Data-Efficient Training of Linear ACE Potentials through Leverage-Guided Subset Selection of ASSYST Structure Pools

Abstract

The construction of machine-learned interatomic potentials (MLIPs) is often limited by the cost of generating large density-functional-theory (DFT) training datasets. For systematically generated structure pools such as ASSYST, a central practical question is how many configurations must be labeled to achieve reliable accuracy. Here we assess geometry-based, label-free subset selection for training linear Atomic Cluster Expansion (ACE) potentials. Using statistical leverage scores and CUR-type sampling, we compare leverage-guided selection against random, energy-based, and force-based baselines under controlled iterative protocols. Elemental Al provides the primary benchmark, with Cu and Al-Cu alloys used for transfer validation. Leverage-guided subsets recover plateau-level energy and force accuracy using substantially smaller labeled fractions (approximately 30-40%) than random sampling, corresponding to an effective 2-3x reduction in DFT labeling for the systems studied. In alloy tests, defect energetics remain comparable across strategies once sufficient chemical diversity is included, while leverage selection maintains competitive accuracy at reduced training size. These results demonstrate that descriptor-space-guided, label-free subsampling can significantly reduce DFT workload for linear ACE models trained on ASSYST structure pools without degrading defect-level fidelity.