Virtual node graph neural network for full phonon prediction
Ryotaro Okabe,Abhijatmedhi Chotrattanapituk,Artittaya Boonkird,Nina Andrejevic,Xiang Fu,Tommi S Jaakkola,Qichen Song,Thanh Nguyen,Nathan Drucker,Sai Mu,Yao Wang,Bolin Liao,Yongqiang Cheng,Mingda Li
DOI: https://doi.org/10.1038/s43588-024-00661-0
Abstract:Understanding the structure-property relationship is crucial for designing materials with desired properties. The past few years have witnessed remarkable progress in machine-learning methods for this connection. However, substantial challenges remain, including the generalizability of models and prediction of properties with materials-dependent output dimensions. Here we present the virtual node graph neural network to address the challenges. By developing three virtual node approaches, we achieve Γ-phonon spectra and full phonon dispersion prediction from atomic coordinates. We show that, compared with the machine-learning interatomic potentials, our approach achieves orders-of-magnitude-higher efficiency with comparable to better accuracy. This allows us to generate databases for Γ-phonon containing over 146,000 materials and phonon band structures of zeolites. Our work provides an avenue for rapid and high-quality prediction of phonon band structures enabling materials design with desired phonon properties. The virtual node method also provides a generic method for machine-learning design with a high level of flexibility.