WorldParticle: one AI model simulates sand, cloth, fluids and folding proteins

Today, a scientist who wants to simulate a new physical process starts by writing a simulator for it. Sand gets one solver, cloth gets another, a folding protein gets a third. MIT and the University of British Columbia think that patchwork is about to collapse into a single model. Their system, WorldParticle, announced on Monday, November 9, 2026, simulates six kinds of physical behavior with one transformer architecture — and, the team says, beats the numerical methods scientists use today on accuracy.

WorldParticle: one architecture, six kinds of matter#

Everything WorldParticle simulates becomes the same thing: thousands or millions of tiny three-dimensional dots. A dune of sand, a sheet of cloth, a drop of oil, a protein molecule — all reduced to particles, in what physicists call a Lagrangian representation. A neural network takes each particle’s 3D position and physical properties and predicts where it will be in the next frame, one second at a time.

The trick is in how it handles forces. Before each step, WorldParticle separates out the external forces acting on the system — an ocean wave slamming into a seawall, say — predicts the particles’ initial states, then folds the forces back in and corrects its predictions. That prediction-correction split is what lets it generalize to materials, forces and environments it never saw in training. “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 demo that lands hardest is the sand. WorldParticle captured a granular pile collapsing naturally, while some baseline methods left sand particles floating in the air. The system also modeled protein structures changing over time about 100 times faster than numerical simulators. The researchers draw their inspiration from world models — AI systems that learn the rules of the physical world to predict how it changes — and applied the idea to simulation itself. For the bigger picture on the field, see our explainer on world models.

The numerical solvers: precise, hand-built, slow#

The incumbent is not a single program but a craft. For each material, a scientist writes down the governing equation that defines how it behaves, converts that equation into a formula a numerical solver can handle, and repeats the whole process for the next material. The results are accurate and trusted — they are the reference WorldParticle is measured against — but the work is bespoke, and the compute is heavy. Protein dynamics, in particular, are expensive: simulating how a protein folds can take a numerical solver far longer than the phenomenon itself.

That is the gap WorldParticle is priced against. A learned simulator does not need the equation converted by hand; it learns the dynamics from data. And once learned, it runs fast — about 100 times faster on proteins, per the team. The trade-off is the usual one for learned models: you are trusting a network’s approximation instead of a solved equation, which is why the accuracy claim matters so much — and why it is framed as beating the numerical methods, not merely matching them.

The earlier neural simulators: fast, but they glitch#

WorldParticle is not the first neural simulator. Earlier systems — Graph Network-based Simulators (GNS), DeepLagrangian and Neural Operator models — already learned physics from data, and they were fast. Their problem was stability. In the WorldParticle paper’s comparisons on fluid, cloth and sand, every baseline showed artifacts on long rollouts: sand particles floating in mid-air, cloth behaving explosively, objects passing through each other.

WorldParticle’s answer is engineering, not magic. The team credits three design choices — the prediction-correction split that factors out known external forces, the particle-merging technique that compresses millions of particles so the model can handle them, and a string of architecture tweaks — for making the same transformer generalize across phenomena. “We needed to do a lot of detailed engineering and employ many tweaks to the architecture for it to be truly generalizable and scalable,” says co-lead author Minghao Guo, an MIT postdoc.

The verdict#

WorldParticle wins this comparison on breadth and speed, and — if the team’s numbers hold up — on accuracy too. No other entry in the table covers six phenomena in one architecture, and none is 100 times faster than the numerical reference on proteins. The use cases the researchers name are the tell: soft-robot manipulation, RNA evolution, even galaxy mergers. A simulator that handles coupled physics — a fluid sloshing against a deformable solid, a soft robot touching sand — is a tool for problems that currently need two solvers stitched together.

The honest caveat sits in the paper itself. The architecture is identical across all six categories, but a separate model is trained for each one; a single shared model that simulates everything at once is still future work. And the accuracy claim comes from the team’s own evaluation, published in a peer-reviewed journal and due to be presented at SIGGRAPH Asia, but not yet independently replicated. The next test is whether other labs can reproduce the numbers — and whether one trained model can really cover all six phenomena, as the team’s “GPT moment for simulation” vision requires. Baidu’s Apollo World shows the same world-model instinct spreading to robotaxis; WorldParticle takes it to physics itself.

Sources: MIT News, “A single AI model simulates the behavior of many substances, from sliding sand to folding proteins” (November 9, 2026), here; Wang et al., “WorldParticle: Unified World Simulation of Lagrangian Particle Dynamics via Transformer,” ACM Transactions on Graphics / arXiv:2605.15305, here. Cover and inline images are AI-generated illustrations, not renders from the WorldParticle paper.