A catboost-based surrogate model for fast prediction of free vibration response in tri-directional functionally graded plates

Abstract

Accurate numerical analysis of tri-directional functionally graded (3D-FGM) plates is computationally intensive, posing a major challenge for design optimization and reliability assessment. To overcome this, we propose an efficient surrogate model based on the CatBoost algorithm and benchmark its performance against a finely tuned Artificial Neural Network (ANN) for rapid prediction of free vibration responses. A high-fidelity dataset comprising 20000 samples was generated using a validated model that integrates Isogeometric Analysis (IGA) with Generalized Shear Deformation Theory (GSDT). Each sample includes eighteen input parameters (material control points) and three outputs: natural frequency, total ceramic volume fraction, and plate mass. The models were systematically evaluated by investigating the influence of hyperparameters and dataset size on prediction accuracy (measured by MSE and MAPE) and computational time. The results demonstrate that the optimized CatBoost model achieves nearly nine-fold lower test MSE and is over 10.8 times faster than the ANN. These findings highlight CatBoost as a highly accurate and efficient surrogate, enabling fast and reliable analysis of complex composite structures for future engineering applications.