High-Throughput Screening of Candidate Substrates for Wide Bandgap Electronic Cooling by Graph Neural Network

Machine learning has demonstrated superior performance in predicting vast materials properties. However, predicting a continuous material property such as phonon density of states (DOS) is more challenging for machine learning due to the inherent issues of data smoothing and sensitivity to peak positions. In this work, phonon DOS of ~5,000 inorganic cubic structures with 63 unique elements from the Open Quantum Materials Database are calculated by high precision density functional theory (DFT). With these training data, we build an equivariant graph neural network (GNN) for total phonon DOS of crystalline materials that utilizes site positions with their atomic masses as input features. The computational cost of training the GNN model is several orders of magnitude cheaper than full DFT calculations. More interestingly, the trained GNN model can predict partial DOS of the constituent atomic species even if such data were not included in the training, which demonstrates GNN’s capability in predicting the species contributions (node-level) of partial DOS from the total DOS predictions without additional computational cost. We then deploy the trained GNN model to predict phonon DOS of >40,000 non-zero bandgap materials to search for thermally conductive substrates for cooling a few representative high electron mobility transistors (HEMT) in terms of high interfacial thermal conductance (ITC). Our results show that high vibrational similarity or phonon DOS overlap is not the only requirement to obtain high ITC. Instead, the average group velocity of heat source and heat sink for the acoustic branches in the phonon DOS overlap region is equally important in determining ITC. Moreover, we highlight that the lattice thermal conductivity of substrates does not always play a significant positive role in determining ITC when cooling HEMT devices. However, higher lattice thermal conductivity substrates indeed cause higher magnitudes of heat flux transporting at the interface. This work demonstrates the power of GNN models and paves the way for high-throughput screening of novel crystalline materials with desirable high ITC for phonon-mediated thermal management of wide bandgap electronics.