Acceleration of Time Resolved X-Ray Imaging using Deep Learning Techniques

Quantifying the three-dimensional (3D) microstructure of materials is vital across various domains, including materials science, condensed matter physics, and semiconductor research. Despite the efficacy of traditional characterization methods such as optical microscopy, electron microscopy, and serial sectioning, these techniques tend to be dimensionally limited, destructive, and require a significant time investment. Over the past few decades, advancements in synchrotron and lab-scale 3D x-ray tomography (XCT) have drastically enhanced researchers' capacity to uncover complex 3D microstructures across numerous materials and fields of research. Specifically, lab-scale XCT systems have facilitated the broader utilization of 3D characterization methods among diverse researchers. Despite their advantages, 3D XCT techniques also present substantial challenges, including but not limited to extended experimentation time, substantially large datasets compared to 2D methods, and high computational demands associated with analyses of the 3D tomography data. These limitations are exacerbated in lab-scale systems, which have flux limitations compared to their synchrotron-scale counterparts. The challenges are further intensified when working with time-resolved 3D tomography or 4D datasets. Applying deep learning, which excels at handling large datasets, offers a potential solution to these challenges. Deep learning can significantly optimize the most time-consuming aspects of lab-scale 4D experimentation—data collection and data analysis or segmentation. In this study, we have utilized Generative Adversarial Networks (GANs) to filter, enhance, and segment low-dose Absorption Contrast Tomography (ACT) and Diffraction Contrast Tomography (DCT) datasets captured using a lab-scale x-ray microscope (XRM). In our GAN-based approach, we adopted the U-Net++ network architecture to improve the quality of the 3D datasets captured at the lab scale. Our study encompassed two model samples: hyper-spherical Aluminum (Al) particles dispersed in a resin matrix and micro-scale Silicon (Si) cubes dispersed within a resin matrix. The Al particles were scanned under low and high x-ray doses (controlled by the exposure time and the number of X-ray projections) and resolutions. The Al particle datasets were used to train our GAN network to filter and enhance the low-dose and low-resolution scans, respectively. The results from this approach demonstrated clear improvements in image quality and significant savings in scan time. Our approach was also carried out on Diffraction contrast tomography (DCT) scans of single-crystal Si cubes to obtain diffraction spots with varying noise levels. The high-dose diffraction spots were manually segmented, a process that is both time-consuming and labor-intensive, and this segmented data was used to train our GAN network to segment the diffraction spots from the low-dose DCT scans directly. The manual and GAN-based segmentation results were used to reconstruct the 3D grain structure of the Si cubes using a forward modeling approach. The reconstructions from the GAN-based and manually segmented datasets were compared to assess the accuracy of the GAN segmentation and quantify the time savings achieved with this approach. The deep learning-based approach outlined in this study helps advance our ability to perform 4D material characterization and analysis at the lab scale. Integrating deep convolutional neural networks into lab-scale 4D material characterization can accelerate data collection and analysis, automate the characterization of materials and open up new avenues for analyses of diverse material classes at an expedited rate.