Inverse identification of cohesive zone parameters for sintered nano-silver joints based on dynamic convolution neural network

The inverse identification of cohesive zone parameters is an important topic in the field of fracture mechanics. In this paper, three data-driven models, including multilayer perceptron (MLP), convolutional neural network (CNN), and dynamic convolutional neural network (DCNN), were presented to predict the cohesive zone parameters of sintered nano-silver end notched flexure (ENF) joints. Based on the construction of experimental and numerical datasets of load versus displacement curves, bilinear cohesive zone model (CZM) parameters of sintered silver joints are adopted as the prediction target. The investigation shows that MLP, CNN, and DCNN are all valid for predicting CZM parameters through load versus displacement curves with reasonable accuracy. However, DCNN has better prediction accuracy and performance than those of CNN and MLP models based on loss analysis, statistical indicator comparison, and K-fold cross-validation. DCNN can be adopted as the suitable surrogate model for CZM parameters inverse identification with high prediction accuracy. Otherwise, DCNN is also not sensitive to the load versus displacement curves data length. However, the computation efficiency of DCNN during the training process is not as high as that of MLP. Those three methods presented in this paper are very hopeful to be adopted for other inverse identifications of CZM parameters for various kinds of adhesive joints.