Algorithm Development in Materials Science and Engineering: Microscale Experiments and Machine Learning
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Integrated Computational Materials Engineering Committee, TMS: Phase Transformations Committee, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Mohsen Asle Zaeem, Colorado School of Mines; Garritt Tucker, Colorado School of Mines; Charudatta Phatak, Argonne National Laboratory; Bryan Wong, University of California, Riverside; Mikhail Mendelev, NASA ARC; Bryce Meredig, Travertine Labs LLC; Ebrahim Asadi, University of Memphis; Francesca Tavazza, National Institute of Standards and Technology
Thursday 8:30 AM
February 27, 2020
Room: 31C
Location: San Diego Convention Ctr
Session Chair: Noah Paulson, Argonne National Laboratory; Vahid Tari, ATI - Allegheny Technologies Incorporated
8:30 AM Cancelled
A Machine Learning Platform for the Accelerated Emulation of Quantum Mechanical Computations: Rampi Ramprasad1; 1Georgia Institute of Technology
Materials simulations are dominated by either quantum mechanics (QM) based methods—which are time-intensive, but accurate and versatile—or semi-empirical/ classical methods—which are fast but are significantly limited in veracity, versatility and transferability. Machine learning (ML) methods have the potential to bridge the chasm between the two extremes and can combine the best of both worlds. We have created a ML platform, trained on accurate QM reference data, for the rapid prediction of properties such as potential energy, atomic forces, stresses, charge density, and the electronic density of states, at a minuscule fraction of the QM cost. The ML models can also be progressively improved in quality by periodically (or on-demand) exposing them to fresh QM data in regions of poor performance. Here, we demonstrate the power and versatility of this new platform in correctly capturing electronic, thermodynamic, mechanical, and diffusive properties for a variety of systems.
8:50 AM
New Workflow for High-throughput Feature Extraction of Deforming Open Cell Foams: Steve Petruzza1; Attila Gyulassy1; Samuel Leventhal1; John Baglino1; Michael Czabaj1; Ashley Spear1; Valerio Pascucci1; 1University of Utah
Metallic open-cell foams are promising structural materials with applications in multifunctional systems such as biomedical implants and energy absorbers. There is a high demand for means to understand and correlate the design space of material performance metrics to the material structure attributes such as ligament, void and node properties. Modern X-ray Computed Tomography (CT) scans are used to segment and characterize those features presenting several challenges and days of manual labor. We present a new workflow for analysis of open-cell foams that combines a new density measurement to identify nodal structures, and topological approaches to identify ligament structures between them. Furthermore, we experiment the workflow measuring and tracking properties over time using a image sequences of foams being compressed. Our approach allows researchers to study larger and more complex foams and enables the high-throughput analysis needed to predict future foam performance.
9:10 AM
Microstructure Image Analysis using Deep Convolutional Neural Networks: Bo Lei1; Elizabeth Holm1; 1Carnegie Mellon University
In quantitative microscopy, microstructure image analysis lays the foundation for modeling the data and understanding the results. However, conventional image analysis methods are inefficient in handling complicated microstructure images and often require sophisticated and particular processing pipelines. We demonstrate that deep convolutional neural networks (DCNN) can be trained and achieve great performance and generality in challenging microstructure segmentation and analysis tasks. We evaluated and compared two DCNN models PixelNet and UNet with different training configurations to optimize the results on different datasets. The ability to segment complex microstructures enables a variety of new and high-throughput analysis methods. We also find that the quality of ground truth labels has a strong impact on performance, and reliable approaches to create ground truth labels are discussed.
9:30 AM Cancelled
Development of Virtual Resonant Ultrasound Spectroscopy Methods for use in Quantifying Defect Content: John Graham1; Ricardo Lebensohn1; Boris Maiorov1; Paul Lafourcade2; Laurent Capolungo1; 1Los Alamos National Laboratory; 2Commissariat a l'Energie Atomique
Resonant Ultrasound Spectroscopy (RUS) is one of many non-destructive evaluation techniques used to analyze bulk properties of materials, including defect content. However, the underlying theory used to develop the relationship between elastic waves and defects often makes unrealistic simplifications, such as dislocations being perfectly straight. In this work, we set out to develop a method of simulating virtual RUS experiments, in conjunction with our generalized discrete defect dynamics (GD3) code, to elucidate a better understanding of the interactions of defects with elastic waves, and to develop new and more accurate models for quantifying the defect content in a sample, based on measured RUS spectra.
9:50 AM
Relating 2D Experimental Information to 3D Simulations using Surface Structure Conserving 3D Microstructure Generation: Theron Rodgers1; Coleman Alleman1; Hojun Lim1; 1Sandia National Laboratories
Many lab-scale analysis methods used in in-situ mechanical testing are capable of capturing fine surface microstructural details and their evolution during deformation. However, the subsurface evolution of the microstructure is unknown. This complicates direct comparison with crystal plasticity simulations except in the simplest cases of single crystals or oligocrystals. Several methods exist for generating statistically equivalent 3D microstructures from 2D surface data, however they typically do not retain the exact surface microstructure. Here, we introduce a Potts Monte Carlo based method to generate 3D microstructures from 2D surface data that retains the exact surface structure. The method allows for the direct comparison between experiments and 3D simulations for equiaxed, polycrystalline microstructures. Many instantiations of 3D microstructures will be generated for a surface structure to quantify the variability introduced by subsurface features. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525k
10:10 AM Break
10:30 AM
Microstructure Reconstruction of Additive Manufactured Metallic Materials with Markov Random Fields: Arulmurugan Senthilnathan1; Pinar Acar1; Marc De Graef2; 1Virginia Tech; 2Carnegie Mellon University
In this work, we present a Markov Random Field (MRF) approach to reconstruct 2D polycrystalline microstructures. The MRF algorithm predicts the large-scale evolution of 2D experimental microstructural maps that are routinely obtained over small spatial domains with diffraction and optical techniques. The coloring of pixels is obtained by computing the conditional probability density using the known states of neighboring pixels in the input experimental images. In this work, we use MRF-based reconstruction to predict the spatial evolution of different metallic microstructures that are forged and additively manufactured. The reconstructed samples are expected to provide equivalent microstructural features to the experimental data. Therefore, a second order moment invariant is used to compute a feature number that characterizes the average size and shape of a grain. The grain sizes and shapes of the reconstructed microstructures are compared to the original forged and additively manufactured samples through these feature numbers.
10:50 AM
Real-time Analysis of Diffraction Data for Enabling In-situ Measurements: Anup Pandey1; John Andrew Redwig Catillo Castilo2; Surya Khalidindi2; Reeju Pokharel1; 1Los Alamos National Laboratory; 2Georgia Tech
Macroscopic properties of materials are governed by the spatial distribution of microstructures. Recent advancements in experimental techniques such as electron backscatter diffraction (EBSD), high-energy X-ray diffraction microscopy (HEMD), etc. have enabled the characterization and evolution of microstructures with better resolution. Specifically, HEMD being non-destructive technique enables 3D microstructures characterization and allows one to probe material dynamics subjected to various thermo-mechanical conditions. However, these experimental techniques generate terabytes of data per sample. We have developed efficient methods in analyzing these large experimental datasets using state-of-the-art techniques based on artificial intelligence and diffraction simulations, thereby opening an avenue for real-time analysis of in-situ measurements.
11:10 AM
Material Parameters Identification, Modeling and Experimental Verification of the New Smart Material Vacuum Packed Particles: Piotr Bartkowski1; Robert Zalewski1; 1Warsaw University of Technology
In this work the material parameter identification, simulations and experimental verification of the new smart material Vacuum Packed Particles is presented. The considered material is a structure composed of granular material inside the plastomer coating. When the pressure inside is equal or higher than atmospheric the system behaves like a liquid, otherwise like an elasto-plastic solid with a temperature and strain rate dependence. By changing the underpressure inside the coating it is possible to control the mechanical properties of the structure in a real time. In this work, the machine learning approach to model parameter estimation is presented. Additionally, the proposed model is implemented into LS-DYNA code Finite Element Method. In order to verify the simulation results a corresponding experimental tests were conducted using optical strain measurement system. The comparison of the outcomes showed a good correlation.
11:30 AM
Machine Learning Exploration and Optimization of Flame Spray Pyrolysis: Noah Paulson1; Joseph Libera1; Marius Stan1; 1Argonne National Laboratory
Materials structure and properties are sensitive to the tuning of processing conditions. Optimal material performance often represents a small region of process parameter space. Traditional synthesis methods struggle to characterize these spaces due to the high cost of test experiments. One such method is flame spray pyrolysis (FSP), where a plume of atomized solution combusts to produce nanoparticles for applications such as catalysis and chemical energy storage. In FSP, particle geometry and chemical/phase makeup are nonlinearly related to variables including solution chemistry, and liquid and gas flow rates. In this work, we employ Bayesian optimization (BO) to explore and optimize the processing space of FSP of silica (SiO2) nanoparticles based on in-situ optical emission spectroscopy and particle size distribution measurements. BO enables the discovery of interesting phenomena and optimizes parameter settings resulting in good performance with minimal expense.