Simulations of silicon-graphene anodes using machine-learning-based interatomic potentials

Abstract

Silicon-based anodes hold great promise for next-generation lithium-ion batteries (LIBs) due to their high energy density. However, their practical application is hindered by significant volume changes during cycling, leading to numerous issues, such as mechanical degradation and capacity fading. Carbon composite, such as graphene or amorphous carbon, incorporation into silicon anodes has emerged as a potential solution to these challenges, but a complete simulation of the silicon-carbon systems with a full lithiation-delithiation cycle remains unexplored. This research approaches the aim by developing a Machine Learning Interatomic Potential (MLIP) using the Gaussian Approximation Potential (GAP) framework. The study represents the first comprehensive effort to model the full electrochemical cycle of a silicon-graphene anode system, aiming to provide atomic-scale insights into structural stability, lithium transport, and the interactions within the silicon-lithium-carbon system. The research methodology relies on Density Functional Theory (DFT), Ab Initio Molecular Dynamics (AIMD), and Machine Learning (ML) techniques to construct accurate and scalable GAP interatomic potentials. By systematically increasing the complexity of the model, this study investigates pure silicon, lithiated silicon, and silicon-graphene and silicon-amorphous carbon composite interface systems. Key findings include the successful simulation of silicon melt-quench transitions, lithiated silicon cycling, and the role of graphene and amorphous carbon in stabilising silicon structures. The newly developed potentials, including the Si-Li-GAP22-P1 and Si-Li-C-GAP23 models, demonstrate good agreement with relevant experimental and theoretical data and provide detailed analyses of formation enthalpies, voltage profiles, silicon clustering, coordination states, and lithium diffusion properties. In bulk systems, simulations revealed crucial insights into the structural and kinetic mechanisms during lithiation and delithiation, with the Si-Li-GAP22-P1 potential accurately capturing the aggregation and fragmentation of silicon clusters. For surface system simulations with Si-Li-C-GAP23, the incorporation of carbon-based layers, especially pristine graphene, was shown to improve cycling stability compared to other carbon-based layers, such as amorphous carbon, by more effectively mitigating void formation and providing superior mechanical support. Comparative analyses of Si-Li-C-GAP23 with universal potentials like MACE-MP-0 underscored the importance of domain-specific training, as general-purpose models did not fully capture critical behaviours in silicon interfaces. This work provides a robust computational framework for studying silicon-graphene-based anodes, bridging the gap between theoretical models and experimental observations. The insights gained are vital for optimising silicon-graphene composites and advancing LIB technology. Future research will explore further refinements of MLIPs and their applications to more complex battery materials.