Prediction of Failure Mode and Buckling Strength of Ring-Stiffened Cylindrical Shells Using Machine-Learning Techniques

This study utilizes the concept of machine learning (ML) to predict the failure mode and associated buckling strength of ring-stiffened cylindrical storage tanks under uniform external pressure. A database comprising 180 ring-stiffened cylindrical specimens was generated from the Ansys Workbench 2023 software platform for use in various ML models; however, the numerical simulation of seven ring-stiffened cylindrical specimens was initially substantiated using widely accepted theoretical equations. To identify the best prediction model, four ML models—k-nearest neighbor (KNN), naïve Bayes, decision trees, and support vector machine (SVM)—were developed and evaluated for the classification of buckling failure modes. The SVM classifier model achieved the highest overall accuracy with 96% for the test data set and exhibited the highest precision and recall values compared to the other ML algorithms. After successfully identifying the buckling failure modes for the given data set, the buckling strength for stiffened cylindrical shells can be easily determined based on the failure mode by using the theoretical equations; however, the buckling strength for stiffened cylindrical shells is also evaluated by using the random forest algorithm. The predicted buckling strength obtained from the proposed random forest model was compared to the numerical buckling strength. The comparison indicated excellent performance of the random forest model for both training and testing data, achieving high accuracy with R2 values of 0.987 and 0.962, respectively. Additionally, the associated mean squared error (MSE) values were low, measuring 0.009 for the training data and 0.22 for the testing data. Regression plots illustrate a strong alignment between the actual and predicted buckling strength while the error histograms indicate a concentrated distribution close to zero, demonstrating that the model accurately predicts most of the training data set and produces strong generalization of new unseen data.