MIT, UBC Researchers Build WorldParticle to Simulate 6 Physical Systems, Running Protein Models 100 Times Faster
Updated
Updated · MIT News · Nov 9
MIT, UBC Researchers Build WorldParticle to Simulate 6 Physical Systems, Running Protein Models 100 Times Faster
2 articles · Updated · MIT News · Nov 9
Summary
WorldParticle uses one deep-learning architecture to simulate materials from sand and fluids to proteins, replacing separate hand-coded solvers with a particle-based model that predicts motion one step at a time.
The system learns underlying dynamics instead of fixed governing equations, then adds external forces back into each step so it can generalize to unseen materials, environments and impacts.
To handle objects split into millions of particles, the researchers compressed particles during processing, making the framework scalable without losing fine-grained behavior.
In tests across 6 physical phenomena, WorldParticle beat numerical simulation methods on accuracy and modeled protein structural changes about 100 times faster, while better capturing effects such as collapsing granular piles.
The team plans to extend the model to tasks including soft-robot manipulation, RNA evolution and galaxy mergers, aiming for a general-purpose simulation pipeline for scientific discovery.
If WorldParticle simulates protein dynamics 100 times faster, what hidden physical anomalies might emerge when AI guesses the laws of nature?
Could a single AI architecture truly replace centuries of hand-coded physics equations to simulate everything from protein folding to galaxy mergers?
How does tokenizing the physical world into a neural network challenge our fundamental reliance on traditional mathematical solvers for scientific discovery?
WorldParticle: The 2026 Breakthrough Unifying Scientific Simulation with Transformer-Based Physics Foundation Models
Overview
WorldParticle, launched in May 2026 by MIT CSAIL, UBC, and partners, is a breakthrough equation-agnostic physics foundation model that unifies simulation across diverse physical domains. Its architecture uses progressive token merging to efficiently handle millions of particles, enabling fast and scalable simulations. By advancing particles with explicit prediction steps before learned corrections, WorldParticle achieves stable, long-horizon rollouts and mitigates error accumulation. The model is open-source, eliminating the need for per-phenomenon solver tuning and supporting direct integration into engineering and robotics tasks. However, it is limited to classical physics and cannot simulate quantum phenomena or guarantee strict boundary constraints, highlighting both its transformative potential and current limitations.