Materials Informatics for Images and Multi-dimensional Datasets: Session I
Sponsored by: ACerS Basic Science Division, ACerS Electronics Division
Program Organizers: Amanda Krause, Carnegie Mellon University; Alp Sehirlioglu, Case Western Reserve University; Daniel Ruscitto, General Electric
Wednesday 8:00 AM
October 20, 2021
Room: A124
Location: Greater Columbus Convention Center
Session Chair: Andrew Hoffman, Catalyst Science Solutions
8:00 AM Invited
Machine Learning and Image Processing Techniques for Materials Evaluation: Roger French1; Benjamin Pierce1; 1Case Western Reserve University
Statistical and machine learning techniques provide researchers with tools to evaluate and quantify materials’ performance. We describe a series of algorithms used to investigate photovoltaic cells via electroluminescence (EL) imaging and current-voltage (I-V) curves. Using convolutional neural networks (CNNs), we classify modules into groups based on which type and degree of degradation. An expansion of this idea uses data integration of IV features to produce predictive and inferential models of power and corrosion from EL images. However, these supervised learning models depend on prelabeled data; whereas our third machine learning algorithm, based on feature extraction and high-dimensional clustering, can mitigate this problem by sorting EL images based on the features detected. We also present a case study of large-scale image analysis of nucleation and growth of AlN crystals from an Al/Ni alloy using a high-performance, distributed computing approach, and the code packages and tooling involved in this image analysis.
8:30 AM
Quantitative Comparisons of 2D Microstructures with the Wasserstein Metric: Ethan Suwandi1; Jeremy Mason1; 1University of California Davis
Comparison of material microstructures is traditionally done using a selection of incomplete and ad-hoc statistics. With the advent of computational materials design, this has encouraged the development of microstructure generation techniques which produce synthetic microstructures that do not replicate a physical process. A more robust method of comparison would be useful both to compare experimental microstructures with emerging standards and to validate synthetic microstructure generation techniques. We propose using a balanced Wasserstein distance to quantify the difference between two micrographs with respect to both geometry and size simultaneously, where the measured is given by an inverted unsigned distance function to the grain boundary network. By finding the best pairwise matching of sets of micrographs taken from two microstructures, an overall similarity metric is established. This method is employed to compare a number of synthetic microstructures generated with similar starting statistics and varying features.
8:50 AM
Spatial and Statistical Representation of Strain Localization as a Function of the 3D Microstructure Using Multi-modal and Multi-scale Data Merging: Marie Charpagne1; J.C. Stinville1; Andrew T. Polonsky2; McLean P. Echlin1; Kelly Nygren3; Dalton Shadle3; Matthew P. Miller3; Tresa M. Pollock1; 1University of California, Santa Barbara; 2Sandia National Laboratories; 3Cornell University
Most structural materials exhibit a localized strain field upon loading. Whereas strain localization occurs in the form of shear bands, slip bands, deformation twins or other forms, it is expected to be highly correlated to the materials microstructure. The intensity and spatial distribution of such deformation structures directly influence most mechanical properties such as strength, ductility and fatigue life. Understanding strain localization processes as a function of the microstructure is therefore of critical importance, in the global aim of improving a materials mechanical properties. A framework for automated multi-modal data merging, involving the combination of digital image correlation captured in the scanning electron microscope and microstructure data collected using 3D electron backscatter diffraction will be presented. The use of computer vision tools and statistical microstructure descriptors enables a quantitative, automated and non-human biased analysis of strain localization patterns. Application examples will be shown in a superalloy and a titanium alloy.
9:10 AM
Building a Database of Fatigue Fracture Images to train a CNN: Katelyn Jones1; Paul Shade2; William Musinski2; Reji John2; Adam Pilchak2; Anthony Rollett1; Elizabeth Holm1; 1Carnegie Mellon University; 2Air Force Research Laboratory
Machine learning and computer vision techniques can be used in materials science to improve and facilitate the analysis of microstructural data and images. Additionally, it can work on large amounts of data and diverse images when enough training data is provided. Convolutional neural networks (CNNs) are a key tool in making connections between fracture images, microstructure, and fatigue characteristics such as stress intensity factor, crack length, and load values. This project collects images from a variety of Ti-6Al-4V fracture surfaces to create a database to train a CNN and identify high stress points, crack initiation sites, and predict values such as stress intensity factor. The images used to develop this model, creation of the CNNs, identified fatigue properties, and fracture characteristics will be presented.