Our work “Virtual Node Graph Neural Network for Full Phonon Prediction,” led by our group members Ryotaro, Abhijatmedhi, and Artittaya, supported by ORNL Scientist YQ and many others, is accepted by Nature Computational Science.
This work overturns the belief that graph node should represent atoms, but rather designs a new machine learning architecture uses virtual nodes to represent collective excitations. The model inputs crystal structures and outputs phonon dispersion (energy-momentum relation, a fundamental but high-dimensional property for materials) in arbitrary complex materials. The prediction accuracy is comparable to computationally costly density functional perturbation theory (DFPT), but the speed is order-of-magnitude faster than the existing fastest method like machine learning interatomic potentials (MLIPs). Big Congratulations!