AI for Big Data Problems in Imaging, Modeling and Synthesis: AI-enabled Materials Characterization
Program Organizers: Mathew Cherukara, Argonne National Laboratory; Badri Narayanan, University of Louisville; Subramanian Sankaranarayanan, University of Illinois (Chicago)

Wednesday 2:00 PM
November 4, 2020
Room: Virtual Meeting Room 41
Location: MS&T Virtual

Session Chair: Mathew Cherukara, Argonne National Laboratory; Badri Narayanan, University of Louisville; Subramanian Sankaranarayanan, University of Illinois (Chicago)


2:00 PM  
Image Characterization of Self-assembled Photonic Crystals and Glasses Using Machine Learning: Bo Lei1; Yen Häntsch2; Gerold Schneider2; Kaline Furlan2; Elizabeth Holm1; 1Carnegie Mellon University; 2Hamburg University of Technology
    Photonic glasses fabricated by self-assembly of polystyrene particles demonstrate unique optical properties known as structural colors. The reflectance properties of such materials are governed by the structural parameters, the degree of order of the particles, and the material refractive index. Synchrotron X-ray tomography can be used for high-resolution structural characterization, but it is costly and time-consuming, thereby not suitable for fast characterization. Here, we propose that image analysis of SEM micrographs can be a very efficient approach for structural characterization. With the help of machine learning methods, we can achieve great accuracy in the classification of materials with different levels of disorder. We also show that it is possible to quantify the local structure and link it to the optical properties using image segmentation and machine learning.

2:20 PM  
A Hybrid EBSD Indexing Method Powered by Convolutional Neural Network (CNN) and Dictionary Indexing (DI): Zihao Ding1; Marc De Graef1; 1Carnegie Mellon University
    Based on Dictionary Indexing (DI) and convolutional neural network (CNN) methods our group developed, we propose a novel hybrid EBSD indexing technique, combining advantages of both approaches. Different from an end-to-end regression CNN, the neural net here provides an efficient classification of orientation interval given an EBSD pattern, while DI precisely determines the final orientation. The classification part is trained by simulated EBSD patterns from the EMsoft-based forward model and reduces the workload of DI; thus, the indexing rate of the whole system is greatly improved. The noise resistance and indexing accuracy of DI are preserved in the hybrid method. Through tests on experimental data, we show that machine learning methods can be applied to accelerate conventional EBSD indexing without a loss of robustness.

2:40 PM  
The Composition-microstructure-property Relationship by Machine Learning: Zongrui Pei1; Michael Gao1; Kyle Rozman1; Tao Liu1; David Alman1; Jeffrey Hawk1; 1National Energy Technology Laboratory
    We present our latest proceeding of machine learning microstructure images of 9-12Cr martensitic/ferritic steels. The variational autoencoder (VAE) models are used to extract the features of Scanning Electron Microscopy (SEM) images. The goal of this study is two folds: (i) prediction of mechanical properties for given images for a type of alloy microstructure; (ii) generation of the microstructure for alloys given their compositions and heat treatment conditions. The two sub-aims are of great importance in design of novel materials. Once realized, the materials design process can be guided by machine learning algorithms. This will render the design process not only more reliable but more efficient as well. In this talk, we will present the machine-learned relation between composition and microstructures, and the relation between microstructures and yield stresses in 2D latent space. These pictures, offered by the VAE models, allow for straightforward demonstrations of the complex relationships among composition-microstructure-property.

3:00 PM  
Instance Segmentation for Autonomous Detection of Individual Powder Particles and Satellites in an Additive Manufacturing Feedstock Powder: Ryan Cohn1; Elizabeth Holm1; 1Carnegie Mellon University
    Materials microstructures often contain multiple instances of a salient feature, and microstructural science involves quantifying these features individually and/or statistically. For example, using a computer vision approach, we can characterize a metal powder by analyzing each individual particle in an image. However, this analysis is challenged when particles touch or overlap. In this study, we take advantage of recent advances in deep learning to perform instance segmentation, in which individual segmentation masks are generated for each occurrence of a feature. For example, in an image of overlapping powder particles, instance segmentation allows individual particles to be extracted for further analysis. When combined with a machine learning classification scheme, we use this approach to measure the satellite content of powder samples, which is not possible with conventional powder characterization or image analysis techniques. This overall approach can be generalized to evaluate repetitive microstructural features across a range of structures.