5th International Congress on 3D Materials Science (3DMS 2021): Data Processing and Machine Learning I
Program Organizers: Dorte Juul Jensen, Technical University of Denmark; Erica Lilleodden, Fraunhofer Insitute for Microstructure of Materials and Systems (IMWS); Scott Barnett, Northwestern University; Keith Knipling, Naval Research Laboratory; Matthew Miller, Cornell University; Akira Taniyama, The Japan Institute of Metals and Materials; Hiroyuki Toda, Kyushu University; Lei Zhang, Chinese Academy of Sciences
Wednesday 12:10 PM
June 30, 2021
Room: Virtual C
Location: Virtual
Session Chair: Peter Voorhees, Northwestern University
Application of Machine Learning to 3D Reconstruction of SOFC Electrodes: Sicen Du1; Scott Barnett2; Katsuyo Thornton1; 1University of Michigan; 2Northwestern University
Solid oxide fuel cells (SOFCs) are one of the promising energy conversion devices with pollution-free and high efficiency under elevated temperature operations. Understanding the source and mechanism of microstructural degradation as well as linking microstructure to performance is vital to optimizing the performance of SOFCs. In recent years, 3D reconstruction is the state-of-the-art approach used in microstructure investigations, within which the image segmentation is the most critical step. The 2D images of SOFC electrodes obtained from FIB-SEM or XCT are grayscale images with unavoidable noises and artifacts, which makes microstructural analyses difficult. In our work, we present an advanced image processing approach to achieve high-quality microstructure reconstruction of a SOFC electrodes enhanced by machine learning techniques. We also generated an artificial dataset with similar noises compared to the experimental data to enable quantification of errors resulting from different algorithms.
Challenges in 3D Correlative Tomography and Microscopy for Materials Science: Bartlomiej Winiarski1; Remco Geurts1; Grzegorz Pyka2; Ali Chirazi1; Daniel Lichau2; 1Thermo Fisher Scientific; 2Université Catholique de Louvain
Multi-scale/multi-modal correlative methods, which involve the coordinated characterization of materials across a range of length scales using various apparatus, allowing solving broad range of scientific problems previously unreachable by the typical experimental operando. However, developing a repeatable/adaptable protocol for multi-scale/multi-instrument investigations is difficult to bring together. Problems arise as sub-samples are dissected from samples and remounted using various specimen holders needed for different apparatus. These generate 3D data clusters having different imaging modalities with range of voxels sizes and each in its own coordinate system. In this work we address these challenges with new, automated and multi-scale correlative workflows. These correlative methods use metrological X-Ray µCT helical scanner, multi-ion Plasma DualBeam and fs-Laser Plasma TriBeam platforms, cross-platform correlative holder kits and integrated image processing and analyses routines. As the practical example of the correlative solutions we explore microstructural features and imperfections of a glass-fiber reinforced polymer composite for automotive applications.
Characterization of 3-D Slip Fields in Deforming Polycrystals: Darren Pagan1; Kelly Nygren1; Matthew Miller2; 1Cornell High Energy Synchrotron Source; 2Cornell University
The interactions of slip systems at grain boundaries have been posited to be a critical factor in stress localization and subsequent nucleation of damage, especially during the dwell fatigue process in titanium alloys. To test these hypotheses, quantitative methods are needed to characterize slip activity in-situ in the bulk of deforming polycrystals. Here we present a new methodology that combines measurements of grain average stresses and lattice orientation fields made using high-energy X-ray diffraction microscopy (HEDM) with crystal plasticity kinematics to reconstruct full 3-dimensional slip activity fields with micron-scale resolution. The utility of the method will be demonstrated through analysis of slip activity in Ti-7Al deformed in uniaxial tension with a focus on analyzing slip mismatch at grain boundaries.
Comprehensive 3D Microstructural Characterization of Nuclear Fuel: Casey McKinney1; Assel Aitkaliyeva1; 1University of Florida
In this work, data sets gathered from focused ion beam (FIB) tomography with complementary electron backscatter diffraction (EBSD) and energy dispersive X-ray spectroscopy (EDS) are used to reconstruct the 3D microstructure of irradiated oxide fuels. The incorporation of EBSD and EDS data sets allows for both microstructural and microchemical characterization, which is particularly important for nuclear fuels where harsh reactor environments (temperature, irradiation, pressure) cause microstructural changes. The large thermal gradients produce a non-uniform microstructure across the pellet radius where regions separated by a few millimeters can have grain sizes differing by a factor of 100. When an atom fissions, it splits into different atoms that agglomerate into new phases and precipitates. These fission products can have different properties compared to the fuel, which can alter fuel performance. Comprehensive 3D characterization provides insight into the evolution of these microstructural aspects that could ultimately compromise the safety of the reactor.
Direct Estimation of Structure Parameters from 3D Image Data without Segmentation: Elise Brenne1; Vedrana Dahl1; Peter Jørgensen1; 1Denmark Technical University
Materials science based on 3D imaging and quantitative analysis is often impeded by the analysis of the image data rather than their acquisition. Human bias and labor-intensive processing of large datasets are of concern. We present a new model for the distribution of intensity and gradient magnitude values in volumetric datasets such as from X-ray computed tomography. In contrast to conventional statistical models such as the Gaussian mixture model, our model accounts for mixed-material interface voxels stemming from blurring inherent to the imaging technique. This results in an improved model fit based on physical parameters the sample and the imaging system. We demonstrate how the model allows for direct extraction of physical sample parameters, like volume fractions and interface areas, without segmenting the data first. The model has potential to improve segmentation accuracy and reproducibility, and with automated maximum-likelihood model parameter estimation, bias from manual segmentation parameter tuning is avoided.
Meaningful Learning in Materials Engineering: Learning through Virtual Reality Learning Environments: Lilian Davila1; Diego Vergara2; Jamil Extremera3; Manuel Pablo Rubio3; 1University of California, Merced; 2Catholic University of Ávila; 3University of Salamanca
The increasing dissemination of virtual reality learning environments (VRLEs) compels the elucidation of how these didactic tools can improve their effectiveness at the formative level. The motivation generated in students by a VRLE is revealed as a key factor in achieving meaningful learning, but such a motivation alone does not guarantee the long-term retention of knowledge. To identify the necessary characteristics of a VRLE to achieve an appropriate level of meaningful learning, this work compares a set of VRLEs created in previous years with a group of recently developed VRLEs, after being used by students. A description of the design process of the both VRLEs groups is included. Analysis of the response of students in a survey reveals how a protocol system helped improve students' knowledge and retention after a year of using a VRLE. This study demonstrates the importance of using modern engines to achieve long-term retention of knowledge.
Using Distance Metrics to Evaluate 3D RVE Size for Micromechanics and Texture: Rachel Lim1; Joseph Pauza1; Matthew Wilkin1; Anthony Rollett1; 1Carnegie Mellon University
A representative volume element (RVE) is the smallest possible volume element of a material which is statistically representative of the macroscopic properties. The extended time required to capture 3D microstructural data and to run full-field micromechanical simulations drive a need for determination and implementation of RVEs. Far-field high energy x-ray diffraction experiments were carried out to study RVE size vs micromechanical response. These results were supplemented by simulation work on the effect of varying texture on volume element size. Distance metrics (e.g. Hellinger distance) were applied to resolved shear stress distributions throughout the work to evaluate RVE size. The work intends to help improve efficiency in both 3D data acquisition as well as efficiency in performing micromechanical simulations.