MIT's WorldParticle: One AI Model Simulates Sand, Cloth, Fluids and Folding Proteins

Researchers at MIT, the University of British Columbia (UBC) and partner institutions have built a single AI system that can simulate six kinds of physical behaviour — from sliding sand to folding proteins — replacing a patchwork of specialist simulators with one unified architecture. The system, called WorldParticle, was announced by MIT News on November 9, 2026, in a paper published in ACM Transactions on Graphics and set to be presented at SIGGRAPH Asia.

Scientific visualization of colorful sand particles cascading and forming a pile

One Model for Six Kinds of Physics

Direct answer: WorldParticle is a single transformer-based system that simulates six kinds of physical behaviour — cloth, elastic solids, Newtonian and non-Newtonian fluids, granular materials and molecular dynamics — using one shared architecture. It represents every material as tiny particles and predicts where those particles will move, one step at a time. In tests it beat existing neural simulators on accuracy across all six categories, and modelled protein motion about 100 times faster than numerical simulators.

Until now, scientists and engineers needed a different simulator for almost every material. A simulator for flowing water is not suited to modelling a deformable solid or the movement of granular material like sand, so teams stitched together separate tools — each with its own equations, assumptions and quirks. WorldParticle replaces that patchwork with a single deep-learning architecture that learns the underlying dynamics instead of being programmed with a separate equation for each case.

The trick is the representation. WorldParticle breaks an object into thousands or millions of tiny three-dimensional dots, or particles, using what is known as a Lagrangian representation. Whether the object is a dune of sand or a protein molecule, the system treats it as a cloud of particles and predicts how each one moves next. Because it learns patterns in how particles move rather than relying on hand-coded rules, it can adapt to materials, forces and environments it never encountered during training.

What WorldParticle Can Do

The research team demonstrated the system across six categories of particle dynamics, comparing it against state-of-the-art neural simulators. In one test, WorldParticle effectively captured how a granular object collapses into a pile, while some baseline methods left sand particles floating in the air. Its method also modelled how protein structures change over time about 100 times faster than numerical simulators, according to co-lead author Minghao Guo, an MIT postdoc.

The practical uses are wide-ranging. Scientists could use the system to model processes that typically require separate simulators, such as a fluid sloshing against a deformable solid. MIT gives the example of sloshing oil impacting an elastic valve inside a storage tank — a calculation that currently needs two different kinds of simulation. It could also help engineers study where and how granular materials like sand or soil move when a soft robot touches them, a question that matters for robotics, construction and planetary exploration.

"We want to have a GPT moment for simulation," says Peter Yichen Chen, a former MIT postdoc who is now an assistant professor at UBC and a co-corresponding author of the open-access paper. The ambition is clear: a general-purpose simulator, the way large language models became general-purpose text tools. Explore related science and tech offers here while this research develops.

3D visualization of a folding protein molecule with glowing ribbon structure

Who Built WorldParticle

The work is led by Peter Yichen Chen, now an assistant professor at UBC, and co-corresponding and co-lead author Minghao Guo, an MIT postdoc. Co-lead authors Caoliwen Wang and Siyuan Chen are UBC graduate students. The team also includes Wojciech Matusik, an MIT professor of electrical engineering and computer science who leads the Computational Design and Fabrication Group within MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), and collaborators at UBC and other institutions.

The paper, "WorldParticle: Unified World Simulation of Lagrangian Particle Dynamics via Transformer," is published in ACM Transactions on Graphics, and the research will be presented at the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia.

An Important Caveat: One Architecture, Not Yet One Model

Readers should note a nuance in the results. The same transformer architecture is used across all six dynamics categories, and it generalizes to unseen materials, boundary configurations, initial conditions and external forces. But for each category, a separate model is trained on data generated by established solvers — Newtonian fluids from DFSPH, cloth from GIPC, granular sand and non-Newtonian fluids from MPM, elastic shapes from VBD, and protein molecular dynamics from OpenMM. A single shared model trained once for everything is still future work.

The model also trades some per-phenomenon accuracy against specialist solvers in places. That is the honest state of the research: a unified architecture that already beats existing neural simulators across six categories and runs protein simulations about 100 times faster than numerical methods — but a research system that will need further testing before it can support practical engineering decisions. Keep following AI research breakthroughs in our Tech & AI coverage on TrendWire Daily.

Frequently Asked Questions

What is WorldParticle?

WorldParticle is a single transformer-based AI system built by researchers at MIT, the University of British Columbia and partner institutions. It simulates six kinds of physical behaviour — cloth, elastic solids, Newtonian and non-Newtonian fluids, granular materials and molecular dynamics — using one shared architecture.

How does WorldParticle work?

It represents every material as tiny particles, using a Lagrangian representation, and predicts where those particles will move one step at a time. It learns patterns in how particles move rather than relying on a separate hand-coded simulator for each material.

How fast is WorldParticle?

According to MIT News, the system modelled how protein structures change over time about 100 times faster than numerical simulators. It also outperformed existing neural simulators on accuracy across all six categories tested.

What can WorldParticle be used for?

Engineers could model processes that need separate simulators today, such as sloshing oil hitting an elastic valve in a storage tank, or study how sand or soil moves when a soft robot touches it. It could also help with drug research by modelling protein motion.

Who created WorldParticle?

The work is led by Peter Yichen Chen, a former MIT postdoc now an assistant professor at UBC, and Minghao Guo, an MIT postdoc, with co-lead authors Caoliwen Wang and Siyuan Chen of UBC, MIT professor Wojciech Matusik of CSAIL, and other collaborators.

Where was the research published?

The paper, "WorldParticle: Unified World Simulation of Lagrangian Particle Dynamics via Transformer," is published in ACM Transactions on Graphics and will be presented at the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia. MIT News announced the work on November 9, 2026.

Search This Blog

TrendWire Daily