Scientists typically rely on distinct, hand-coded simulators for each physical process they wish to model, such as fluid dynamics or protein folding. WorldParticle, a new deep-learning system, unifies these tasks within a single AI architecture. The researchers, including Peter Yichen Chen, a former MIT postdoc now at UBC, and Minghao Guo, an MIT postdoc, drew inspiration from "world models," AI systems that learn the underlying rules of the physical world to predict changes over time.
WorldParticle represents objects as thousands or millions of tiny, three-dimensional particles, using a Lagrangian representation. It then tracks the movement and velocity of these particles across space and time to simulate material dynamics. Unlike traditional methods that use governing equations for each material, WorldParticle employs a neural network to learn and predict how these particles will move, one step at a time. Chen likened this process to how ChatGPT predicts words in a text.
The team demonstrated WorldParticle's capabilities by modeling six different physical phenomena, including deformable objects, granular materials, and proteins. The model exhibited higher accuracy compared to conventional numerical simulation methods. For instance, WorldParticle accurately depicted a granular object collapsing into a pile, a scenario where some baseline methods showed sand particles floating in the air.
Furthermore, WorldParticle modeled protein structure changes approximately 100 times faster than numerical simulators. This speed improvement, coupled with its ability to generalize to unseen materials and conditions, suggests potential for accelerating drug discovery and materials science.
The system's architecture incorporates a prediction-correction design. An explicit predictor first advances particles based on known external forces. A learned corrector then predicts residual position and velocity updates through a particle tokenizer, a super-token encoder, and a super-token decoder. This design allows the model to capture both local interactions and global coupling within a fixed architecture. To manage computational demands, the researchers implemented a merging technique to combine and compress particles while retaining fine-grained detail.
The researchers anticipate that WorldParticle could assist in various applications, including the design of soft robots, the development of new medicines, and the study of astrophysical phenomena. They envision a future where a single foundation model can simulate any scenario, much like large language models operate across diverse text-based tasks. The research paper detailing WorldParticle was published in ACM Transactions on Graphics and will be presented at SIGGRAPH Asia.