Large-scale Atomistic Simulations of Lithium Diffusion in a Graphite Anode with a Machine Learning Force Field

Graphite is the anode material in the vast majority of current commercial lithium-ion batteries (LIBs) due to its excellent electrochemical performance and abundant reserves. However, the development of high-performance LIBs still faces technical bottlenecks including limited rate capability, partly because of the slow Li diffusion in the graphite anode. Experimental measurements of Li diffusivity versus Li contents in graphite show both monotonic and non-monotonic trends with reported Li diffusivity spanning orders of magnitude. Comprehensive insights of the Li diffusion process and its bottleneck in graphite are essential to develop the next-generation LIBs. In this work, we developed a machine learning force field (MLFF) to investigate the Li diffusion in the LixC6 system. We benchmark density functional theory (DFT) functionals and dispersion corrections versus experiment, finding nonlocal van der Waals functional rVV10 shows a good agreement with experimental results in terms of the structural, energetic, and electrochemical properties of the LixC6 system. We then train a MLFF based on our recently developed charge recursive neural network (QRNN) architecture to simulate Li diffusion in graphite at different Li contents, stage structures and temperatures. Our Li diffusion analysis demonstrates a phase-transition dependent Li diffusion in the LixC6 system, which supports the experimental measurements of a non-monotonic relation between Li diffusivity and Li content. This work demonstrates the capability of our QRNN model in carrying out large scale molecular dynamics simulations to identify the Li diffusion bottleneck in a graphite anode.