Data Science and Analytics for Materials Imaging and Quantification: Session I: Data-led Approaches for 2D Characterization & EBSD
Sponsored by: TMS Structural Materials Division, TMS Materials Processing and Manufacturing Division, TMS: Advanced Characterization, Testing, and Simulation Committee, TMS: Integrated Computational Materials Engineering Committee
Program Organizers: Emine Gulsoy, Northwestern University; Charudatta Phatak, Argonne National Laboratory; Stephan Wagner-Conrad, Carl Zeiss Microscopy; Marcus Hanwell, Brookhaven National Laboratory; David Rowenhorst, Naval Research Laboratory; Tiberiu Stan, Asml
Monday 8:30 AM
March 15, 2021
Room: RM 16
Location: TMS2021 Virtual
Session Chair: Emine Gulsoy, Northwestern University
8:35 AM
Computer Vision and Machine Learning for Microstructural Characterization and Analysis: Elizabeth Holm1; Ryan Cohn1; Nan Gao1; Katelyn Jones1; Bo Lei1; Srujana Yarasi1; 1Carnegie Mellon University
The characterization and analysis of microstructure is the foundation of microstructural science, connecting the materials structure to its composition, process history, and properties. Microstructural quantification traditionally involves a human deciding a priori what to measure and then devising a purpose-built method for doing so. However, recent advances in data science, including computer vision (CV) and machine learning (ML) offer new approaches to extracting information from microstructural images. This talk presents CV approaches to numerically encode the visual information contained in a microstructural image, which then provides input to supervised or unsupervised ML algorithms that find associations and trends in the high-dimensional image representation. CV/ML systems for microstructural characterization and analysis include image classification, semantic segmentation, object detection, and instance segmentation. These tools enable new approaches to microstructural analysis, including the development of new visual metrics.
9:00 AM
Microstructure Image Segmentation with Deep Learning: from Supervised to Unsupervised Methods: Bo Lei1; Elizabeth Holm1; 1Carnegie Mellon University
In quantitative microscopy, microstructure image segmentation is essential for image analysis and materials characterization. The rising deep convolutional neural network methods for semantic segmentation in natural images have recently transferred to materials images and demonstrated outstanding performance in complicated microstructure datasets. However, current supervised learning solutions require pixel-level human annotations, which is painstaking, biased and infeasible for some cases. It is worth exploring the possibility to go from fully supervised to semi-supervised and unsupervised methods with the goal of achieving comparable performance while alleviating the annotation cost. Here, we demonstrate our effort in moving from supervised methods to unsupervised methods. Multiple aspects of the strategies including dataset generation, annotation, transfer learning, evaluation, etc., are discussed.
9:20 AM
Improved EBSD Indexing through Non-Local Pattern Averaging: David Rowenhorst1; Patrick Brewick1; 1Naval Research Laboratory
Electron Backscattered Diffraction has become a widely used technique for the characterization of polycrystalline materials. While recent advances in detector technology have greatly increased data collection speeds, the total time for collection can still be a significant limiting factor for many analyses, especially in low signal-to-noise situations. This work takes lessons learned from de-noising algorithms in image processing to develop a Non-Local Pattern Averaging Reindexing (NLPAR) method that can utilize the highly redundant information contained within most EBSD scans to enhance the the pattern quality without losing signal integrity near the interface boundaries, all the while operating on computational timescales that are similar to the traditional Hough based indexing algorithms. A benchmark case study within nickel will be presented as well as examples that in martensitic steels and aluminum alloys which show that data can be collected 2-3x faster than with traditional methods.
9:45 AM
Resolving Pseudosymmetry in Tetragonal ZrO2 Using EBSD with a Modified Dictionary Indexing Approach: Edward Pang1; Peter Larsen1; Christopher Schuh1; 1Massachusetts Institute of Technology
Resolving pseudosymmetry has long presented a challenge for electron backscatter diffraction (EBSD) and has been notoriously challenging in the case of tetragonal ZrO2 in particular. In this work, a method is proposed to resolve pseudosymmetry by building upon the dictionary indexing method and augmenting it with the application of global optimization to fit accurate pattern centers, clustering of the Hough-indexed orientations to focus the dictionary in orientation space, and interpolation to improve the accuracy of the indexed solution. We demonstrate that the proposed method can successfully and efficiently resolve pseudosymmetry in simulated patterns of tetragonal ZrO2 with high degrees of binning and noise as well as experimental datasets. We then explore the effect of crystal orientation and camera distance on the ability to resolve pseudosymmetry using intensity-based indexing approaches.
10:05 AM
Dictionary Indexing of EBSD Patterns Assisted by Convolutional Neural Network: Zihao Ding1; Marc Graef1; 1Carnegie Mellon University
Though accurate and noise-resistant, conventional Dictionary Indexing (DI) on EBSD patterns is a computation-intensive task. For each experimental pattern, the process needs to calculate the dot product of it and all patterns in a large dictionary. We use a convolutional neural network (CNN) to greatly cut down the workload. It predicts a customized smaller dictionary for each experimental pattern before calling DI. The system combines the efficiency of machine learning methods and the advantages of DI. It turns out the total time required is only 15% of pure DI, while the accuracy is at the same level. We also take this opportunity to release Python APIs for EMsoft, which are used in this project. With great flexibility and support for multiprocessing, it allows researchers to construct DI workflow conveniently and efficiently in the study.
10:25 AM
Advancements in EBSD Using Machine Learning: Kevin Kaufmann1; Chaoyi Zhu1; Alexander Rosengarten1; Daniel Maryanovsky1; Tyler Harrington1; Hobson Lane2; 1University of California, San Diego; 2Tangible AI LLC
Electron backscatter diffraction (EBSD) is a powerful tool with the ability to collect diffraction patterns over large areas with relatively small step sizes, thus supporting multi-scale analysis. After EBSD pattern collection, current indexing techniques (whether Hough-based or dictionary pattern matching based) are capable of reliably differentiating between a user selected set of phases, if those phases contain sufficiently different crystal structures. Despite considerable efforts, the challenges of phase differentiation and identification remain. Recent improvements in EBSD detectors allow for unprecedented pattern collection rate and resolution, opening the door for implementing techniques from the data science field. This work demonstrates the application of convolutional neural networks for extracting crystallographic and chemical information from the information rich diffraction patterns and compares the results of this approach with solutions offered in commercial systems. Investigations into the internal mathematical operations of the “black box” algorithm operating on EBSD patterns will also be discussed.