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

Generative artificial intelligence has advanced the design of crystalline materials, but existing evaluation frameworks primarily focus on thermodynamic stability, neglecting dynamical stability, which is essential for material synthesis and persistence. PhononBench is introduced as the first large-scale benchmark for assessing dynamical stability in AI-generated crystals, utilizing the MatterSim interatomic potential for efficient phonon spectrum calculations.

This tool enabled the analysis of 133,838 crystal structures generated by seven leading models. Results showed that the average dynamical stability rate was only 32.15%, with the best-performing model, MatterGen, achieving 45.05%.

Additionally, 32,995 crystal structures were identified as phonon-stable under a strict threshold of -0.001 THz. PhononBench also offers a web-based service for rapid phonon predictions, facilitating further research in materials science.

These findings underscore the need for improved generative models that can produce more stable crystal structures.