Accelerated Discovery and Qualification of Nuclear Materials for Energy Applications: Data Analytics and Machine Learning in Nuclear Energy Applications
Sponsored by: TMS Structural Materials Division, TMS Materials Processing and Manufacturing Division, TMS: Integrated Computational Materials Engineering Committee, TMS: Nuclear Materials Committee, TMS: Additive Manufacturing Committee
Program Organizers: Yongfeng Zhang, University of Wisconsin; Adrien Couet, University of Wisconsin-Madison; Michael Tonks, University of Florida; Jeffery Aguiar, Lockheed Martin; Andrea Jokisaari, Idaho National Laboratory; Karim Ahmed, Texas A&M University
Wednesday 8:30 AM
March 17, 2021
Room: RM 48
Location: TMS2021 Virtual
Session Chair: Dane Morgan, University of Wisconsin; Karim Ahmed, Texas A&M U
8:30 AM Invited
Machine Learning and Atomistic Modeling of Defect Diffusion in Concentrated Ni-Fe Alloys: Wenjiang Huang1; Xian-Ming Bai1; 1Virginia Polytechnic Institute and State University
Single-phase concentrated solid solution alloys including high entropy alloys are promising structural materials for various high-temperature applications including nuclear energy. Defect diffusion and evolution in these non-traditional alloys play central roles in governing their macroscopic properties. Here we use atomistic modeling and artificial neural network based machine learning method to study how the atomic configurations influence the vacancy diffusion in Ni-Fe concentrated alloys in the full composition range. Molecular dynamics are conducted to calculate the vacancy diffusivities in these alloys at different temperatures, alloy compositions, and atomic configurations. Based on many alloy properties obtained from atomistic modeling such as vacancy formation energy distribution, migration barrier distribution, short-range-order parameter, and heat of mixing, a machine learning based model concerning statistical uncertainties is developed to predict the vacancy diffusivities for different atomic configurations. The effects of these alloy properties on the vacancy diffusion are also analyzed from the machine learning model.
9:00 AM
Characterization of As-Fabricated Additively Manufactured Alloy 718 Enhanced by Modern Tools and Machine Learning: Stephen Taller1; Luke Scime1; Kurt Terrani1; 1UT-Battelle
The design of new materials for nuclear energy applications has increased with computing power, and manufacturing has seen revolutionary improvements with advanced manufacturing (AM) techniques. To keep pace with these developments, characterization and post irradiation examination procedures must be improved to accelerate nuclear materials evaluation. This presentation discusses the tools being developed for automated transmission electron microscopy (TEM) image and energy dispersive spectroscopy (EDS) spectra acquisition. These tools are being used to enable operator-free data collection using a FEI F200X Talos scanning TEM. The resulting images are used to train a platform-agnostic artificial intelligence and machine learning (AI/ML) tool with pixel-wise defect-detection algorithms. These techniques have been demonstrated using high-throughput characterization of an as-fabricated AM Ni-Fe-Cr alloy 718. Precipitate morphology, density, and elemental composition were characterized in anticipation of irradiated specimens to determine precipitate stability under irradiation.
9:20 AM Invited
Machine Learning for Accelerating Property Prediction and Materials Characterization in Irradiated Materials: Dane Morgan1; Mingren Shen1; Ryan Jacobs1; G. Robert Odette2; Kevin Field3; 1University of Wisconsin-Madison; 2University of California, Santa Barbara; 3University of Michigan
Machine learning methods have the potential to greatly accelerate nuclear materials development through predicting properties and automating analysis. In this talk, I will discuss our recent efforts to predict hardening in reactor pressure vessel steels, focusing on the challenges of extrapolation to irradiation conditions outside those in the training data. I will also discuss our recent efforts to automate deep learning object detection approaches to find the location and geometry of different defect types in electron microscopy images of irradiated steels. We show that an accuracy comparable to human analysis can be achieved with relatively small training data sets, suggesting a future where defect analysis is more standardized and orders of magnitude faster than today.
9:50 AM
Point Defect Energies in Concentrated Alloys Using Ab Initio Calculations and Machine Learning: Anus Manzoor1; Gaurav Arora1; Dilpuneet Aidhy1; 1University of Wyoming
Concentrated alloys, including high entropy alloys, consist of multiple principal elements randomly distributed on a crystal lattice that causes large variations in point-defect formation and migration energies in a given alloy composition. Statistically capturing the variation requires performing large number of density functional theory calculations. The challenge is compounded due to the exponentially large number of compositions that are possible in these alloys. We solve the problem by leveraging machine learning tools where the defect energies computed from binary alloys are used to train the models to predict energies in multi-element alloys. We demonstrate accurate predictions of migration barriers in FeNiCr, and vacancy formation energies in NiCuAu. A major benefit of this approach is that once the binary database is built and the model is trained, defect energies can be easily predicted thereby bypassing the need to perform large number of calculations every time a new composition is discovered.
10:10 AM Invited
Machine Learning Perovskites in the Quest for Improved Scintillators: Anjana Talapatra1; Christopher Stanek1; Blas Uberuaga1; Ghanshyam Pilania1; 1Los Alamos National Laboratory
Inorganic scintillator-based detector materials find a wide variety of applications. These materials essentially convert a fraction of the total energy deposited by incident gamma rays or X-rays into visible or near-visible range of the spectrum. A "good" scintillator would exhibit high light output, fast response time, and emission at suitable wavelengths, among many other application-specific desired characteristics. However, no single scintillator is ideal for all uses. To accelerate the discovery of customized optimal scintillator materials with targeted properties and performance, efforts are ongoing to develop a closed-loop machine learning driven adaptive design framework based on data from literature, in-house experiments and quantum mechanical calculations. This talk will present an overview of this framework, focusing on the screening of complex perovskite and double perovskite chemistries with high band-gaps and other favorable electronic structure features to yield custom scintillation properties.
10:40 AM
An Integrated Approach for Coupling Experimental Data, Physics-based Models, and Machine Learning Algorithms for Predicting the Effective Thermal Conductivity of U-based Fuels: Karim Ahmed1; Fergany Badry1; 1Texas A&M University
We present a new approach for coupling experimental data, physical models, and machine learning algorithms to predict the effective thermal conductivity of UO2-BeO, UO2-Mo, and U-10Zr nuclear fuels. First, a physics-based mesoscale model that takes into account the effect of underlying microstructure, temperature, and interface thermal resistance is developed. Then the model is verified and validated against available experimental data. The validated model is then utilized to generate more data at conditions different from experiments. Existing experimental data and the additional model results are then used to train and test non-linear regression models in open-source machine learning software. The final outcome of this framework is validated, physics-based reduced order models of the effective conductivity of these fuels. This novel approach reduces the required number of experiments and at the same time minimizes the computational cost necessary for qualifying new materials.
11:00 AM
Deep Learning for Automated Analysis of Cavities in Transmission Electron Microscopy Images: Chun Yin Wong1; Xing Wang2; Zhe Fan3; Karren More4; Sergei Kalinin4; Maxim Ziatdinov4; 1University of Tennessee; 2The Pennsylvania State University, Oak Ridge National Laboratory; 3Lamar University, Oak Ridge National Laboratory; 4Oak Ridge National Laboratory
The formation of cavities, including voids and bubbles, is a major threat to the integrity of materials under irradiation. Transmission electron microscope (TEM) is widely used for characterizing the size and density of cavities. Statistical analysis requires measuring hundreds of cavities, which is a time-consuming task if conducted manually. A robust and universal framework for automated cavity measurement can substantially accelerate the analysis. Here, such a framework was developed based on a custom convolutional neural network. The framework, trained using only five labeled images, was able to identify the locations and sizes of cavities in similar TEM images with an intersection-over-union (IoU) of 0.80 and 60 times faster than manual labeling. The framework was further extended to identify overlapping cavities via the Laplacian of Gaussian and Hough transform. Finally, the universality of the framework has also been demonstrated by applying to cavity analysis in TEM images acquired from different materials.