Leveraging machine learning for robust parameter identification of Yld2004-18p

This study proposes an experiment-efficient ANN-based framework for characterizing the anisotropy of AA6022-T4 and calibrating the Yld2004-18p anisotropic yield function from a single hole expansion (HE) test and evaluates its predictions against two texture-driven meso-scale plasticity approaches, namely crystal plasticity finite element method (CPFEM) and elasto-visco-plastic self-consistent (ΔEVPSC). The ANN framework exploits the heterogeneous deformation field generated during the HE test, with training data obtained from finite element (FE) simulations using randomly assigned Yld2004-18p parameters. Two distinct ANN frameworks are developed: one predicts anisotropic properties (stress ratios and r-values), while the other directly estimates Yld2004-18p parameters. Three ANN input selection strategies are explored, i.e., using all strain components from 21 extraction locations in the HE simulation, local stress-based selection, and a variance-based approach. In the crystal plasticity approaches, CPFEM and ΔEVPSC simulations based solely on the measured initial crystallographic texture provide virtual uniaxial and equibiaxial responses that are subsequently fitted to Yld2004-18p. Comparative assessment shows that the ANN framework, as well as the CPFEM- and ΔEVPSC-based approaches, reproduces the anisotropic response, yield locus, and HE thinning profile of the material with generally good agreement with experiments and with the conventional calibration, while substantially reducing experimental time, cost, and material consumption.