Gaussian process regression-driven deep drawing blank design method

This research introduces a machine learning (ML)-based methodology for the optimal blank design of components manufactured through the deep drawing process, considering the interplay among material, process, and geometric parameters. The proposed blank design mapping function (BDMF) leverages a Gaussian process regression (GPR) ML model in conjunction with a radial basis function (RBF) kernel. This combination allows for correlating predictions with their standard deviations, capturing the estimations' quality effectively. The GPR model was trained using the results from a three-dimensional adaptive mesh-based finite element analysis (FEA) model, characterized by a fixed node count of 23 input parameters, an explicit solution scheme, and an average computational time of 270 s. Laboratory-scale experiments on an R47.5 mm flanged cup constructed from AISI-304 steel and AA5754 aluminum alloy served to validate the FEA models and the proposed BDMF. The comparison between experimental outcomes and FEA results revealed maximum deviations of 13.3 % in the drawing force and 0.35 % for the earing profile over a 90° segment. The comparison between experimental data and BDMF predictions for the sheet metal blank indicated average deviations of 0.015 mm (or 1.3 %) in estimating thickness and 0.12 mm (or 0.25 %) in predicting the outer radius. Application of the BDMF to four additional flanged geometries with varying shapes demonstrated its reliability and generality; the maximum and average deviations for the earing profile were 4.3 % and 3.1 %, respectively, and for post-forming sheet thickness, they were 6.7 % and 3.1 %, respectively.