Data-Driven Modeling to Predict Bending Angles in Laser Bending Process of AA6061 Sheets: A Hybrid FEM-Artificial Neural Network Approach
Pouya Pashaie, Aliakbar Asgharpour, Mohammad Bakhshi-Jooybari and Hamid Gorji
This study investigates and predicts the bending angle in the laser bending process by combining finite element modeling (FEM) and artificial neural networks (ANN). To develop a robust predictive model, a training and testing dataset was designed using the Response Surface Methodology (RSM) with a Central Composite Design (CCD). The finite element simulation results were utilized to create the training dataset, which was designed using the CCD method. The ANN model was trained using 80% of the generated data, while 20% was used for testing, achieving a mean square error (MSE) of 0.0004 and an R² value of 0.98. The ANN demonstrated high prediction accuracy, showing near-perfect agreement in most cases (e.g., 0.632° vs. 0.628°, and 0.61° vs. 0.609°). The method enhanced bending angle prediction accuracy while achieving sub-second computation speeds, demonstrating the efficacy of the FEM-ANN integration for laser bending modeling.
Keywords: Laser bending process, bending angle prediction, Finite Element Modeling (FEM), Artificial Neural Network (ANN), Central Composite Design (CCD)