Inverse estimation of tensile shear strength from fracture surface images using deep learning

Abstract

To improve the long-term durability of adhesive joints, it is important to analyse the fracture surface and identify the fracture factors. Although conventional optical observation methods are simple and widely used, they were limited to qualitative discussions. In this study, an inverse estimation method utilising deep learning was investigated to clarify the quantitative relationship between the tensile shear strength and the fracture surface of single lap joints immersed in water. The deep learning analysis revealed that the tensile shear strength could be estimated from the fracture surface images with very high accuracy, indicating a strong quantitative correlation between the fracture surface images and the residual tensile shear strength. Grad-CAM indicated that the deep learning model could estimate the residual tensile shear strength by observing from the topology and colour of the remaining adhesive.