6th International Congress on 3D Materials Science (3DMS 2022): 3D Data Processing III: Statistical Analyzes and Machine Learning
Program Organizers: Dorte Juul Jensen, Technical University of Denmark; Marie Charpagne, University of Illinois; Keith Knipling, Naval Research Laboratory; Klaus-Dieter Liss, University of Wollongong; Matthew Miller, Cornell University; David Rowenhorst, Naval Research Laboratory
Tuesday 9:30 AM
June 28, 2022
Room: Columbia A&B
Location: Hyatt Regency Washington on Capitol Hill
Session Chair: Keith Knipling, Naval Research Laboratory
9:30 AM Invited
Imposing Equilibrium on Measured 3-D Stress Fields Using Helmholtz Decomposition and FFT-based Optimization: Ricardo Lebensohn1; Hao Zhou2; Peter Reischig3; Wolfgang Ludwig4; Kaushik Bhattacharya2; 1Los Alamos National Laboratory; 2Caltech; 3InnoCryst Ltd; 4MATEIS, INSA Lyon
We present a methodology to impose micromechanical constraints to arbitrary voxelized stress fields obtained, e.g. by x-ray diffraction. The method consists in finding the equilibrated stress field closest to the measured field, posed as an optimization problem. The extraction of the divergence-free (equilibrated) stress is performed using the Hodge/Helmholtz decomposition of a symmetric matrix field. The combination of the latter with the Euler-Lagrange equations of the optimization gives an expression that contains the bi-harmonic operator and the curl operator acting twice on the measured stress field. These high-order derivatives are efficiently performed in Fourier space. The method is applied to filter: a) synthetic piecewise constant stress fields, b) synthetic equilibrated fields perturbed with noise, c) measured stress fields in Gum Metal, a beta-Ti alloy. In cases a) and c), the largest corrections were obtained near grain boundaries. In case b), the filter was able to partially de-noise the perturbed field.
10:00 AM
Automated 4D Reconstruction and Data-mining of Dislocation
structures From In-situ TEM Experiments: Synthetic Training Data
generation and Customized Deep Learning Strategies: Kishan Govind1; Marc Legros2; Stefan Sandfeld1; 1Institute for Advanced Simulation; 2CEMES-CNRS
Quantitative TEM imagingof dislocations helps to understand the structure-property relations directly from experimental data. Such studies are, up to now, typically based on coarse, geometrical estimates of the curvature of dislocations for a few snapshots in time. Generalizing this to thousands of frames and/or many dislocation is currently not possible. In this presentation we show how Deep Learning can be an important tool for data mining of TEM images of dislocation. We demonstrate how TEM images can be automatically segmented, and how individual dislocations can be tracked during time. For the training we generate artificial images of dislocation microstructure using a para- metric model. We demonstrate how such a new, physical data augmentation approach can be used in orderto overcome the common problem of “never enough training data”. This approach is used for the automated data-mining of dislocation motion in a high-entropy alloy.
10:20 AM
Using Deep Learning to Reconstruct Grains from Simulated Far-field Diffraction Data: Ashley Lenau1; Yuefeng Jin2; Ashley Bucsek3; Stephen Niezgoda1; 1Ohio State University; 2University of Michigan; 3University of Michigan
Far-Field High Energy Diffraction Microscopy (ff-HEDM) is invaluable for quantifying the orientation and elastic strain within the bulk of a 3D polycrystalline sample. However, it has limited ability to capture morphology or orientation and strain gradients. In this presentation, we demonstrate a deep learning framework that reconstructs the 3D grain shape given diffraction spots from a single grain. The deep learning framework is based on Pix2Vox, which uses an encoder-decoder structure to convert multiple 2D images of an object into a 3D volume render. Unlike standard Pix2Vox, which uses a single encode-decoder for all 2D images, our network utilizes an independent encoder for each diffraction spot. This network is demonstrated on synthetic 3D datasets. The ground truth grain shapes are generated via DREAM.3D and the simulated ff-HEDM data is generated by a virtual diffractometer. While still in the nascent stages of development, here we demonstrate high-fidelity 3D grain reconstruction.
10:40 AM Break