Machine learning-based prediction of CZM parameters for mode-I/II fractures in adhesive joints

Accurate identification of cohesive zone model parameters (CZMPs) is essential for reliable fracture mechanics simulations of crack initiation and propagation. A neural network-based machine learning framework is developed to identify CZMPs that govern fracture behavior in adhesively bonded joints. Three data-driven models, including a convolutional neural network (CNN), Long Short-Term Memory (LSTM), and a hybrid (CNN-LSTM), were implemented within a two-stage sequential prediction framework. The comprehensive load-displacement curves collected from experimental and numerical sources serve as input. The bilinear cohesive zone model parameters were selected as the prediction target. These models were trained using the hyperparameters optimized through Bayesian tuning. The results show that the CNN–LSTM achieves the highest prediction accuracy with coefficients of determination (R2) 0.997 and 0.989 outperforming LSTM (R2 = 0.996 and 0.956), and CNN (R2 = 0.985 and 0.914) corresponding to Mode I and Mode II cohesive strengths, respectively. However, the improved performance of the CNN–LSTM model comes at a higher computational cost, with training times of 480 and 342 min, compared to 30 and 28 min for CNN and 205 and 164 min for LSTM in normal and shear mode, respectively. For a balanced approach, LSTM-based architectures offer an optimal balance between prediction accuracy and computational cost.