Most physics simulators are built one phenomenon at a time: a solver for fluids, another for sand, another for soft tissue, another for proteins. A team at MIT's CSAIL and the University of British Columbia has built one neural network that learns the underlying dynamics instead, and demonstrated it on six different physical regimes.
WorldParticle, to be presented at ACM SIGGRAPH Asia and described in an open-access paper, represents any material as thousands to millions of 3D particles — a Lagrangian representation — and predicts where each particle goes next. 'It is a similar architecture to ChatGPT,' says Peter Yichen Chen, a former MIT postdoc now at UBC and a co-corresponding author. 'But instead of predicting one word at a time to write a piece of text, we are predicting how these particles will move, one step at a time.'
The usual workflow requires deriving a governing equation for each material and converting it into something a numerical solver can handle — repeatedly, for every new substance. WorldParticle instead trains a network to solve governing equations in general. It also separates external forces before each prediction step and integrates them back afterwards, correcting its own output, which is how it generalises to materials, forces and environments it never saw in training.
Scale is the hard part, and the team's fix is unglamorous: merge and compress particles while preserving fine detail, so a million-particle object does not have to be evaluated a million times.
The demos: deformable objects, granular materials and proteins among six phenomena, with higher accuracy than numerical baselines. In one comparison, a collapsing granular pile stayed a pile, where baseline methods left sand grains floating in mid-air. For protein dynamics, the model ran about 100 times faster than numerical simulators, according to co-lead author Minghao Guo.
The pitch is a foundation model for simulation — 'We want to have a GPT moment for simulation,' Chen says — with applications in soft robotics, RNA evolution and astrophysics, including galaxy mergers. That ambition is also the honest limit: six phenomena on a research codebase is a proof of concept, not a general simulator, and the paper's comparisons are against classical solvers on selected tasks.
Still, the direction is notable. World models are usually discussed as a way to give robots and agents an intuition for the physical world. This is the same idea pointed at science: instead of hand-coding the physics, learn it once and reuse it across domains that currently require separate software, separate expertise and separate validation.
The work involved researchers from UBC, Georgia Tech, Inria, Meta, the University of Pennsylvania, the University of Utah, UCLA and the University of Hong Kong, alongside Wojciech Matusik's Computational Design and Fabrication Group at CSAIL.



