Characterization: Structural Descriptors, Data-Intensive Techniques, and Uncertainty Quantification: Structural Descriptors Enabling PSP Linkages
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, DEVCOM Army Research Laboratory; Tomoko Sano, U.S. Army Research Laboratory; James Hogan, University of Alberta; Srikanth Patala, SASA Institute; Oliver Johnson, Brigham Young University; Francesca Tavazza, National Institute of Standards and Technology
Monday 2:30 PM
February 24, 2020
Room: Theater A-3
Location: San Diego Convention Ctr
Session Chair: James Hogan, University of Alberta; Tomoko Sano, CCDC Army Research Laboratory
2:30 PM Introductory Comments
2:35 PM Invited
Artificial Intelligence Approaches to Microstructural Science: Elizabeth Holm1; 1Carnegie Mellon University
The process of scientific inquiry involves observing a signal (data) and interpreting it to generate information (knowledge). Artificial intelligence (AI) – a broad term comprising data science, machine learning (ML), neural network computing, computer vision, and other technologies – opens new avenues for extracting information from high-dimensional materials data. This presentation will focus on AI applications in the context of multimodal image-based data, including experimental and simulated atomic structures, defect structures, and microstructures. The visual information contained in these images is numerically encoded using black box computer vision (CV) methods as well as feature-based representations. ML tools are then selected based on the characteristics of the data set and the desired outcome. Results range from advanced methods for microstructural segmentation and characterization to prediction of microstructural evolution and material properties. The ultimate goal is to develop AI as a new tool for information extraction and knowledge generation in materials science.
3:05 PM Invited
Methods for the Correction of Epistemic Resolution Error through Data Collection Process Simulations: Lori Graham-Brady1; Noah Wade1; 1Johns Hopkins University
The collection of high resolution 3D serial sectioned data sets has become increasingly important for investigation of material structures and properties. However, despite improvements in imaging technology, collecting high resolution information over statistically significant volumes requires significant experimental and computational resources. Often this results in a tradeoff between imaging resolution and sample volume size. Methods for examining the propagation of epistemic resolution error can provide valuable information for addressing this problem. By simulating the data collection process at progressively coarser resolutions on a small high resolution volume the propagation of epistemic resolution error can be quantified and predicted. Using these predictions, correction methods can be formulated and applied to lower resolution data sets to reduce error. This results in statistical properties from the low resolution data set which are significantly closer to those from the high resolution data set, ultimately leading to more accurate prediction of material properties.
3:35 PM
Determination of Representative Volume Elements for Small Cracks in Heterogeneous Domains via Convolutional Neural Networks: Karen Demille1; Ashley Spear1; 1University of Utah
Microstructurally small crack (MSC) behavior is strongly dependent on the microstructural features near the crack front. The characterization of microstructural features, through data-intensive characterization techniques such as x-ray imaging and serial sectioning, enables the relationship between MSC behavior and local microstructure to be studied. In such studies, a balance between representing enough microstructural volume and maintaining tractability must be achieved. Quantitative microstructural descriptors, used in conjunction with a convolutional neural network (CNN), can provide insights into materials characterization in terms of the minimum requisite volume for analyzing MSCs. Microstructural descriptors are obtained by sampling microstructural features at points around the crack front and using a CNN to reduce the dimensionality of the sampled features. These microstructural descriptors are used to predict convergence trends of crack-front parameters (e.g. J-integrals) with respect to heterogeneous-domain size. The convergence trends are used to inform the selection of the minimum requisite volume for analyzing MSCs.
3:55 PM Break
4:15 PM
Machine Learning Approaches to Image Segmentation of Large Materials Science Datasets: Tiberiu Stan1; Zachary Thompson1; Bo Lei2; Elizabeth Holm2; Peter Voorhees1; 1Northwestern University; 2Carnegie Mellon University
Modern imaging techniques generate an increasing amount of data that must be accurately analyzed to extract materials parameters. We have trained a variety of machine learning convolutional neural network (NN) architectures to perform semantic segmentation of large materials science datasets such as x-ray computed tomography, serial sectioning optical microscopy, and scanning electron microscopy. The images contain diverse microstructural features, length scales, and artifacts which make segmentation challenging. Many NN architectures have fundamentally different encoder and decoder networks, thus some architectures perform better on certain datasets than others. Ways to increase NN performance using limited training data, general best practice NN training methods, and NN transferability are discussed. Fully trained NNs can accurately segment images nearly 1000 times faster than humans and sematic segmentation is becoming a powerful tool for analysis of large datasets.
4:35 PM
Predicting Crack Location Using a Radial Distribution Function as a Unique Descriptor of Pore Networks: John Erickson1; Ashley Spear1; Aowabin Rahman1; 1University of Utah
Additive manufacturing (AM) has opened a wide range of new capabilities in the field of manufacturing. However, due to defects resulting from the build process (namely, porosity), the resulting mechanical behavior of AM specimens can be unpredictable. Our research introduces a new method of uniquely characterizing pore networks using a radial distribution function (RDF), which is then used to predict crack location. The RDF indicates/signals interconnectivity of pores by taking into account pore location, size, and distance to free surface. Using a finite-element-modeling framework, 120 tensile specimens with statistically similar pore networks were virtually tested to failure. The pore networks were characterized by the RDF, which was then compared to the nominal location of fracture. The RDF signal predicted crack locations within a 5% error for 88% of the cases and proved to be a more reliable indicator for fracture than fraction porosity, reduced-cross-section percent, and largest pore diameter.
4:55 PM
Investigating the Effect of Solute Segregation to Grain Boundaries in Nanocrystalline Alloys Toward Stability and Strengthening: Ankit Gupta1; Gregory Thompson2; Garritt Tucker1; 1Colorado School of Mines; 2University of Alabama
Nanocrystalline (NC) materials offer an improvement in several mechanical properties such as increased strength/hardness. However, poor stability of grain boundaries at such smaller length scales can lead to grain growth at higher temperatures and during mechanical loading, thereby, severely affecting the mechanical properties of NC materials. Addition of alloying elements have shown to drastically improve their stability. In this study, Interfacial solute segregation behavior of P in nanocrystalline (NC) Ni alloys is investigated on an atomic scale using cross-correlative PED-APT measurements and atomistic simulations. An inhomogeneous P-segregation in grain boundaries (GBs) of NC Ni is observed. The interfacial solute excess is characterized as a function of initial GB energy, volume and misorientation. It is shown that the solute segregation behavior of GBs, especially in NC structures, cannot be properly captured with average measures such as GB misorientation, energy. Improved indices for quantifying GB solute segregation behavior are developed.
5:15 PM
Predicting Compressive Strength of Consolidated Solids from Features Extracted from SEM Images: T. Yong Han1; 1Lawrence Livermore National Lab
Application of computer vision, machine learning, and deep learning in materials science can provide powerful tools to analyze and automate scientific data analysis. Here, we explored the application of computer vision and machine learning to quantify materials properties based on SEM images of materials microstructure. We showed that it is possible to train machine learning models to predict materials performance based on SEM images alone, demonstrating this capability on predicting uniaxially compressed peak stress of consolidated molecular solid samples. We explore two complementary approaches to this problem: (1) a traditional machine learning approach using state-of-the-art computer vision features and (2) an end-to-end deep learning approach, where features are learned automatically from raw images. We demonstrated that random forest performs best in the “small data” regime in which many real-world scientific applications reside, whereas deep learning outpaces random forest in the “big data” regime, where abundant training samples are available.
5:35 PM
Utilizing Convolutional Neural Networks for Prediction of Process and Material Parameters from Microstructural Images: Richard Couperthwaite1; Levi McClenny1; Jaylen James1; Vahid Attari1; Raymundo Arróyave1; Ulisses Braga Neto1; 1Texas A&M University
Microstructural images contain a wealth of information that can be used to determine both heat treatment conditions and mechanical properties. While standard methods for characterizing microstructures, such as grain size, and volume fraction of phases are useful, convolutional neural networks are capable of featurizing images in more complex ways. This ability can be utilized to generate feature information that can be used to fit regression models. A further step is that it is possible to terminate convolutional neural networks in a single node that is capable of providing a regression result directly from the neural network. Utilizing an image set of 10000 phase field microstructures, the current work considers the effect of different structures and secondary regression methods on the effectiveness of predicting various parameters used in the production and analysis of the microstructures.