Decoding Defects: Failure Analysis Using X-ray CT

Several factors need to be carefully considered for effective X-ray CT analysis of batteries:

• X-ray energy and image contrast: The choice of X-ray energy (either bremsstrahlung or characteristic radiation) significantly influences the image contrast. Higher X-ray energies (higher kV) are generally better for larger and denser samples like full battery cells (e.g., 18650 batteries), as they result in less absorption and better visibility of internal features. Lower X-ray energies, like those produced by copper or molybdenum targets, are more suitable for imaging smaller, lower-density samples such as raw battery materials (e.g., cathode or separator materials), as they provide higher absorption and contrast for these delicate structures.
• Spatial resolution and field of view: Spatial resolution, determined by the "voxel" size, dictates the smallest features that can be resolved. A general rule of thumb is that the optimal voxel size should be 1/5 or smaller than the smallest feature being imaged. However, there's a trade-off with the field of view: smaller voxels (higher resolution) result in a smaller field of view and significantly larger file sizes, which can be computationally challenging. A common strategy is to first collect a low-resolution scan of the entire sample to identify regions of interest, and then zoom in to collect higher-resolution data on those specific areas.
• Signal-to-noise ratio (SNR): Sufficient SNR is crucial for accurate data analysis and segmentation. Lower SNR leads to grainy images and unclear feature boundaries, making quantitative analysis difficult. Ways to improve SNR include increasing scan time, shortening the source-to-detector distance (to reduce X-ray absorption by air), and binning pixels (though this reduces resolution).
• Experiment type: The chosen experiment type depends on the research goal. "Ex situ" experiments involve imaging isolated battery materials. "In situ" experiments involve imaging intact battery cells, often repeatedly after charge/discharge cycles. "Operando" experiments take this a step further by collecting data during charge and discharge cycles, requiring very fast scan times.

X-ray CT provides both qualitative and quantitative insights into battery performance and degradation. It allows researchers and manufacturers to understand structural changes that directly impact battery efficiency and lifespan.

At the micrometer to nanometer scale (e.g., electrode materials):

• Quantitative: Porosity, particle size distribution, particle sphericity/shape (length-width measurements), permeability, path of lithium ions, tortuosity, and surface area.
• Qualitative: Anode and cathode designs, separator design, electrolyte distribution, and the behavior of electrodes and internal materials in small battery cells.

At the micron to meter scale (e.g., whole battery):

• Quantitative: Angular deflection of electrodes, electrode thickness, void porosity, delamination volume and thickness, and uniformity.
• Qualitative: Detection of deflected electrodes (e.g., from manufacturing defects or swelling), burrs, internal impurities, and changes in casing due to venting mechanisms.

Ultimately, this data helps link performance loss (e.g., capacity reduction) to structural changes like deformation, delamination, and crack formation. Qualitative information about failure analysis can be extracted from the data.

There are three main approaches to examining battery function with X-ray CT:

• Ex situ experiments: These involve imaging isolated or raw battery materials, such as an electrode layer, outside of a complete battery cell. This is useful for initial characterization of materials.
• In situ experiments: In this approach, an intact battery cell is mounted on the instrument, imaged, then removed for charge/discharge cycles, and subsequently re-imaged. This allows researchers to observe structural changes that occur after specific usage patterns without disassembling the battery.
• Operando experiments: This is the most dynamic type, where data is collected while the battery undergoes charge and discharge cycles. This provides real-time insights into the structural evolution during operation. However, it requires very fast scan times, as the scan must be faster than the process being studied.

After collecting 2D projection images and reconstructing them into a 3D volume, the next crucial step for quantitative analysis is segmentation. This process involves assigning every voxel (3D pixel) in the reconstructed volume to a specific material phase (e.g., anode, cathode, separator, void space, casing). Software tools like VG Studio Max or Dragonfly (which includes deep learning modules for complex segmentations) are used for this.

Once segmented, various quantitative metrics can be extracted:

• Electrode analysis: Curvature of electrodes (e.g., anode overhang), exit angles, thickness.
• Material characterization: Particle size distribution, particle sphericity or shape, porosity, permeability, surface area, and tortuosity (path of lithium ions).
• Defect quantification: Volume and thickness of delamination, void spaces, and other structural defects.

For example, to determine particle size distribution, individual particles within a segmented volume can be separated using techniques like "watershed transformation," allowing for analysis of each particle's dimensions.

The main challenge lies in the inherent trade-off between spatial resolution (voxel size) and the field of view. While smaller voxels (higher resolution) are ideal for resolving fine details (e.g., individual fibers or small particles), imaging an entire large battery at very high resolution (e.g., 1 micron for a 100 mm battery) would result in an impractically enormous file size (e.g., 2,000 terabytes) that current computers and software cannot handle.

Conversely, a larger field of view (to encompass an entire battery) necessitates a larger voxel size (lower resolution), which means smaller features may not be resolved accurately or at all. For instance, a 100-micron voxel size for a 100 mm battery reduces the file size to a manageable 16 GB, but smaller features within the battery might not be clearly visible. This necessitates a compromise, often involving initial low-resolution scans of the whole sample to identify areas of interest, followed by higher-resolution scans of those specific regions.

Yes, deep learning models can be effectively applied in X-ray CT analysis, particularly for complex segmentation tasks. Software packages like Dragonfly include deep learning modules where users can "paint" small areas to label different material phases (e.g., anode, cathode, delamination space). This labeled data serves as "ground truth" to train a deep learning model. Once trained, this model can be applied to the entire dataset to perform accurate and efficient segmentation.

The main benefits of using deep learning models are:

• Automation and efficiency: Deep learning can automate complex and time-consuming segmentation tasks, especially for intricate structures or large datasets.
• Improved accuracy: Well-trained models can often achieve higher accuracy in identifying and segmenting different phases compared to manual or traditional thresholding methods, particularly in challenging cases.
• Consistency and bias reduction: Once a model is trained, it can be applied consistently across multiple images, including those collected as a function of time, temperature, or from different samples. This eliminates potential human bias that might otherwise be introduced during manual segmentation.
• Time-resolved analysis (4D CT): Deep learning models can facilitate the analysis of time-resolved (4D) CT scans by segmenting corresponding features across different time points, allowing for tracking of critical performance parameters and structural evolution over time.

To ensure optimal performance and comparability when applying a trained model across multiple images, it's recommended to normalize histogram intensities by calibrating the CT scan intensities and ideally collect all data with the same settings (e.g., voxel size).