A Machine Learning Model for AA7075-T6 Based on Anisotropic Elastoplastic Constitutive Theory

In this study, a machine learning model for the anisotropic mechanical behavior of AA7075-T6 was successfully developed by combining classical plasticity theory with neural network algorithms. First, a yield criterion incorporating strength asymmetry and orientation dependence was constructed based on a fourth-order linear transformation of invariants, and a stress update algorithm was implemented using the return mapping and implicit integration schemes. Subsequently, a dataset comprising experimental results from tension, compression, and shear tests at multiple orientations, as well as theoretically generated data from additional strain paths, was established to train a genetic algorithm-optimized two-hidden-layer neural network. Plastic-stage assessments indicated that, for the equal-biaxial path, the RMSE values of the stress-strain curves predicted by the machine learning model relative to the constitutive implementation were 31.87 and 39.24 MPa. It should be noted that the stress-strain response under equal-biaxial loading represents an additional prediction case. The trained model exhibited satisfactory predictive performance when compared against experimental data for tension, compression, and shear and was capable of reproducing the theoretical stress paths derived from the classical constitutive model. This approach leverages the physical interpretability of conventional constitutive modeling and the high efficiency of data-driven methods, providing a viable solution for efficiently predicting the anisotropic mechanical response of metallic materials under complex stress states.