ICME 2023: AI/ML: Microstructure I
Program Organizers: Charles Ward, AFRL/RXM; Heather Murdoch, U.S. Army Research Laboratory
Monday 9:50 AM
May 22, 2023
Room: Boca I-III
Location: Caribe Royale
Session Chair: Shankarjee Krishnamoorthi, ATI Specialty Materials
9:50 AM Invited
HPC+AI@Edge Enabled Real-time Materials Characterization: Mathew Cherukara1; 1Argonne National Laboratory
The capabilities provided by next generation light sources such as the APSU along with the development of new characterization techniques and detector advances are expected to revolutionize materials characterization (metrology) by providing the ability to perform scale-bridging, multi-modal materials characterization under in-situ and operando conditions. For example, providing the ability to image in 3D large fields of view (~mm3) at high resolution (<10 nm), while simultaneously acquiring information about structure, strain, elemental composition, oxidation state, photovoltaic response etc. However, these novel capabilities dramatically increase the complexity and volume of data generated by instruments at the new light sources. Conventional data processing and analysis methodologies become infeasible in the face of such large and varied data streams. The use of AI/ML methods is becoming indispensable for real-time analysis, data abstraction and decision making at advanced synchrotron light sources such as the APS. I will describe the use of high-performance computing (HPC) along with AI on edge devices to enable real-time analysis of streaming data from x-ray imaging instruments at the APS.
10:20 AM
Microstructure-sensitive Materials Design via Efficient Uncertainty Propagation and Process-structure-property Linkages: Vahid Attari1; Danial Khatamsaz1; Allison Kaye Ituralde Arabelo1; Douglas Allaire1; Raymundo Arroyave1; 1Texas A&M University
The field of Integrated Computational Materials Engineering (ICME) combines a broad range of methods to study materials’ responses over a spectrum of length scales. A relatively unexplored aspect of microstructure-sensitive materials design is uncertainty propagation and quantification (UP/UQ) of materials' microstructure, as well as establishing process-structure-property (PSP) relationships. An accurate UP technique built on the idea of changing probability measures for microstructure-based problems is proposed. Probability measures are used to represent microstructure space, and Wasserstein metrics are used to test the efficiency of the method. By using Variational AutoEncoder (VAE), we identify the correlations between the material/process parameters and the thermal conductivity of heterogeneous microstructures. Through high-throughput screening, UQ/UP, and deep-generative learning methods, PSP relationships that are too complicated/complex can be revealed through high-throughput exploitation of the materials' design space with an emphasis on microstructures.
10:40 AM
Deep Learning Enabled Additive Manufacturing (AM) Lattice Segmentation: Michael Juhasz1; Nick Calta1; 1Lawrence Livermore National Laboratory
Ex-situ computed tomography (CT) analysis of Additive Manufacturing (AM) produced parts is commonplace as a means of Non-Destructive Evaluation (NDE) quality assurance. Most CT examinations focus on porosity, both from keyholing or entrained gas. With the recent acceleration in image processing enabled through Deep Learning/Machine Learning (ML/DL), this presentation suggests expanding CT analysis of AM parts to extend beyond porosity analysis to encompass the study of other requirement-driven, critical geometries. As a case study, AM produced lattices underwent CT and were subsequently segmented into component pieces. These segmented components can be examined individually or as a larger subset where statistics can be taken, with an eye to drawing comparison and correlation to both ex-situ mechanical properties and in-situ process signatures. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.
11:00 AM
A Deep Learning Approach for Phase Detection in 2D-XRD Patterns of Ti-6Al-4V
: Weiqi Yue1; Pawan Tripathi1; Nathaniel Tomczak1; Gabriel Ponon1; Zhuldyz Ualikhankyzy1; Matthew Willard1; Vipin Chaudhary1; Roger French1; 1Case Western Reserve University
Beamline 2-D X-ray diffraction (XRD) diffractograms play an important role in determining properties associated with crystal structure such as crystalline phase, texture, and chemistry. Deep learning techniques help create an efficient and accurate solution capable of processing large amounts of image data. Herein, we analyzed patterns of a Ti-6Al-4V (Ti-64) alloy that was heat treated throughout capture of the diffraction patterns. We designed a convolutional neural network (CNN) to predict the titanium beta phase volume percentage during the temperature fluctuations. Images were pre-processed to remove bias in the XRD patterns and prevent loss of information. The 2-D XRD patterns were input directly into the CNN model instead of traditional 1-D XRD peak patterns. The model was trained using an experimental dataset that contains 3,012 XRD images and was tested with another 1,102 experimental XRD images, achieving a mean square error (MSE) of 0.076%.
11:20 AM
Understanding Grain Growth Using a Physics-regularized Interpretable Machine Learning Model: Joseph Melville1; Vishal Yadav1; Michael Tonks1; Amanda Krause2; Joel Harley1; 1University of Florida; 2Carnegie Mellon University
Physics-based mesoscale models of grain growth have been unable to accurately represent the grain growth behavior of real materials. We are developing the Physics Regularized Interpretable Machine Learning Microstructure Evolution (PRIMME) model that can learn to predict grain growth directly from experimental data. The PRIMME algorithm uses a multi-level neural network to predict grain growth in a voxelated domain. It uses a regularization function that encourages evolution that decreases the number of nearest neighbor voxels assigned to different grains. PRIMME helps to interpret and understand its learned grain growth behavior by determining the likelihood of a voxel changing to the grain of neighboring voxels. PRIMME was originally trained using data from 2D isotropic simulation results. It is now being extended to 3D isotropic and 2D anisotropic behavior. It will begin to be trained using experimental data rather than just simulation results in the near future.