Learning Textures using Deep Neural Networks

Abstract

Properties of highly ordered crystalline materials, like strength and ductility, are dependent on the preferred orientations of the grains within the material, that is, the texture. When a material is processed and microstructural transformations occur, the texture of the material changes drastically. These texture evolutions are presently simulated using Crystal Plasticity Finite Element Methods (CPFEM). While these simulations are precise, they are computationally expensive and slow.

A surrogate for such simulation methods that can benefit from a data-driven approach could be deep learning through artificial neural networks. The aim of this study is to leverage a suitable deep learning neural network model and assess its ability to capture the complexity of texture. As such, Normalizing Flows (NF), a generative deep learning model, is employed to learn textures by capturing the multi-modal distribution of discrete sets of orientation matrices that represent those textures. The performance of the model is assessed for two types of data, matrices that belong to a fixed texture (unconditional modelling) and matrices paired with simple conditions (conditional modelling) that produce textures based on the conditions. The training data for the neural network is synthetically generated in a microstructure modelling software. The model is trained on this synthetic data and subsequently used to generate samples on the distribution it has learned. These samples and ground truth data are plotted in pole figures to compare the distribution of the generated data and the ground truth data, respectively, to visualize the model’s generative capability.

It is found that the model closely approximates the distribution of a given texture, capturing the structure of the distribution, albeit having a reduced density in the modes of the distribution. The conditional model is also capable of generating the approximate texture relevant to the condition given to it during evaluation. These findings indicate that the model is capable of learning texture, and further improvements in the model’s architecture could make the model highly robust.