Quality-Weighted Determinantal Point Processes for Sample Efficient DFT Surrogate Models

Density functional theory (DFT) is widely used to compute electronic structure and materials properties, but it is expensive to run. When compute is limited, practitioners must choose which candidate systems to evaluate. We study this selection problem on JARVIS-DFT, a large standardized dataset of DFT labeled crystal structures and a realistic testbed for training machine learning models to predict expensive DFT computed properties. We propose quality-weighted determinantal point process (QW-DPP), a label-free coreset selection method for training DFT surrogate models. QW-DPP selects candidates that are both diverse through an RBF kernel and representative through a density-based quality score from unlabeled structural fingerprints. On JARVIS-DFT, QW-DPP reduces test MAE compared with random sampling by up to 8.9% at a 1% labeling budget across formation energy and bulk modulus. This improvement holds across Ridge, Bayesian Ridge, and Gradient Boosting. We find that QW-DPP is most useful at very small budgets because it improves worst case coverage of feature space.