Researchers at the Massachusetts Institute of Technology (MIT), the University of British Columbia (UBC) and other institutions have developed an artificial intelligence system that can simulate the movement and behaviour of materials ranging from collapsing sand and flowing liquids to deformable objects and changing protein structures.
Called WorldParticle, the system uses a deep-learning architecture to represent different materials as collections of tiny particles and predict how they move over time. The researchers report that it modelled changes in protein structures using time steps 100 times larger than those used in conventional numerical simulations.
The research was announced by MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) on November 9. The team plans to present its findings at the ACM SIGGRAPH Asia conference on computer graphics and interactive techniques.
The project addresses a limitation in conventional scientific simulation, where researchers often need separate computational tools to study different physical processes. Modelling flowing water, deformable materials, granular substances such as sand and molecular structures typically requires specialised numerical solvers built around the equations governing each system.
WorldParticle uses a common architecture to model these different processes. It breaks an object into thousands or millions of tiny three-dimensional points, known as particles, and tracks their positions, velocities and physical properties. A neural network then predicts how these particles will move in the next stage of the simulation.
Peter Yichen Chen, a former MIT postdoctoral researcher and now an assistant professor at UBC, compared the system's architecture to ChatGPT.
“It is a similar architecture to ChatGPT. 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,” Chen said.
How does the system work?
Conventional numerical simulations rely on governing equations that describe how a particular material behaves under different conditions. Researchers translate these equations into mathematical forms that computers can process, often developing a separate solver for each physical system.
WorldParticle uses a neural network trained to predict particle movement and interactions. It takes the three-dimensional positions and physical properties of particles as inputs and estimates their next positions and states. The system also accounts for external forces, such as a push acting on an object or a wave hitting a barrier, before correcting its predictions.
The researchers drew inspiration from world models, a category of AI systems designed to learn how physical environments change in response to actions and conditions.
Minghao Guo, an MIT postdoctoral researcher and co-corresponding author, said combining different physical processes within a common framework could allow researchers to study interactions that previously required separate simulation tools.
“If you study deformable objects and fluids, you would need two simulators, one for each physical regime. But if you unify them into one model, maybe you could couple the physics of these two processes and gain insights into new problems?” Guo said.
From collapsing sand to protein structures
The researchers tested WorldParticle across six physical phenomena, including granular materials, deformable objects, fluids and protein dynamics. In one demonstration, the system simulated a granular object collapsing into a pile. Some comparison methods produced unrealistic results, leaving individual sand particles suspended in the air.
The team also used the system to model changes in protein structures. According to Guo, WorldParticle handled these simulations approximately 100 times faster than numerical simulators.
The paper describes this improvement in terms of simulation time steps. WorldParticle used a 50-femtosecond step for protein simulations, compared with the 0.5-femtosecond steps used in traditional simulations. A femtosecond is one quadrillionth of a second.
The distinction matters because a larger time step allows a simulation to cover a longer interval with fewer computational steps. The reported 100-fold figure refers to this difference in time-step size, rather than an independently established 100-fold reduction in total computing time.
To manage the computational demands of representing complex objects as millions of particles, the researchers developed a technique to merge and compress particle information while retaining details needed to predict their movement.
Where could it be used?
The researchers see possible applications in engineering, materials science, biology and robotics. One potential use involves studying how fluids interact with deformable structures, such as oil moving against an elastic valve inside a storage tank. Another involves predicting how sand or soil shifts when manipulated by a soft robot.
The team also plans to investigate applications involving RNA evolution and astrophysical processes such as the merging of galaxies.
However, these remain proposed areas of application. The current research demonstrates the system across six physical phenomena, while further work is required to extend it to additional systems and build a complete general-purpose simulation platform.
The paper also makes an important distinction between a shared architecture and a single model trained across all materials. The researchers currently train separate sets of model parameters for different categories of physical behaviour. Developing one set of parameters that can handle every category remains a future goal.
“We want to have a GPT moment for simulation,” Chen said, describing the team's ambition to develop a foundation model that could simulate different physical processes without requiring researchers to build a new system from scratch each time.
The research team includes scientists from MIT, UBC, Georgia Tech, Inria, Meta, the University of Pennsylvania, the University of Utah, the University of California, Los Angeles, and the University of Hong Kong.