ICME 2023: AI/ML: Microstructure III
Program Organizers: Charles Ward, AFRL/RXM; Heather Murdoch, U.S. Army Research Laboratory
Wednesday 9:40 AM
May 24, 2023
Room: Boca I-III
Location: Caribe Royale
Session Chair: Victoria Miller, University of Florida
9:40 AM
Chemistry and Processing Prediction for Targeted Microstructure Morphology: Mahmood Mamivand1; Amir Abbas Kazemzadeh Farizhandi1; 1Boise State University
Engineering a microstructure morphology has been a long-lasting challenge in materials science. Historically, forward-based models, including experimental and high-fidelity models, based on trial and error, have been used to engineer a specific microstructure. In this work, we have developed a novel fused-data deep learning algorithm that is able to predict the required chemistry and processing to reach specific microstructure morphologies. FeCrCo permanent magnets are the model alloy in this work. The model input is the Fe distribution morphology and it predicts the Cr and Co concentrations and processing time and temperature for that particular morphology. The model analysis shows that shallow networks can predict chemistry well. However, deep networks are required to predict the processing time and temperature. We validated the model against a TEM micrograph and while the model is trained with synthetic data it performs reasonably well in chemistry and processing prediction for a TEM micrograph.
10:00 AM
A Data-driven Approach for Estimating Three-dimensional Microstructural Features of Bainitic Steels Using Phase-field Simulation Results: Dhanunjaya Kumar Nerella1; Ingo Steinbach1; 1Ruhr University Bochum
A comprehensive understanding of multiple materials phenomena requires description of three-dimensional (3D) microstructures and their significant features. Most of the experimental observations are limited to two dimensions (2D). Accessing 3D microstructural information using methods like serial sectioning is expensive and time-consuming. Frequently applied phenomenological models, like phase-field (PF) method has proven to be more realistic with the experimental observations. They directly provide information on 3D microstructures and their salient features. With the results of PF simulations, a data-driven approach is presented to estimate the 3D microstructural features from 2D (taken in the form of slices) morphological information in bainitic steels and theircorrelation with the strength of material.
10:20 AM
Predicting Laser Powder Bed Fusion Microstructures Using Machine Learning: Gregory Wong1; Anthony Rollett1; Gregory Rohrer1; 1Carnegie Mellon University
The ability to predict as printed microstructures is essential for use in modeling mechanical performance in metal additive manufacturing. The computational expense of existing methods leads to the option of employing rapid machine learning based methods. This talk covers work using conditional generative adversarial networks (cGANs) to generate synthetic microstructures corresponding to metal additive parts made of cubic metals. A set of training data has been developed using existing methods and varying the parameters used in a laser powder bed fusion additive process (laser power, laser velocity, hatch spacing, etc.). The cGAN model is trained using 2D slices of the 3D model output that have been labeled with the printing parameters used in each simulation for conditioning. Microstructure images used for training alongside corresponding cGAN model outputs will be presented.
10:40 AM
Application of Deep Learning Object Detection and Image Segmentation Code Such as YOLO and U-Net for Detection of Helium Bubbles and Voids in Nuclear Reactor Materials: Shradha Agarwal1; Sydney Copp1; July Reyes1; Steven Zinkle1; 1University of Tennessee and Oak Ridge National Laboratory
Analysing micrographs of microstructural features using transmission electron microscopy is key for predicting the performance of structural materials in nuclear reactors. Analysing micrographs is often a very tedious manual process, therefore recently many researchers have tried to automate the process by using various types of neural network, however, application of these networks still require lot of manual work. This paper compares two state-of-the-art neural networks, YOLO and U-Net to maximize the automation of tasks such as counting of microstructural features like helium bubbles and voids. To better understand the accuracies, performance and limitation of each model, we conducted robust hyperparameter validation test including suite of random splits and datasetsize-dependent and domain-targeted cross-validationtests.
11:00 AM
ICME for DNA-templated Dye Aggregate Design for Quantum Information Applications: Lan Li1; 1Boise State University
Organic molecules, known as dyes, which can absorb and emit light, are potential candidates for quantum computing owing to their unique properties, including exciton delocalization and coherence features when dyes are aggregated. Importantly, exciton delocalization and coherence can occur at ambient temperature. These novel applications are controlled by dye properties, requiring high extinction coefficient, high transition dipole moment, good aggregation ability, and high exciton exchange energy. Dye aggregate networks via deoxyribonucleic acid (DNA) templating exhibit exciton delocalization, energy transport, and fluorescence emission. DNA nanotechnology provides scaffolding upon which dyes attach in an aqueous environment. To better control the process and optimize the properties, we applied machine learning-driven multiscale modeling techniques to identify candidate dyes and reveal their dye aggregate-DNA interactions and the dye orientations. Those structural features were found to have a strong impact on the resultant performance of the DNA-templated dye aggregates. The computational results were validated with experiments.