Machine-learning interatomic potentials: A new avenue towards modelling of energy-storage materials

点击翻译按钮获取中文摘要

The objective of this project is to develop machine learning interatomic potentials for the ternary Li-Si-O system and apply them in molecular-dynamics simulations to improve the understanding of the microstructure and material properties of silicon monoxide with particular focus on the (de)lithiation behaviour. In doing so, we will benchmark different classes of machine learning (ML) potentials, evaluate their computational speed, numerical accuracy and transferability (extrapolation capabilities and finally classify them, accordingly. In a first step, we will extend our recently-developed Gaussian approximation potential (GAP) for SiO2 to the full binary Si-O system (SiO_X ). We will use and extend our training database obtained from density functional theory (DFT) calculations. Using this extended dataset we will train alternative ML potentials, such as neural network potentials (NNP), moment tensor potentials (MTP) and atomic cluster expansion potentials (ACE). This will allow for a non-biased comparison between different ML approaches and guide our selection for the ternary system. Moreover, the comparison will provide a non-biased comparison of the various potential classes performance- Using the this novel potential for the Si-O subsystem, we will then generate different atomistic models for pristine silicon monoxide. The goal is to better understand the micro-/nanostructure of these materials and related mechanical and thermal properties. The full Li-Si-O potential will then be used to study (de)lithiation. Specifically, the simulations will allow us to identify Li diffusion pathways and mobility, reversible Li storage sites and irreversible side reactions leading to formation of secondary phases. Understanding these processes may eventually help guiding experiments to improve cycle life and rate capability by targeted design of silicon-monoxide anode nanostructure.