| description abstract | Abstract. In ocean engineering, spherical shells often serve as common pressure vessels. Under external pressure, these shells tend to suffer from buckling damage. To ensure the stability of spherical shells and optimize their structural design, it is necessary to predict their buckling load under external pressure. However, current predictions based on the knockdown factors are relatively conservative, and predictions based on common machine learning methods lack physical orientation. Based on the energy barrier method, a physics-informed one-dimensional-convolutional neural networks (1D-CNN) model is proposed to address this problem. These newly developed physics-informed features and loss functions improve the accuracy of the model. Both metallic and nonmetallic spherical shells datasets are used to establish the models. Then, we determine the optimal hyperparameters using K-fold cross-validation and the optuna optimization framework and compare their predictive performance against models like decision trees (DT), random forests (RF), and artificial neural networks (ANN). The results indicate that our model outperforms the others in predictive accuracy. Additionally, the potential of the model, guided by the energy barrier method, is demonstrated by comparing various loss functions. For the actual spherical shell, the trained model can be used for buckling analysis and the optimized design of spherical shells in engineering practice. | |
