Simulating a physical process often means choosing a tool built for that specific job. Researchers at MIT, the University of British Columbia (UBC), and other institutions have developed WorldParticle, a deep-learning system intended to model a wider range of materials with one framework. It represents matter as particles and predicts how they move over time.
One framework for different physical systems
Many existing simulators are built around equations for a particular material or process. Scientists must turn those equations into forms a numerical solver can handle, and repeat the work when they want to simulate a different kind of system.
WorldParticle takes a different route. It uses a neural network that learns to solve governing equations in general, rather than requiring a new model and solver for each case. The researchers drew on the idea of world models: AI systems that learn rules about how objects or environments change in response to actions and conditions.
The system represents an object as tiny, three-dimensional particles, a method known as a Lagrangian representation. It uses each particle’s position and physical properties to predict its state at the next step, building a picture of how the material moves through space and time.
How WorldParticle predicts motion
The network receives the particles’ positions and physical properties, then predicts their positions and properties for the next frame. The approach is somewhat like a language model predicting one word at a time, but here the next prediction concerns particle movement.
WorldParticle also accounts for forces from outside the system. Before a simulation step, it separates those forces from the system’s behavior. After making its initial particle predictions, it adds the forces back and adjusts the predictions to reflect their effects.
This design is intended to help the model work with materials, forces, and environmental conditions it did not encounter during training. It also makes it possible to simulate interactions that might otherwise call for separate tools, such as a fluid moving against a deformable solid.
Scaling to complex materials
Representing a complex object can involve thousands or millions of particles, creating a heavy computational workload. To manage that demand, the researchers developed a merging technique that compresses groups of particles while preserving fine-grained details of their behavior.
The team demonstrated WorldParticle on six different physical phenomena, including deformable objects, granular materials, and proteins. They report that it achieved higher accuracy than numerical simulation methods in those demonstrations.
In one example, the system captured a granular object collapsing into a pile, while some baseline methods left particles suspended in the air. The researchers also modeled how protein structures change over time at about 100 times the speed of numerical simulators, Guo adds.
Potential uses and next steps
A unified simulator could help researchers investigate connected physical processes. For example, it could model how sloshing oil affects an elastic valve in a storage tank, or how sand or soil moves when touched by a soft robot. Those applications depend on representing materials that behave differently within a shared simulation.
The researchers say the framework could also be applied to soft robot manipulation, RNA evolution, and astrophysical phenomena such as merging galaxies. These are possible directions for the method; the team says it wants to adapt WorldParticle to simulate these and other systems.
The longer-term goal is a general-purpose pipeline that helps scientists explore questions existing simulation methods cannot effectively address. The research will be presented at the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia.