Some crystals behave like a switch: nudge the temperature or pressure, and they turn from a conductor into an insulator. This is called a metal–insulator transition (a change in the ability to conduct electricity). These chameleon materials are needed for smart electronics, neuromorphic chips (which mimic the brain), and tunable optics. But they are hard to find — only a handful of confirmed examples exist.

A team of researchers proposed not to blindly screen millions of crystal lattices, but to entrust the search to generative artificial intelligence. But here, the AI doesn't just generate stable structures: it's trained to think like a physicist. The approach is modular: a hypothesis about a new material must explain why two phases (metal and insulator) compete, how sharp the property jump is, whether the transition path is feasible, and whether it can be controlled in practice.

The screening proceeds step by step, like assembling a puzzle. First, they assess phase competition — which phase wins. Then they look at the contrast in electronic properties: how strongly the conductivity changes. Next, the plausibility of the transition path, to avoid a "bridge to nowhere." Finally, controllability: can the transition be triggered by pressing a button, not by a lightning strike? This approach is especially valuable when labeled data are scarce and discoveries can't wait.