Characterization: Structural Descriptors, Data-Intensive Techniques, and Uncertainty Quantification: 3D Microstructure Descriptors & Uncertainty
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
Tuesday 2:00 PM
February 25, 2020
Room: Theater A-3
Location: San Diego Convention Ctr
Session Chair: Francesca Tavazza, National Institute of Standards and Technology; James Hogan, University of Alberta
2:00 PM Introductory Comments
2:05 PM
X-Ray Computed Tomography of 3D Crack Lattices in Advanced Ceramics and their Effect on Mechanical Response: James Hogan1; Calvin Lo1; Haoyang Li1; Brendan Koch1; Tomoko Sano2; 1University of Alberta; 2CCDEVCOM Army Research Lab
Brittle material fails via crack growth leading to coalescence and fragmentation, and then pulverization into powder. For seismically damaged geomaterials or multi-hit armor applications, material can have significant internal cracks but require reloading to completely fragment. Inducing these damage states produce complex internal structures with significant variability. To address this, computed x-ray tomography obtains a wealth of data on the internal crack networks, allowing for the quantification of damage metrics such as crack volume, orientations, and lengths. Compressive loading to failure obtains the changes in material properties such as stiffness and Poisson’s ratio. Variability in induced microstructure requires repeated characterization and testing, but intensive data collection allows for correlating physical damage to mechanical response. The data intensity of these experiments also allows us to model intact material microstructure as low damage material, transforming uncertainty in intact material properties into an indirect measure of internal defects.
2:25 PM
3D Morphological Characterization of Porous Cu by Vapor Phase Dealloying Zn-Cu Alloys: Qingkun Meng1; Kai Wang1; Changhang Zhao2; Lijie Zou2; Yibin Ren3; Mingyuan Ge4; Xianghui Xiao4; Wah-Keat Lee4; Yu-chen Karen Chen-Wiegart2; 1China University of Mining and Technology; 2Stony Brook University; 3Shenyang Ligong University; 4Brookhaven National Laboratory
Vapor phase dealloying (VPD) is a versatile and environmentally friendly method to fabricate bi-continuous open porous materials. Compared to chemical or metallic dealloying, it uniquely takes advantage of the saturated vapor pressure difference between constituent elements for selectively removing one of the components from an alloy. As an emerging method, the processing-structure relationship in the porous metals fabricated by VPD remains unexplored, which is critical to effectively design new porous metal by VPD. In our work, the three-dimensional (3D) morphology of porous Cu dealloyed from Zn-Cu precursor alloys, was visualized and quantified by synchrotron X-ray nano-tomography. The 3D geometric factors such feature size, shape, tortuosity and curvatures were measured to study the characteristics of the porous structures. We will discuss the mechanisms associated with the porous formation and 3D morphological evolution in VPD. Such understanding will facilitate future development of novel porous materials and applications by this emerging new method.
2:45 PM
Automated Anomaly Detection in Unlabeled Computed Tomography Images: Donald Loveland1; Hyojin Kim1; T. Yong-Jin Han1; 1Lawrence Livermore National Laboratory
X-ray computed tomography (CT) offers a non-destructive characterization method to analyze 3-dimensional structures. However, this process creates an abundance of unlabeled data that can be difficult to efficiently analyze. With advances in machine learning (ML), automated analysis has become an important area of research to expedite this process. That said, a common vulnerability of ML models come from anomalies that may create inaccurate predictions during inference time. In CT data for granular composite materials, anomalies can exist, such as vacancies, non-material object inclusions, and inconsistent contrasting, which prove difficult for trained scientists to identify, let alone ML methods. Our work explores clustering and pixel-wise signal processing to capture anomalies, finding that signal processing approaches perform better, especially in highly homogeneous cases. We also show that this technique is capable of providing clear visual feedback with a means of accurately quantifying anomalies within a given sample.
3:05 PM
Deep Convolutional Networks for Image Reconstruction from 3D Coherent X-ray Diffraction Imaging Data: Mathew Cherukara1; Henry Chan1; Subramanian Sankaranarayanan1; Youssef Nashed1; Ross Harder1; 1Argonne National Laboratory
Coherent X-ray diffraction imaging (CDI) is a powerful technique for operando characterization. Visualizing defects, dynamics, and structural evolution using CDI, however, remains a grand challenge since state-of-the-art iterative reconstruction algorithms for CDI data are time-consuming and computationally expensive, which precludes real-time feedback. Furthermore, the reconstruction algorithms require human inputs to guide their convergence, which is a very subjective process. I will describe our work in the use of deep convolutional networks (CDI NN) in accelerating the analysis of, and potentially increasing the robustness of image recovery from 3D X-ray diffraction data. Once trained, CDI NN is hundreds of times faster than traditional phase retrieval algorithms used for image reconstruction from coherent diffraction data, opening up the prospect of real-time 3D imaging at the nanoscale.
3:25 PM
Uncertainty Quantification of Far-field HEDM Measurements: Rachel Lim1; Joel Bernier2; Anthony Rollett1; Paul Shade3; 1Carnegie Mellon University; 2Lawrence Livermore National Laboratory; 3Air Force Research Laboratory
Far-field high energy x-ray diffraction microscopy (ff-HEDM) is a synchrotron-based x-ray technique which can be employed to study micromechanics in polycrystalline materials. A diffraction simulator was built to model the experimental setup with the variety of different parameters in order to quantify the uncertainty coming from experimental equipment and setup on the measurements center-of-mass, grain-averaged orientation, and grain-averaged elastic strain tensor. Additionally, the representative volume element (RVE) to get statistically significant data from ff-HEDM is studied for several different properties including elasticity and plasticity. It has been found that the RVE for elasticity increases with increasing macroscopic strain while the RVE for plasticity remains constant as macroscopic strain increases.
3:45 PM Break
4:05 PM
Uncertainty Quantification Techniques Applied to Ductile Damage Predictions in the 3rd Sandia Fracture Challenge: James Sobotka1; John McFarland1; 1Southwest Research Institute
The 3rd Sandia Fracture Challenge involved blind predictions of ductile damage in a unique 3D geometry built using a laser powder-bed fusion process. Here, efficient uncertainty quantification techniques are needed due to high computational costs per analysis and increased variability of material properties in additively manufactured alloys. This presentation describes a modeling and simulation framework with ensemble studies as the fundamental unit of analysis. This framework supports uncertainty quantification by rigorous statistical methods and fast-running surrogate models. In this framework, results from computational designs of experiment provide input for Gaussian process models operating on principal components that define the quantities of interest. Results shown in this presentation exercise the fast-running surrogate models to calibrate material property distributions and to predict load-displacement curves and load-strain curves at expected, 20th percentile, and 80th percentile levels. Measurements from corresponding experiments are shown for comparison.
4:25 PM
Uncertainty Propagation in a Multiscale CALPHAD-reinforced Elastochemical Phase-field Model: Vahid Attari1; Pejman Honarmandi1; Thien Duong1; Daniel Sauceda1; Douglas Allaire1; Raymundo Arroyave1; 1Texas A&M University
In any materials design framework, uncertainties across the chain of models alter the final outcome considerably. The quantification of these uncertainties across the chain is often a sobering task, requiring 1) extensive computational resources, 2) systematic automation of propagation processes, 3) defining/designing proper descriptors, and 4) rigorous analysis of large amount of results. In this work, a framework to propagate uncertainties in a chain of models that involve CALPHAD, microleasticity, and phase-field models is utilized to investigate the uncertainty in microstructure of Mg_2(Si_xSn_{1-x}) thermoelectric materials. First, Markov Chain Monte Carlo-based inference of the CALPHAD model parameters are carried out, and then advanced sampling schemes are used to propagate uncertainties across the model input space. High throughput phase-field simulations resulted in approximately 200,000 time series of synthetic microstructures. Moreover, machine learning techniques are employed to differentiate between the parameter space that induces phonon scattering versus mass scattering for better thermoelectric response.
4:45 PM
Machine Learning Reinforced Crystal Plasticity Modeling of Titanium-Aluminum Alloys under Uncertainty: Pinar Acar1; 1Virginia Tech
We present a two-step computational approach to achieve a robust crystal plasticity model for Titanium-Aluminum alloys by considering experimental uncertainty and using machine learning techniques. The two-step solution involves the validation of the global (component-scale) and local (grain-level) features with the experimental data. The first step identifies the lower and upper bounds of the unknown crystal plasticity parameters through an inverse problem that aims to match the computations with the stress-strain response test data. The second step validates the local features with surrogate-based optimization by minimizing the difference between the simulated and experimental microstructural textures. The Artificial Neural Network is used to generate the computationally efficient surrogate representation of the crystal plasticity model to simulate the microstructural texture. With this machine-learning reinforced two-step solution approach under uncertainty, we achieve significant improvements on the accurate representation of global and local microstructural features, over the previous deterministic solutions in the literature.
5:05 PM
Predicting Microstructure-sensitive Fatigue-crack Path in 3D Using a Machine Learning Framework: Kyle Pierson1; Aowabin Rahman1; Ashley Spear1; 1University of Utah
Service lives of structural components are often significantly influenced by initiation and evolution of microstructure-sensitive fatigue cracks; however, the dependence of crack propagation on microstructural features can be complex and difficult to model. In this talk, we present a convolutional neural network (CNN)-based framework to approximate the underlying function relating crack path to the relevant microstructural and micromechanical features. The key components of the framework include: (i) a feature-selection scheme to determine a lower-dimensional representation of spatially varying features; (ii) a CNN model to compute the local kink angle of the crack using two different parameterization strategies; and (iii) a dropout technique to compute the model uncertainties associated with the CNN predictions. In general, the CNN model performs comparatively better than other ML algorithms in predicting crack path – even when micromechanical fields are not available as inputs, as the CNN can account for the spatial distribution of microstructural features