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, GE Research
Wednesday 8:00 AM
October 12, 2022
Room: 310
Location: David L. Lawrence Convention Center
Session Chair: Amanda Krause, Carnegie Mellon University
8:00 AM
FeCrAl Alloy Design Utilizing Literature, Experiments, High Throughput Characterization, and Machine Learning: Sandipp Krishnan Ravi1; Andrew Hoffman1; Rajnikant Umretiya1; Bojun Feng1; Subhrajit Roychowdhury1; Sayan Ghosh1; Raul Rebak1; 1GE Research
Iron-Chromium-Aluminum (FeCrAl) alloys are considered as lead Accident Tolerant Fuel Cladding (ATF) candidate because of their ability to form an effective passive Al film during high temperature exposure. FeCrAl alloys also exhibit good hydrothermal corrosion in light water reactor (LWR) operating conditions due to their Cr content. Though there is a significant amount of industrial data available on FeCrAl alloy behavior at high temperature environments (e.g. catalytic converter), there is a need of generating oxidation behavior data at relevant lower temperature conditions pertinent to LWR operating conditions. GE Research is conducting experiments looking at the phase stability, corrosion behavior, and mechanical properties of FeCrAl alloys with varying compositions and microstructures. A material discovery endeavor (based on alloy chemistry optimization) is undertaken through the framework of Bayesian active learning and probabilistic machine learning to develop FeCrAl alloys for LWR applications. Results from the experiments and models will be presented and discussed.
8:40 AM
Multimodal Data of Fatigue Fracture Surfaces for Analysis in a CNN: Katelyn Jones1; Elizabeth Holm1; Anthony Rollett1; 1Carnegie Mellon University
This work seeks to collect SEM, BSE, and Scanning White Light Interferometry (SWLI) images of Ti-6AL-4V fatigue fracture surfaces and apply Convolutional Neural Networks (CNNs) to identify high stress points, crack initiation sites, and predict values such as stress intensity factor and crack growth rate. SEM images are the standard for studying the topography of fracture surfaces, but BSE images and SWLI data offer the addition of compositional and surface height information as well. Computer Vision and Machine Learning were developed for optical images but have been successfully applied to electron images and a variety of other media. CNNs have been used to make successful classification and predictions of fracture surfaces. The images used to develop this model, creation of the CNNs, identified fatigue properties, and fracture characteristics will be presented.
9:00 AM
Establishing PSP Relationships with Microstructure Features Quantified Using Machine Learning: Joshua Stuckner1; 1NASA Glenn Research Center
Neural network encoders were pre-trained on over 100,000 micrographs from NASA and the literature to learn robust microstructure representations. The pre-trained encoders were applied through transfer learning and individually fine-tuned to segment and extract features from micrographs of different materials. The extracted features were then linked to processing and/or property data in order to establish quantitative processing-structure-property relationships. The presentation will demonstrate the technique on several materials including: Ni-superalloys where precipitate morphology and matrix channel width are quantified, and environmental barrier coatings where a thermally grown oxide is segmented and its roughness, thickness, porosity, and inter-crack spacing are quantified and related to processing.
9:20 AM
Process-Structure-Property Relationships from Variational Autoencoders: Michael White1; N.H. Gowtham2; Christopher Race1; Philip Withers1; Bikramjit Basu2; 1University of Manchester; 2Indian Institute of Science
Microstructure provides the link between processing and properties. A key aim of metallurgy is to understand how microstructural features influence properties and how processing determines microstructure. Although microstructure is well understood in this context, we often use qualitative descriptions to make comparisons between materials. A variety of metrics can be applied to form a quantitative description, but the metrics that are suitable depend on the material and any processing it has undergone – there is no catch-all method for describing microstructure quantitively. Since the rise of machine learning in materials informatics, many new approaches to constructing quantitative descriptions of image data have become available.Here, we explore the use of variational autoencoders (VAEs) for constructing encoded representations of microstructure from image data for titanium alloys. We then use these encoded representations as inputs to a variety of tasks, including prediction of processing parameters and mechanical properties, highlighting process-structure-property relationships.
9:40 AM
Polycrystal Graph Neural Network: Minyi Dai1; Mehmet Demirel1; Xuanhan Liu1; Yingyu Liang1; Jia-Mian Hu1; 1University of Wisconsin-Madison
Graph Neural Network (GNN) has recently emerged as a powerful machine learning model for predicting the properties of molecular and crystal structures, but its application to 3D, topologically complex polycrystalline microstructures still remains scarce. Here, we develop a Polycrystal Graph Neural Network (PGNN) model that permits an accurate prediction of the properties of polycrystalline microstructures by considering the physical features and interactions of both grain and grain boundaries. Trained with 4000 data points, our PGNN model achieves a property prediction error of ~1.5%, which is significantly lower than baseline machine learning models such as ResNet (error ~4%). We also show that such trained PGNN model can be transferred to accelerate and improve the prediction of other physical properties with smaller available data. Our accurate, and transferable PGNN model is well suited for harnessing large-scale datasets of 3D polycrystalline microstructures, which is crucial for realizing accelerated design of polycrystalline materials.
10:00 AM Break
10:20 AM
Automated Defect Identification for Tristructural Isotropic Fuels: Joseph Oncken1; Nancy Lybeck1; Jeffrey Phillps1; Scott Niedzialek2; Justin Coleman1; 1Idaho National Laboratory; 2BWX Technologies
During the manufacture of tristructural-isotropic-coated nuclear fuel particles, internal fissure defects can form in the uranium oxycarbide (UCO) kernels. These fissures result in a defective fuel particle that can fracture during subsequent fuel processing. Therefore, it is necessary to detect fissured kernels in a batch to determine if the batch meets specification prior to blending with other batches and upgrading processes. Previous attempts at identifying fissures involved the manual inspection of micrographs of UCO fuel kernel cross sections. This process is tedious, time-consuming, and may introduce counting errors, making it a good candidate for automation. This work presents an automated detection method for fissures in UCO kernels, using image segmentation to extract relevant features from micrographs, which then serve as the input to a convolutional neural network used to automatically distinguish between fissured and non-fissured kernels.