One AI Model Simulates Sand, Fluids and Proteins Alike

"We want to have a GPT moment for simulation," Chen said. "Every time people want to build a simulator for a particular physical phenomenon, they need to code it up from scratch. Our vision is to have a foundation model that works out of the box for any scenario you want to simulate. This is a first step in that direction."

Note the hedging. "A first step" is doing honest work in that sentence.

How does the model actually work?

The researchers drew on the idea of world models — AI systems that learn the rules of the physical world and predict how environments change over time in response to actions or conditions.

WorldParticle breaks an object into thousands or millions of tiny 3D dots, called particles, using what's known as a Lagrangian representation — in plain English, it follows each particle's position and velocity through space and time. Conventionally, scientists feed these particles a governing equation, the mathematical law dictating a material's behavior, then convert that equation into a formula a numerical solver can handle. Repeat for every new material.

WorldParticle skips that bespoke step. A neural network learns to solve governing equations in general. It takes the particles' 3D positions and physical properties as input, then outputs their positions for the next frame, predicting motion in one-second increments.

"It is a similar architecture to ChatGPT," Chen said. "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."

Before each step, the system separates external forces — an impulse, say, when an ocean wave slams a seawall — then reintegrates them afterward, correcting its particle predictions. That trick lets it generalize to unseen materials, changing conditions, and varied forces.

What are the limits?

Plenty, as the researchers themselves concede. Complex objects can shatter into millions of particles, a computational burden that would overwhelm the model outright. The team's answer was a merging technique that compresses particles while preserving fine-grained behavioral detail.

"We needed to do a lot of detailed engineering and employ many tweaks to the architecture for it to be truly generalizable and scalable," Guo said. That's a candid admission: generalization here is engineered, not magic.

What does this change for scientists?

Right now, not much in daily practice. Six demonstrated phenomena, however impressive, is a demonstration — not a deployed tool. The long-term ambitions are bigger: soft robot manipulation, RNA evolution, even galaxy mergers. The team wants to build toward a full, general-purpose simulation pipeline for scientific discovery.

"With a unified model, we could unlock many discoveries that we can't even imagine," Guo said.

Maybe. But the promise rests on preliminary demonstrations, and the researchers say so themselves. Watch this space — carefully.