PhononBench:A Large-Scale Phonon-Based Benchmark for Dynamical Stability in Crystal Generation
Researchers have introduced PhononBench, a new benchmark designed to evaluate the dynamical stability of AI-generated crystalline materials. This benchmark utilizes the MatterSim interatomic potential for efficient phonon calculations, enabling analysis of over 133,000 crystal structures. Findings indicate that current generative models struggle with dynamical stability, with an average rate of only 32.15% across generated structures, highlighting a significant limitation in AI-driven materials discovery. AI
IMPACT Highlights a critical gap in AI-driven materials science, potentially guiding future model development towards more practically viable crystal structures.