AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Session I
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
Monday 8:30 AM
March 20, 2023
Room: Cobalt 520
Location: Hilton
Session Chair: Darren Pagan, Pennsylvania State University
8:30 AM Keynote
RVE, SERVE and Digital Material Volumes for Design and Engineering: Dennis Dimiduk1; Somnath Ghosh2; David Furrer3; 1BlueQuartz Software LLC; 2Johns Hopkins University; 3Pratt & Whitney
The concept of representative volume elements (RVE) has been around for at least 60 years, yet materials, processes, and structures engineering (MPSE) make little use of it. Engineering implementations not only require establishing a statistical approach to quantify microstructures (i.e., hierarchical material structure), but also establishing a means to define the structure spatially throughout the volume of a component. Further reasons are incomplete science, underdeveloped experimental capability, gaps in modeling and simulation tools, uncertainty quantification and validation methods, data systems, and a lack of industry standards. However, for the last two decades methods for producing and interrogating digital material volumes significantly advanced—lending new promise for engineering implementation. This presentation treats aspects of these issues, beginning with a historical summary, followed by more detailed discussion of current day research on statistically equivalent RVE (SERVE). Following these, some necessary steps for MPSE design implantation are shown. Finally, key limitations are highlighted.
9:00 AM
A Framework to Solve the Inverse “Process-Structure” Problem of Identifying Process Parameters to Produce a Target Microstructure: Dung-Yi Wu1; Todd Hufnagel1; 1Johns Hopkins University
Although the “forward” problem of predicting what microstructure will result from a given processing schedule is solvable using deterministic models, the “inverse” problem of determining what processing is required to produce a desired microstructure is much more challenging, typically requiring multiple iterative cycles of experimentation and simulation". There is a need for a smarter way to navigate through a high-dimensional processing parameter space to find the combination of parameters to achieve a microstructure that satisfies multiple, quantitative descriptors.In this work, we apply a Bayesian Optimization framework based on Gaussian Process Regression to efficiently explore parameter space in the context of heat treatment of aluminum alloys to maximize resistance to spall failure. We demonstrate the utility of our approach through a case study that controls the volume fraction of Al7Cu2Fe second-phase particles and aluminum grain size distribution in a commercial aluminum 7085 alloy.
9:20 AM
A Hybrid Gaussian Random Field – Deep Learning Model for Statistically Controllable Synthetic Microstructure Generation: Andreas Robertson1; Conlain Kelly2; Michael Buzzy1; Surya Kalidindi1; 1Georgia Institute of Technology; 2Georgia Tech
Recently, we proposed a Multi-Output Gaussian Random Field (GRF) model for statistically controllable synthetic microstructure generation. Because first and second order microstructure statistics are directly incorporated into the model’s construction, the model can be used to generate synthetic microstructures from arbitrary microstructure processes. Although it can stably extrapolate to previously unobserved statistics without the need for training data, its ability to generate realistic microstructures is limited by its rigid higher-order statistics. In this talk, expanding on this original model, we develop a hybrid deep-learning model that combines the first and second order statistical parameterization of the GRF model with the expressive higher order statistics of deep learning models. Importantly, we demonstrate that using this hybrid approach, the model can be effectively trained even in Materials Informatics’ characteristic data-starved learning environments. After presenting the model structure and necessary algorithms, we explore several important case studies to emphasize its strengths and weaknesses.
9:40 AM
Data-driven Surrogate Models for Predicting Microstructural Evolution: Peichen Wu1; Kumar Ankit1; Ashif Lquebal1; 1Arizona State University
Phase-field (PF) models are one of the most powerful tools to simulate microstructural evolution in metallic materials, polymers, and ceramics. However, for accuracy, existing PF approaches rely on rigorous mathematical model development, sophisticated numerical schemes, and high-performance computing. Although recently developed surrogate models demonstrate the viability of deep-learning techniques in predicting microstructural evolution, such approaches require enormous amounts of high-fidelity training data while primarily relying on principal component analyses for microstructure representation, where spatiotemporal information is lost in the pursuit of dimensionality reduction. Given these limitations, we present a novel data-driven emulator (DDE) for predicting microstructural evolution which combines an image-based convolutional and recurrent neural network (CRNN) with tensor decomposition, while leveraging previously obtained PF datasets for training. The efficacy of our microstructure emulation technique will be discussed in the context of modeling microphase separation at the mesoscale.
10:00 AM
Predicting Grain Boundary Properties Using Strain Functional Descriptors and Supervised Machine Learning: Avanish Mishra1; Sumit Suresh1; Khanh Dang1; Saryu Fensin1; Edward Kober1; Nithin Mathew1; 1Los Alamos National Laboratory
Properties of grain boundaries (GBs) are controlled by the local atomic environments at the boundary. Although it is possible to estimate properties for well-defined GBs, extrapolating these understanding to real GBs is a formidable task. A significant challenge in establishing desired structure-property relationships is the absence of physically meaningful set of descriptors. Existing descriptor sets can estimate selected properties; however, they fail to provide any physical insights. Strain functional descriptors (SFD), a complete and symmetry-adapted set of descriptors, are ideal for characterizing the local atomic environment at GBs and provides the capability to decompose the atomic arrangement in terms of physically meaningful descriptions such as strains, strain-gradients, and higher order deformations. The presentation will show the application of SFDs for predicting various properties, such as grain boundary energy and atomic energy density on a large database of metastable GBs in Cu and application on polycrystals, using supervised machine learning models.
10:20 AM Break
10:35 AM
Statistical Generation of Three-Dimensional Dislocation Microstructures with Graph Neural Networks: Dylan Madisetti1; Jafaar El-Awady1; Christopher Stiles2; 1Johns Hopkins University; 2Johns Hopkins Applied Physics Laboratory
Building upon recent success in using Graph Neural Networks (GNNs) to model protein folding, we explore machine learning graph-based techniques for predicting three-dimensional (3D) dislocation networks in metals from a variety of material signals (e.g., X-Ray diffraction patterns, plastic response). Dislocation networks are represented as directed spatial graphs, where the directed edges represent dislocation lines (with metadata such as Burgers vector) connected at vertices (dislocation nodes). Adjacency matrices, constructed from sub-volumes of the dislocation network, are used as inputs to a GNN, which encodes connectivity into a latent vector. Material signals are then associated with this latent vector, allowing for fast microstructure similarity lookup. Additionally, the latent vector is used to produce new, statistically similar microstructures. This dual ability allows for the generation of representative dislocation networks from experiments, enabling a deeper understanding of dislocation evolution during loading with forward simulation in 3D Discrete Dislocation Dynamics
10:55 AM
Comparing Microstructure Representations for Machine Learning Models Predicting Material Properties: Akhil Thomas1; Ali Durmaz1; Harald Sack2; Chris Eberl3; 1Fraunhofer IWM; 2FIZ Karlsruhe / KIT Karlsruhe; 3University of Freiburg
For making effective and efficient machine learning (ML) models that predict material properties, it is necessary to identify the right kind of representations of its input material structure and/or process, besides optimizing over different types of ML models. In the presented work, we take the example of initial fatigue damage prediction using a multi-modal High-Cycle-Fatigue dataset and explore various input microstructure representations for the task. In particular, we examine and compare various representations of Electron Backscatter Diffraction (EBSD) data and other modalities registered to EBSD. Various input representations allow different types of ML models to be used, which are also presented.
11:15 AM
Inferring Topological Transitions in Pattern-forming Processes via Self-supervised Learning: Marcin Abram1; Keith Burghardt1; Greg Ver Steeg1; Remi Dingreville2; 1University of Southerm California; 2Sandia National Laboratories
The identification of transitions in microstructural regimes in pattern-forming processes is critical for understanding and fabricating microstructurally precise materials. Motivated by the universality principle for dynamical systems, we use a self-supervised approach to solve the inverse problem of predicting process parameters from observed microstructures using neural networks. We show that the difficulty of performing this prediction task is related to the goal of discovering transitions in microstructure regimes, because qualitative changes in microstructural patterns correspond to changes in uncertainty for our prediction problem. We demonstrate our approach for two distinct pattern-forming processes: the spinodal decomposition of a two-phase mixture and the formation of concentration modulations of binary alloys during physical vapor deposition of thin films. This approach opens a promising path forward for discovering unseen or hard-to-detect transition regimes. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525. SAND2022-3365 A
11:35 AM
What Does a Computer Vision Model Trained to Classify Material Microstructure Images Actually Understand?: Colby Wight1; Henry Kvinge1; Davis Brown1; Keerti Kappagantula1; 1Pacific Northwest National Laboratory
Deep learning-based computer vision models are increasingly being incorporated into research pipelines designed to explore and analyze microstructure images. Within other domains, state-of-the-art explainability techniques have begun to illuminate “reasons” behind model predictions. This understanding is critical in high-consequence areas where domain experts need to have confidence in their models. In this work, we apply interpretability methods to classification models for SEM images of AA7075 tubes manufactured by shear assisted processing and extrusion (ShAPE). We explore how these models respond to some of the features (e.g., grain size distribution, precipitate morphology, void topology) that are used when analyzing microstructure images in typical research process flow. Through this effort, we gain insight into what features the model is sensitive to and identify new features used by the classification models. For example, feature visualization in temper classification models, reveals models behave in peculiar ways only somewhat aligned with human intuition.