Investigation of the Conductivity of Some Lithium (Thio)Boracite Materials Using the Allegro Machine Learned Interatomic Potential Package

Previously [1] [2], we have investigated several lithium (thio)boracite solid-state electrolytes using first principles methods to determine their ground state structures, phonon spectra, and stability with respect to chemical decomposition and in contact with a lithium reservoir.  These materials are promising lithium-ion conductors as they have large void regions in their framework structures which facilitate lithium-ion diffusion.  We have qualitatively seen that the chemical variations among these materials may improve ionic conductivity.  Typical quantitative methods of approximating ionic conductivity involve running Molecular Dynamics (MD) simulations for very long times to perform the needed ensemble averaging [3].  For simulations with large numbers of atoms per unit cell, First Principles MD falls short, achieving on the order of one ps of simulation time per day.  Machine Learned Interatomic Potentials (MLIPs) promise near FPMD level accuracy with massive increase in simulation speed.  Here we investigate using MLIP models trained using the Allegro package [4] to approximate the ionic conductivities of several of the lithium thio(boracite) materials.

[1] Li, et al. 2022. DOI: 10.1103/PhysRevMaterials.6.025401

[2] Lynch, et al. 2024. DOI: 10.1103/PhysRevMaterials.8.065401

[3] Marcolongo, et al. 2017. DOI: 10.1103/PhysRevMaterials.1.025402

[4] Musaelian, et al. 2023. DOI: 10.1038/s41467-023-36329-y