Development of a Deep Learning Model for Capturing Plastic Anisotropy-Texture Linkage

Traditional methods for characterizing plastic anisotropy in metal alloys require iterative experiments or high-fidelity computational simulations. In order to avoid expensive anisotropy characterization procedures, this work has developed a novel data-driven anisotropy prediction model from a large dataset of crystal plasticity (CP) calculations. Specifically, a deep learning (DL) model was trained on a total of 54,880 analytical CP calculations of normalized yield stresses and plastic strain increments along every 15 degrees in three principal planes with varying initial crystallographic textures. The validity and accuracy of the DL model were assessed by comparing its prediction results to an additional 20,000 CP calculations, on which the model was not trained (i.e., a validation dataset). Quantitative comparisons of CP and DL predictions of yield stresses and lateral strain ratios for different loading directions demonstrated that the DL model accurately and efficiently links material's initial crystallographic texture to plastic anisotropy. Moreover, DL-based predictions were further validated by performing finite element simulations of the cup drawing using the non-quadratic yield function parameterized from the DL anisotropy predictions. The DL-based simulation showed excellent agreement with those performed using yield functions parameterized from CP and CP-FEM, demonstrating that accurate anisotropy information was obtained without the need to perform expensive high-fidelity simulations.

Automated summary

This study develops a data-driven method for predicting plastic anisotropy in materials using a deep learning model. The model accurately and efficiently predicts the relationship between the initial crystallographic texture and plastic anisotropy by training on a large dataset of crystal plasticity calculations.