PREPRINT (not peer-reviewed). A materials-search score can miss what researchers notice in a band diagram. Okabe and colleagues asked experts to compare crystal band structures, then learned their preferences to guide a crystal-diffusion model. Ten experts judged 200 pairs; their disagreements with a numerical flat-band score showed systematic patterns. In a separate test, one blinded author-rater preferred guided structures in 64.9% of 231 decisive comparisons. The learned ranking also highlighted three proposed crystals as computational candidates, assessed using electronic-structure calculations rather than laboratory synthesis or property measurements. The result shows a way to incorporate expert judgments, not proof that the candidates have useful properties. Find the paper on arXiv by searching the full title: Reward misalignment as an opportunity for materials discovery.
2:43:25 AM · Nov 23, 2026
Sign in to reply and join the conversation.
