WorldParticle: A new AI model capable of simulating everything from sand to proteins
Researchers from MIT and UBC have developed a new AI architecture called WorldParticle, which can simulate a wide variety of substances—from grains of sand to protein folding—by learning physical dynamics instead of relying on traditional equations.
Researchers from MIT and UBC have introduced an AI model called WorldParticle that integrates different physical processes into a single structure. Unlike traditional methods that require specific equations for each substance, this system represents objects as small particles and predicts their movements step by step.
WorldParticle uses a Lagrangian representation to learn the physical dynamics of objects through a neural network. By separating external forces from the simulation, the model can generalize across previously unseen materials and environments. The researchers stated that the model provides faster and more accurate results than traditional numerical simulation methods in tasks such as granular collapse and protein structure changes.
The MIT and UBC team is preparing to present their work at the ACM SIGGRAPH Asia Conference. Designed as a foundation model for scientific discovery, WorldParticle is intended for use in fields such as soft robotics, RNA evolution, and astrophysics.
Sources
- news.mit.edu · Link