Publication
Machine learning has demonstrated superior performance in predicting vast materials properties. However, predicting a spectral-like continuous material property such as phonon density of states (DOS) is more challenging for machine learning. In this work, with phonon DOS of 4994 inorganic structures with 62 unique elements calculated by density functional theory (DFT), we developed a crystal attention graph neural network (CATGNN) model for predicting total phonon DOS of crystalline materials. The computational cost of training the CATGNN model is several orders of magnitude cheaper than full DFT calculations. We find that high vibrational similarity or phonon DOS overlap is not the only requirement to obtain high interfacial thermal conductance (ITC) instead, the average acoustic group velocity of heat source and heat sink for the acoustic branches in the phonon DOS overlap region is equally important in determining ITC. Pearson correlation analysis yields a few simple material descriptors that are strongly but negatively correlated with ITC. These easy-to-calculate material features combined with the proposed high average acoustic group velocity and phonon DOS overlap predicted by CATGNN model offer a new reliable and fast route for high-throughput screening of novel crystalline materials with desirable high ITC for phonon-mediated thermal management of wide bandgap electronics.