AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Image Characterization
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Saurabh Puri, VulcanForms Inc; Dennis Dimiduk, BlueQuartz Software LLC; Darren Pagan, Pennsylvania State University; Anthony Rollett, Carnegie Mellon University; Francesca Tavazza, National Institute of Standards and Technology; Christopher Woodward, Air Force Research Laboratory
Tuesday 8:00 AM
March 1, 2022
Room: 256A
Location: Anaheim Convention Center
Session Chair: Marat Latypov, University of Arizona
8:00 AM Invited
Physics-informed Data-driven Surrogate Modeling for Advancing Experiments and the Study of Novel Materials: Anup Pandey1; Reeju Pokharel1; 1Los Alamos National Laboratory
With the advancements in experimental facilities like high-energy X-ray sources, there is an enormous increase in the experimental data, which has imposed challenges in their timely analysis and interpretation. In recent years, more practical applications of state-of-the-art machine learning (ML)algorithms are emerging as a powerful tool in accelerating the computation time, providing real-time feedback during the experiments, and accelerating the data collection and reconstruction process. I will discuss the use of physics-informed data-driven surrogate modeling in providing real-time feedback and accelerate the reconstruction process during the experiments, such as high energy X-ray diffraction microscopy and electron backscattering diffraction. On the other hand, the molecular dynamics based atomistic study of materials such as multicomponent alloys is limited to fewer candidates due to the lack of accurate potentials. I will discuss using ML force field trained on data obtained from quantum mechanics calculations in atomistic modeling of a wide range of alloys.
8:30 AM
Learning from 2D: Data-driven Model Predicting Bulk Properties Based on 2D Microstructure Sections: Marat Latypov1; 1University of Arizona
Microstructure--property relationships are central to design of structural materials. Advances in computational methods made it possible to directly simulate bulk properties using 3D microstructure-based models. 3D representative volumes of microstructures required as input to these models are typically obtained either from 3D characterization experiments or digital reconstruction based on 2D microstructure information. In this work, we present a machine learning model that predicts bulk properties directly from 2D microstructural sections. The model is trained on a specially designed dataset that contained microstructure features quantifying 2D sections of diverse and synthetically generated 3D microstructures and their corresponding properties obtained from 3D simulations. Upon training, the model allows predicting properties from 2D sections (whose experimental acquisition is more accessible than 3D characterization) and without the need in additional computations (and underlying assumptions) involved in digital reconstruction of 3D microstructures.
8:50 AM
Investigation of Microstructure Image Segmentation via Deep Learning with Limited Data Annotations: Bo Lei1; Elizabeth Holm1; 1Carnegie Mellon University
In quantitative microscopy, microstructure image segmentation is essential for image analysis and materials characterization. The rising deep convolutional neural network methods for semantic segmentation in natural images have recently been transferred to materials images and demonstrated outstanding performance in complicated microstructure datasets. However, typical deep learning solutions require a considerable amount of data annotations to reach good performance and they are especially difficult to obtain for materials images. Here, we investigated the possibility to significantly reduce the amount of human annotations while achieving comparable results with the help of transfer learning and data augmentation. We introduced bag-of-words method for training image selection and made adjustments to the deep network model for better generality.
9:10 AM
Synthesizing Realistic Images of Material Microstructures Using Convolutional Neural Networks: Stephen Baek1; H.S. Udaykumar2; WaiChing Sun3; Phong Nguyen1; 1University of Virginia; 2University of Iowa; 3Columbia University
Synthetic images of microstructures facilitate the design of functional materials by enabling the exploration of design space. Numerical experiments can be conducted on synthetically generated microstructures with controlled morphologies, to assimilate linkages between microstructural morphology and physical quantities of interest. However, traditional ways of generating synthetic microstructures are limited to some idealized shapes, such as spheres and polyhedrals, and significantly lack realism, which limits the use of synthetic microstructures to simple trend analyses. To this end, machine learning approaches have rapidly emerged as feasible alternatives. Especially, the unprecedented capacity of convolutional neural networks (CNN) in machine cognition and image synthesis has created exciting new frontiers in materials research. In this talk, we present some of our recent case studies on the use of CNNs for microstructure image synthesis, focusing on the materials-by-design of energetic materials.
9:30 AM Break
9:50 AM
Density-based Monte Carlo Consensus Clustering (DMC3) for Feature Extraction from Atom Probe
Tomographs: Evan Still1; Daniel Schrieber2; Peter Hosemann1; 1University of California Berkeley; 2Pacific Northwest National Laboratory
Clustering algorithms, such as DBSCAN, OPTICS, and maximum separation serve as the basis for microstructural feature identification from Atom Probe Tomography (APT) point clouds. While meta-heuristics exist for identification of ideal hyperparameters for these algorithms APT data routinely violates their operational assumptions, due to high solute densities within the matrix, non-convex features, and the presence of features of differing length scales and shapes, such as grain boundaries and precipitates, within one dataset. In this work, we demonstrate that a modification of Monte Carlo reference-based consensus clustering (M3C) for DBSCAN referred to as Density-based Monte Carlo Consensus Clustering (DMC3) identifies near optimal, as measured by adjusted mutual information, epsilon parameters for a given order parameter over a wide range of feature sizes, and solute concentrations. This new method provides an improved approach for reproducible and more objective analyses of clustering behavior within APT data.
10:10 AM
Deep Learning-based Algorithms for X-ray Microtomography Analysis: Unravelling Challenges for 4D Experiments: Hamidreza T-Sarraf1; Hanyu Zhu1; Swapnil Morankar1; Amey Luktuke1; Sridhar Niverty1; Nikhilesh Chawla1; 1Purdue University
3D characterization is used to understand the relationships between materials microstructure and function. X-ray microtomography is an important 3D characterization technique due to its non-destructive nature which provides time-dependent (4D) information. However, image processing and segmentation of 4D tomographic data is extremely time intensive. Moreover, factors such as phase transformation or defect propagation during a time-evolved tomography experiment limits the scan time and/or number of scan iterations. Thus, a robust algorithm needs to be established that can render x-ray datasets accurately and efficiently. In this talk, we describe the application of Deep Convolutional Neural Network algorithms for X-ray image quality enhancement and segmentation. Using a modified Generative Adversarial Network algorithm we provide a workflow to transform low quality x-ray tomograph acquired by a fast scan to a high quality dataset. Our results point to the ability to drastically reduce x-ray data acquisition times,thereby opening a window for efficient 4D experiments.