Lithium metal is a promising anode material for batteries due to its high theoretical capacity and thermodynamic potential. However, its instability leads to side reactions, solid electrolyte interphase (SEI) formation, and dendrite growth, hindering the development of solid-state lithium batteries (SSBs). Previous research on SSBs often focused on small-scale observations of dendrite growth. This study aims to provide a more comprehensive understanding of plating and stripping dynamics in SSBs at a larger scale (micron-scale) to improve battery performance. Operando X-ray μCT offers the necessary spatial resolution, but analyzing the resulting large datasets presents a significant challenge. Manual labeling is infeasible, and the low X-ray attenuation of lithium makes accurate segmentation difficult using traditional thresholding methods. This research leverages the power of machine learning, specifically a deep learning architecture called U-Net, to address these limitations. U-Net, originally developed for biomedical image segmentation, has been adapted for 3D applications and shown effectiveness in segmenting various structures in materials science. This paper uses operando X-ray μCT to track lithium dynamics in a solid-state Li-metal symmetric pouch cell during one cycle, collecting high-resolution 3D image volumes at multiple time steps. An iterative Residual U-Net-based segmentation algorithm is employed to classify components. This automated technique allows for quantitative analysis of lithium morphology, providing crucial information for battery design and quality control.

Existing literature highlights the challenges associated with lithium metal anodes in solid-state batteries, including dendrite formation and SEI layer growth. Operando X-ray μCT has been used to study these processes, but previous studies often involved small active areas, leading to edge effects and artifacts. Machine learning, especially deep learning, has emerged as a powerful tool for analyzing large-scale imaging data, with U-Net architectures demonstrating success in biomedical and materials science applications. However, applications to large-scale, operando μCT data of solid-state batteries for detailed lithium morphology segmentation remain relatively unexplored. This study builds upon this existing work by applying a sophisticated deep learning model to address the unique challenges presented by operando X-ray μCT data of Li-metal batteries.

A Li foil/polyelectrolyte/Li symmetric pouch cell with a 0.5 cm² active area was cycled using a galvanostatic intermittent titration technique (1.5 mA cm⁻² for 1 hour, followed by a 20-minute rest period). Operando X-ray μCT scans were collected during the rest periods (25 scans total) at Beamline 2-BM of the Advanced Photon Source, utilizing a monochromatic synchrotron beam (27.5 keV) and achieving a 1.33 μm pixel size. The raw data underwent preprocessing, including alignment and registration, followed by region of interest (ROI) selection. An iterative Residual U-Net-based segmentation algorithm was trained to segment the images into five classes: dendrites, pits, deposited Li (during charging), redeposited Li (during discharging), and background. The iterative training involved an initial model trained on a small set of hand-labeled data, with subsequent iterations refining the segmentation using predictions from previous models. The final model utilized 42 hand-labeled cross-sectional slices across 15 scans, achieving 98.8% testing accuracy and 95.5% overall intersection over union (IoU). Quantitative analysis was performed on the segmented data, focusing on volume and thickness changes of the various lithium components, electrode thickness, and spatial relationships between different components. Data processing and visualization were performed using Fiji® and ORS Dragonfly software.

Operando X-ray μCT and the batternYNet model revealed dynamic changes in lithium morphology during battery cycling. Dendrites formed during charging, while deposited Li appeared on the plating side and pits formed on the stripping side. During discharging, deposited Li decreased, redeposited Li filled pits, and some dendrites remained, categorized as 'dead Li' (1.11 x 10⁷ µm³). The volume of dendrites increased during charging and decreased during discharging, showing more dendrites at the start of discharge than at the end of charging. Analysis showed a net Li volume change of 16.5 × 10⁻³ mm³ during the cycle, with a variation of 1.94 × 10⁻³ mm³, possibly due to segmentation errors related to porous Li structures. The effective thickness of electrodes changed as expected, with the plating electrode growing and the stripping electrode shrinking, but net changes were small. The rate of increase/decrease in electrode thickness varied, potentially due to the integration of newly plated Li into the existing porous structure and the accumulation of dead Li creating a tortuous Li+ pathway. 2D spatial distribution analysis revealed that almost all deposited Li disappeared after one cycle, suggesting conductivity; dendrites occupied a significant portion of the ROI area (30% at half cycle, 12% after full cycle). Pits were formed during charging and filled with redeposited Li during discharging, but the filling was not compact due to void formation between the redeposited Li and electrode. Analysis of void space after a full cycle showed a significant unoccupied area (53.9%), indicating limited contact between the electrode and solid electrolyte. Finally, the spatial location of deposited Li and dendrite volume decrease did not correlate with the location of maximum redeposited Li increase, suggesting that Li-ions do not travel in straight lines but follow more complex pathways.

The findings of this study directly address the limitations of previous research by providing a high-resolution, large-scale observation of lithium plating dynamics in solid-state batteries. The development of batternYNet allows for the automated and quantitative analysis of operando X-ray μCT data, overcoming the challenges of manual segmentation and enabling a more comprehensive understanding of the complex morphological changes during cycling. The quantitative data on lithium volume and electrode thickness changes, along with the spatial relationships between different components, provide critical insights into battery performance and failure mechanisms. The discovery of significant void formation between the redeposited Li and electrode highlights a potential factor limiting battery cycling life and efficiency. This detailed analysis is crucial for designing improved Li-metal batteries with enhanced performance and longer lifespan.

This research successfully demonstrated the application of a machine learning-powered segmentation algorithm, batternYNet, to analyze operando X-ray μCT data of a solid-state lithium-metal battery. The detailed quantitative analysis of lithium morphology provided crucial insights into the dynamic processes occurring during charging and discharging, including dendrite formation, Li plating/stripping, and void formation. This work contributes significantly to the understanding of Li-metal battery performance, informing future designs and quality control strategies. Future research could focus on extending this methodology to different battery chemistries and operating conditions, as well as exploring more advanced machine learning techniques to improve segmentation accuracy and efficiency.

The study focused on a single type of solid-state battery and cycling conditions. The observed morphological changes and quantitative measurements might vary with different electrolytes, electrode materials, and cycling protocols. While batternYNet demonstrated high accuracy, the segmentation may still be influenced by noise and artifacts in the μCT data. The assumption of flat surfaces in calculating electrode thickness is a simplification, potentially impacting accuracy. Finally, the study was limited to one full cycle, and further investigation is needed to observe long-term behavior.