What is Matbench Discovery?

Matbench Discovery is a benchmark framework that evaluates how well machine learning models can screen large numbers of hypothetical crystal structures and predict which of them correspond to stable, physically realizable materials. It tests candidate models against a curated, held-out set of known material structures and thermodynamic stability data, giving researchers a standardized way to compare different materials-discovery models rather than relying on ad hoc accuracy claims from each individual research paper.