Characterization: Structural Descriptors, Data-Intensive Techniques, and Uncertainty Quantification: Microscopy & Machine Learning
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS Extraction and Processing Division, TMS: Computational Materials Science and Engineering Committee, TMS: Materials Characterization Committee
Program Organizers: Shawn Coleman, CCDC Army Research Laboratory; Tomoko Sano, U.S. Army Research Laboratory; James Hogan, University of Alberta; Srikanth Patala, North Carolina State University; Oliver Johnson, Brigham Young University; Francesca Tavazza, Nist
Wednesday 2:00 PM
February 26, 2020
Room: Theater A-3
Location: San Diego Convention Ctr
Session Chair: Oliver Johnson, Brigham Young University; Francesca Tavazza, National Institute of Standards and Technology
2:00 PM Introductory Comments
2:05 PM Invited
Neural Networks for Real-time Processing of Scanning Transmission Electron Microscopy Data: James LeBeau1; 1Massachusetts Institute of Technology
In this talk, I will discuss our recent work implementing deep convolutional neural networks to autonomously quantify electron diffraction and image data. I will discuss how the network was trained to automatically calibrate the zero-order diffraction disk size, locate the center, and determine pattern rotation without the need for other data pretreatment. The performance of the network to measure sample thickness and tilt will be explored as a function of a variety of variables including thickness, tilt, and dose. The processing speed will also shown to outpace a least squares approach by orders of magnitude. We will also discuss the generality of the method to other materials/orientations as well as a hybrid approach that combines the features of the neural network with least squares fitting for even more robust analysis.
Application of Machine Learning to Microstructure Quantification and Understanding: Ryan Noraas1; Greg Levan1; Asa Fry1; Iuliana Cernatescu1; 1Pratt & Whitney
Microstructure quantification typically involves measurement of features in 2D or 3D to establish averages and in certain cases a distribution of feature parameters. Limits in dynamic properties of materials e.g. fatigue and fracture are often related to the tails of microstructural feature distributions and assemblages. While it is important to understand structure – property relationships relative to physical features within a microstructure, not all critical features can be determined a priori or reduced to single point metrics like average grain size. This requires new approaches to guide characterization and understanding of microstructures. The use of machine learning tools is opening up new approaches to statistically quantify microstructures and to identify controlling features within microstructures. Examples of these new approaches for the development and control of new and legacy materials will be presented.
Adversarial Networks for Microstructure Generation and Modeling Phase Transformation Kinetics: Wufei Ma1; Elizabeth Kautz2; Arun Devaraj2; Saumyadeep Jana2; Vineet Joshi2; Daniel Lewis1; Bulent Yener1; 1Rensselaer Polytechnic Institute; 2Pacific Northwest National Laboratory
Micrograph quantification is essential to studying kinetics of phase transformations in several metallic systems. Machine learning has previously demonstrated success in image recognition tasks across several disciplines, however, sufficient image data available for model training is critical for success. In materials science studies, original image data can be limited, and is thus a hurdle to overcome when developing machine learning for microstructure image recognition tasks. Here, we develop a conditional generative adversarial network architecture (CGAN) for the purpose of generating synthetic microstructure images that can help when original image data is limited for phase transformation studies. In this work, a uranium-molybdenum alloy that undergoes a discontinuous precipitation reaction during thermo-mechanical processing is studied. We hypothesize that by understanding how synthetic images are processed by the CGAN we can gain insight into phase transformation kinetics, particularly what features of the microstructure image data is of most interest to the transformation process.
Monte Carlo Studies of EBSPs Spectroscopy: Elena Pascal1; Patrick Callahan2; Saransh Singh3; Marc De Graef1; 1Carnegie Mellon University; 2Naval Research Laboratory; 3Lawrence Livermore National Laboratory
Pushing the angular resolution of electron backscatter diffraction (EBSD), commonly known as high angular resolution EBSD (HR-EBSD), has been yielding access to small elastic lattice strains information. However, the cross-correlation based approach to HR-EBSD assumes, and is limited to, a uniform electron energy distribution over the detector. Since the reflection geometry used in EBSD is not selected for this condition, the question of the true distribution and its impact on the accuracy limit of conventional indexing become critical. The energy and direction distribution of electrons in the SEM is commonly predicted from single scattering Monte Carlo models where the Bethe slowing down approximation renders the model easy to implement and fast to compute. Since computational power is no longer the limiting factor in explicit scattering modelling we will compare the similar predictions of the standard approach with the direct simulation model and the dielectric function MC implementations for EBSD geometry.