New AI Model Predicts Physics Better by Tracking How Forces Flow

AI models that simulate physics usually treat the world as a grid of points connected by lines, called a graph. Think of atoms in a molecule or particles in a fluid. Older models passed simple numbers along those lines. The new model — ESNN — passes richer "directional" information, like arrows showing which way a force pushes. This lets it understand movement more naturally, without needing complicated math that slows everything down.

A key trick is "equivariance": if you rotate the whole scene, the model's answer rotates too, instead of falling apart. The researchers proved that, when the only clue is relative position, the model automatically separates motion into straight-line (radial) and spinning (tangential) parts. That's like how a spinning top moves differently from a falling ball — the AI learns both on its own.

The team also built a way to "break" symmetry when needed. For example, if a system has a special direction — like gravity pulling down — the model can learn that direction from data, then still revert to full symmetry when the direction isn't there. In tests, ESNN beat older models at predicting particle dynamics, handling mesh simulations, and guessing molecular properties. It also stayed accurate on rotations it had never seen.

Why does this matter for you? Better physics models mean more precise weather forecasts, more stable robotic movements, and faster drug discovery. Because ESNN does this with simpler computations, it could run on smaller devices or in real time. The catch: it's still research software, not yet in your phone. But it points to a future where AI understands the physical world without needing a room full of servers.