A Systematic Approach to Crystal Structure Prediction Following Imaginary Phonon Modes Combined With Polynomial Machine Learning Potentials

T Naruse and A Seko and I Tanaka, JOURNAL OF THE PHYSICAL SOCIETY OF JAPAN, 94, 074601 (2025).

DOI: 10.7566/JPSJ.94.074601

Machine learning potentials (MLPs) have become powerful tools for predicting stable and metastable crystal structures. General crystal structure prediction approaches generate numerous local minimum structures, including both dynamically stable and unstable configurations. By following imaginary phonon modes (FIPM) in dynamically unstable structures, it is possible to rationally derive dynamically stable structures. In this study, we apply a global structure search procedure based on the FIPM approach, accelerated by polynomial MLPs, to systematically predict stable and metastable structures of elemental Si under various pressure conditions. Additionally, we automate a complex recursive workflow consisting of geometry optimizations, lattice dynamics calculations, and updates to crystal structures based on phonon modes with imaginary frequencies. The results of this study demonstrate that FIPM can be a useful tool for discovering inorganic compounds and crystalline polymorphs.

Return to Publications page