Researchers develop AI system to simulate material movement
The WorldParticle system uses a deep-learning architecture to model everything from collapsing sand to changing protein structures.

Researchers at the Massachusetts Institute of Technology (MIT), the University of British Columbia (UBC), and other institutions have developed an artificial intelligence system called WorldParticle. The system can simulate the movement and behaviour of materials including flowing liquids, collapsing sand, deformable objects, and changing protein structures.
WorldParticle uses a deep-learning architecture to represent different materials as collections of tiny three-dimensional points, known as particles. A neural network tracks the positions, velocities, and physical properties of these particles to predict how they will move in the next stage of a simulation. The system also accounts for external forces, such as a wave hitting a barrier or a push acting on an object, before correcting its predictions.
The research was announced by MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) on November 9. Peter Yichen Chen, an assistant professor at UBC and former MIT postdoctoral researcher, compared the architecture to ChatGPT, noting that instead of predicting words, the system predicts how particles move one step at a time.
In tests involving protein structures, WorldParticle handled simulations approximately 100 times faster than numerical simulators in terms of time steps. While traditional simulations used 0.5-femtosecond steps, WorldParticle used a 50-femtosecond step. The team also developed a technique to merge and compress particle information to manage the computational demands of representing millions of particles.
Minghao Guo, an MIT postdoctoral researcher and co-corresponding author, said combining different physical processes within one framework could allow researchers to study interactions that previously required separate simulation tools. The team plans to present its findings at the ACM SIGGRAPH Asia conference.
The researchers see possible applications in biology, engineering, materials science, and robotics. Future work includes investigating RNA evolution, astrophysical processes like the merging of galaxies, and developing a single set of parameters that can handle every category of physical behaviour.
Key facts
- WorldParticle uses a neural network to predict the movement of particles in a simulation.
- The system modeled protein structures using time steps 100 times larger than conventional numerical simulations.
- The research team includes scientists from MIT, UBC, Georgia Tech, Inria, and Meta.
Three perspectives
Neutral
Researchers have developed an AI system called WorldParticle that uses a deep-learning architecture to simulate the movement of various materials.
Positive
The system could allow researchers to couple different physical processes to gain insights into new problems.
Negative
Current research requires separate sets of model parameters for different categories of physical behaviour.
Your week in news, every Wednesday
The biggest stories of the week for your edition, in your inbox every Wednesday morning, your time. Free, and you can unsubscribe at any time.