MIT and UBC unveil WorldParticle, one AI model for many physical materials

Overview

Updated 57 minutes ago

For decades, simulating a physical process has meant building a dedicated program: one solver for water, another for cloth, another for how a protein folds. On Nov. 9, researchers at MIT and the University of British Columbia showed a single neural network that handles all of them.

WorldParticle breaks matter into millions of tiny particles and predicts how they move one step at a time, an approach its authors compare to a language model predicting the next word. If it holds up, engineers could simulate interactions that today require separate programs, like oil sloshing against a flexible valve or a soft robot pressing into sand.

Why it matters

A single model for fluids, solids, and proteins could cut the cost and time of engineering design and molecular research.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

Physical phenomena modeled

The paper reports six dynamics categories, including cloth, elastic solids, fluids, granular materials, and molecular dynamics.

100×

Faster protein modeling

The team reports protein structure changes modeled about 100 times faster than numerical simulators.

Transformer architecture

A single model replaces separate hand-coded solvers for each phenomenon.

Millions

Particles per object

Objects are broken into up to millions of three-dimensional points.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Peter Yichen Chen

Assistant professor at the University of British Columbia; former MIT postdoc; co-corresponding author

Co-leading the WorldParticle project

Minghao Guo

MIT postdoc; co-corresponding and co-lead author

Co-leading the WorldParticle project

Wojciech Matusik

MIT professor of electrical engineering and computer science; leads the Computational Design and Fabrication Group at CSAIL

Senior author on the WorldParticle paper

Caoliwen Wang

Graduate student at the University of British Columbia; co-lead author

Co-lead author on WorldParticle

Organizations Involved

MIT Computer Science and Artificial Intelligence Laboratory

Research laboratory

Home institution of several WorldParticle authors

MIT's Computer Science and Artificial Intelligence Laboratory, where the Computational Design and Fabrication Group is based.

University of British Columbia

Research university

Co-leading institution on WorldParticle

A public research university in Vancouver, British Columbia.

ACM SIGGRAPH Asia

Conference

Venue for the WorldParticle paper presentation

The Asia edition of the ACM SIGGRAPH computer graphics conference.

Timeline

May 2026 November 2026

2 events Latest: Today

November 2026
MIT and UBC announce WorldParticle
Today Announcement

MIT News reports a single transformer model that simulates sand, fluids, cloth, and proteins.
Today
May 2026
WorldParticle preprint posted to arXiv
Research

The paper describing the unified particle simulator appears as an open-access preprint.
May 2026

Scenarios

WorldParticle anchors a new class of general physics simulators

Possible Resolves by Nov 9, 2028

Discussed by: The MIT and UBC research team, in MIT News and the paper

The team's stated goal is a general simulator that works out of the box. If other labs adopt the architecture and extend it across materials, it could become a shared base model for physics simulation, the way language models became a base for text tasks.

A commercial simulation vendor ships a WorldParticle-based product

Possible Resolves by Nov 9, 2029

Discussed by: Industry observers; no vendor has announced a product

Engineering software firms could integrate the architecture into released products. The trigger would be a proven speed or accuracy advantage on industrial workloads. None has been demonstrated publicly on commercial tasks.

Independent benchmarks expose limits, and numerical solvers stay standard

Possible Resolves by Nov 9, 2027

Discussed by: Simulation researchers who have not yet tested the model

The claims rest on the authors' benchmarks. If independent labs find the model breaks on new materials or forces, adoption could stall, and hand-coded solvers would remain the standard for safety-critical engineering.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

2020

AlphaFold2 (2020)

At the 2020 CASP14 assessment, DeepMind's AlphaFold2 predicted protein structures with accuracy rivaling experimental methods. It showed a neural network could learn biological structure from data rather than simulate it from first principles.

Then

AlphaFold2 was released in 2021 and quickly became a standard tool for biologists.

Now

Protein structure prediction shifted from a laborious experimental task to a computational one, and DeepMind later released predicted structures for nearly all known proteins.

Why this matters now

WorldParticle applies a similar learned-dynamics idea to protein motion over time, and its authors aim for a comparable shift in simulation.

November 2023

GraphCast (2023)

DeepMind's GraphCast, published in Science, outperformed the European Centre for Medium-Range Weather Forecasts' operational high-resolution model on most medium-range predictions. It ran on a single Google tensor processing unit in under a minute.

Then

Weather agencies began testing learned models alongside physics-based ones.

Now

The result established learned models as credible competitors to numerical simulation in a major operational field.

Why this matters now

GraphCast is a clear precedent for a learned model beating a hand-coded numerical simulator, the claim WorldParticle makes across materials.

2020

Graph Network Simulator (2020)

DeepMind researchers published the Graph Network Simulator, which learned particle dynamics for fluids and granular materials from example simulations. It could roll out predictions but struggled with long-term stability and new scenarios.

Then

It became a baseline for learned physics simulation.

Now

Later work, including WorldParticle, built on particle-based learning while trying to fix generalization and stability.

Why this matters now

WorldParticle compares against the Graph Network Simulator and similar models, claiming more stable long rolls and better generalization.