Multi-Scale Anisotropic Yield Function Based on Neural Network Model

The increasingly complex form of traditional anisotropic yield functions brings difficulties to parameter calibration and finite element application, and it is necessary to establish a unified paradigm model for engineering applications. In this study, four traditional models were used to calibrate the anisotropic behavior of a 2090-T3 aluminum alloy, and the corresponding yield surfaces in σxx,σyy,σxy and α,β,r spaces were studied. Then, α and β are selected as input variables, and r is regarded as an output variable to improve the prediction and generalization capabilities of the fully connected neural network (FCNN) model. The prediction results of the FCNN model are finally compared to the calibration results of the traditional model, and the reliability of the FCNN model to predict the anisotropy is verified. Then, the data sets with different stress states and loading directions are generated through crystal plasticity finite element simulation, and the yield surface is directly predicted by the FCNN model. The results show that the FCNN model can accurately reflect the anisotropic characteristics. The anisotropic yield function based on the FCNN model can cover the characteristics of all traditional models in one subroutine, which greatly reduces the difficulty of subroutine development. Moreover, the finite element subroutine based on the FCNN model can model anisotropic behaviors.