Scientists who want to simulate a physical process usually need a separate computer tool for each material. One solver handles flowing water, another handles folding proteins, and each is hand-coded around the equations that describe that specific process. Researchers at MIT and other institutions have now built a single deep-learning system, called WorldParticle, that can simulate many kinds of matter with one architecture. The work is described in an MIT News article and in an open-access paper that will be presented at an ACM SIGGRAPH conference.
WorldParticle represents every material as a cloud of tiny three-dimensional particles, an approach called a Lagrangian representation, which tracks how each particle moves over time. A neural network takes the positions and physical properties of the particles and predicts where they will be one step later, in increments of one second. Instead of being programmed with the governing equation of each material, the network learns the underlying dynamics from data. Before each step, the system separates external forces and adds them back to correct its predictions. This helps it handle materials and conditions it never saw during training.
Built to scale up
Because a complex object may be split into millions of particles, the computation can become too heavy. The team used a merging technique that combines and compresses particles into a smaller set while keeping fine detail. The researchers tested the system on six kinds of physical phenomena, including deformable objects, granular materials such as sand, and proteins. This was more accurate than numerical simulation methods. For example, it correctly showed sand collapsing into a pile while some baseline methods left grains floating in the air, and it modeled how protein structures change over time about 100 times faster than numerical simulators.
This matters because simulation is a basic tool of science and engineering, and today's tools are fragmented. A unified model could simulate coupled processes, like oil sloshing against an elastic valve in a storage tank, or how sand and soil shift when a soft robot touches them. The researchers say the same framework could extend to soft robot manipulation, RNA evolution, and even the merging of galaxies, and they hope it becomes part of a general-purpose simulation pipeline for scientific discovery.