Extracting dislocation microstructures by deep learning

Abstract

The microstructure of dislocations can be accessed by the total density of dislocations or the density of geometrically necessary dislocations (GND). The total dislocation density determines the flow strength of a crystal, which, in the case of high dislocation contents, is a quantity very difficult to measure accurately. On the other hand, related to crystal rotations, the GND densities are conveniently measured from electron diffraction experiments or calculated via simulations. Here, a novel and modern approach is proposed to understand the microstructures of dislocations based on deep learning, which estimates the total density of dislocations from a given density of GND distributions. In this method, the convolutional neural networks (ConvNets) are applied to extract the hidden information in the GND distribution maps to understand the microstructures of dislocations. It is demonstrated that the pre-trained ConvNets can be used to predict the distribution of total dislocation density from a small GND density map. Moreover, this technique is further developed to post-process real EBSD images for α-Fe to estimate the average total dislocation density, which corresponds to stress increments from a Taylor hardening assumption that is in good agreement with experimental values. Compared with previous methods involving much effort to track individual dislocations or other quantities, the present machine learning method is quick and convenient to use.