Answers are grounded in this paper and cite their sources.
Oxygen vacancies control conduction, catalysis and ion transport in oxides, and an error of 0.3 eV in their neutral formation energy changes their concentration at 1000 K thirty-fold. Graph networks predict this energy from structure alone, but whether electronic descriptors of the pristine host from density-functional theory help is untested at matched training oxides. We compared a structure-only network (S) with one also given 22 such descriptors (D) on 1726 vacancy sites in 818 non-magnetic oxides of the Kumagai database, testing on unseen formulas. S reached 0.297 eV, the published accuracy. With 654 training oxides and equal tuning, D lowered the site error by 0.033 [0.021, 0.045] eV (95% interval), inside a 0.05 eV tolerance (1.8-fold in concentration). Tuned alone in a wider search, D* gained 0.069 [0.060, 0.078] eV, above it, but 0.039 [0.030, 0.048] eV over default-initialised S. The descriptors remove 11 to 23% of the error of S, depending on tuning. Within oxides the gain is concentrated in hosts whose cations have an empty valence s shell: empty-shell post-transition hosts gain 0.075 [0.050, 0.102] eV more than lone-pair hosts. Descriptors predicted from structure never surpass S, so the gain requires computed descriptors; where these are unavailable, S suffices for neutral vacancies.
No takes yet. Share an insight, caveat, or question.
Gang et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: