Transfer learning on universal interatomic potential embeddings improves generalization in structure-property defect models

Oxygen vacancy formation energies govern the performance of metal oxides across energy conversion, catalysis, and electronics, yet predicting them accurately without density functional theory calculations for novel chemical systems remains challenging. Here, we show that frozen 256-dimensional embeddings extracted from a pretrained MACE universal machine-learning interatomic potential encode information sufficient to predict vacancy formation energies obtained from density functional theory without requiring supercell construction, vacancy creation, or geometry optimization. A lightweight multilayer perceptron trained on these embeddings (MACE-dGNN) achieves an element-wise cross-validation mean absolute error of 0.38 eV, halving the error of a baseline graph neural network trained from scratch (0.72 eV) and offering a modest advantage over direct MACE relaxation calculations (0.51 eV). Because the embedding approach decouples representation from the prediction task and generalizes well with limited data, it will (1) extend naturally to fine-tuning and prediction of properties (e.g., charged defects) that are inaccessible to generic energy-and force-output interatomic potentials and (2) be trainable on necessarily small datasets, potentially enabling the use of more expensive but accurate tools for generating training data (e.g., hybrid functional calculations) for which training or fine-tuning of interatomic potentials will be difficult. We demonstrate practical impact by revisiting a prior screening of thermochemical water-splitting materials, where improved generalization alters ∼45 % of candidate classifications.