A single AI model simulates the behavior of many substances, from sliding sand to folding proteins.
Researchers at MIT, the University of British Columbia and other institutions built WorldParticle, a single deep-learning system meant to simulate materials from sand to proteins without a separate hand-coded simulator for each.
Scientists who want to simulate a physical process today usually have to choose a tool built for that one job, such as a solver for flowing water or one for protein folding. Each is hand-coded around the equations that govern the process. Researchers at MIT, the University of British Columbia (UBC) and elsewhere have now described a single system, called WorldParticle, that is meant to replace that patchwork with one unified AI architecture.
WorldParticle represents a material, whether a dune of sand or a protein molecule, as thousands or millions of tiny three-dimensional particles. A neural network takes the particles' positions and physical properties as input and outputs them for the next frame, predicting motion one step at a time. Because the model learns the underlying dynamics instead of being given an equation, the researchers say it can adapt to materials, forces and environments it did not see in training. Before each step, the system separates out external forces, then folds them back in to correct its predictions.
A complex object can break into millions of particles, which makes the computation heavy. To keep it manageable, the team used a merging technique that combines and compresses particles while preserving fine-grained detail about how they behave.
The team tested WorldParticle on six physical phenomena, including deformable objects, granular materials and proteins, and reports higher accuracy than numerical simulation methods. In one example, the model showed a granular object collapsing into a pile, while some baseline methods left sand particles floating in the air. Co-lead author Minghao Guo said it modeled changes in protein structure about 100 times faster than numerical simulators.
The researchers suggest uses such as sloshing oil acting on an elastic valve in a storage tank, or sand moving as a soft robot touches it. They also say the framework could extend to RNA evolution and galaxy mergers, though they still need to adapt it to those systems. “We want to have a GPT moment for simulation,” said Peter Yichen Chen, a former MIT postdoc who is now an assistant professor at UBC. The work will be presented at ACM SIGGRAPH Asia.
At a Glance
- WorldParticle was tested on six physical phenomena, including deformable objects, granular materials and proteins.
- Guo said it modeled protein structure changes about 100 times faster than numerical simulators.
- It represents materials as thousands or millions of 3D particles and predicts their motion one step at a time.
- The research will be presented at the ACM SIGGRAPH Conference and Exhibition in Asia.
Sources
This brief was written by Brightside Report from the reporting linked above. Spot an error? Request a correction.