Simulations enabled by machine learning potentials predict sulfide solid electrolyte with covalent network structure and glassy behavior

Machine learning (ML) accelerated simulations were used to explore how the isovalent substitution of oxygen by sulfur in LiGa(SeO3)2 alters its ionic conductivity. Molecular dynamics (MD) driven by a custom-trained neural network (NN) potential were scaled to simulate 5,120 atoms for 100 nanoseconds in order to identify and characterize a unique structural prototype for glassy sulfide solid electrolytes with the stoichiometry LiGa(SeS3)2, featuring GaS4 tetrahedra interconnected by bridging Se-S covalent chains. A phase transition from a crystalline phase to a structurally-disordered phase of LiGa(SeS3)2 was identified in the simulations between 325 and 350 K, whereby the molar volume and ionic conductivity significantly increase, while the vibrational density of states spectra for Li+ becomes characteristically broadened, similar to superionic conductors. The ionic conductivity is calculated to be around 10^-2 S/cm at 350 K. Furthermore, this glassy sulfide solid electrolyte exhibits an anomalously low bulk modulus of 0.7 GPa, which may be beneficial for accommodating plasticity in the solid electrolyte of an all solid-state inorganic battery. The electrochemical stability window is predicted to be relatively narrow since the band gap is conservatively 1.45 eV, but the stability window could potentially be tuned by doping to form an oxysulfide.