Ming Hu1
University of South Carolina1
Ming Hu1
University of South Carolina1
Existing machine learning potentials for predicting phonon properties of crystals are limited to either small amount of training data or a material-to-material basis, primarily due to the exponential scaling of model parameters with the number of atomic species or elements. This renders high-throughput infeasible when facing large-scale new materials. Unlike previous machine learning models with inherently limited training on small data and global properties, we develop Elemental Spatial Density Neural Network Force Field (Elemental-SDNNFF) with abundant atomic level environments as training data. Benefiting from this innovation, we integrate ~30 million atomic data to train a single deep neural network without increased expensive ab initio simulations. The effectiveness and precision of the Elemental-SDNNFF approach is demonstrated on a set of ~100,000 inorganic crystals spanning 63 elements in the periodic table by prediction of complete phonon properties. Our algorithm achieved a competitive force root mean square error of less than 10 meV/Å and a speed-up of 3 – 4 orders of magnitude in comparison to first principles. Self-improvement schemes such as active learning and data augmentation techniques are also incorporated allowing human-free refinement of the force accuracy on arbitrary combinations of chemistries. As case studies of our predicted phonon properties, deep insight into the ultralow lattice thermal conductivity (<1 W/mK) of predicted structures are gained by p-d orbital hybridization analysis and charting the thermal conductivity data in the bonding-antibonding map, which offers a quick approach for future screening of crystals with strong intrinsic phonon anharmonicity. Our algorithm is promising for accelerating discovery of novel phononic crystals for emerging applications, such as heat dissipation in electronics, thermoelectrics, and coherent phonons for quantum information technology.