Indexing of EBSD Patterns Using a Convolutional Neural Network
Talk, TMS 2020, San Diego, California
We propose a new convolution neural network (EBSD-CNN) with residual block and separable convolution to realize high accuracy and near real-time indexing of EBSD patterns. The integrated output of unit quaternions and a disorientation loss function are implemented to adapt the neural net for crystallographic orientation indexing. In addition to validating on simulated EBSD patterns, data from a series of experiments on Nickel with various exposure time have also been tested to study the network’s robustness against pattern noise. The results suggest that a CNN can provide an alternative indexing method to the commercial Hough-transform-based indexing with comparable accuracy and indexing rate. To gain insight into the model, we provide for a visualization of the filters as well as intermediate output in the network. As more features are extracted during the process, the approach also shows potential to measure other material properties that are encoded inside the EBSD patterns.