Finite Element and Artificial Neural Network-Based Buckling Load Analysis of Cylindrical: Impact of Training Data Proportions

The accurate prediction of buckling loads in thin-walled cylindrical shells is essential for structural stability. This study employs artificial neural networks (ANNs) to predict the buckling load of cylindrical shells, comparing models trained with $70 \%$ and $80 \%$ of the dataset. Finite element analysis (FEA) using ABAQUS serves as the benchmark for evaluating ANN accuracy. Design of Experiment (DoE) was used to setup the input and output parameters. Performance metrics reveal that the $70 \%$ training model achieves a lower mean squared error (MSE) of 0.4907 and a higher regression coefficient $(R=0.9730)$ on the training set, compared to the $80 \%$ training model with MSE $=1.4409$ and $\mathbf{R}=\mathbf{0 . 9 4 8 7}$. Furthermore, in validation and test phases, the $\mathbf{7 0 \%}$ model outperforms the $\mathbf{8 0 \%}$ model, with test MSE values of 6.1525 and 7.4103, respectively. A one-tailed t-test confirms that these differences are statistically not significant. The close alignment between ANN and FEA predictions highlights the effectiveness of machine learning for structural buckling analysis, offering a computationally efficient alternative to conventional numerical simulations.