PhononBench: A Large-Scale Benchmark Reveals the Dynamical Stability Challenge in AI-Generated Crystals (IMAGE)
Caption
Figure 1. Systematic evaluation of dynamical stability in crystal-generation models: (a) workflow of the study. Seven generation models, including CrystaLLM, MatterGen, DiffCSP, InvDesFlow-AL, CrystalFlow, and CrystalFormer, generated 252,437 novel structures. After deduplication and CIF validation and post-processing, 133,838 successfully relaxed crystals underwent full phonon-spectrum calculations using MatterSim combined with Phonopy. (b) MatterSim-v1 was adopted as the unified potential for subsequent phonon calculations to ensure consistent evaluation standards. (c) Benchmark comparison of dynamical stability on the MP20 dataset. Under the −0.1 THz phonon-frequency threshold, MatterGen (45.05%), DiffCSP (43.94%), and CrystalFlow (40.00%) achieve the highest stability rates, while the best LLaMA2-70B configuration reaches 32.50%.
Credit
Xiao-Qi Han#, Ze-Feng Gao#, Peng-Jie Guo and Zhong-Yi Lu from Renmin University of China.
Usage Restrictions
Credit must be given to the creator.
License
CC BY