AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Session III
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Saurabh Puri, VulcanForms Inc; Francesca Tavazza, National Institute of Standards and Technology; Dennis Dimiduk, BlueQuartz Software LLC; Darren Pagan, Pennsylvania State University; Kamal Choudhary, National Institute of Standards and Technology; Saaketh Desai, Sandia National Laboratories; Shreyas Honrao, Aionics; Ashley Spear, University of Utah; Houlong Zhuang, Arizona State University
Tuesday 8:00 AM
March 21, 2023
Room: Cobalt 520
Location: Hilton
Session Chair: Ashley Spear, University of Utah
8:00 AM
Adversarial Hierarchical Variational Autoencoder: A Novel Autoencoder Architecture for Microstructure Synthesis and Feature Extraction: Simon Mason1; Mengfei Yuan2; Ashley Lenau1; Octavian Donca1; Dennis Dimiduk3; Steve Niezgoda1; 1Ohio State University; 2Ping An Insurance; 3BlueQuartz Software LLC
Because of the high cost of experimental 3D microstructure data, computational methods of microstructure synthesis have been established that allow large-scale dataset generation, including phase-field models, DREAM.3D, and neural network architectures. The present work explores development of a novel autoencoder, the adversarial hierarchical variational autoencoder (AHVAE), that combines the benefits of the traditional VAE network, the hierarchical nature of the Nouveau VAE’s latent spaces, and the improved image generation capabilities of adversarial training. Along with its use as a microstructure generator, the AHVAE can also be used as a feature extractor through the utilization of the information-dense latent space. After training the full network, the encoder can be used independently to generate low-dimensional latent space representations of microstructures. These latent spaces can then be used within separate, specifically-trained decoders to extract a variety of desired material features, including physical parameter extraction, microstructure topology analysis, and material property prediction.
8:20 AM
Automated Classification of Powder X-ray Diffraction Data Using Deep Learning: Jerardo Salgado1; Zhaotong Du1; Samuel Lerman1; Chenliang Xu1; Niaz Abdolrahim1; 1University of Rochester
Emerging X-ray scattering techniques provide the capability to probe the correlations between structure and material properties with sensitivities approaching the single-molecule level and picosecond resolution. However, temporal diffraction images at extreme temperatures and pressures generate terabytes of data, making it impossible for experts to analyze. With the increased demand for new insights at these conditions and big data process capabilities, our goal is to apply artificial intelligence to characterize inorganic materials accurately. We trained deep learning models with simulated Xrds – synthetic profiles from existing verified materials - then used the models to make predictions on experimental data. Our convolutional neural network and dense neural network models have shown state-of-the-art performance on synthetic data, achieving testing accuracies of 97% for crystal structure classification. The model’s ability to adapt to experimental domains was also studied, where accuracies have reached 70% and 55% for crystal system and space group classification respectively.
8:40 AM
Comparison of U-Net and Mask R-CNN Neural Network for Detection of Helium Bubbles and Voids in Nuclear Reactor Materials: Shradha Agarwal1; Sydney Copp2; July Reyes2; Steven Zinkle1; 1University of Tennessee and Oak Ridge National Laboratory; 2University of Tennessee
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, these networks still require lot of manual work. This paper compares two state-of-the-art neural networks, Mask-RCNN 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-validation tests.
9:00 AM
Computer Vision Assisted Automated Grain Segmentation and High-Throughput Composition Analysis with Scanning Electron Transmission Microscopy: Doruk Aksoy1; Jenna Wardini1; Timothy Rupert1; William Bowman1; 1University of California, Irvine
High-throughput data-driven methods offer rapid screening tools for microstructural characterization, an otherwise slow and laborious task. In this work, annular dark field, and bright field images of an electrically conducting polycrystalline oxide are obtained with scanning transmission electron microscopy. First, grain boundaries are marked manually on acquired images to create masks for semantic image segmentation. Then, these images and masks are vectorized and processed using a data augmentation pipeline to simulate image acquisition deficiencies, in addition to avoid overfitting. Augmented images and masks are fed into a hyperparameter optimized convolutional neural network-based architecture to obtain high-fidelity grain segmentation maps. These maps are utilized as guides for obtaining spatially resolved spectroscopic data. Ultimately, leveraging computer vision in experimental image acquisition enables fast and detailed interrogation of the local composition and chemistry of grain boundary populations in polycrystalline materials.
9:20 AM
Weakly-Supervised Segmentation of Microstructure Images with Deep Convolutional Neural Networks: Bo Lei1; Elizabeth Holm1; 1Carnegie Mellon University
Deep convolutional neural networks have demonstrated outstanding predictive capability in complicated microstructure segmentation tasks. However, typical deep learning solutions require a considerable amount of dense pixel-level annotations for model development. These annotations are especially difficult and time-consuming to obtain for materials images. Here, we focus on a weakly-supervised method that only uses hundreds of pixel annotations per image and reaches comparable performance to fully-supervised method. The method is developed in an active learning manner where pixel annotations are queried from users incrementally, making it possible to be integrated with interactive programs. This approach can significantly reduce the annotation efforts and would mostly benefits rapid microstructure characterization in common materials research and development scenarios.
9:40 AM Break
10:00 AM
Utilizing and Understanding Deep Learning for 3D Microstructure Synthesis: Neal Brodnik1; Devendra Jangid1; McLean Echlin1; B. S. Manjunath1; Samantha Daly1; Tresa Pollock1; 1University of California Santa Barbara
Nonlinear machine learning tools such as neural networks are promising ways to rapidly infer complex material relationships. However, this ability depends on both sufficient data for training and an effective means of output evaluation, each of which presents its own set of challenges. Here, we present the application of deep learning to 3D microstructure recognition and generation in ways that facilitate the learning of broad materials concepts and creation of more robust datasets. These network-based approaches are connected back to fundamental materials science through the incorporation of physics into learning metrics and network architectures, as well as with evaluation approaches centered on the principles of materials development. Physics-based metrics and architecture enable faster learning of microstructural principles with lower data burden. Finally, well-structured output evaluation allows for network capabilities to be quantified in terms of emergent microstructural properties and efficacy in the materials development pipeline.
10:20 AM
Physics-Based Deep Learning Methods for Enforcing Stress Equilibrium in GAN Generated Stress Fields: Ashley Lenau1; Dennis Dimiduk2; Stephen Niezgoda1; 1Ohio State University; 2BlueQuartz Software LLC
Deep learning (DL) is an increasingly growing field in computational materials science with advancements in modeling the structure-property relationships of materials, material design, and microstructure generation. However, large amounts of data are typically required to sufficiently train a DL network, and acquiring this amount can be a difficult task in materials science. Incorporating physics-based regularization methods into DL algorithms can possibly speed up the time and reduce the data needed to train a DL network. In this work, an image translation generative adversarial network is used to translate the elastic stress fields from a high-contrast composite image. Various physics-based regularization methods are used to capture high-frequency features and to enforce a stress equilibrium constraint on the elastic stress fields. In this presentation, the time and data needed to adequately train the DL model having physics-based regularization methods are shown in comparison to the DL model without these methods.
10:40 AM
Generation of 3D Synthetic Polycrystalline Microstructures using Gaussian Random Fields and Two Point Spatial Correlations: Michael Buzzy; Andreas Robertson1; Surya Kalidindi1; 1Georgia Institute of Technology
The generation of synthetic microstructures is critical for many tasks such as the creation of artificial datasets, microstructure sensitive design, and process optimization. Current frameworks for synthetically generating polycrystalline microstructures are limited to enforcing mean field statistics, such as grain size distributions and orientation distribution functions. We propose a new method to generate polycrystalline microstructures which incorporates higher order spatial correlations. These added statistics enable generative models to better express a desired grain morphology and capture spatial patterns of orientations. Our method also allows for easy interpolation between polycrystalline structures allowing for the creation of large datasets from limited experimental data. We will discuss the necessary theoretical and computational development as well as showcase potential applications.
11:00 AM
Synthetic Data Development towards Automated Defect Detection of Irradiated Materials: Matthew Lynch1; Priyam Patki1; Ryan Jacobs2; Steven Chen1; Gabriella Bruno1; Dane Morgan2; Kevin Field1; 1University of Michigan - Ann Arbor; 2University of Wisconsin - Madison
Machine Learning (ML) vision algorithms have attracted significant attention recently in enabling rapid labeling and analysis of data from transmission electron microscopy (TEM) experiments. However, recent efforts have exclusively focused on costly human-based labeling processes for ML training data sets. Here we demonstrate an effective synthetic ML database generation process using simplified physics-based phase contrast simulations. Individual cavity simulations are integrated into unirradiated microscopy images to produce a hybrid synthetic image. This novel process enables nearly instant and unlimited generation of automatically labeled training data, devoid of human bias. ML models trained on this data can perform similarly to those trained on traditional data, and there is a possible synergistic effect when combining datasets. This process can be applied to various defect types and used to simulate rare defect structures, generating data that is challenging to experimentally obtain.