(192i) Data-Driven Design of Solid-State Electrolytes

Solid state batteries provide many safety and performance advantages by replacing the combustible organic liquid electrolyte in a lithium-ion battery with a ceramic solid-state electrolyte (SSE). However, the ionic conductivity in these SSEs is often several orders of magnitude lower than in their liquid counterparts. Predicting a material’s ionic conductivity directly from its crystal structure is limited by data availability and incomplete material descriptors. First-principles (i.e., with density functional theory, or DFT) molecular dynamics (MD) is an established approach to calculate and study ionic conductivity but is limited in the number and type of materials that can be simulated due to the high computational cost. To this end, we leverage advanced machine learning (ML) algorithms to more efficiently calculate ionic conductivity and optimize material composition.

We accelerated the calculation of forces and energies in MD with an artificial neural network force field, which scales linearly and enables the calculation of ionic conductivity at experimentally relevant compositions. While these force-fields are system specific, we show the valence electron density is an effective universal descriptor, allowing direct calculation of ionic conductivity from the crystal structure with a partial least squares algorithm. To augment the limited number of training samples, we trained additional ML models to predict the lithium probability density. Our novel 3D material segmentation network provides both quantitative and qualitative insight into the topology of diffusion pathways to accelerate SSE design. From these models, we predict diffusion activation barriers in materials which are inaccessible with traditional DFT methods and identified several new promising classes of solid-state electrolyte candidates to have conductivities greater than 16 mS/cm as verified by DFT-MD simulations.

The proposed combinations of first principles simulation data with ML models will greatly accelerate the rate of materials design/discovery and can automate calculating structure-property relationships for other solid-state systems with high accuracy and without sacrificing interpretability.