Machine learning-assisted free vibration modeling of variable angle tow laminated composite plates

Purpose The purpose of this paper is to develop an artificial neural network (ANN) model for the free vibration analysis of variable angle tow (VAT) laminated composites, which are challenging to model due to their spatially varying fiber orientations. Design/methodology/approach Training data is generated by finite element method (FEM) simulations based on first-order shear deformation theory, which establish a relationship between geometric parameters and fiber angles and the fundamental natural frequency. The Bayesian regularization (BR) and Levenberg–Marquardt (LM) methods are used in MATLAB to train the ANN, and transfer functions (tan-sigmoid and log-sigmoid) are chosen for best results. Findings The ANN approach substantially lowers computational effort while accurately predicting natural frequencies and maintaining strong consistency with FEM results. Using the log-sigmoid activation function, the LM algorithm yields prediction errors of approximately 2%–4%, whereas the BR method typically limits errors to about 1%. Owing to effective regularization and improved generalization, Bayesian-trained ANN models with both tan-sigmoid and log-sigmoid functions closely match FEM results for all aspect ratios. Conversely, ANN models trained using the LM algorithm show comparatively higher deviations at aspect ratios of 1.25, 2.25 and 2.50, mainly due to overfitting and increased sensitivity to nonlinear behavior. Originality/value The presented ANN framework is a computationally efficient alternative to FEM-based VAT composite vibration analysis. It offers accurate frequency prediction with minimum error, allowing for more efficient design and optimization of complex composite structures in engineering applications.