BP neural network-based springback predictions during creep-age forming of aluminum alloy 2219

Abstract

To address the challenge of accurately predicting springback during creep-age forming (CAF) of aluminum alloy 2219, a rapid prediction method that integrates a creep-aging constitutive model, finite element analysis (FEA), and a back-propagation neural network (BPNN) was developed. First, a constitutive model was established for the 2219 aluminum alloy using uniaxial thermal tensile creep tests. Taking single-curvature bent components fabricated by CAF of aluminum alloy 2219 as the subjects of the study, a corresponding finite element model was constructed and validated against physical experiments that were conducted under identical conditions; this model exhibited a relative error of 6.77%. Using mixed-level orthogonal experiments and range analysis, the effects of key process parameters, including the sheet length, sheet width, sheet thickness, sheet curvature radius, aging time, and applied stress, on the springback rate were assessed. Second, Latin hypercube sampling (LHS) was used to generate various parameter combinations, and the optimization objective was calculated using FEA. Based on the training samples, a BPNN was constructed to map the non-linear relationships between the process parameters and the springback rate. Third, an improved particle swarm optimization (PSO) algorithm was utilized to optimize the process parameters, thereby minimizing the springback rate. The results indicated that a relative error of only 0.91% existed between the results of a numerical simulation that used the optimal parameters and the BPNN predictions. These results demonstrated that the proposed hybrid model can effectively predict springback during CAF of aluminum alloy 2219.