Deep learning-based prediction of delamination growth in composite structures: bayesian optimization and hyperparameter refinement

In this study, a novel method for detecting the growth of delamination in sandwich structures has been proposed. To this end, we suggested hybridizing the Deep Learning techniques (DL) and Finite Element Method (FEM) for predicting the growth of delamination in this structures. A dataset of simulated delamination growth under different delamination sizes has been produced using the FEM method. Then, a DL model has been trained using this dataset to precisely predict the growth of delamination. This study focused on predicting delamination growth using a tuned and optimized deep learning based regressor. Therefore, to find the ideal set of hyperparameters, the Bayesian optimization algorithm has been used for selecting the best structure and enhancing the regressor performance. Afterward, the model was evaluated and multiple processes were conducted to improve its behavior and solve its stability and overfitting issues. Particularly, an inconsistency between validation loss and training loss has been initially detected in the behaviour of the model, which may indicate overfitting. To tackle this issue, dropout regularization has been added, which improved the consistency between the loss functions but results in less smooth convergence from the expectations. So, in a third study, dropout and L1 regularization has been combined to improve the stability of the model. This combination achieved a consistent and smooth convergence between the validation and training loss functions. The findings highlight the importance of hyperparameter optimization and regularization techniques in improving regression model performance. The study shows the efficiency of Bayesian optimization in hyperparameter tuning and the iterative optimization of a regression model. Furthermore, the outcomes show that the suggested method can identify and predict delamination growth with high accuracy.