Deep learning based automated fracture identification in material characterization experiments

[Display omitted] •Automated fracture identification in material testing is investigated.•The transfer learning potential of a wide range of CNN architectures is evaluated.•Deep learning, high accuracy and low numerical cost CNN models are identified.•All-in-one deep learning designs with accuracies above 99% are reported.•The automated crack localization and profile characterization is elaborated. In the current work, the automated fracture identification in material testing experiments is investigated through deep learning convolutional neural network (CNN) techniques. Three widely employed material characterization experiments are considered, namely uniaxial tensile, punch, and shear experiments. Initially, the surface crack identification performance of well-known, minorly modified CNN architectures is investigated in the context of transfer learning. In particular, the transfer learning performance of VGG, ResNet, and Inception CNN architectures is analyzed. It is found that moderate-size filters and deeper rather than wider network architectures yield significantly higher validation accuracies in all cases, with validation scores that approach and exceed 95% accuracies. Moreover, low computational cost models that can run on minimal storage and computing power devices and deliver superior performance for all experimental testing setups are identified through greedy network searches. Thereupon, CNN deep learning architectures that allow for validation accuracies above 99% are reported, delimiting low-performing model counterparts. Moreover, CNN-based deep learning «all-in-one» material fracture models are elaborated, furnishing automated, computer vision-based fracture characterization techniques. Furthermore, feature map insights are provided, highlighting the salient-feature identification potential of the networks, while the automated crack localization and profile characterization are demonstrated through Local Interpretable Model-agnostic Explanations (LIME) machine learning (ML) methods and image subset creation techniques. The results are expected to serve as a basis for material testing automation with the support of deep learning techniques, providing reliable analysis methods in labor-intensive engineering tasks.