Sub-Trillion-Scale Atomic Data Integrated Deep Learning Approach for Predicting Thermal Transport Properties of 60,000 Inorganic Perovskites

Although first-principles based anharmonic lattice dynamics method coupled with the phonon Boltzmann transport equation has been developed to obtain the phonon properties including lattice thermal conductivity at highest accuracy ever, the costly and time-consuming nature of the required interatomic force constants calculations renders high-throughput infeasible when facing tens of thousands of new materials. Here, a high-dimensional multi-element deep neural network with sub-trillion-scale atomic data is trained, dubbed Elemental Spatial Density Neural Network Force Field (Elemental-SDNNFF), achieving a competitive force root mean square error and a speed-up of 4 to 5 orders of magnitude in comparison to first principles. The effectiveness and precision of the Elemental-SDNNFF approach is demonstrated on a set of 150,000 inorganic crystalline structures spanning 63 elements in the periodic table by prediction of complete phonon properties such as phonon dispersions and lattice thermal conductivity. We then use our trained neural network model to predict and screen full phonon properties of 60,000 single and double perovskite structures. Due to the inherent structural feature of dynamical disorder, the perovskite structures have shown diverse thermodynamic and phonon transport properties. Dynamic stability of all 60,000 perovskite structures is predicted by screening negative frequencies in the Brillouin zone. Four-phonon scatterings and two-channel thermal transport are also analyzed and screened. The underlying mechanism is analyzed in deep at the electronic level. This study demonstrated that our algorithm is very powerful for predicting phonon properties of large-scale inorganic crystals and is also promising for accelerating high-throughput search of novel phononic materials for emerging applications.