Comparison of training approaches for machine-learned interatomic potentials to study diffusion in solid-state electrolytes: application to Li₉S₃N

As society moves towards decarbonization, the rising demand for safe and efficient energy storage devices asks for new solutions beyond current Li-ion technology, among them solid-state electrolytes. Realistic numerical simulations of ionic diffusion in solid-state batteries require large simulation boxes and/or very long timescales which rapidly become intractable for first-principle (FP) calculations. To lift this limitation, we turn to machine-learned interatomic potentials (MLIP) that combine the accuracy of FP calculations and the numerical efficiency of empirical force fields.

Here, we present a comparative study of different training approaches to generate MLIP describing the potential energy surface of Li₉S₃N and assess their performance for predicting the activation energy and diffusion mechanisms. We first compare the Moment Tensor Potentials (MTP) [1] and SchNet models [2] passively trained on configurations extracted from a FP molecular dynamics simulation. We then investigate whether the built-in active learning algorithm of the MTP framework improves the passively trained potentials and compare with a fully actively trained MTP model. Our work provides insight on how to optimize the numerical cost of generating training sets for complex systems.

[1] Mach. Learn.: Sci. Techn. 2 025002 (2021)

[2] J. Chem. Theo. Comp. 15 448-455 (2019)