AI/Data informatics: Design of Structural Materials: AI/ML Frameworks; Grain Growth and Simulation Integration
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS Structural Materials Division, TMS: Mechanical Behavior of Materials Committee, TMS: Computational Materials Science and Engineering Committee, TMS: Integrated Computational Materials Engineering Committee
Program Organizers: Jennifer Carter, Case Western Reserve University; Amit Verma, Lawrence Livermore National Laboratory; Natasha Vermaak, Lehigh University; Jonathan Zimmerman, Sandia National Laboratories; Darren Pagan, Pennsylvania State University; Chris Haines, Ccdc Army Research Laboratory; Judith Brown, Sandia National Laboratories

Wednesday 2:00 PM
March 17, 2021
Room: RM 32
Location: TMS2021 Virtual


2:00 PM  Invited
Data Science Approaches for Microstructure-property Connections in Structural Materials: Elizabeth Holm1; bo Lei1; Katelyn Jones1; Ryan Cohn1; Nan Gao1; 1Carnegie Mellon University
    Optimizing the performance of structural materials often involves engineering the microstructure through composition and process control. Because the microstructural parameter space is high-dimensional and complex, the tools of data science and informatics offer promising approaches for understanding and optimizing the microstructure-properties relationship. Materials properties can be divided into those that depend primarily on the distribution of microstructural features in the microstructure as a whole (i.e. yield strength) and those that are governed by relatively rare, critical features (i.e. fracture initiation). Case studies, drawing from the design of high-temperature alloys, fatigue failure of superalloys, alloy processing, and steel metallurgy, will address the application of data science methods, including computer vision and machine learning, to each property class. In all cases, understanding the characteristics of the data set, designing an information-rich feature representation, and selecting appropriate machine learning schemes are essential to achieving the most useful results.

2:30 PM  Invited
Physics-informed Data-driven Machine Learning Approach for Mesoscale Materials Science: Reeju Pokharel1; Anup Pandey1; Alexander Scheinker1; 1Los Alamos National Laboratory
    Macroscale properties and performance of structural materials are influenced by mesoscale interactions between crystallographic grain boundaries, defects, and dislocations. Advanced experimental techniques developed in the last several decades have provided destructive or non-destructive characterization of mesoscale microstructures, micro-mechanical fields and their evolution under various thermo-mechanical conditions. One of the major challenges faced by the advanced imaging and diffraction techniques is the slow analysis and interpretation of experimental data. This has limited our ability to use experimental observations for informing and validating mesoscale deformation and damage models. To address this, we are working on developing a hybrid data-driven physics-informed machine learning approach to speed up data collection and reconstruction beyond the current-state-of-the-art, which will enable accelerated design and deployment of new materials.

3:00 PM  Invited
Combined Statistical and Energetic Approach to Understand Grain Boundary Embrittlement for Segregation Engineering: Doruk Aksoy1; Remi Dingreville2; Douglas Spearot1; 1University of Florida; 2Sandia National Laboratories
    Sulfur segregation to grain boundaries in polycrystalline nickel adversely affects fracture behavior in the form of embrittlement. Both the structure and the distribution of grain boundaries in a Ni polycrystal are important; however, it is difficult to separate the relative importance of these characteristics towards segregation induced embrittlement. In this work, molecular statics calculations are performed, using an embedded-atom method interatomic potential developed specifically for studying embrittlement, to provide grain boundary segregation energy and substitutional site embrittling potency populations using 378 different symmetric tilt grain boundaries. To account for both embrittlement energetics and the statistics associated with S segregation to specific Ni grain boundaries within a polycrystal, a new grain boundary metric is proposed: the embrittling estimator. Ultimately, this combined statistical and energetic approach may provide a tool to assist the engineering and design of grain boundaries and polycrystalline microstructures against segregation induced embrittlement.

3:30 PM  
Machine Learning Approach to Understanding Abnormal Grain Growth: Ryan Cohn1; Megna Shah2; Adam Pilchak2; Eric Payton2; Anthony Rollett1; Elizabeth Holm1; 1Carnegie Mellon University; 2Air Force Research Laboratory
    Abnormal grain growth occurs in metals and ceramics when one grain achieves a significant size advantage over its neighbors due to a faster growth rate during grain growth. Controlling abnormal grain growth, either to limit or encourage it, is desirable for many structural and functional materials. However, the abnormal growth initiation process is not well understood. Previous studies employed Kinetic Monte Carlo methods to simulate abnormal grain growth in polycrystals with nonuniform grain boundary mobility. The results successfully replicated the statistics of abnormal grain growth, and the simulations often agreed with experiments as well. In this work, we apply machine learning and data science techniques to these data to discover how local grain environments influence the abnormal growth process. Machine learning enables both prediction of the growth mode on a grain-by-grain basis as well as determination of the geometric, topological, and crystallographic factors that encourage or inhibit abnormal growth.

3:50 PM  
Machine Learning for the Recognition and Synthesis of Polycrystalline Metal Microstructures: Neal Brodnik1; Devendra Jangid1; Amil Khan1; Michael Goebel1; McLean Echlin1; B. S. Manjunath1; Samantha Daly1; Tresa Pollock1; 1University of California Santa Barbara
    One major challenge in the development of materials is establishing representative behavior from different processing conditions, which often demands extensive physical testing. However, computational models can replace or augment physical tests and process modeling to save time and cost. This work explores the use of convolutional, adversarial, and graph neural networks to recognize and generate polycrystalline metal microstructures based on prior experimental information. Networks are trained on microstructural morphologies and arrangements gathered using a 3D serial sectioning technique known as the Tribeam. This information is then used to produce new microstructural features that are distinct from the ground truth data while still bearing similarity in a physical and statistical capacity. These approaches can also be used to mitigate imaging artifacts and explore relationships between microstructure and mechanical response. With sufficient fidelity, network generated microstructures could be used to supplement experimental approaches and greatly accelerate materials development.

4:10 PM  Invited
Using Machine Learning for Targeted Alloy Design in High Entropy Composition Spaces: Tanner Kirk1; Richard Couperthwaite1; Guillermo Vazquez1; Daniel Sauceda1; Pejman Honarmandi1; Prashant Singh2; Raymundo Arroyave1; 1Texas A&M University; 2Ames Laboratory
    Alloy discovery in the large composition spaces associated with High Entropy Alloys can be a daunting task given the combinatorial explosion of potential compositions. However, a variety of machine learning techniques can reduce the design process to a tractable problem. These techniques are demonstrated to find suitable high strength alloys in each of two high entropy alloy spaces: the refractory, largely BCC, W-Mo-Nb-Ta-V-Al system and the largely FCC Fe-Mn-Cr-Co-Ni-V-Al system. High throughput CALPHAD modeling as well as analytical property models are compared to design requirements to identify feasible alloys. Dimensionality reduction techniques like t-distributed Stochastic Neighbor Embedding (t-SNE) can visualize the location of the feasible region in the total composition space. K-medoids clustering produces a representative subset of feasible alloys for more expensive modeling, like DFT, or experimentation. After characterization, models are updated and Batch Bayesian Optimization suggests further experiments based on design preferences, eventually arriving at the optimal alloy.

4:30 PM  
Unsupervised ML to Bridge Molecular Dynamics and Phase field Simulations: Sukriti Manna1; Henry Chan1; Subramanian Sankaranarayanan1; 1Argonne National Laboratory
    A machine learning approach is demonstrated to bridge atomistic molecular dynamics and mesoscopic phase-field method. We use a representative grain growth simulation to demonstrate the efficacy of our approach. Our molecular dynamics simulation evaluates the materials properties such as grain boundary energy, grain boundary width, and grain boundary mobility. An unsupervised ML is used to seamlessly feed materials properties from MD to evaluate our phase-field model parameters and back-mapping atoms on to the phase-field model. An efficient information exchange between atomistic molecular dynamics and mesoscopic phase-field method enables us to leverage the best attributes of molecular dynamics and phase-field simulations for mesoscale simulations.