Machine Learning Based Prediction of Buckling in Composite Cylindrical Shells Under External Pressure Using Design Code and FEM Analysis

Abstract Composite materials, known for their strength and weight advantages, find extensive use in the maritime industry. Nevertheless, predicting the buckling pressure of composite cylindrical shells, especially with specific layup patterns, is a challenging task. This study aims to enhance buckling pressure prediction accuracy through machine learning. The model is trained using data from NASA design codes and finite element analysis in ANSYS. Twenty-four-layer composite cylindrical shells with a thickness of 2.52mm, constructed from carbon-epoxy prepreg tape (USN-125), were subjected to external pressure. Stacking patterns, denoted as [θ/θ]12, , were varied at 2-degree or 5-degree intervals for each method, ranging from 0 to 90 degrees. This variation resulted in a dataset comprising 2378 data points. A Random Forest machine learning algorithm, which demonstrated higher prediction accuracy compared to other algorithms, was trained on this dataset., with weights assigned to NASA design equations and FEM results at a 1:2 ratio. Predictions using random angle combinations were compared to those solely based on the design equation, demonstrating higher accuracy. This study represents an initial step in integrating machine learning to enhance buckling pressure prediction accuracy. Future enhancements may involve incorporating post-buckling test data and fine-tuning variable weights in the algorithm to further improve precision.