Deep Learning Analysis of Solid-Electrolyte Interphase Microstructures in Lithium-Ion Batteries

Abstract

A comprehensive understanding of the solid-electrolyte interphase (SEI) in lithium-ion batteries is crucial for improving energy efficiency, battery performance, and safety. In this study, a transformer-based instance segmentation framework, integrating deep convolutional neural networks is introduced with a feature pyramid network (FPN), to quantitatively analyze High-Resolution Transmission Electron Microscopy (HRTEM) images and explain the complex microstructural features of the SEI. The model is trained on a dataset of simulated HRTEM images generated using Density Functional Theory (DFT)-optimized grain boundary (GB) structures and calibrated with experimental microscope parameters. The model achieves robust segmentation performance, with training and validation mean intersection over union (mIOU) values of 0.98 and 0.96, respectively. On unseen test data, the model attains mean area match (AM) scores of 91.4% for GBs, 92.3% for Li2CO3, 91.7% for LiF, 88.7% for LiOH, and 88.6% for Li2O. These quantitative results highlight the model's high fidelity and its ability to capture subtle variations in crystallographic orientations and material contrasts. By enabling detailed, statistically grounded segmentation of SEI components, the approach offers valuable insights into ion transport and degradation mechanisms, paving the way for more resilient and efficient energy storage solutions.