AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Session VI
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
Wednesday 2:00 PM
March 22, 2023
Room: Cobalt 520
Location: Hilton
Session Chair: Anh Tran, Sandia National Laboratories
2:00 PM Cancelled
Graph Attention Networks for Microstructural Understanding: Ryan Cohn1; Elizabeth Holm1; 1Carnegie Mellon University
Polycrystalline microstructures can be represented as graphs, which capture both grain geometry and connectivity. A number of machine learning (ML) algorithms operate on graph data, raising the question of whether they can be used to predict microstructural evolution in polycrystals. Operating on a data set of Monte Carlo grain growth simulations, we find that a simple graph convolution network outperforms a computer vision approach for predicting the occurrence of abnormal grain growth (AGG) in a model polycrystalline system. Based on this successful proof-of-concept, we extend the data set and enhance the data structure. A graph attention network significantly outperforms simple graph convolution, achieving a 20% reduction in error rate. In addition, feature importance analysis identifies the grain characteristics associated with AGG. Taken together, these results show the promise of ML for both predicting microstructural outcomes and supporting microstructural science.
2:20 PM
Accelerating Microstructurally Small Crack Growth Predictions in Three-dimensional Microstructures Using Deep Learning: Vignesh Babu Rao1; Brian Phung1; Bjorn Johnsson1; Ashley Spear1; 1University of Utah
The ability to rapidly predict the growth behavior of microstructurally small cracks (MSCs) has the potential to significantly advance fracture-based designs and structural prognosis. The experimental and numerical difficulties associated with characterizing or predicting MSC growth preclude the applicability of such techniques in industrial design approaches, despite their potential benefits. Here, we propose a framework to accelerate the MSC growth predictions using deep-learning algorithms such as convolutional neural networks (CNNs). The primary research aim is to train CNNs to predict the rules governing MSC growth and to subsequently apply them to make rapid forward predictions of local crack extension given microstructural neighborhood information along a crack front. The training data are acquired from a large number of “virtual” MSC growth observations enabled by high-fidelity finite-element-based simulations. The MSC-growth-simulation framework, data-extraction strategies, and application of deep-learning algorithms for data-driven model development will be presented, and the resulting advantages will be demonstrated.
2:40 PM
Adaptive Latent Space Embedding for Real-Time 3D Diffraction Data Analysis: Alexander Scheinker1; Reeju Pokharel1; 1Los Alamos National Laboratory
In-situ 3D characterization of defects and interfaces and their evolution at the mesoscale (nm-μm) are required to develop microstructure-aware physics-based models and to design advanced materials with tailored properties. There are two major challenges faced by the most advanced 3D imaging techniques. The first challenge is the long measurement time (0.5-8 hours) limiting the number of samples that can be imaged during a given beam time at a light source and also limits the temporal resolution of the dynamics being studied to quasi-static measurements. The second challenge is the long reconstruction time (days-weeks on HPS clusters) which makes it impossible to provide real-time feedback based on reconstructions during an ongoing experiment. Our work studies the use of adaptive tuning of a low-dimensional latent space embedding within 3D convolutional neural networks to speed up measurements (by reducing the number of measurements required) and reconstructions of 3D coherent diffraction imaging (CDI) methods.
3:00 PM
Prediction of Slip Localization and Transmission in Polycrystalline HCP Metals via Incorporation of Micromechanical Modeling and Machine Learning: Behnam Ahmadikia1; Adolph Beyerlein2; Irene Beyerlein1; 1University of California Santa Barbara; 2Clemson University
Slip bands (SBs) play a significant role in deformation of HCP metals since they accommodate a large share of slip while occupying a small volume fraction. Intense stress fields developed by SBs act as potential sites for crack nucleation or cause SBs to transmit across the grain boundaries, leading to long SB chains that percolate across the microstructure. While slip localization in HCPs is reported to be influenced by many factors, including material, texture, and microstructural features, the significance of each has not yet been investigated. We use an explicit SB full-field FFT-based model to verify the variations in intensity of slip localization and its transmission propensity under different circumstances. Then, we apply a Random Forest classifier and a Neural Network to a dataset of 3D microstructures developing SBs to identify the most influential factors and discuss the design parameters that hinder serial transmission of SBs in HCP polycrystals.
3:20 PM
Denoising of Electron Back Scatter Patterns for Improved EBSD Characterization Using Deep Learning: Mani Krishna Karri1; Radhakrishnan Madhavan1; Mangesh Pantawane1; Ramniwas Singh1; Narendra Dahotre1; 1University of North Texas
Electron back scatter diffraction (EBSD) enables in-depth characterization of the microstructures by indexing experimental electron back scatter patterns (EBSPs). The success of the EBSD depends on the quality of the EBSPs. In case of samples with poor EBSPs (highly deformed, nano crystalline, faceted or in-situ experiments) EBSD characterization is quite challenging and existing automated EBSP indexing methods largely fail.In this work, a novel machine learning (ML) method of refining noisy EBSPs is proposed. For this, conditional generative adversarial neural networks (c-GAN) have been employed. The ML model was trained using 10000 EBSPs acquired under different settings and samples (additively manufactured FCC, BCC and HCP alloys) ensuring enough diversity and complexity in training data set. The trained model has brought out an improvement of more than 3 times in EBSD indexing success rate on test data, accompanied by betterment of indexing accuracy.
3:40 PM Break
4:00 PM
Examining the Effects of Grain Boundary Structure Variability, Solute Atoms, and Interatomic Potential on the non-Arrhenius Migration of Incoherent Twin Grain Boundaries in Nickel: Akarsh Verma1; Eric Homer1; Oliver Johnson1; Shigenobu Ogata2; Gregory Thompson3; 1Brigham Young University; 2Osaka University; 3University of Alabama
Migration of incoherent twins is unique in comparison with other grain boundaries because some incoherent twins migrate at relatively high velocities and don’t follow expected Arrhenius temperature dependence. This antithermal behavior, slower migration at higher temperatures, can emerge from a combination of thermally activated processes. However, since a given incoherent twin in a real material will have variability in structure and impurities, we examine the effects structural variability and solute content have on the observed migration behaviors. Additionally, it is known that empirical potentials cannot replicate the exact behaviors of a given material, so we compare various interatomic potentials for variability of the measured response. Uncertainty quantification methods are used in assessing the boundary migration velocity. The variability in response to metastability, solute content, and potential are not quantified but discussed qualitatively to gain insight into how they might impact the behavior in real materials.
4:20 PM
Modelling Nucleation in Crystal Phase Transition from Machine Learning Metadynamics: Qiang Zhu1; Pedro Santos-Florez1; Howard Yanxon1; Yansun Yao2; 1University of Nevada, Las Vegas; 2University of Saskatchewan
In this work, we present an efficient framework that combines machine learning potential (MLP) and metadynamics to investigate solid-solid phase transition. We have developed a scalable MLP model to warrant an accurate interpolation of the energy surface where two phases coexist. Applying the framework to the metadynamics simulation of B4-B1 phase transition of GaN under 50 GPa with different model sizes, we observe the sequential change of phase transition mechanism from collective modes to nucleation and growths. With a small size, the nucleation and growth appear to follow a preferred direction. At larger sizes, the nucleation tends to occur at multiple sites simultaneously and grow to microstructures by passing the critical size. The observed change of atomistic mechanism manifests the importance of statistical sampling with large system size. The combination of MLP and metadynamics is likely to be applicable to a broad class of reconstructive phase transitions at extreme conditions.
4:40 PM
Data Assimilation for Microstructure Evolution in Kinetic Monte Carlo: Anh Tran1; Theron Rodgers1; Yan Wang2; 1Sandia National Laboratories; 2Georgia Institute of Technology
Modeling grain growth has been a subject of interest in computational material science, as it occurs in thermal-based processing methods such as annealing and sintering. Kinetic Monte Carlo with Potts model is often used as an integrated computational materials engineering (ICME) grain growth model and can generate high-fidelity synthetic microstructures. In this talk, we offer a data-driven stochastic calculus perspective on the kinetics of grain growth and model the microstructure evolution through the lens of stochastic differential equations, based on Langevin dynamics and Fokker-Planck equation to forecast the grain size distribution. We demonstrate that our proposed approach agrees reasonably well with the hybrid Potts-phase field model using SPPARKS in forecasting the long-term evolution of grain size distribution.
5:00 PM
How to Lead R&D Digital Transformation in a Chemical Corporation: Yoshishige Okuno1; Shimpei Takemoto1; 1Showa Denko K.K.
We demonstrate how we lead R&D Digital Transformation at Showa Denko, a Japanese chemical corporation. Successful data-driven R&D requires the establishment of processes for data collection, storage, analysis, and decision-making and an IT infrastructure to support these processes. Improving material developers' data literacy and enabling analytical decision-making is also important. We have established data pipelines to collect experimental data from electronic lab notebooks. The collected data is automatically transformed into structured data and stored in a relational database. Machine learning models for predicting material properties are automatically generated based on the database and deployed to a web application system. Material developers can effortlessly search, visualize, and analyze data on GUI. Machine learning model predictions are used for the forward and inverse design of novel materials. MLOps for efficiently managing the machine learning models and the web application system have also been introduced.