Physics adapted generative AI for metal insulator transition materials under label scarcity

Abstract

Metal-insulator-transition (MIT) materials are promising candidates for switchable electronics, neuromorphic hardware, and reconfigurable photonics, yet experimentally verified examples remain limited and the underlying mechanisms are often complex. We argue that generative AI for MIT discovery should move beyond the search for stable crystal structures alone and instead prioritize mechanism-informed phase-transition hypotheses. A modular physics-adapted framework, combined with staged verification of phase competition, electronic contrast, transition-pathway plausibility, and control accessibility, can guide credible discovery under severe label scarcity.