PhononBench:A Large-Scale Phonon-Based Benchmark for Dynamical Stability in Crystal Generation

PhononBench is introduced as the first large-scale benchmark for assessing dynamical stability in AI-generated crystals, leveraging the MatterSim interatomic potential for efficient phonon calculations. It evaluates 133,838 crystal structures generated by seven leading models and reveals that the average dynamical-stability rate is only 32.15%, with the best model, MatterGen, achieving 45.05%. This benchmark is crucial for practitioners as it highlights the limitations of current generative models in producing synthetically viable materials and provides a web-based service for rapid phonon predictions.