Characterizing the microstructure of materials is crucial for understanding their properties and performance. Techniques like SEM, EBSD, and micro-CT provide valuable microstructural information, but often face a trade-off between resolution and field of view. High-resolution imaging captures fine details but limits the imaged volume, potentially leading to statistically unrepresentative data due to material heterogeneity. Low-resolution imaging covers larger volumes but misses crucial details. This is particularly relevant in Li-ion batteries, where multi-scale heterogeneity in electrodes necessitates large field of views to capture representative volumes of particles. Fine features like cracks (<500 nm) within electrode particles require both high resolution for accurate identification and large fields of view for statistically meaningful results. Super-resolution techniques offer a potential solution by enhancing the resolution of low-resolution images, increasing the detail available from larger-area images. Previous work has applied GANs and other deep learning methods to microscopy images, but often relied on synthetically downsampled images for training, which may not accurately model real experimental data. This paper investigates the use of SRGANs trained on experimentally-measured low- and high-resolution image pairs to super-resolve SEM images of aged LiNiₓMnᵧCo₂O₂ (NMC) cathode particles, focusing on the improved quantification of cracks.

The authors review existing literature on machine learning applications in materials science, specifically focusing on image classification, segmentation, and synthesis using CNNs and GANs. U-nets, a modified version of CNNs, are highlighted for their successful application in biomedical and materials science image segmentation. The use of GANs for generating digital twins of real microstructures and for supervised and unsupervised super-resolution is also discussed. The authors note that while various machine learning methods for super-resolution exist, there is a challenge in accurately modeling experimentally measured low-resolution images using only synthetically downsampled high-resolution images for training. Prior work applying GANs to super-resolve SEM images of nanoparticles is mentioned, highlighting the need for an approach that better handles the nuances of real experimental data.

The study uses a modified version of the SRGAN architecture from Ledig et al. (2017), trained on experimentally measured paired low- and high-resolution SEM images of aged NMC cathode particles. The high-resolution images are denoised before training. The GAN architecture consists of a generator (G) and a discriminator (D). The generator takes a low-resolution image as input and produces a super-resolved high-resolution image. The discriminator aims to differentiate between the generated high-resolution image and the actual high-resolution image. The generator is trained to minimize a perceptual loss function (PL) and to “trick” the discriminator. The discriminator aims to maximize the difference in outputs between the true and generated high resolution images. The authors modify the original SRGAN by replacing PReLU layers with ReLU layers and omitting Batch Normalization layers to improve computational efficiency and accuracy. The training process uses the Adam optimizer with a learning rate of 10⁻⁴. The authors also train a second GAN architecture (CinCGAN) based on Yuan et al. (2018), suitable for scenarios where matched pairs of low- and high-resolution images are unavailable. The authors compare the results obtained using SRGAN with those obtained by trained U-NetGAN and SRResNet1 architectures, taking into account differences in architecture and loss functions. Furthermore, they train an SRResNet2 architecture solely on synthetically downsampled high-resolution images to highlight the potential limitations of such an approach. The performance of the trained networks is quantitatively assessed on a test dataset using metrics like MSE, PL (for VGG16 and VGG19), and MSSIM. Finally, the authors perform crack segmentation using a modified Westhoff et al. (2018) method on the high-resolution, upsampled, and super-resolved images, evaluating the results using the Jaccard index, the relative error in crack density and a distance metric between the probability densities of the area-equivalent diameters of the cracks.

Visual and quantitative analysis of the super-resolution results shows that SRGAN outperforms the other four networks (U-NetGAN, SRResNet1, SRResNet2, and CinCGAN) in terms of MSE, PL, and MSSIM. SRGAN achieves a significantly improved Jaccard index (0.679) for crack segmentation compared to upsampling the low-resolution image using bilinear interpolation (0.556). The relative error in crack density for the SRGAN super-resolved image is 0.036 compared to 0.136 for the bilinearly upsampled image. The distribution of crack sizes is more accurately captured with the SRGAN super-resolved images, reflecting in lower values of the distance metric between the probability density of area-equivalent diameters of the cracks. The CinCGAN architecture, trained in an unsupervised setting without matched image pairs, performs better than SRResNet2 (trained only on synthetic low-resolution images), demonstrating its suitability for scenarios where paired data are not available. This suggests that GANs trained with actual experimental low-resolution data, regardless of the availability of matched high-resolution pairs, provide superior super-resolution performance than networks solely trained on synthetically generated low-resolution data.

The findings demonstrate the effectiveness of SRGANs for super-resolving SEM images of Li-ion battery electrodes, significantly improving the accuracy of fine-feature quantification, particularly for crack detection. The superior performance of SRGAN compared to other networks highlights the benefits of using experimentally measured low-resolution data during training, and the advantages of the perceptual loss function for super-resolution tasks. The unsupervised approach using CinCGAN proves useful when paired datasets are unavailable or impractical to collect. The improved crack segmentation results obtained using super-resolved images compared to simply upsampled images confirm the value of the super-resolution method for accurately characterizing material degradation. The ability to reliably estimate crack density and size distribution using SRGAN opens up new opportunities for studying battery aging processes and other materials science applications.

This study successfully demonstrates that SRGANs offer a powerful approach to super-resolve microscopy images, particularly SEM images of Li-ion battery electrodes. The use of experimentally measured low- and high-resolution image pairs in training leads to superior performance compared to methods using only synthetically downsampled data. The improved accuracy in crack detection and quantification highlights the potential of this technique for analyzing material degradation and for enabling more accurate microstructure characterization in various materials science applications. Future work could explore the integration of this method with other advanced image analysis techniques and the application of this super-resolution strategy to other types of microscopy and material systems.

The current study focuses on a specific type of Li-ion battery cathode material (NMC532) and a specific aging process. The generalizability of the SRGAN model to other materials or aging mechanisms needs further investigation. The computational resources required for training GANs can be significant. The accuracy of crack segmentation depends on the quality of the initial high-resolution image data and the performance of the crack segmentation algorithm. The current segmentation algorithm might have limitations which could be improved by using alternative segmentation methods.