Support boundaries for reference transfer of oxygen vacancy energetics in ABO₃ perovskites

Abstract Predicting oxygen-vacancy formation energies from known compounds requires references with compatible chemistry. We examine when this condition is met in ABO₃ perovskites, using 840 PBE+U vacancy-energy records for 812 compositions with explicit A- and B-site assignments. Three compounds at the corners of an A–B composition rectangle estimate the fourth when the A- and B-site contributions are approximately additive. A four-point energy difference gives the prediction error exactly. We select references by bulk structure, nominal A-site charge, B-site element family, and chemical diversity. Across ten validation splits that keep each composition together, the strict support criterion covers 12.6% of compositions and gives a mean absolute error (MAE) of 0.069 eV per removed oxygen relative to the source DFT energies. A local additive model gives 0.077 eV, showing that chemical localization accounts for most of the accuracy. Using reference transfer where supported and Extra Trees regression elsewhere reduces the MAE over all compositions from 0.470 to 0.457 eV, with improvement in every split. This benefit persists when the d-series families are merged. On 38 external SCAN+U targets, nominal-valence matching reduces the reference-transfer MAE from 0.962 to 0.475 eV compared with structure-only matching, giving accuracy comparable to external Extra Trees under a looser support rule. The two datasets support valence matching as a useful chemical criterion, while the available references determine where local transfer can improve prediction.