Detecting lithium plating dynamics in a solid-state battery with operando X-ray computed tomography using machine learning

Li metal is a promising candidate as a battery anode material due to its high theoretical capacity (3860 mAh g⁻¹) and high magnitude of thermodynamic potential (−3.06 V vs. SHE). In practical use, Li metal is generally not stable, leading to various side reactions, the formation of solid electrolyte interphase (SEI), and dendrite formation. Solid electrolytic materials with solid-state Li conductors can form a more stable interface with Li metal. Li-metal solid-state batteries (SSBs) have higher energy density compared to commercial Li-ion batteries, reducing battery weight and size. However, dendrite formation caused by heterogeneous Li metal plating hinders development of SSBs. Prior research largely examined dendrites or their growth in small nanoscale fields of view, but a micron-scale, larger field-of-view understanding of plating and stripping dynamics is needed to improve performance. X-ray μCT has the spatial resolution to capture dendrite formation. Operando CT has been used to study Li metal plating in symmetric Li polymer electrolyte cells, initiation and plating in Li-graphite cells with polymer electrolyte, and void formation during stripping at Li/solid-electrolyte interfaces. However, many prior studies employed small active-area cells and plastic housings (e.g., PEEK) to reduce X-ray absorption, which can introduce edge effects, artifacts, shading, and variable airtightness. Here, the authors develop a multi-resolution synchrotron X-ray μCT method with in-line cloud computing, enabling high-throughput operando imaging while maintaining electrochemical performance comparable to larger, practical cells. A major analysis challenge is the size of operando CT datasets (tens of gigabytes per scan; multi-time stacks exceeding 26 billion voxels), making manual labeling infeasible. Additionally, accurate segmentation of Li is difficult due to Li’s low X-ray attenuation and minimal contrast between Li voids and other Li components, rendering thresholding methods ineffective. Deep learning, particularly U-Net-based architectures, can achieve high-quality segmentations in low-contrast, noisy images by leveraging multi-resolution spatial context; 3D extensions are widely applied in medical and materials imaging. Machine learning has been increasingly used to accelerate CT analysis of batteries, mostly focused on electrodes/cathodes. In this work, operando X-ray μCT tracks Li dynamics in a solid-state Li-metal symmetric pouch cell across one cycle. Using a monochromatic synchrotron beam, 25 time-resolved 3D volumes (1.33 μm pixel) of the same SSB were collected. An iterative Residual U-Net-based approach segments each volume into five classes: dendrite, pit, deposited Li (during charging), redeposited Li (during discharging), and background. The authors analyze temporal changes of each Li-related component and spatial correlations among components, providing an automated, scalable segmentation and quantitative morphology analysis pipeline with applications to battery quality control.

Cell and cycling: A Li foil/polymer electrolyte/Li symmetric pouch cell (active area 0.5 cm²) was assembled. Free-standing Li metal foil (~100 μm) and a polymer electrolyte membrane (~140 μm thickness; research sample from Ionic Materials) were used. A 50 μm shim defined a 0.8 cm diameter region; two 1.1 cm diameter polymer electrolyte disks were placed on each side of the shim and sandwiched by two Li electrodes. External tabs were attached; the pouch (1.5 × 3.5 cm²) was vacuum-sealed. The cell was cycled at 1.5 mA cm⁻² (3.0 mAh cm⁻²), with current applied for 1 h segments and ~20 min rest periods when CT scans were acquired. No external pressure was applied. Operando μCT imaging: Experiments were conducted at APS 2-BM. Imaging used a 20 μm LuAG scintillator, X-ray lenses, and a sCMOS PCO Edge camera. Pixel size was 1.33 μm; field of view 34 mm. A multilayer monochromator selected 27.5 keV X-ray energy. Three overlapping fields of view were recorded and stitched to give a vertical height >3 mm. Per scan: 1500 projections over 180° with 100 ms exposure. Tomographic reconstructions used TomoPy with Gridrec; algorithm parameters followed prior work. Data volume and acquisition schedule: The same pouch cell was imaged at 25 time steps during one full charge–discharge cycle. Scans were taken during the rest periods to reduce concentration-gradient effects and enable Li dissolution/deposition between current pulses. Preprocessing and region-of-interest selection: Raw scans were preprocessed (alignment, registration, stitching), and regions of interest (ROIs) were defined focusing on the electrode–electrolyte interfaces. The workflow encompassed alignment/registration, ROI selection, segmentation, and quantitative analysis. Component definitions for segmentation and analysis: Five phases were targeted: dendrites (Li protrusions exceeding the uniformly plated layer), pits (voids left by nonuniform stripping on the stripping side), deposited Li (uniform plating on the plating side during charge), redeposited Li (Li filled into pits on the stripping side during discharge), and electrodes (bulk Li). Redeposited Li and pits appear only on the stripping side; deposited Li only on the plating side. Iterative Residual U-Net training: Due to the dataset scale and low contrast for Li, an iterative labeling and training approach was adopted. A small set of manual labels initiated training to segment pits, Li, and background; predictions seeded further hand-labeling. To improve efficiency, phases were labeled in separate binary tasks (e.g., deposited Li, redeposited Li, electrode), then combined into multi-class labels. Early labels were made in 2D in-plane slices near electrodes (morphologically diverse), while some distinctions (e.g., electrode vs electrolyte) benefited from cross-sectional context. A cross-sectional model was refined using 42 hand-labeled cross-sectional slices across 15 of the 25 scans, improving classification in ambiguous regions. Model training environment and procedure: Training and data processing were conducted at NERSC on a single 4-GPU Perlmutter node using PyTorch distributed data parallel. Early networks used partial, imperfect labels; data splits were 85/15% train/validation. Training was time/plateau limited (stopped after 1 h or when validation accuracy failed to improve for 10 iterations). Final multi-class labels (25 cross-sectional slices across 1 of the 25 scans, refined from previous predictions) were split 72/13/15% into train/validation/test, producing the final model with testing accuracy 98.8% and overall IoU 95.5% (dendrites and deposited Li had lower IoU due to class imbalance). Post-segmentation processing and visualization used Fiji and ORS Dragonfly (2022.1). Quantitative analyses: From segmented volumes, the team measured component volumes, effective electrode thicknesses (excluding dendrites/deposited or redeposited layers as appropriate), spatial distributions via 2D projections of flattened 3D data, electrode displacement (edge tracking), and spatial correlations/colocalization between plating- and stripping-side features. Traveling Li volume from imaging was compared with electrochemical estimates via Faraday’s law. Battery swelling was inferred from electrode gap changes.

- Developed an iterative Residual U-Net pipeline (batternYNet/batteryNET) that semantically segments large operando μCT datasets into five classes (dendrites, pits, deposited Li, redeposited Li, electrodes) with high accuracy (testing accuracy 98.8%; overall IoU 95.5%). - Temporal evolution: During charging, dendrites grew and deposited Li formed on the plating side while pits formed on the stripping side. During discharge, deposited Li on the plating side decreased while redeposited Li filled pits on the stripping side. - Dead Li quantification: After a full cycle, some dendrites remained; Li volume of dead Li was 1.11 × 10⁷ μm³. - Electrode thickness dynamics: Plating-side (electrode + deposited Li) increased from 107.0 to 140.1 μm during charging, with growth slowing after 60 min and nearly stopping after ~80 min (indicative of plated Li integrating into a porous structure unresolved at ~1 μm resolution). Stripping side (stripping electrode during charge) decreased from 112.9 to 101.7 μm at ~0.93 μm per 10 min charging segment; during discharge, the stripping side (electrode + redeposited Li) increased at ~2.47 μm per 10 min discharging segment, ending at 130.5 μm, exceeding the pristine 112.9 μm. - Dendrite/deposited Li spatial coverage: Deposited Li coverage area dropped from 98.4% (post-charge) to 0.16% (post-discharge), indicating most deposited Li remained conductive and moved. Dendrite area occupied ~30% of the ROI at half cycle and ~12% after a full cycle, indicating partial persistence/activation through microstructural connections. - Electrode displacement and swelling: The gap between plating and stripping electrodes increased by ~39.3 μm after charging, evidencing cell swelling (likely due to interphase oxidation or void formation). Little displacement occurred during discharge. - Pit filling and contact: After one full cycle, nearly the entire pit area was filled by redeposited Li; however, voids between redeposited Li and the stripping electrode occupied 53.9% of the ROI, implying only 46.1% electrode–electrolyte contact area. - Transport pathways: No strong spatial correlation existed between regions of largest decrease in deposited Li/dendrites on the plating side and largest increase in redeposited Li on the stripping side, suggesting tortuous Li-ion transport likely influenced by membrane features/defects. - Imaging vs electrochemistry discrepancy: Traveling Li estimated from electrochemistry exceeded that from μCT segmentation; the missing capacity could be >70% after one full cycle, attributed to segmentation limits and unresolved porosity.

The study demonstrates that operando μCT combined with iterative deep learning segmentation can resolve and quantify the dynamic evolution of Li-related morphologies in a realistic, sealed solid-state pouch cell across a full cycle. By separating dendrites, deposited Li, pits, and redeposited Li, the analysis reveals expected charge/discharge trends and provides spatially resolved evidence of mechanisms that limit performance. Dendrite formation during charging and their partial persistence after discharging (dead Li) reflect heterogeneity in plating and incomplete reconnection during stripping. The large reduction of deposited Li coverage upon discharge indicates that much of the plated Li remains electrochemically connected, moving back to the stripping side, whereas persistent dendrites contribute to inactive Li. Thickness trends—growth of the plating side and shrinkage of the stripping side during charge, with the inverse during discharge—confirm net Li transfer, but the final over-thickness of the stripping side coupled with significant voiding indicates noncompact redeposition and poor interfacial contact. The lack of spatial correlation between plating-side depletion and stripping-side accumulation, along with the increased electrode gap and significant void fraction at the interface, point to tortuous Li-ion transport pathways and interfacial degradation that reduce effective active area. These insights explain observed polarization behavior and mass-transport-limited kinetics during intermittent galvanostatic operation. Overall, the segmentation-driven quantification directly addresses the research need to understand Li plating/stripping dynamics at micron scales in larger-format cells, informing designs and operating conditions that mitigate dendrites, improve contact, and reduce swelling.

This work introduces an operando μCT and iterative Residual U-Net segmentation pipeline that enables high-throughput, accurate, multi-class semantic segmentation of large 3D datasets from a solid-state Li-metal pouch cell. The method quantitatively tracks dendrites, deposited Li, pits, and redeposited Li over a full cycle, uncovering key behaviors: tortuous transport decoupling plating and stripping regions, significant voiding at the redeposition interface (reduced contact area), battery swelling (electrode gap increase), and persistence of dead Li. These findings provide actionable insights for SSB design, interfacial engineering, and quality control, and the segmentation approach is transferable to other operando datasets. Potential future directions include: improving spatial resolution and contrast to resolve sub-micrometer porosity; expanding labeled datasets to boost performance on rare classes (dendrites, thin deposits); integrating physics-informed constraints to reduce imaging–electrochemistry discrepancies; and correlating segmentation with varied electrolytes, pressures, and cycling protocols to generalize the observed dynamics.

- Imaging resolution (~1.33 μm) cannot resolve sub-micrometer porosity, affecting segmentation accuracy and volume/thickness estimates. - Low X-ray attenuation of Li yields low contrast; traditional thresholding fails and deep learning still faces ambiguity in thin/rare classes (lower IoU for dendrites and deposited Li). - Segmentation errors likely contribute to discrepancies with electrochemical estimates (missing capacity potentially >70% after one cycle). - Training labels were limited and iteratively refined; class imbalance impacts accuracy for less prevalent phases. - Operando scans were acquired during rest periods; dynamics during current flow are inferred indirectly. - Specific cell construction (no external pressure, particular polymer electrolyte) may limit generalizability to other SSB chemistries or mechanical boundary conditions.