Deep Learning Multiclass Segmentation of Intra- and Interlaminar Microdamage in Synchrotron CT of Advanced Composites

In this study, we report the development and examination of deep learning approaches for automated, high-throughput, and scalable multiclass segmentation of intra- and interlaminar progressive damage mechanisms in advanced composite laminates, imaged using datasets from in situ synchrotron radiation micro-computed tomography (SRCT). The SRCT dataset includes six different composite specimen configurations that were each scanned in a loading rig at four tensile load steps ranging from unloaded to near failure, totaling 24 scans (>10^10 voxels). The aerospace-grade composites consisted of unidirectional plies of carbon (micro) fiber reinforced epoxy stacked in a cross-ply sequence prone to delamination. Additionally, laminate design variables including interlaminar nanoreinforcement and effective ply thickness were varied across the six specimens, collectively leading to visualization of a relatively broad range of characteristic progressive damage profiles (e.g., small dispersed polymer cracking vs. large concentrated cracking) dominated by intralaminar (0° and 90° plies) and interlaminar (delamination) polymer damage classes. Automated quantification of these damage classes via traditional computer vision algorithms was precluded by myriad challenges (e.g., highly interconnected, sparse damage networks, large crack-opening displacements, and SRCT artifacts), and thus deep learning (DL) segmentation was investigated. DL models were developed using ground truth segmentation results from a state-of-the-art semi-automatic rule-based approach (time-intensive and subjective) by a trained human, which was performed for 11 scans (>22,000 slices) across the specimen/load set and required human labor times of >20 h (up to ~70 h) for each scan. Leveraging best practices from recent related literature, which in contrast to the present study did not include interlaminar damage or background as distinctly learned classes, a hyperparameter study was conducted and the best-performing DL model was used to infer all scans. The inference results exhibited excellent qualitative and quantitative agreement with the human ground truth results, considering the practical bias toward loaded scans in the training set and unloaded scans (more sensitive to ground truth errors and DL model corrections) in the validation set. Moreover, inference results for the remaining 13 scans (not human labeled) demonstrated excellent consistency and generalizability, especially in high-quality delineation of intralaminar vs. interlaminar damage at local boundaries and interfaces. These findings expand the limited understanding of DL segmentation efficacy applied to complex SRCT of advanced composites, including successful extraction of previously unexplored feature targets, and can significantly accelerate access to rich damage-linked insights into emerging composite reinforcement mechanisms.