ICME Gap Analysis in Materials Informatics: Databases, Machine Learning, and Data-Driven Design: Poster Session
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee, TMS: Integrated Computational Materials Engineering Committee
Program Organizers: James Saal, Citrine Informatics; Carelyn Campbell, National Institute of Standards and Technology; Raymundo Arroyave, Texas A&M University
Tuesday 5:30 PM
February 25, 2020
Room: Sails Pavilion
Location: San Diego Convention Ctr
L-21 (Digital): Deep Learning Image Analysis for Lattice Material Qualification: Ben White1; Anthony Garland1; Brad Boyce1; Bradley Jared1; David Saiz1; Michael Heiden1; Matthew Roach1; David Moore1; 1Sandia National Laboratories
Additively manufactured lattice metamaterials expand the range of accessible material properties. The mechanical properties of lattices are highly dependent on the additive manufacturing process so that a variation in the process settings such as laser power or scan rate can radically change bulk lattice material properties. Non-destructive evaluation of lattices is needed before they can be used as functional parts. In this talk, we show how deep learning image analysis can be applied to lattice metamaterial qualification. Using a single image of a lattice before physical testing, the quality of the print and thereby the properties of the lattice can be estimated. Transfer learning and data augmentation enable successful training and validation of our machine learning model with only 48 lattice samples. This demonstration of non-destructive evaluation shows how deep-learning can provide valuable insight to material engineers. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.
L-18 (Invited): Multi-fidelity Surrogate Assisted Framework for Prediction and Control of Meltpool Geometry in Additive Manufacturing Processes: Sudeepta Mondal1; Nandana Menon1; Daniel Gwynn1; Asok Ray1; Amrita Basak1; 1Pennsylvania State University
Thermal gradients in the meltpool play an important role in determining the final microstructure during Additive Manufacturing (AM). Physics-based models predicting the thermal field during AM processes are often computationally expensive, and it becomes prohibitive to run adequate simulations for a grid search over the process parameter space in order to generate a look-up table for choosing desirable operating conditions. The problem becomes more critical in the presence of a hierarchy of multi-scale multi-physics models, where budget limitations constrain the number of queries from the more expensive higher fidelity models. In order to surmount the existing gap, we propose a multi-fidelity (MF) surrogate assisted framework that encapsulates the statistical information in the varied fidelity levels via MF Gaussian Processes and searches for optimal process parameters via Bayesian Optimization strategies aimed at reaching the optima with limited exploitation of higher fidelity models.
L-19: Data-driven Hard-magnetic Materials Selection for AC Applications by Multiple Attribute Decision Making: Sunny Pinnam1; Alex Paul1; Tanjore Jayaraman1; 1University of Michigan-Dearborn
Hard magnetic materials are ubiquitous and, are used in a myriad of applications including, but not limited to computers, green energy technologies, and defense systems. Over the years, a variety of hard magnetic materials were developed to cater to the immanent technological demands. In the recent past, materials informatics has been an essential component of materials discovery, design, and development. We present a methodology that combines various multiple attribute decision-making methods, hierarchical clustering, and principal component analysis for data-driven hard magnetic materials selection. Shannon’s entropy model evaluated the relative weights of various properties followed by the ranking of the hard magnetic materials by the various multiple attribute decision making methods. Akin to Ashby charts, two-dimensional plots were developed to provide a visual presentation, based on the decision-making models, clustering, and component analysis followed by the assessment of the predictive capability of the data-driven model.
L-20: Data Driven Prediction of Crystallographic Attributes of Small Molecules Using Various Molecular Fingerprints: Piyush Karande1; Peggy Li1; Soo Kim1; Joanne Kim1; Hyojin Kim1; Donald Loveland1; T. Yong-Jin Han1; 1Lawrence Livermore National Laboratory
Crystal structures of organic molecules dictate several key properties that are critical in various industrial applications. Conventional methods to predict crystal structure rely heavily on expensive physics based computational models and simulations. As a potential alternative, here we propose a data driven approach to predict several different crystallographic attributes such as density, symmetry, and crystal packing motifs. We use the 3D structure of a subset of molecules from the Cambridge Crystal Structure Database and investigate the feasibility of various molecular fingerprinting and machine learning methods to predict the attributes. We present results using a) 3D Convolutional network, b) Graph convolutional networks, and c) Extended 3D Fingerprint with support vector machines and random forests, to predict these quantities. Each of these methods produce promising results and provides an insight into the information embedded in the 3D structure of these organic molecules.
L-22: Effect of Microtextured Regions on the Deformation Behavior of Titanium Alloys Submitted to Monotonic and Cyclic Loadings Investigated using FFT-EVP Simulations: Azdine Nait-Ali1; Samuel Hemery1; 1Ensma
Titanium alloys potentially exhibit micro textured regions (MTRs), also called macro-zones, which have a detrimental influence on the performance under cyclic loadings, but recent studies revealed a significant influence on the monotonic tensile behavior as well. Numerical simulations of the macroscopic response of large 3D aggregates containing MTRs of different shapes and sizes were performed using a crystal plasticity model based on fast Fourier transforms. The macroscopic response as well as the stress and strain fields at the microscopic scale were found to be affected by the presence of MTR. The importance of an explicit modeling of these MTRs is highlighted, in particular for cyclic loading conditions.
Cancelled
L-23: Identification of Fatigue Crack Nucleation Mechanics Using Bayesian Inference: Maxwell Pinz1; George Weber1; Somnath Ghosh1; 1Johns Hopkins University
A mechanics-based, data-driven approach to identify the underlying mechanics behind crack nucleation in Ni-based superalloy Rene88-DT is presented. Accurate estimates for the state variables of experimentally observed micrographs are obtained by simulating the observed microstructures with a multi-scale crystal plasticity model. These microstructures are embedded within a homogenized constitutive model to alleviate inherent errors due to boundary condition effects. By virtually loading experimental specimens, a data set of state variables is generated at each material point. A machine learning, Bayesian crack nucleation model is developed that optimally selects the most reliable state variable predictors of crack nucleation. The result of this approach is twofold. First, this process generates a probabilistic model to predict microstructure-sensitive cracks. Secondly, it selects form a list of candidate state variables the best predictors of crack nucleation. This approach leverages statistical analyses and machine learning to enhance our understanding of the mechanics behind complex phenomena.
Cancelled
L-24: Mapping Depletion Zone of High Nitrogen Stainless Steel Cr2N Using STEM-EDS: An Application of Multi-variate Statistical Analysis: Juyoung Kim1; Jee-Hwan Bae2; Hong Kyu Kim2; Jaeyoung Hong2; Gyeung Ho Kim2; Kyu Hyoung Lee1; Dongwon Chun2; 1Yonsei University; 2Korea Institute of Science and Technology
Multi-variate statistical analysis (MSA) method has been demonstrated that can reveal EDS and EELS signals which can be isolated as components arising from interfacial region. Moreover, these methods can also enable noise filtering of spectra which can lead to improved detection of small signals. The PCA [1] and ICA [2], which are unsupervised machine learning algorithms among MSA, were used to reduce dimension and individual elements extraction, respectively. And these algorithms are applied to Cr2N EDS mapping result on High nitrogen stainless steel (HNS) specimen. In this research, we have obtained Cr2N EDS mapping result through STEM-EDS analysis. As the result, we have succeeded in analyzing depletion zone around Cr2N precipitation phase using MSA. Also, we confirmed availability of MSA process by confirming depletion region through EELS with higher detection limit than STEM-EDS. Reference [1] Karl Pearson F.R.S., Philosophical Magazine 2 (1901)[2] Pierre Comon, Signal Processing 36 (1994)
L-25: Multi-class Inclusion Identification via Machine Learning of Multilevel Image Features: Nan Gao1; Mohammad Abdulsalam1; Bryan Webler1; Elizabeth Holm1; 1Carnegie Mellon University
Multilevel image features including contrast and morphology are investigated via computer vision (CV) and machine learning (ML) for inclusion classification. In the steel industry, distinguishing inclusions from steel substrates relies heavily on Energy Dispersive X-Ray Spectroscopy (EDS) equipped on a Scanning Electron Microscope (SEM). But more efficient, timely and cost-effective analysis methods are still needed since EDS is time-consuming for element analysis. Considering the capability of pulling out high level features and superfast processing speed via CV and ML, these techniques offer us opportunities to solve this issue. State-of-the-art pretrained CNNs are utilized to capture morphologic features. Additionally, color features using histogram and color moment representations are incorporated into features vectors for the formation of multilevel features. Fewer classification errors were observed especially for inclusions with similar chemical composition. This study can be used to explore the potential of using CV and ML instead of SEM/EDS for element analysis.
L-26: Prediction of Temperature after Cooling in Coils Using Machine Learning and Finite Element Method: Hyeok Jae Jeong1; Seonghwan Kim1; Nam Hoon Goo1; 1Hyandai steel
The Integrated Computational Materials Engineering (ICME) method provides chances for developing new materials and improving conventional processes. It is important to minimize the computational cost and bottlenecks to utilize ICME approach in the field. Machine learning is useful technique to reduce the computing time. Only a few seconds are required to obtain the outputs using the machine learning because most of computational cost is consumed during the training. In this study, the neural network was trained to predict the temperature distribution during the cooling process of the hot rolled coil. The training dataset was generated using the finite element method (FEM). Several material and process parameters were considered as labels in the neural network.
L-27: Uncertainty Quantification in Metallic Additive Manufacturing Through Physics-informed Data-driven Modeling with Experimental Validation: Lei Chen1; Zhuo Wang2; Zhen Hu1; Sankaran Mahadevan3; 1University of Michigan-Dearborn; 2Mississippi State University; 3Vanderbilt University
The complicated metal-based additive manufacturing (AM) process involves various uncertainty sources, leading to variability in AM products. To achieve a comprehensive uncertainty quantification (UQ) study of AM processes, we present a physics-informed data-driven modeling framework with experimental validation, in which multi-level data-driven surrogate models are constructed based on extensive computational data obtained by experimentally validated multi-scale multi-physics AM models. It starts with computationally inexpensive surrogate models for which uncertainty can be readily quantified, followed by a global sensitivity analysis for a comprehensive UQ study. Using AM fabricated Ti-6Al-4V components as examples, this study demonstrates the capability of the proposed data-driven UQ framework for efficiently investigating uncertainty propagation from process parameters to material microstructures, and then to macro-level mechanical properties through a combination of advanced AM multi-physics simulations and data-driven surrogate modeling. The model correction and parameter calibration for the constructed surrogate models using limited amount of experimental data is discussed.