Neural Network-based Simulations of Solid Electrolyte Interfaces

The solid-electrolyte interface (SEI) is a crucial component of electrochemical devices, but its computational description remains challenging due to the many chemical reactions that lead up to its formation. In this paper, we use a neural network forcefield, NeuralIL, to carry out molecular dynamics simulations of the organic layers of a model SEI based on ethylcarbonate and different metal cations, namely lithium, sodium and potassium. The need for an explicit interface is bypassed by introducing reduced electrolyte molecules into the system, whose presence starts a subsequent reaction chain. A series of reaction pathways are identified, leading to a predicted composition of the SEI that is in very good agreement with experimental data from the literature. Moreover, analysis of the structural and dynamic properties reveals key differences between the different metals, such as the presence of additional reactions and chemical species, shifts in the charge distribution of the different molecules, changes in the solvation shells of the metal cations and their impact on the diffusion coefficients.