Automated characterisation of resin-rich regions in sheet moulding compounds via deep learning-based segmentation of 3D CT images

Resin-rich regions are important microstructural features in carbon fibre thermoplastic sheet moulding compounds (CFRTP-SMCs), where they occur as irregular three-dimensional fibre-deficient zones associated with the random arrangement of discontinuous fibre tapes. Their identification in X-ray computed tomography (CT) images is challenging because they cannot be defined by greyscale intensity alone and require interpretation of the surrounding fibre architecture. In this study, an automated nnU-Net-based 3D segmentation pipeline was implemented to directly identify resin-rich regions from greyscale CT volumes. Quantitative evaluation showed strong agreement with the reference annotations, with a mean Dice coefficient of 0.905, mean IoU of 0.828, region detection rate of 100%, mean signed volume error of 0.99%, and mean absolute volume error of 2.25%. Independent orthogonal-plane annotations further supported the model’s ability to capture complex local resin-rich-region geometries. The proposed method also substantially outperformed a conventional texture-based approach, while fibre-tracing analysis showed spatial correspondence between predicted resin-rich regions and fibre-deficient zones and fibre-bundle interfaces. The trained model was successfully applied to a substantially larger CT volume from a different specimen, supporting scalability beyond the model-development sub-volume size. Finally, an ash-test-based consistency check provided independent specimen-scale support for the physical relevance of the predicted resin-rich-region volume fractions. These results demonstrate a reliable, automated, and scalable approach for quantitative three-dimensional characterisation of resin-rich regions in CFRTP-SMCs.