New MIT AI simulates physical reality – a step toward physical models of the real world

However, MIT has now developed an AI system capable of simulating a wide range of materials and their physical behavior. MIT calls the program WorldParticle. Instead of modeling each material using its own physical equations and a custom-coded simulation method, the AI breaks down different objects into large numbers of tiny three-dimensional particles.

Each particle has a spatial position, velocity, physical properties, and interactions with other particles and/or external forces. While ChatGPT, for example, predicts the next word, WorldParticle predicts the particles' next state of motion.

So far, the model can already simulate six different physical phenomena: granular materials such as sand or soil, liquids, deformable solids (such as elastic objects), and even proteins (modeling how molecular structures change over time). The simulation of protein movements, for example, is said to have been around 100 times faster than the conventional methods used for comparison.

This is particularly relevant when it comes to unifying physical simulations across virtually every field. In the future, there may no longer be a need for specialized software for individual material simulations. A single AI model could simulate a wide variety of materials. This would increase simulation speeds while lowering the barriers to testing and reducing development time.

Potential beneficiaries include drug development and the development of products such as cars, smartphones, video games, and just about anything else. This is nothing less than an important step toward Foundation Models for Physics. Such models could make dynamic processes in the real world easier to predict and test.