Beyond response surface methodology: High-fidelity prediction of residual stresses in Al6061 deep drawing using Artificial Neural Networks

Abstract

Residual stresses, resulting from the complex cyclic bending-unbending strain history in deep drawing, are fundamental to the structural integrity and fatigue life of formed Al6061 components. Accurate characterization of these stresses is often challenging due to the highly non-linear interaction between material anisotropy and evolving contact conditions. This study presents a high-fidelity hybrid framework designed to minimize residual stresses and thickness variations through an integrated Finite Element (FE) and Artificial Neural Network (ANN) approach. Eight governing parameters, including tool geometry, blank holder force, and localized friction coefficients, were investigated. To ensure mechanical rigor, the FE model was grounded in experimental true stress-strain data and the Hill’48 anisotropic yield criterion. Numerical predictions were validated using the semi-destructive hole-drilling method (ASTM E 837-99) and ultrasonic thickness gauging, showing a strong correlation with a maximum discrepancy of 4.5%. A comparative analysis demonstrates that while parametric Response Surface Methodology (RSM) adequately models monotonic responses like punch force (R² = 0.89), it fails to resolve the history-dependent nature of residual stress fields (R² = 0.72). In contrast, the developed MLP neural network exhibits superior generalization (R = 0.903), effectively mapping the non-linear strain paths. The optimized configuration achieved a peak residual stress reduction to 41.25 MPa. The results establish ANN-driven surrogate modeling as a robust, computationally efficient tool for advanced strain analysis and process optimization in precision sheet metal forming.