Deep learning-based segmentation of lithium-ion battery microstructures enhanced by artificially generated electrodes

The performance of lithium-ion batteries (LIBs) is strongly influenced by the morphology of their constituent materials. The pore structure and the distribution of the carbon black-binder domain (CBD) around active particles significantly affect ionic and electronic resistance, impacting capacity, rate capability, and cycle life. Accurate 3D representations, where different material phases are segmented and labeled, are crucial for material selection, process optimization, and parameter tuning. X-ray-based tomographic analysis offers a versatile approach for 3D reconstruction, but challenges remain due to the varying length scales, low contrast between components, and low attenuation of carbon-based materials. Active materials like Li(Ni,Mn,Co)O2 provide good contrast, but graphitic materials commonly used in negative electrodes do not, hindering accurate segmentation. High-resolution techniques offer detailed information but have limited fields of view, making it difficult to obtain statistically relevant data on the electrode scale. This study addresses these challenges by using supervised deep learning for semantic segmentation of high-resolution volumetric image data of LIB electrodes.

While deep learning has been applied to battery state-of-health assessment, microstructural design improvement, and defect detection, microstructure segmentation still often relies on basic filtering and thresholding. Although machine learning is beginning to be used for segmentation, crack detection, and particle detachment, deep learning approaches are mainly used in high-contrast systems (cathodes), and algorithms like those used in medical imaging for full-body scans have not been applied to LIB electrodes. X-ray tomographic microscopy (XTM) offers a good balance between field of view, resolution, and acquisition time for LIB component imaging but still presents challenges for accurate segmentation due to low contrast and diverse material phases.

This research utilizes the 3D U-Net architecture for semantic segmentation of volumetric images of graphite-silicon composite negative electrodes obtained using XTM. A key challenge is the need for sufficient high-quality learning data, typically comprising real, experimentally obtained and labeled image pairs (input and segmented output images). However, obtaining such data requires multimodal imaging techniques, which are time-consuming and may still lack sufficient contrast for perfect segmentation. To address this, a hybrid learning dataset is proposed, combining limited real datasets from multimodal imaging (PXCT and XTM) with computationally generated synthetic datasets. Synthetic electrode structures are created based on known volume percentages, size, and shape distributions of the components. A CycleGAN algorithm is used for image-to-image translation, converting basic structures into realistic microstructures, using real segmented data as templates. Tomographic reconstructions of these synthetic structures are then simulated, incorporating beamline and measurement effects (energy, resolution, artifacts, noise). This synthetic data is used to train the 3D U-Net along with the real datasets. The trained network is then applied to segment the XTM datasets of graphite-silicon composite electrodes in pristine and cycled states (2, 5, and 8 cycles). Segmentation quality is evaluated using various metrics, including the Dice coefficient. To further enhance the accuracy of carbon black-binder domain (CBD) and pore space segmentation, k-means clustering is implemented to determine an intensity threshold from the gray value histogram, followed by thresholding applied to the neural network predictions. Finally, microstructural analysis is performed on the segmented data to investigate the evolution of the structure during cycling.

Conventional segmentation methods, such as thresholding and random walker algorithms, proved inadequate for the low-contrast XTM images. The 3D U-Net model trained solely on the limited real dataset showed some ability to distinguish silicon particles but failed to reliably separate pore space from graphite particles and accurately segment fine features of the CBD. The Dice coefficients for active particles and pore space were 0.6-0.7, while the CBD Dice coefficient was only 0.38. Integrating the synthetic data significantly improved segmentation accuracy. With hybrid learning data (real and synthetic), the Dice coefficients improved to 0.69 (pore space), 0.77 (graphite), 0.82 (silicon), and 0.58 (CBD). The combination of hybrid data training with subsequent thresholding further boosted the CBD Dice coefficient to 0.72. Analysis of the segmented electrodes revealed that the volume fractions of the different phases matched well with the expected values. The surface coverage of graphite particles with the pore phase remained consistently high (97%), while the silicon surface coverage with the pore phase increased from 53% in pristine samples to 65% after 2 cycles, indicating CBD detachment during cycling. Numerical diffusion simulations, using the segmented structures, indicated an effective transport coefficient of 0.14, reduced to 0.19 when neglecting the CBD. This points towards the CBD's significant contribution to reducing effective transport, through decreased porosity and increased tortuosity. Interestingly, this effective transport coefficient remained relatively constant throughout cycling. A location-specific analysis revealed differences in CBD distribution around graphite and silicon particles, with a significant detachment and repositioning of the CBD around silicon particles during cycling, while CBD distribution around graphite remained largely unchanged. This suggests that the main morphological changes to the CBD structure occur locally around silicon particles.

The study successfully demonstrated the value of incorporating synthetic data into deep learning for accurate segmentation of low-contrast battery electrode images. The improved segmentation enabled detailed analysis of the microstructural evolution during cycling, revealing key insights into the role of the CBD in lithium transport and its dynamics around different active materials. The relatively constant effective transport coefficient throughout cycling, despite local changes around silicon particles, is a notable finding. The methodology developed here has significant potential for accelerating battery research and development, facilitating more detailed microstructure-performance correlations. The differences in CBD distribution around different active materials suggest a need for more sophisticated models of composite electrode structures.

This work presents a novel deep learning approach for segmenting complex, low-contrast battery electrode images, combining real and synthetic data for enhanced accuracy. This approach overcomes limitations of conventional techniques, providing statistically meaningful analysis of microstructural evolution. The findings highlight the importance of considering CBD distribution and its dynamic changes during cycling in optimizing battery performance. Future work could focus on refining the synthetic data generation process, exploring other deep learning architectures, and applying the method to a wider range of battery chemistries and cycling protocols.

The study focused on a specific graphite-silicon composite anode and a particular cycling protocol. The conclusions might not be directly generalizable to all battery chemistries or cycling conditions. The resolution of the XTM imaging limited the ability to fully resolve the nanoscale porosity within the CBD. The relatively small sample size for imaging introduces uncertainty in the representativeness of the electrode structure, particularly for the analysis of local changes near silicon particles. The ex-situ nature of the study limits the ability to observe the dynamic changes during cycling directly.