Abstract

ML energy models have made significant leaps over the past 3 years. I will present Matbench Discovery, a benchmark designed to measure how useful ML actually is in guiding prospective materials discovery, quantify the kind of acceleration we can expect in future discovery efforts, and chart progress over time to determine which ML method performs best at thermodynamic stability prediction from unrelaxed crystal structures. I will also present preliminary results which suggest that phonons from foundational force fields (esp. equivariant ones) can help us go beyond thermodynamic stability and predict dynamic stability across material space very cheaply and at unprecedented scale, paving the way for future high-throughput discovery with increased fidelity. I will point out unresolved problems stemming from over-softened potential energy surfaces (PES) in these foundation models and how we might address them.