design and learning from real-world robot manipulation data[2]. The researchers see uses in soft robots working in sand or soil, RNA evolution and even merging galaxies.

World models have mostly been about video and robot actions. WorldParticle applies the same idea to physics itself: learn the rules once and reuse them across materials. If it holds up beyond the six benchmark categories, it could cut the solver engineering that slows simulation-heavy work in robotics, materials and drug discovery. The authors call it a first step towards a general-purpose simulation pipeline for scientific discovery[1].