Minimizing data requirements for chemical and material property prediction via cluster-based training set selection.
Abstract
Machine learning is increasingly used to predict chemical and material properties, and when data are scarce, data acquisition is often necessary to support these efforts. Despite the practical advantages of minimizing acquisition costs, it remains unclear which data acquisition strategies are the most efficient across diverse property prediction tasks or what principled basis underlies their relative performance. Here we survey 100 space-filling and active learning algorithms across 60 property prediction tasks, finding that cluster-based training set selection consistently outperforms active learning when choosing from a predefined candidate library, a result that runs counter to prevailing practice and is robust to the dimensionality of material representations, the choice of fixed or learned representations, and training set size. A theoretical framework grounded in Gaussian process variance establishes cluster-based selection as the principled optimum in the low-data limit, providing a unified basis for understanding training set quality across contexts. Overall, this work establishes cluster-based training set selection as a broadly effective and principled strategy for data-efficient, machine-learning-enabled materials design.