Researchers at MIT, the University of British Columbia and several partner institutions have developed a single deep-learning system, named WorldParticle, that can simulate a wide range of physical phenomena — from flowing granular material to folding proteins — using one architecture, according to WebWire.
The work, described in an open-access paper, was presented at the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia. WebWire reports that the team includes collaborators from Georgia Tech, Inria, Meta, the University of Pennsylvania, the University of Utah, UCLA and the University of Hong Kong.
Why it matters
Today, anyone who wants to model a physical process generally has to build a dedicated tool for it. A solver written for moving water is not the same as one written to capture how a protein folds, and each is coded around the specific equations that describe its target process.
WorldParticle proposes to replace that collection of separate tools with one system. According to WebWire, the researchers say engineers could use it to model situations that normally call for more than one simulator at once — for example, a liquid pushing against a solid that can deform, such as oil sloshing against an elastic valve inside a storage tank. It could also help researchers study how granular material like sand or soil shifts when a soft robot makes contact with it.
Peter Yichen Chen, a former MIT postdoc now an assistant professor at UBC and a co-corresponding author of the paper, framed the ambition in terms of a widely used language model. "We want to have a GPT moment for simulation," he said, according to WebWire. He added that people currently have to write a simulator from scratch whenever they want to model a particular phenomenon, and that the team's goal is a foundation model that works out of the box for whatever scenario a user wants to simulate. He described the current work as a first step toward that.
Particles instead of equations
The system borrows from the idea of world models — AI systems that learn the rules governing physical environments so they can forecast how objects will change in response to actions or conditions. The researchers applied that concept to build one architecture spanning multiple physical regimes.
WorldParticle represents every material it handles, whether a dune of sand or a protein molecule, as thousands or millions of tiny three-dimensional points, an approach known as a Lagrangian representation. It then tracks where those points move and how fast, step by step, to reproduce the material's behavior.
Conventionally, scientists working with Lagrangian particles rely on a governing equation that defines the laws controlling how a material behaves across space and time, then convert that equation into a form a numerical solver can process — a cycle repeated for each material and process. The new system instead uses a neural network trained to solve governing equations in general, which the researchers say removes the need to design a fresh model and solver for each one.
Once an object has been converted into particles, the network takes their 3D positions and physical properties as input and produces their positions and properties for the following frame, predicting motion in one-second increments. Chen compared the setup to the architecture behind ChatGPT, with one difference: rather than predicting the next word in a piece of text, the model predicts how particles will move, one step at a time.
Before each step, WorldParticle isolates the external forces acting on the system — such as the impulse felt by an ocean wave striking a seawall. After the network predicts the particles' initial states, those forces are folded back in, correcting the predictions according to how the forces would affect the object. This is what allows the system to handle materials, environmental conditions and forces it did not see during training.
Minghao Guo, an MIT postdoc and co-corresponding and co-lead author, explained the appeal of combining regimes. Studying deformable objects and fluids would normally require two simulators, one per physical regime, he said, according to WebWire. Unifying them in a single model, he suggested, might let researchers couple the physics of both processes and gain insight into new problems.
Engineering for scale
Splitting a complex object into millions of particles quickly becomes too demanding computationally. To keep processing manageable, the team used a merging technique that combines and compresses particles while still preserving fine-grained detail in their behavior. Guo said the work required extensive engineering and numerous adjustments to the architecture before it became genuinely generalizable and scalable.
The researchers demonstrated WorldParticle on six different physical phenomena, including deformable objects, granular materials and proteins, and reported higher accuracy than numerical simulation methods.
The paper's co-lead authors are Caoliwen Wang and Siyuan Chen, both UBC graduate students. Wojciech Matusik, an MIT professor of electrical engineering and computer science who leads the Computational Design and Fabrication Group within the Computer Science and Artificial Intelligence Laboratory, is also among the authors.
Over the longer term, the team hopes the approach will help scientists uncover new materials and examine physical processes that existing simulation methods cannot study effectively.
Watch for how the research community responds to the accuracy claims and whether the architecture extends beyond the six demonstrated phenomena toward the general-purpose simulator the authors describe.
