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; Kristen Brosnan, General Electric Research; Alp Sehirlioglu, Case Western Reserve University
Monday 2:00 PM
November 2, 2020
Room: Virtual Meeting Room 42
Location: MS&T Virtual
2:00 PM
Introductory Comments: Materials Informatics for Images and Multi-dimensional Datasets: Amanda Krause1; 1University of Florida
Introductory Comments
2:05 PM Invited
Neural Networks and Community Driven Software for Scanning Transmission Electron Microscopy: James LeBeau1; 1MIT
I will highlight a number of scanning transmission electron microscopy (STEM) developments that have provided new insights into material properties and have the potential to dramatically accelerate materials characterization. I will present a core component to enable the autonomous electron microscope, the Universal Scripting Engine for Transmission Electron Microscopy (USETEM). We will show that this scripting engine is widely applicable and simplifies scripting to enable high-throughput atomic-level imaging of materials. As a first step towards this vision, a deep convolutional neural network is demonstrated that can be used to automate convergent beam electron diffraction pattern analysis. The process enables, for example, autonomous determination of sample thickness to within 1 nm and tilt to within a fraction of a milliradian, at real-time speeds. Automating the electron microscope using artificial intelligence will address data size, bias, and documentation concerns, providing improved inputs for machine learning algorithms for faster discovery of new materials.
2:35 PM
Developing Granular Dielectrics Based on Reconstructed Micro-CT Images: Kevin Hager1; Christina Wildfire2; Edward Sabolsky1; Terence Musho1; 1West Virginia University; 2National Energy Technology Laboratory
This research is focused on developing statistically equivalent finite element (FE) geometries of granular dielectrics from 3D micro-CT scans. The reconstructed geometries were being used in an electromagnetic FE solver to predict and develop new granular dielectrics. In this study, the dielectric material of interest was a coal char based material. The approach taken involves determining the particle statistics using ImageJ based on 3D micro-CT scans, reconstructing the geometry using a discrete element method (DEM), post-processing in Paraview, and exporting as a CAD neutral file to COMSOL. Particle statistics of interest include statistical distribution of Feret diameter and particle count throughout the entire stack of CT images. The DEM was used to provide a realistic deposition of the particles, where the volume fraction and packing of particles influence the effective dielectric properties.
2:55 PM
Feature Characterization of Electron Backscatter Patterns from Rotating Lattice Single Crystals Using Machine Learning: Evan Musterman1; Joshua Agar1; Volkmar Dierolf1; Himanshu Jain1; 1Lehigh University
Crystallographic information with high spatial resolution can be acquired in the scanning electron microscope through electron backscatter diffraction (EBSD) techniques. Diffracted electrons create Kikuchi bands across backscatter patterns which are fit to a particular crystal phase and orientation. These patterns, acquired on a pixel-by-pixel basis, create large multidimensional datasets which are generally reduced to a few parameters with conventional EBSD analysis. Using Sb2S3 rotating lattice single (RLS) crystals in glass for their novel crystallography, we demonstrate a novel example of unsupervised machine and deep learning, such as convolutional neural networks, to identify and visualize latent features in the EBSD datasets beyond the conventional analysis. These models exhibit the ability to distinguish crystal from glass and identify crystal rotation. A comparison of this analysis is made for an RLS crystal vs. a polycrystalline sample.
3:15 PM
Automated Optical Microscopy for Rapid Defect Screening: Andrew Kitahara1; Elizabeth Holm1; 1Carnegie Mellon University
Quality control processes in manufacturing settings may utilize optical microscopy as a primary product screening tool for quickly identifying obvious defects in surface microstructure. We present the outcomes of a project to automate this process in a manufacturing setting for high-throughput screening by utilizing a pre-trained convolutional neural network that was fine-tuned for defect classification. The classifier works in real-time using the microscope camera video output, and the motorized stage is used to manipulate the sample such that entire specimens can be analyzed with minimal human interaction. The primary goal is to reduce the human expert’s time expense on an easily automated task to allow more time to address more significant challenges. But more than this, we propose that this platform can be applied in academic research settings as a rapid data acquisition tool to complement the growing interest in materials informatics research disciplines.
3:35 PM
Assessment of the Ability of Laboratory Accelerated Corrosion Tests to Accurately Predict On-road Corrosion of 6xxx Al Alloys: Dadi Zhang1; Jayendran Srinivasan1; Jenifer Locke1; 1The Ohio State University
AA6061-T4 and T6 and AA6022-T4 were exposed to both laboratory accelerated corrosion tests and on-road exposure. The laboratory corrosion test methodologies examined include an immersion test, ASTM G110, and salt spray tests of ASTM B117, ASTM B368 (CASS), ASTM G85-A2 (MASTMAASIS), a cyclically modified version of ASTM B117, and GMW14872. On-road exposure was conducted for up to 2 years on The Ohio State University campus bus system. A combination of pitting and intergranular corrosion was observed on all alloys after on-road exposure and laboratory tests using acidified solutions. The resulting corrosion morphology was quantitatively characterized by fractal analysis using the box-counting method, the ratio of corrosion feature boundary length to the corroded area, and by use of an open-source convolutional neural network (GoogLeNet). It was found that laboratory tests utilizing acidic solutions generally outperformed other tests regarding ability to simulate corrosion morphologies after on-road exposure across all tested alloys.
3:55 PM
Towards Smart Categorization of Growth Morphology by Machine Learning: Kimberly Gliebe1; 1Case Western Reserve University
This project examines growth kinetics during thin film deposition by pairing the analysis technique reflection high energy electron diffraction (RHEED) with machine learning. The first phase of this project underlines the necessary descriptors within RHEED videos to distinguish crystals of varying growth mode while the second phase utilizes this descriptor in order to understand how the kinetics of growth for several materials compare. RHEED videos were broken into frames so that the length, width, and intensity of the diffraction patterns could be analyzed over time via a self-made R program. In the first phase of the research a support vector machine model classified SrRuO3 samples by growth condition using the descriptor of RHEED spot length with time. In the second phase, materials LiLaTiO3, SrTiO3, LiNdTiO3, and LaAlO3 were added and compared via unsupervised learning to better understand the relationship between their growth modes.