Deep learning-based autonomous morphological fracture analysis of fiber-reinforced composites

Abstract

Morphological assessment of fractured fiber-reinforced composites (FRCs) is crucial for understanding the failure mechanisms at the microscopic level. This assessment is typically conducted manually using scanning electron microscopy (SEM) images to identify the cause of damage and its specific failure mode. Such information is essential to improve material design, optimize manufacturing processes, and enhance the performance and durability of FRCs. However, manual morphological assessment is time-consuming, prone to human error, requires excessive domain knowledge, and lacks the efficiency needed for material scientists who during material design must repeatedly analyze morphology. Therefore, an autonomous morphological assessment is proposed using deep learning to improve accuracy and efficiency in analyzing fractured FRCs. A comprehensive ablation study is performed by evaluating six state-of-the-art deep learning models, namely DenseNet, GoogleNet, ResNet, VGG-16, VGG-19, and Xception. Moreover, since deep learning models require large datasets for effective training, data augmentation, and transfer learning concepts are utilized to overcome the limited data issues. The results based on numerous evaluation metrics demonstrated that the ResNet-DA model can efficiently perform autonomous morphological assessment of FRCs, achieving an accuracy and f1-score of 96.85 % and 96.84 %, respectively. This highlights the potential of the proposed approach in assisting the material scientists to improve the material design process by using suitable proportions of fibers and matrix ensuring only desired failure modes.