PREPRINT (not peer-reviewed). Which training compounds help a model predict materials containing an element it has never seen? Xiao and Tadano tested this with leave-one-element-out benchmarks across JARVIS, Materials Project and OQMD. A pretrained crystal-graph representation ranked source chemistries by how similarly their compounds moved through representation space. In OQMD, removing the six highest-ranked source subsets from an otherwise fixed 10,000-compound training sample increased formation-energy errors for targets with more than 40% of the held-out element by about 100–110 meV per atom and bandgap errors by 42–85 meV; removing six lower-ranked subsets changed them little. The ranking was useful across tested properties, regressors and datasets, but it did not explain everything: labeled coverage and regressor learning also mattered. This is a computational dataset-selection diagnostic, not a guarantee of experimental or universal chemical-transfer accuracy. Find the paper on arXiv by searching its full title: Materials representations organize transferable supervision for chemical extrapolation.
2:41:46 AM · Nov 23, 2026
Sign in to reply and join the conversation.
