Simulating Diffusion Properties of Solid-State Electrolytes via a Neural Network Potential: Performance and Training Scheme^†

Graphical Abstract

That's electric: We develop a scheme to train artificial neural networks (ANN) for molecular dynamics (MD), avoiding overfitting and reducing to a minimum the number of configurations used. An initial approximate model is trained on a small set of configurations and used to start the training loop. The training set is then iteratively augmented till the desired property becomes stationary. We test our scheme on solid-state electrolyte materials for battery applications, focusing on the evaluation of the diffusion coefficient.

Abstract

The recently published DeePMD model, based on a deep neural network architecture, brings the hope of solving the time-scale issue which often prevents the application of first principle molecular dynamics to physical systems. With this contribution we assess the performance of the DeePMD potential on a real-life application and model diffusion of ions in solid-state electrolytes. We consider as test cases the well known Li₁₀GeP₂S₁₂, Li₇La₃Zr₂O₁₂ and Na₃Zr₂Si₂PO₁₂. We develop and test a training protocol suitable for the computation of diffusion coefficients, which is one of the key properties to be optimized for battery applications, and we find good agreement with previous computations. Our results show that the DeePMD model may be a successful component of a framework to identify novel solid-state electrolytes.