Data-Driven Fast Prediction of Superionic Conductors from Phonon Features

Superionic conductors are one of the most important components for the development of all-solid-state batteries (ASSBs) with improved safety and performance. High-throughput computational approaches, which involve screening large databases of diverse and vast compounds, can accelerate the discovery of new materials with descent ionic conductivity. However, directly calculating ionic conductivity of large-scale structures from first principles has not been implemented, because of the unbearable computational costs. Phonons, the lattice vibration in crystalline materials, are ubiquitous and at the center of materials science such as solid-state physics, quantum communication, phase stability, as well as other emerging applications in modern science and technology. Lattice vibrations are the hosting media for many transport properties, including superionic conductors. Despite the central role of phonons, to date little research has systematically assessed the role and correlation between phonon features and superionic conductivity. With recently developed universal machine learning potentials, we first performed high-throughput computation of diffusion coefficients of large amount of potential superionic conductors screened from existing material databases. We identified some strong correlations between diffusion coefficients and both basic and derived phonon features, including but not limited to acoustic cutoff frequencies, mean square displacements. More importantly, those phonon features can be easily trained by the state-of-the-art machine learning models, such as graph neural network (GNN). With further trained GNN models, screening potential superionic conductors can be accelerated by orders of magnitude as compared to traditional first principles calculations.