Physicists build AI that obeys conservation laws by design, not by luck

Researchers created a new type of generative AI that builds physical laws directly into its structure, so it can't break them even by accident. Instead of learning energy and momentum conservation as approximate rules, the AI enforces them at every step — meaning the models are more reliable and easier to interpret when used on particle physics data.

For years, physicists have used generative AI to model particle collisions, but the models would sometimes produce physically impossible results because they learned conservation laws only roughly. This is like training a calculator that mostly gets the right answer but occasionally violates basic arithmetic — you can't trust the output when the stakes are high. By building the constraints directly into the model's architecture, this approach guarantees physically valid results, which matters for experiments where you need to know whether your AI is predicting real physics or hallucinating. The next step is testing whether this approach makes the models actually useful for analyzing real detector data, not just simulated collisions.