<p>In this paper, the possibilities of deep learning approaches for sandwich structure delamination detection and prediction have been investigated. The research was divided into three parts: a validation study, a classification analysis of delamination shape, and delamination location forecasting. The best mesh configuration in terms of computational time and accuracy was chosen during the validation study by comparing a COMSOL MULTIPHYSICS model’s results with those obtained in a previously conducted related study. The natural frequencies of sandwich beams with various delamination forms and locations were collected from COMSOL MULTIPHYSICS and used in the classification and regression studies. Deep learning (DL) models were trained to predict the location of delamination and detect the shape of delamination using the obtained dataset. To improve accuracy and generalizability, we proposed tuning the hyperparameters of the model to choose the best DL structures for achieving both regression and shape detection tasks with as high performance as possible. The Bayesian algorithm was used to select the best combination of values for multiple hyperparameters used in the regression and classification models. Furthermore, several optimization phases have been held to explore the hyperparameter space carefully. In the optimization study, various regression and classification metrics were used to determine the optimum model. The outcomes demonstrate that both delamination location prediction and shape detection were accomplished with excellent performance by the optimized deep learning models. Particularly, a mean squared error value of 0.0000043303 was achieved, which allows regression analysis to identify the location of the delamination accurately. Likewise, 0.9944 of accuracy value has been achieved for the shape detection analysis, which allows the model to identify delamination shapes precisely.</p>