5th International Congress on 3D Materials Science (3DMS 2021): Data Processing and Machine Learning II
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

Thursday 2:10 PM
July 1, 2021
Room: Virtual C
Location: Virtual

Session Chair: McLean Echlin, University of California Santa Barbara


A Surface-mesh Gradation Tool for Generating Optimized Tetrahedral Meshes of Microstructures with Defects: Brian Phung1; Junyan He1; Ashley Spear1; 1University of Utah
    A meshing tool for generating tetrahedral meshes of polycrystalline microstructures with an optimal element-size distribution, determined by underlying defects (e.g., cracks and voids), is presented. The meshing tool provides users the capability to tune the element-size distribution by specifying parameters such as desired edge lengths and refinement zone radii around defects. Surface meshes can be obtained from 3D image data generated from experiments or tools such as DREAM.3D. Remeshing is performed directly on the original surface mesh via edge splitting and collapsing operations. The modified surface mesh is used in conjunction with a volume mesher to create a gradated volume mesh optimized for the defect(s). Relative to an unmodified uniform mesh, the framework is shown to significantly reduce element count and improve computational efficiency while ensuring sufficient mesh refinement in regions with high gradients. The meshing capabilities are demonstrated with examples involving crack propagation within polycrystalline microstructures.

Advanced Acquisition Strategies for Laboratory Diffraction Contrast Tomography (LabDCT): Jette Oddershede1; Jun Sun1; Florian Bachmann1; Erik Lauridsen1; 1Xnovo Technology
    Imaging the three-dimensional grain microstructure of polycrystalline material nondestructively is key to better understanding of the material performance. Recent advances of Laboratory Diffraction Contrast Tomography (LabDCT) allow to record and reconstruct larger representative volumes seamlessly. We will present and discuss different acquisition strategies with emphasis on how to approach a given acquisition problem inherent to the sample. Combining the 3D grain microstructure, i.e. grain morphology and crystallographic orientation, together with traditional absorption contrast tomography gives unprecedented insights into materials structure. Time resolved studies of the response of the material under investigation to external stimuli in- or ex-situ can be conducted.

Automatic Segmentation of 3D Microstructures of Steel Using Machine Learning Methods: Hoheok Kim1; Yuuki Arisato1; Junya Inoue1; 1The University of Tokyo
    Microstructure greatly influences mechanical and chemical properties, so many endeavors have been made to characterize it. Especially, 3D microstructural characterization of steel is very challenging due to the existence of various phases as well as the high cost for delicate 3D observations. Recently, researches using machine learning methods have shown good performances in classifying 3D microstructures. However, those approaches are generally based on supervised learning algorithms that require preparation of labeled datasets, which is not only time-consuming but also difficult even for experts. In this study, we propose an unsupervised algorithm that performs 3D microstructure segmentation without the need for labeled datasets. The new method, which is a combination of feature extraction with filters for texture analysis and a clustering algorithm, is applied to a 3D image obtained with a serial sectioning technique and optical microscope. The segmentation result shows that 3D microstructures are well divided into constituent phases.

Formation and Annihilation of Thin Twins in Hexagonal Close Packed Materials: Hamidreza Abdolvand1; Karim Louca1; Charles Mareau2; Marta Majkut3; Jon Wright3; 1The University of Western Ontario; 2Arts et Metiers ParisTech; 3European Synchrotron Radiation Facility
    Slip and twinning contribute significantly to plastic deformation in hexagonal close packed (HCP) materials. In this study, zirconium and magnesium specimens with HCP crystals are deformed in-situ to study formation and annihilation of twins using three-dimensional synchrotron X-ray diffraction. Methods are developed for accurate determination of grain properties, results of which are compared to those from electron backscatter diffraction measurements. New methods are developed to find the twins that form with loading and determine their corresponding parents. The measured initial microstructures are mapped into crystal plasticity models to study the evolution of grain resolved stresses. It is shown that twins are stressed when they are thin but relax with further loading. While a sign reversal is observed for the resolved shear stress (RSS) acting on the twin habit plane in the parent, the sign of RSS within the majority of twins stays unchanged until twin annihilation during the load reversal.

Predicting Microstructure-dependent Mechanical Properties in Additively Manufactured Metals Using Machine- and Deep-learning Methods: Carl Herriott1; Ashley Spear1; 1University of Utah
    The efficacy of machine-learning (ML) and deep-learning (DL) models to predict microstructure-sensitive mechanical properties in metal additive manufacturing (AM) is assessed using results from high-fidelity, multi-physics simulations as training data. Build domains exhibiting vastly different microstructures of AM SS316L were generated using the physics-based framework. Microstructural subvolumes and corresponding homogenized yield-strength values (~7700 data points) were then used to train two types of ML models (Ridge regression and XGBoost) and one type of DL model (convolutional neural network, CNN). The ML models require substantial pre-processing to extract volume-averaged microstructural descriptors; whereas, 3D image data comprising basic microstructural information are input to the CNN models. Among all models tested, the CNN models that use crystal orientation as input provide the best predictions, require little pre-processing, and predict spatial-property maps in a matter of seconds. Results demonstrate that suitably trained data-driven models can complement physics-driven modeling by massively expediting structure-property predictions.

Reconstruction of Microstructure and Defects in an Alpha+Beta Processed Ti-6Al-4V Plate Product Using High-energy X-ray Diffraction Microscopy and DREAM.3D: Krzysztof Stopka1; Jun-Sang Park2; Hemant Sharma2; Andrew Chuang2; Peter Kenesei2; Yan Gao3; Thomas Broderick4; William Musinski4; Paul Shade4; Sean Donegan4; Michael Jackson5; Jonathan Almer2; David McDowell1; 1Georgia Institute of Technology; 2Argonne National Laboratory; 3GE Global Research Center; 4U.S. Air Force Research Laboratory; 5BlueQuartz Software LLC
    In situ near-field (NF) and far-field (FF) high-energy x-ray diffraction microscopy (HEDM) and micro-computed tomography (μCT) were conducted on an alpha+beta processed Ti-6Al-4V specimen subject to cyclic loading to non-destructively characterize the evolution of microstructure associated with short fatigue crack growth (SFCG) and micromechanical response. A novel reconstruction and data fusion approach was developed to obtain a detailed 3D microstructure using DREAM.3D; the NF-HEDM was used to reconstruct the alpha grains while μCT distinguished the alpha and beta regions in the volume of interest. Combined with appropriate boundary conditions and specification of initial micromechanical state, the ensuing virtual microstructure will be used in micromechanical simulations. These simulation results can be compared to the FF-HEDM data to understand the SFCG process in Ti-6Al-4V.

Supervised Texture-based Classification for 3DEM: Alexander Hall1; Remi Blanc1; Jeffrey Caplan2; Madesh Muniswamy3; 1Thermo Fisher Scientific; 2University of Delaware; 3University of Texas Health, San Antonio
    Dense 3DEM data recovered by Serial Block Face Imaging (SBFI), Serial Section Transmission Electron Microscopy (ssTEM), and Focused Ion Beam Scanning Electron Microscopy (FIB-SEM) present challenging image segmentation needs. Traditional image processing may struggle to faithfully follow features. The variance in electron density between materials provides ample signal for machine learning methods, though. Here, we demonstrate supervised image classification based on image texture as implemented in Amira-Avizo Software. One provides training data from manual segmentation or automated methods. The program then trains a classifier using two statistical categories: features based on co-occurrence matrices and features based on intensity statistics. We tested our implementation on a 10 Gb SBFI mouse heart tissue block collected using an Apreo VolumeScope SEM. When followed by a small amount of image processing, the texture-supervised classification was able to fully label this data set with less than 10 minutes of user effort.