wisdom of crowds business intelligence market study tm ... · capability development for the...

LA-UR-19-22653 Approved for public release; distribution is unlimited.

Title: Computational Data Science Approaches for Materials

Author(s): Shea, Kelly Edna

Intended for: Computational Data Science Approaches for Materials 2019 Conference, 2019-04-08/2019-04-10 (Los Alamos, New Mexico, United States)

Issued: 2019-03-25

Disclaimer: Los Alamos National Laboratory, an affirmative action/equal opportunity employer, is operated by Triad National Security, LLC for the NationalNuclear Security Administration of U.S. Department of Energy under contract 89233218CNA000001. By approving this article, the publisherrecognizes that the U.S. Government retains nonexclusive, royalty-free license to publish or reproduce the published form of this contribution,or to allow others to do so, for U.S. Government purposes. Los Alamos National Laboratory requests that the publisher identify this article aswork performed under the auspices of the U.S. Department of Energy. Los Alamos National Laboratory strongly supports academic freedomand a researcher's right to publish; as an institution, however, the Laboratory does not endorse the viewpoint of a publication or guarantee itstechnical correctness.

LA-UR #

Dir ac M at er ials - Repor t . docx

Computational Data

Science

Approaches for

Materials Sponsored by the

Institute for Materials Science, National Security Education Center

Los Alamos National Laboratory

Computation Data Science Approaches for Materials 2019

Organizers: T. Ahmed, Los Alamos National Laboratory D.J. Luscher, Los Alamos National Laboratory

Co-Organizers: Alejandro López-Bezanilla, Los Alamos National Laboratory Roxanne Tutchton, Los Alamos National Laboratory

Workshop Associates: Filip Ronning, Los Alamos National Laboratory Kelly Shea, Los Alamos National Laboratory

Introduction Background: In today’s world, new computational data science tools are continuously being developed every day, impressively keeping up with the high speed real –time data acquisition methods and techniques. Much of this data science effort is dedicated towards improving the predictive capabilities ranging from social media to materials and physical properties. Some methods have proven to be quite versatile and easily adaptable between different applications. This naturally started vigorous collaborations between data scientists and materials science communities by developing and sharing efficient algorithms and data. “Machine Learning” (ML) and “Data Mining” are now common research tools in materials science. There are countless examples of success (as well as limitations) in predicting materials properties using state-of-the-art data science (e.g. ML) methods. In the last half decade, we have noticed numerous publications as well as workshops and summer schools dedicated to Materials Informatics. We have also noticed an explosive growth of open source Databases, focused primarily on materials properties whose scope and scale vary from electronic-structure to structural properties and beyond. Besides these academic and open-source resources, there are many specialized and restricted databases for addressing materials related problems, many of which are dedicated to mission and national/global security applications. We have also realized that these two research communities can benefit by a more effective and a closer communication, and by sharing many general methodologies and algorithms without breaching data confidentiality. We believe our workshop will bring two communities together in one place with the expectation to reinforce new collaborations and to avoid ‘reinventing- the- wheel‘ for developing many of the existing computational predictive models on both sides. .

Institute for Materials Science National Security Education Center Los Alamos National Laboratory

2


Purpose /Goal: Unlike past academic conferences on materials informatics, our workshop is designed and focused to improve connections between the two sectors: “the open source materials-informatics community” and “the data science community dedicated towards programmatic applications.” Only through a better understanding of the needs, requirements, sensitivity and capabilities of these two apparently independent research sectors, can we bridge the gap, and progress towards the successful development of predictive capabilities for the next generation materials. Our workshop will invite people from both the academic and programmatic research with expertise in machine learning and data science methods to solve materials related problems. This workshop will provide a unique environment for capability reviews of various teams nationwide, and also initiate effective and realistic collaborations between them. This is schematically summarized in the following viewgraph:

Outcome and Mission Impact: The main outcomes are the following:

1) Knowledge update on existing computational data science methods and data base capabilities in various sectors of materials research.

2) Formulate effective ways of creating complementary knowledge (e.g. algorithm, methods, data) transfer without breaching data confidentiality or security in different sectors.

3) Reinforce collaborations between academia and national laboratory (LANL).

This long-needed workshop in materials science will address the outstanding problems identified by DOE Office of Science’s ASCR and SciDAC programs in the area of advanced computing for materials. The workshop also aligns very well with the Basic Research on Materials Informatics and an initiative of DoD. We envision other potential collaborations and opportunities on the capability development for the ‘Materials for the Future’ pillar within LANL.


3

-

-

-


Conference Schedule Monday 8th

Note: Poster Session

Monday April 8th

MORNING I

Opening Remarks from Associate Laboratory Director (SC) 8:45am – 9:00am Irene Qualters, ALD (SC, LANL) Session 1A: Session 1A: Materials Discovery by Design from

Electronic Structures using Data-Science 9:00am – 9:45am Andrew Rappe (UPenn) 9:45am – 10:15am Sorelle Friedler (Haverford College) 10:15am - 10:30am Jack Shlachter (T-DO, LANL) 10:30am – 10:45am Thom Mason, Director (LANL)

COFFEE BREAK (10:45am 10:55am) Monday April 8th

MORNING II

Session 1B: Session 1B: Fundamental Understanding of Materials from Electronic Structures for ‘Materials by Design’.

10:55am – 11:30am Arunima Singh (Arizona State) 11:30am – 11:55am Antia Sanchez Botana (Arizona State) 11:55am - 12:15pm Jianxin Zhu (T-4, LANL)

LUNCH BREAK (12:15pm 1:30pm) Monday April 8th

AFTERNOON I

Session 1C: Session 1C : Applying Data Driven Approaches for Molecular and Nano-Scale Materials

1:30pm – 2:15pm Subramanian Sankaranarayanan (Argonne) 2:15pm – 2:40pm Sergei Tretiak (T-1, LANL) 2:40pm - 3:00pm Alexandrov Boian (T-1, LANL)

COFFEE BREAK (3:00pm 3:15pm) Monday April 8th

AFTERNOON II

Session 1D: Session 1D : Data Driven Methods for Predicting Micro-Structures in Solids and Fluids

3:15pm – 4:00pm Steve Waiching Sun (Columbia University) 4:00pm – 4:25pm Juan Saenz (XCP-4, LANL) 4:25pm - 4:45pm Arvind Mohan (T-4, LANL)

4:45pm – 6:00pm POSTER SESSION (will stay up for 3 days in the lobby)

Tuesday, April 9th

Note: Banquet

Tuesday April 9th

Session 2A: Session 2A: Data-Science Tools Development for Materials including Databases and Visualization Techniques


4

-

-

-

-

-


MORNING II 8:45am – 9:30am Olexandr Isayev (UNC, Chapel Hill) 9:30am – 9:55am James Ahrens (CCS-7, LANL) 9:55am - 10:15am Mike McKerns (CCS-3, LANL)

COFFEE BREAK (10:15am 10:30am) Tuesday April 9th

MORNING II

Session 2B: Session 2B: General Data Science Method Development

10:30am – 11:15am Ying Wai Li (ORNL) 11:15am – 11:40am Christine Anderson-Cook (CCS-6, LANL) 11:40am - 12:00pm Kipton Barros (T-1, LANL)

LUNCH BREAK (12:00pm 1:30pm) Tuesday April 9th

AFTERNOON I

Session 2C: Session 2C : Data-Driven Approaches for Materials and Engineering Problems

1:30pm – 2:15pm Congwang Ye (LLNL) 2:15pm – 2:40pm Brian Giera (LLNL) 2:40pm - 3:00pm David Mascarenas (NSEC, LANL)

COFFEE BREAK (3:00pm 3:15pm) Tuesday April 9th

AFTERNOON II

Session 2D: Session 2D : Dataminingand Machine Learning to Guide Experiments and Device Technology

3:15pm – 4:00pm Matthew Rever(LLNL) 4:00pm – 4:25pm Ghanshyam Pilania (MST-8, LANL) 4:25pm - 4:45pm Nikolay Makarov (UbiQD) 6:30pm Banquet (Location: TBA)

Wednesday, April 10th

Note: Last Day

Wednesday April 10th

MORNING II

Session 3A: Session 3A: Data-Driven Methods for Mechanical and Mesoscale Structural problems of Materials.

8:45am– 9:30am Surya Kalidindi (Georgia Tech) 9:30am – 9:55am Turab Lookman (T-4, LANL) 9:55am - 10:15am Reeju Pokharel (MST-8, LANL)

COFFEE BREAK (10:15am 10:30am) Wednesday April 10th

MORNING II

Session 3B: Session 3B: Data-Driven Approach for Radiation-Matter Interaction

10:30am – 11:00am Ricardo Lebensohn (T-3, LANL) 11:00am – 11:30am Hari Viswanathan (EES-16, LANL) 11:30am - 12:00pm Eddy Timmermans (XCP-5, LANL)

LUNCH BREAK (12:00pm 1:30pm)


5

-


Wednesday April 10th

AFTERNOON I

Session 3C: Session 3C : Data-Driven Approach for High Energetic Materials

1:30pm – 2:00pm Edward Kober (T-1, LANL) 2:00pm – 2:30pm Marc Cawkwell (T-1, LANL) 2:30pm - 3:00pm Chris Biwer (CCS-7, LANL)

COFFEE BREAK (3:00pm 3:15pm) Wednesday April 10th

AFTERNOON II

Session 3D: Session 3D : Data-Driven Approaches for Nuclear Materials and Nuclear Data

3:15pm – 4:00pm Bobby Sumpters (ORNL) 4:00pm – 4:25pm Denise Neudecker (XCP-5, LANL) 4:25pm - 4:55pm Closing Remarks: TBD

End of Agenda


6


Invited and Contributed Talks

Invited Talks

Interpretable Machine Learning for Scientific Discovery

Sorelle Friedler – Haverford College

Abstract: Machine learning algorithms are increasingly useful for modeling and predicting the outcome of scientific experiments. But accurate predictions are not the main goal such scientific exploration. In this talk, we will explore how

interpretability techniques can be used to understand machine learning models and support the development of scientific hypotheses. By using post-hoc interpretability methods, the power of complex models can be paired with an examination of the influence of features on the outcome to allow for a deeper understanding of the scientific landscape behind the models.

Biography: Sorelle Friedler is an Assistant Professor of Computer Science at Haverford College and an Affiliate at the Data & Society Research Institute. Her research focuses on the fairness and interpretability of machine learning algorithms, with applications from criminal justice to materials discovery. She is a co-founder of the ACM Conference on Fairness, Accountability, and Transparency (FAT*) and its former program committee co-chair. Sorelle is the recipient, along with chemistry professors Josh Schrier and Alex Norquist, of a DARPA contract and two NSF Grants to apply data mining techniques to materials chemistry data to speed up materials discovery, using interpretable machine learning techniques to inform scientific hypotheses. Their work on this topic was featured on the cover of Nature and was covered by The Wall Street Journal and Scientific American. Sorelle holds a Ph.D. in Computer Science from the University of Maryland, College Park.


7


Machine Learning Based Monitoring of Advanced Manufacturing Brian Giera –Lawrence Livermore National Laboratory

Abstract: As with most advanced manufacturing (AM) systems, analysis of AM sensor data currently occurs post-build, rendering process monitoring and rectification impossible. Supervised machine learning offers a route to convert sensor data into real-time assessments; however, this requires a wealth of labeled sensor data that traditionally is too time-consuming and/or expensive to assemble. In this work, we solve this critical issue in a variety of AM systems. We develop

and implement machine learning (ML) algorithms for the purposes of automated quality assessment and, in some cases, rectification. We discuss ML-based algorithms capable of automated detection in a host of AM technologies such as Laser Powder Bed Fusion and Direct Ink Write and also microfluidic platforms that are used for feedstock production. The common thread within these systems is that routinely collected sensor data (e.g. high-speed video, pressure gauges, etc.) contains pertinent information about the state of the system that can be converted into actionable information in real-time via ML. Successful implementation of these machine learning algorithms will reduce time and cost during process by automating quality assessment and lead to process control.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.

Authors: Brian Giera*, Bodi Yuan, Albert Chu, Phillip DePond, Gabe Guss, Du Nguyen, Congwang Ye, Will Smith, Nik Dudukovic, Sara McMains, and Manyalibo Matthews

Biography: I am Principle Investigator on the Lab Directed R&D Exploratory Research Project “Rapid Closed-Loop Control of Advanced Manufacturing with Machine Learning.” More broadly, my research focus is to develop and apply traditional (molecular dynamics & continuum/theory) and data driven (machine learning & computer vision) models to address materials processing and characterization problems within advanced manufacturing systems.


8


Accelerating design of inorganic materials with machine learning and AI Olexandr Isayev – University of North Carolina, Chapel Hill

Abstract: Historically, materials discovery is driven by a laborious trial-and-error process. The growth of materials databases and emerging informatics approaches finally offer the opportunity to transform this practice into data- and knowledge-driven rational design—accelerating discovery of novel materials exhibiting desired

properties. The Materials Genome Initiative (MGI) has transformed Materials Science into a data-rich discipline. These developments open exciting opportunities for knowledge discovery in materials databases using informatics approaches to inform the rational design of novel materials with the desired physical and chemical properties. Statistical and data mining approaches have been successfully employed in both chemistry and biology leading to the development of cheminformatics and bioinformatics, respectively. However, until recently their application in materials science has been limited due to the lack of sufficient body of data, methods and reliable infrastructure.

In this work we showcase a pilot materials informatics platform capable of (i) instantaneously query and retrieve the necessary material information in the form of web app and RESTful API; (ii) identify, visualize and study important data patterns, and (iii) generate experimentally-testable hypotheses by building predictive Machine Learning (ML) models based on materials fingerprints and descriptors. Our computational approach relies on cheminformatics methodologies that one of our groups has developed and employed successfully to enable rational design of organic compounds with desired properties (e.g., drug candidates). By using data from the AFLOW repository (www.aflow.org) for high-throughput ab-initio calculations, we have generated ML models to predict many critical material properties like superconductivity, Debye temperature, Seebeck coefficient, bulk modulus, and band gap energy.

Biography: Olexandr Isayev is an assistant professor at UNC Eshelman, School of Pharmacy, University of North Carolina at Chapel Hill. His group's research focuses on the design of new molecules and materials with machine learning and artificial intelligence. Before joining UNC in 2013, Oles was a post-doctoral research fellow at Case Western Reserve University and research scientist with the U.S. Army. In 2008, he received his Ph.D. in computational chemistry from Jackson State University. His honors include the 2017 emerging technology


9

www.aflow.org


award from the American Chemical Society, the 2016 Eshelman Institute for Innovation award, and the GPU computing award from NVIDIA Corp.

Materials Innovation Driven by Data and Knowledge Systems Surya Kalidindi – Georgia Institute of Technology

Abstract: Current approaches to exploring materials and manufacturing (or processing) design spaces in pursuit of new/improved engineered structural materials continue to rely heavily on extensive experimentation, which typically demand inordinate investments in both time and effort. Although

tremendous progress has been made in the development and validation of a wide range of simulation toolsets capturing the multiscale phenomena controlling the material properties and performance characteristics of interest to advanced technologies, their systematic insertion into the materials innovation efforts has encountered several hurdles. The ongoing efforts in my research group are aimed at accelerating materials innovation through the development of (i) a new mathematical framework that allows a systematic and consistent parametrization of the extremely large spaces in the representations of the material hierarchical structure (spanning multiple length/structure scales) and governing physics across a broad range of materials classes and phenomena, (ii) a new formalism that evaluates all available next steps in a given materials innovation effort (i.e., various multiscale experiments and simulations) and rank-orders them based on their likelihood to produce the desired knowledge (expressed as PSP linkages), and (iii) novel higher-throughput experimental assays that are specifically designed to produce the critically needed fundamental materials data for calibrating the numerous parameters typically present in multiscale materials models. I will present and discuss ongoing research activities in my group.

Biography: Surya Kalidindi is a Professor in the Woodruff School of Mechanical Engineering at Georgia Institute of Technology, Georgia, USA with joint appointments in the School of Materials Science and Engineering as well as the School of Computational Science and Engineering. Surya earned a Ph.D. in Mechanical Engineering from Massachusetts Institute of Technology in 1992, and joined the Department of Materials Science and Engineering at Drexel University as an Assistant Professor. After twenty years at Drexel University, Surya moved into his current position at Georgia Tech. Surya’s research efforts have made seminal contributions to the fields of crystal plasticity, microstructure design, and materials informatics. Surya has been elected a Fellow of ASM International,


10


TMS, and ASME. In 2016, he and his group members have been awarded the top prize as well as one of the runner-up prizes in the national Materials Science and Engineering Data Challenge sponsored by the Air Force Research Lab in partnership with the National Institute of Standards and Technology and the U.S. National Science Foundation. He has also been awarded the Alexander von Humboldt Research Award, the Vannever Bush Faculty Fellow, the Government of India’s Vajra Faculty Award, and the Khan International Award.

Combining machine learning and Monte Carlo methods for the studies of finite temperature materials properties

Ying Wai Li – Oak Ridge National Laboratory

Abstract: Classical Monte Carlo methods are proven to be robust computational techniques to study thermodynamics of matter. When combined with high fidelity Hamiltonian models, they are able to obtain finite temperature properties like phase transitions and stabilities to a practical accuracy. However, the

huge computational cost has hindered its general application in practice. We investigate the use of machine learning techniques to improve and accelerate Monte Carlo simulations from different perspectives: (1) the construction of accurate model Hamiltonian from experimental data; (2) the training of surrogate models from first principles calculations to simultaneously predict multiple physical observables; and (3) the training of generative models to reproduce canonical distributions. We will illustrate how these techniques accelerate the modeling of physical systems ranging from spin models, spin ice, to binary alloys.

Biography: Dr. Ying Wai Li is a R&D staff scientist in the National Center for Computational Sciences at Oak Ridge National Laboratory (ORNL), U.S. Her research interests lie at the interface between condensed matter physics and high performance computing (HPC). Her expertise includes state-of-the-art classical and parallel Monte Carlo algorithms in statistical mechanics, numerical and first principles methods (density functional theory and quantum Monte Carlo) in the study of electronic structures of materials, and machine learning methods applied to materials. She is also interested in HPC scientific software development, algorithm design, performance analysis and optimization for


11


massively parallel, scalable scientific software for leadership-class computers such as Titan and Summit at ORNL.

Optimization Challenges in Applications of Visible-to-Near-Infrared CuInSeS/ZnS Quantum Dots

Nikolay S. Makarov – UbiQD, LLC

Abstract: Not-toxic CuInSeS/ZnS quantum dots (QDs) offer unique opportunities for various applications thanks to their bright emission tunable across the visible and near-infrared (NIR) spectral range and low-cost scalable synthesis. Each of these applications requires certain parameters to be optimized, not just during the synthesis of the nanomaterials, but also in the device

architecture, and even decisions of whether to use the QDs in certain parts of the devices (such as whether to use solar windows in all facades of the tall buildings). Here we describe several optimization challenges and show some of the outcomes. In particular, we first demonstrate that detailed calculations can predict benefits of the solar windows for variously-oriented sides of the tall buildings in different cities across the US. Next we show the reasons why optimization is required for light coupling from the laminate window to the solar cells. Further, we formulate optimization challenge for agriculture products. Finally, we show steps towards optimization of UbiQD’s medical fiber-coupled device. Overall, between the material synthesis, product idea, and product implementation, there are numerous of steps requiring optimization in order to achieve the best device performance at a minimal price. This presentation outlines some of the optimization challenges being currently solved by UbiQD, and points towards potential needs of the Company.

Biography: Nikolay S. Makarov was born in St. Petersburg, Russia, in 1980. He received his Ph.D. degree from Montana State University, Bozeman in May 2010. After two postdoc positions (at Georgia Institute of Technology and Los Alamos National Lab), he is currently a Director of Applied Physics at UbiQD, Inc. His main research interests are in quantitative and ultrafast spectroscopy including multi-photon absorption in organic molecules and colloidal quantum dots for optical 3D storage, biophotonics, and solar applications, and stimulated Raman scattering. He has co-authored ~60 papers and participated in numerous international meetings.


12


Modeling Heterogeneous Electrocatalysis on Realistic Surfaces from First-Principles, Thermodynamics, and Machine Learning

Andrew M. Rappe – University of Pennsylvania

Abstract: Theoretical and computational approaches have made important contributions to heterogeneous electrocatalysis. Such approaches, however, assume that the catalytically-active surface for a particular class of materials is the same. We have discovered that the electrocatalytic activity of nickel phosphides

and CaMnO3 toward the H2 and O2 evolution reactions, are governed by aqueous surface equilibria, i.e. the most catalytically-active surfaces are aqueous surface reconstructions. In addition to the determination of realistic surfaces for computational studies of heterogeneous electrocatalysis, we are also developing combined first-principles and machine learning techniques for the automated discovery of catalytic descriptors. Using these techniques, we discovered that the Ni-Ni bond length is an excellent descriptor for the hydrogen evolving activity of Ni2P and that it can be modulated via nonmetal surface doping, which induces a chemical pressure effect.

For decades, ab initio thermodynamics has been the method of choice for computationally determining the surface phase diagram of a material under different conditions. The surfaces considered for these studies, however, are often human-selected and too few in number, leading both to insufficient exploration of all possible surfaces and to biases toward portions of the composition-structure phase space that often do not encompass the most stable surfaces. To overcome these limitations and automate the discovery of realistic surfaces, we combine density functional theory and grand canonical Monte Carlo (GCMC) into "ab initio GCMC." We demonstrate ab initio GCMC for the study of oxide overlayers on Ag(111), which, for many years, mystified experts. Ab initio GCMC rediscovers the surface phase diagram of Ag(111) with no preconceived notions about the system. Using nonlinear, random forest regression, we discover that Ag coordination number with O and the surface O-Ag-O bond angles are good descriptors of the surface energy. Additionally, using the composition-structure evolution histories produced by ab initio GCMC, we deduce a mechanism for the formation of oxide overlayers based on the Ag3O4 pyramid motif that is common to many reconstructions of Ag(111). Ab initio GCMC is a promising tool for the discovery of realistic surfaces that can then be


13


used to study phenomena on complex surfaces such as heterogeneous catalysis and materials growth, enabling reliable and insightful interpretations of experiments.

Biography: Andrew M. Rappe is the Blanchard Professor of Chemistry and Professor of Materials Science and Engineering at the University of Pennsylvania. He received his A. B. in “Chemistry and Physics” summa cum laude from Harvard University in 1986, and his Ph. D. in “Physics and Chemistry” from MIT in 1992. He was an IBM Postdoctoral Fellow at UC Berkeley before starting at Penn in 1994. Andrew received an NSF CAREER award in 1997, an Alfred P. Sloan Research Fellowship in 1998, and a Camille Dreyfus Teacher-Scholar Award in 1999. He was named a Fellow of the American Physical Society in 2006. He received the Humboldt Research Award in 2017. Andrew is one of the two founding co-directors of the VIPER honors program at Penn, the Vagelos Integrated Program in Energy Research. Andrew has published more than 250 peer-reviewed articles. In recent years, he has become a leader in the theory of hybrid organic-inorganic perovskites and of topological materials. He has championed the use of the bulk photovoltaic effect for solar energy harvesting, and he has made seminal contributions to the theory of ferroelectric materials.

Feedstock Optimization Using Computer Vision and Machine Learning Techniques Matthew A. Rever – Lawrence Livermore National Laboratory

Abstract: The feedstock optimization project at Lawrence Livermore National Laboratory (LLNL) aims to accelerate the materials design and optimization process by combining computer vision (CV), machine learning (ML), data analytics, and experimental validation to pinpoint critical material attributes

needed to obtain desired properties and performance. This presentation will cover some advances in image feature extraction as well as regression methods for predicting key performance metrics for the feedstock materials. In addition, techniques for augmenting limited data for use with deep-learning will be discussed. Development of these tools may significantly shorten the research and development time for materials optimization and integration by potentially providing insights and inferences based on experimental and computational data.


14


Biography: Matthew Rever (PhD, EE, University of Michigan) is an image-processing, computer vision, and machine learning researcher at Lawrence Livermore National Laboratory. He has developed a multitude of automated analysis and control packages for various applications. In addition to working on material science applications, he is also currently investigating machine learning for improving multi-physics simulation robustness.

Rational design for low dimensional magnetism

Antia Sanchez-Botana – Arizona State University

Abstract: The term “rational design” has recently come to signify a closer coupling of computational modeling with synthesis and characterization, versus the high-throughput examination of selected properties (i.e. band gap, structural energy) offered by publicly available databases. In this talk, I

will show how to integrate rational design into the materials design & discovery process by performing an efficient selection of the appropriate structures and compositions for the targeted property. The focus will be on low dimensional magnetism using mineralogical databases.

Biography: Antia Botana is an assistant professor in the Department of Physics at Arizona State University. Prior to joining ASU, she was a postdoctoral fellow at Argonne National Lab and at the University of California, Davis. Her research employs density functional theory to direct the computational design of materials with novel functionalities. She works on topics ranging from superconductivity to frustrated magnetism, thermoelectricity, and confinement effects in nanostructures.

Accelerating Materials Discovery and Design using AI and Machine Learning Subramanian Sankaranarayanan – Argonne National Laboratory

Abstract: The ever-increasing power of modern supercomputers, along with the availability of highly scalable atomistic simulation codes, has begun to revolutionize predictive modeling of materials. Molecular dynamics (MD), in particular, has led to breakthrough advances in diverse fields, including tribology, energy storage, catalysis, sensing. Furthermore, recent integration of MD simulations with X-ray characterization has demonstrated promise in real-


15


time 3-D atomistic characterization of materials. The popularity of MD is driven by its applicability at disparate length/time-scales, ranging from ab initioMD (hundreds of atoms and tens of picoseconds) to all-atom classical MD (millions of atoms and tens of nanoseconds), and coarse-grained (CG) models (microns and tens of micro-seconds). Nevertheless, a substantial gap persists between AIMD, which is highly accurate but restricted to extremely small sizes, and those based on classical force fields (atomistic and CG) with limited accuracy but access to larger length/time scales. The accuracy and predictive power of classical MD is dictated by the empirical force fields, and their capability to capture the relevant physics.

In this talk, I will present some of our recent work on the use of machine learning (ML) to seamlessly bridge the electronic, atomistic and mesoscopic scales for materials modeling. Our automated ML framework aims to bridge the significant gulf that exists between the handful of research groups that develop new interatomic potential models (often requiring several years of effort) and the increasingly large user community from academia and industry that applies these models. Our ML approach showed marked success in developing force fields for a wide range of materials from metals, oxides, nitrides, hetero-interfaces to two-dimensional (2-D) materials6and even water (arguably the most difficult system to capture from a molecular perspective)7. This talk will also briefly discuss our ongoing efforts to integrate such cheap yet accurate atomistic models with (a) AI techniques to perform inverse design and construct metastable phase diagrams of materials (b) Deep learning to improve spatiotemporal resolutions of ultrafast X-ray imaging.

Biography: Subramanian Sankaranarayanan is a Scientist in the Nanoscale Science and Technology Division at Argonne National Laboratory and a Senior Fellow at the Institute of Molecular Engineering at University of Chicago. Prior to joining Argonne, Subramanian was a post-doctoral fellow at the School of Engineering and Applied Sciences at Harvard University. His research focuses on the use of machine learning to bridge the electronic, atomistic and mesoscopic scales for accelerated materials discovery and design. He is using supervised machine learning techniques to develop first-principles based force fields for simulating reactive and mesoscopic systems. Other programmatic efforts include development and use of AI algorithms for inverse design of materials and deep learning for integrated X-ray imaging of ultrafast energy transport across solid-solid and solid-liquid interfaces. His interests span a diverse range of applications from energy storage, tribology, corrosion to


16


neuromorphic computing and thermal management. He is a co-inventor on 4 patents and has co-authored more than 130 journal publications including several high impact publications in Science, Nature, Nature Materials, Nature Communications, Proceedings of National Academy of Sciences, ACS Nano, Nanoletters, and Physical Review Letters to name a few.

High-Throughput Screening of Substrates for Synthesis and Functionalization of 2D Materials

Arunima K. Singh – Arizona State University

Abstract: Less than 5 % of the >1500 theoretically predicted and promising two-dimensional (2D) materials have been experimentally synthesized. In this talk I will present a density-functional theory based framework that can be used to identify

suitable substrates that enable growth and functionalization of as-yet-hypothetical 2D materials. We have applied this formalism to identify substrates for 2D group-III-V materials,1,2,3 validated it against experimental synthesis of 2D MoS2

4 and integrated the results from the framework with phase-field models to predict the micro-structure of graphene on various metallic substrates.5

We are currently applying this strategy for a high-throughput screening of substrates for the thousands of as-yet hypothetical 2D materials. In order to automate the steps associated with the search such as generation of heterostructures, creation of input files, submission of runs on computing resources, post-processing of simulations, error management and curation of key properties in an open-source database, we have developed an open-source python-package MPInterfaces.6 The high-quality electronic structure data and physio-chemical properties of 2D materials-substrate interface emerging from this study will provided an invaluable theoretical input to 2D materials growers as well as serve as a critical atomic-scale input for multi-scale studies leading to an accelerated growth and functionalization of 2D materials.

1. Singh et al., Phys. Rev. B 2014, 89, 245431.

2. Singh et al., Appl. Phys. Lett. 2014, 105, 051604.

3. Zhuang, Singh et al., Phys. Rev. B 2013, 87, 165415


17


4. Singh et al., Appl. Phys. Lett. 2015, 107 (5), 053106.

5. Smirman, Taha, Singh et al., Phys. Rev. B 2017, 95 (8), 085407.

6. Mathew, Singh, et al., Comp. Mater. Sci. 2016, 122, 183.

Biography: Arunima K. Singh is an assistant professor in the Department of Physics at Arizona State University and a graduate faculty of the Materials Science and Engineering department at ASU. Prior to joining the ASU faculty, Singh was a postdoctoral associate at the Lawrence Berkeley National Laboratory, in the Materials Project team, from 2017-2018 and at the National Institute of Standards and Technology, in the Materials Genome team, from 2014-2016. She received her doctorate in 2014 from Cornell University. Her research focuses on accelerating materials discovery, synthesis and application using first-principles computations. She is particularly interested in physical phenomena occurring at surfaces and interfaces of materials.

Toward a physical intelligence of functional materials Bobby G. Sumpter – Oak Ridge National Laboratory

Abstract: Recent technical advances in the area of nanoscale imaging, spectroscopy and scattering/diffraction have provided tremendous capabilities for investigating materials structural, dynamical and functional characteristics. At the same time, advances in computational algorithms, including deep learning approaches, and computer capacities that are orders of

magnitude larger and faster, have enabled extreme-scale simulations of materials properties starting with nothing but the identity of the atomic species and the basic principles of quantum and statistical mechanics and thermodynamics. Along with these advances, enormous amounts of data have emerged (both experimental and theoretical). This confluence of capabilities/advances and the information bound in large volumes of data, offers new opportunities for advancing materials and chemical sciences. In this talk I will discuss how we are probing in-situ, important aspects of chemical reactions and hierarchical assembly as a modality for direct feedback to the experiment in order to precisely impart directed energy (electrons, ions, photons) that sculpts a material at the nanoscale. This approach is enabled via the dual capability of high-resolution imaging for near real time atomistic imaging co-directed energy delivery for atomic sculpting, leading to data rates and volumes that provide a deep learning


18


framework ability for efficient identification of structures and dynamics. We have found that the approach enables efficient mapping of solid-state reactions and transformations, and that such advances are rapidly transforming our ability to “direct matter”.

This research was conducted at the Center for Nanophase Materials Sciences, which is a US Department of Energy Office of Science User Facility.

Biography: Bobby G. Sumpter received his Bachelor of Science in Chemistry from Southwestern Oklahoma State University (1983) and a Ph.D. in Physical Chemistry from Oklahoma State University in 1987. Following postdoctoral studies in Chemical Physics at Cornell University 1987-1988 and in Polymer Chemistry at the University of Tennessee, Bobby joined the Chemistry Division at Oak Ridge National Laboratory in the Polymer Science group. He is currently the Deputy Director of the Center for Nanophase Materials Sciences, group leader for the Computational Chemical and Materials Sciences group, director of the Nanomaterials Theory Institute, and interim group leader for Macromolecular Nanomaterials at Oak Ridge National Laboratory. Dr Sumpter’s research is focused on the fundamental understanding of nanoscale self-assembly processes, interactions at interfaces, structure and dynamics of molecular-based materials, including multicomponent polymers and composites, and the physical, chemical, mechanical, and electronic properties of nanostructured materials. His research groups pursue forefront nanoscience using high-performance computing at scale alongside rational synthesis and materials characterization.

A cooperative multi-agent game for automated generations and validation of predictive elastoplastic models with AI-guided experimentation

Steve WaiChing Sun – Columbia University

Abstract: Many machine learning models used in mechanical simulations employ make predictions with little or no interpretability and demand significant amount of experimental data that could be costly to obtain. This disconnection with human knowledge makes it difficult to

assess and manage risk for engineering applications. In this talk, we present a new framework in which multiple AIs assigned to different roles (experimentalist, modeler) are participating and interacting in the scientific discovery process. In


19


this meta-modeling framework, all data accessible by the modeler AI is modeled as labeled vertices (e.g. porosity, permeability) and all possible ways to connect those vertices are modeled as directed edges (e.g. mathematical expression, neural nets, support vector machine). Therefore, all possible actions of both agents are mathematically represented by a labeled directed multi-graph. The learning process is, therefore, re-cast as the procedure to find a directed graph that links the input and output data with the optimal set of edges that maximize the reward (the objective function that measures the prediction qualities) with given constraints (material frame indifference, thermodynamics laws). Based on the performance of the modelers, the experimentalist AI then design experiments most likely enhance the blind prediction quality. This autonomous research cycle repeated sequentially until no further action can improve the predictions. With well-defined agents, reward, action space, and environment, both AI will then improve the estimation of Q values such that their decision of making actions will improve. The products generated from the cooperative game are (1) a combinatorically optimized constitutive law that represent the most plausible relations among physical quantities and (2) the minimum amount of data required to achieve the threshold performance. By leveraging the ability of the AI modeler to repetitively practice and improve its modeling skills, we demonstrate that the proposed algorithm is able to both discover new hidden hierarchical structures of mechanics knowledge and rediscover well-known mechanics knowledge without any human intervention.

Biography: Professor Sun obtained his B.S. from UC Davis (2005); M.S. in civil engineering (geomechanics) from Stanford (2007); M.A. degree from Princeton (2008); and Ph.D. in theoretical and applied mechanics from Northwestern (2011). From 2011 to 2014, he worked at the Mechanics of Materials department at Sandia (Livermore), first as postdoc, then SMTS. Sun’s research focuses on theoretical, computational and data-driven mechanics of porous media and geological materials. He is the recipients of NSF CAREER Award (2019), the EMI Leonardo da Vinci Award (2018), the Zienkiewicz Numerical Methods Engineering Prize, AFOSR Young Investigator Program Award (2017), Dresden Fellowship (2016), ARO Young Investigator Program Award (2015), and the Caterpillar Best Paper Prize (2014), among others.


20


Design and Production of Microcapsules, and What’s Next? Congwang Ye – Lawrence Livermore National Laboratory

Abstract: This talk will first showcase the use of microfluidics to achieve innovative microencapsulation of chemicals and species that would otherwise be difficult to use as-is, and then I will demonstrate an example of how machine learning has impacted the usual production procedure. Microfluidics has been utilized for many applications ranging from life science research to lab-on-a-

chip. This technique has been developed over the past decade to produce high fidelity capsules and particles with low material consumption requirements, making it an ideal method for fast screening and analysis. By controlling fluid properties and the operational parameters, we can create monodisperse core-shell capsules and matrix particles, with a broad size range from sub µm to 5 mm. At LLNL, microcapsules have been designed and produced for energy, biology, and pharmaceutical applications. The encapsulated form broadens the options for practical application, increases the chemical shelf life, and improves the performance efficiency. Machine learning, as a tool to enable automated sample detection and process control, has been applied in our capsule production to reduce the labor requirement. Seeing the benefit of applying ML, we would like to expand the use of it to many other aspects. This is an essential step towards industrial level sample development and production in order to achieve technology deployment in practical applications.

Authors: Congwang Ye*, Will Smith, Albert Chu, Du Nguyen, Brian Giera, Josh Stolaroff, and Roger Aines

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS-765840

Biography: Congwang Ye received his Ph. D. in Materials Engineering from Purdue University in 2015. He was hired as a post-doctoral research staff at Lawrence Livermore National Laboratory after graduation and was then promoted to a research & development engineer in 2018. His research focuses on using microfluidic and related technologies to produce functional microcapsules for specific applications. At LLNL, the major part of his work is the development of microfluidic-based capsule design and the subsequent scale-up production for industrial carbon capture and flue gas cleanup – a global mission to reduce CO2

emission. His work also includes using similar techniques to design and produce


21


particles/capsules for a broad range of applications including cell immobilization, catalyst reaction, self-healing composite, flexible display, and laser target design. With years of experience to work closely with companies and institutions, Dr Ye also leads the effort to commercialize these technologies to impact people’s daily lives. In order to achieve that, he’s currently exploring the use of machine learning and data science to help material discovery, process automation, and sample characterization.

Contributed Talks

Data Science for Material Science - A Database, Data-driven Modeling and Visualization Approach James Ahrens

Abstract: Success in material science rests on accurately capturing, understanding and reusing information about material properties. In this talk, I will describe an evolving data science approach that incorporates the use of relational databases, to store discrete material science information, data-driven statistical modeling, to represent this stored information in a continuous manner, and visualization, for scientists to visually understand this information. I will give an overview of each approach, how we envision the separate approaches working together. Visual demonstrations of exploring a high-dimensional material science parameter space from our LANL LDRD entitled “Real-time Adaptive Acceleration of Dynamic Experimental Science (ASSIST)” will be presented.

Innovations in Design of Experiments Christine Anderson-CookAbstract: After describing the two major alternative approaches for designed experiments based on current understanding of the process being studied, several innovations for how to approach data collection will be presented. When a response surface methodology approach toapproximate the underlying relationship between input factors and responses is suitable, Pareto front methods allow for several user-specified criteria to be simultaneously considered for the construction of an optimal design. When no model form is being assumed, a space-filling design has frequently been used. A modification of this approach is presented which allows non-uniform space-filling throughout the input factor space that is adapted to match the experimenter’s objectives.


22


Machine Learning of Interatomic Potentials Kipton Barros

Abstract: Machine learning is emerging as a powerful tool to emulate electronic structure calculations. Deep neural networks can now predict atomic interactions with accuracies exceeding density functional theory, and approaching that of coupled cluster theory, at a tiny fraction of the computational cost. I will discuss recent methods for building interatomic potentials relevant to chemistry, materials science, and biophysics applications. A key idea is active learning, in which the training dataset is generated on-the-fly, to fill in gaps of the machine learning model, and to achieve a surprising level oftransferability.

Unsupervised phase mapping of X-ray diffraction data by nonnegative factorization integrated with custom clustering Alexandrov Boian

Abstract: Analyzing x-ray diffraction (XRD) of large datasets is a key step in high-throughput mapping of the compositional phase diagrams of combinatorial materials libraries. We have developed a new unsupervised machine learning method for pattern analysis of such datasets that can be used for phase extraction. The method expands the nonnegative factor analysis combined with custom clustering and cross-correlation algorithms. It is capable of a robust determination of the number of basis patterns presented in the data, which, in turn, enables straightforward identification of any possible peak-shifted “spurious” patterns. Such peak-shifted patterns arise due to small lattice variations of the same structure and they are caused by chemical alloying and are ubiquitous in XRD datasets of combinatorial thin-film libraries. Successful exclusion of the peak-shifted patterns permits to quantify and classify the contribution of all distinct structures to each data point, which can be used to determine the compositional phase diagram of the studied system.

Bayesian emulation of the response of materials under dynamic loading Chris Biwer

Abstract: The ability to predict the response of a material enables capabilities such as sensitivity analyses that characterize the influential parameters in a model and inference with data from dynamic loading experiments. The rate-


23


limiting steps in these analyses are the significant computational time of detailed simulations and the parameterization of constitutive models of a material, such as strength and plasticity, traditionally being done by inspecting and iteratively tuning simulation results into agreement with dynamic loading experiments. We demonstrate a Bayesian analysis that can accelerate the prediction of diffraction images and provide a framework for inferring constitutive model parameters from dynamic loading experiments.

Optimized DFTB models for molecular dynamics simulations of energetic materials Marc Cawkwell

Abstract: Density functional tight binding (DFTB) theory is a parameterized, semi-empirical electronic structure method that provides an accurate, physics-based description of interatomic bonding and charge transfer in organic materials. I will discuss strategies for generating parameterizations for DFTB that provide close-to-DFT accuracy through a global optimization of the Slater-Koster bond integrals and repulsive potentials to databases of DFT calculations. The lanl31 parameter set for molecules containing C, H, N, and O, which was developed using this approach, displayed outstanding transferability to the 133k molecule QM-9 database despite being trained to only 68 small organic molecules. I will describe recent applications of lanl31 to simulations of shock-induced chemistry, equations or state, and the development of reduced-order chemical kinetics models.

Learning the mechanical and chemical response of high explosives Edward M. Kober

Abstract: High explosive crystals tend to have complex physical and chemical structures which can frustrate the analysis and formulation of reduced models for their behavior. Two methods for overcoming this will be discussed. The mechanical response is difficult to characterize because the crystal structures tend to have low symmetries, and methods developed for the analysis of metals often rely on assumed high symmetry structures. A general geometry analysis method, Strain Functional Analysis, overcomes this restriction, and enables the classification of deformation processes. An example of the plastic deformation in RDX will be shown. The chemical response is even more difficult to characterize because there are hundreds of possible reactions and intermediates. Here, a Coordination Geometry Analysis approach is used that classifies each atom in terms of its bonded neighbors. This defines a countable basis that tracks the


24


oxidation state changes of the elements and from which net reaction pathways can be inferred. Analyzing the correlations in these changes using Non-negative Matrix Factorization produces chemical wave structures that can be interpreted with Arrhenius type models. Examples of RDX and HMX will be shown.

Machine learning from physics-based spectral polycrystal plasticity models Ricardo Lebensohn and Laurent Capolungo

Abstract: Crystal plasticity (CP) models are increasingly used in engineering applications to obtain microstructuresensitive mechanical response of polycrystalline materials. Full-field CP formulations based on Fast Fourier Transforms (FFT) are an efficient alternative to CP Finite Elements, therefore well adapted for forwardmodelling of a micromechanical experiment, and for combination with Machine Learning (ML) to obtain surrogate models for data reduction/inversion/interpretation. Also, since FFT-based models operate directly on voxelized microstructural images, they are ideally suited for combination with emerging characterization methods in Experimental Mechanics that typically produce such images. After presenting the main characteristics of these physically-based FFT-based models, we will show two Computational Solid Mechanics ML applications of the latter.

The first one [1] is a combination of: 3-D in-situ synchrotron X-ray experiments done at European Synchrotron Research Facility (ESRF), LANL’s FFT-based micromechanical modelling, and Bayesian network (BN) analysis. This combined methodology enabled learning, based on a fusion of experimental and simulated data, of a new material-specific indicator to predict the propagation path and growth-rate of a fatigue crack in a bcc Ti alloy. For this, cycle-by-cycle experimental data of a crack propagating in a beta metastable Ti alloy, using phase (PCT) and diffraction contrast tomography (DCT) measured at ESRF was used as input of FFT-based simulations to supplement the data, by including the micromechanical fields ahead of the crack tip. It was first demonstrated, using supervised BN analysis, that existing fatigue indicator parameters (FIPs) proposed in the literature for other materials and loading conditions were not particularly predictive (correct less than 50% of the time). Next, unsupervised BN analysis, driven by the PCT/DCT/FFT multimodal dataset, was used for ML of a new, material-specific FIP. The spatial correlation of the identified FIP showed much better agreement (correct 60-80% of the time) with the experimentally determined crack path.


25


The second application concerns the dislocation density quantification via the use of X-ray peak profile analysis. A FFT-based mechanical solver was implemented and used to accelerate Discrete Dislocation Dynamics (DDD) simulations. The gain in computational efficiency resulting from the FFT use allowed DDD methods to rapidly generate a database of deformed single crystal microstructures. This was done in the cases of bcc Fe and fcc Al, for which more than 100 virtual dislocation microstructures were generated. Next, virtual line profile analysis, considering only kinematic diffraction, was performed for each simulated microstructure. Data analytics methods were then used to derive an analytical model relating dislocation content within a microstructure to its diffraction signature. The results obtain thus far—although limited, for the time being, to single crystal microstructures—appear to be more accurate than those obtained with standard whole line profile methods, based on analytical theories for idealized dislocation configurations.

Reference:

[1] A. Rovinelli, M.D. Sangid, H. Proudhon, Y. Guilhem, R.A. Lebensohn and W. Ludwig: “Predicting the 3-D fatigue crack growth rate of short cracks using multimodal data via Bayesian network: in-situ experiments and crystal plasticity simulations”, Journal of the Mechanics and Physics of Solids 115, pp. 208-229 (2018).

TBD Turab Lookman

The development of Hybrid Engineer/Materials Science tools for Addressing Challenges at LANL David Mascareñas

Additional Authors: Bridget Martinez, Yongchao Yang, Alex Marchi, Alessandro Cattaneo, Michelle Lockhart, Garrett Kenyon, Thomas Lienert, Charles Farrar

Abstract: The LANL Engineering Institute focuses on basic research in cross-disciplinary engineering. Our primary focus has traditionally been in the fields of structural health monitoring and non-destructive evaluation. We have specifically focused energy on the application of machine learning to structural health monitoring as well as the development of wireless sensor networks and robotics for structural inspection. However, we have increasingly been finding application for these techniques in fields such as non-proliferation, global security and smart


26


materials. As a result we have diversified our collaborations with researchers lab-wide. This presentation will provide an overview of these research efforts. Research of interest to the data science/material science community will include the development of semi-automatic video-based structural health monitoring technologies, monitoring of liquid metal melt pools for welding and additive manufacturing, video-analysis of the structural dynamics of cancer cells and beating heart cells, the combination of material science and signal processing for next-generation tamper-evident seals, and the suitability of Barkhausen noise as a unique fingerprint of ferrous components. We have also performed significant work on the use of Augmented Reality for nuclear criticality safety. We hope this overview will spark new ideas for possible collaborations between the engineering institute and other parts of the laboratory.

TBD Michael McKerns

Wavelet-Convolutional LSTM: An Efficient Deep Learning Paradigm for High Fidelity Turbulence Arvind Mohan

Abstract: High fidelity simulation of turbulence with DNS/LES presents many challenges for real-world flows due to their prohibitive computing costs. Recently, deep learning based approaches such as Convolutional LSTM (ConvLSTM) neural networks have shown considerable promise in modeling 2D spatio-temporal datasets for weather and medical imaging applications. We recently proposed a variant of this method Compressed ConvLSTM (CC-LSTM) to model 3D turbulence, which is orders of magnitude cheaper than DNS. In this work, we propose a novel paradigm called wavelet-ConvLSTM, where the complementary strengths of wavelet transforms and ConvLSTM are leveraged to learn dynamics of 3D turbulence at a much lower cost than current high dimensional deep learning approaches. The accuracy of predicted flow from the wavelet-ConvLSTM model is analyzed with physical metrics of turbulence, such as a) Energy spectra with 5/3 law comparison, b) PDF of velocity gradient and c) Normalized Q-R plane dynamics of a local Lagrangian velocity-gradient volume, coarse grained by a scale r. This allows us to study dynamics of small, inertial and large scales individually. The results show excellent agreement in large and inertial scales of turbulence, with some improvements possible in small scales by suitable wavelet thresholding. Results indicate that wavelet-ConvLSTM predicts stable, long-term temporal realizations and is an order of magnitude faster


27


compared to primitive flow variables in CC-LSTM, due to the superior entropy/information gain in wavelet bases. Its adaptive, low dimensional nature leads to comparatively scarce memory utilization for wavelet thresholding, enabling higher resolution models to be built at lower cost. Further directions would also be discussed, with the eventual goal of modeling real-world flows.

Using Machine Learning to Highlight Potential Shortcomings in Nuclear Data Denise Neudecker

Authors: D. Neudecker, P. Grechanuk, M. Grosskopf, W. Haeck, M. Herman, M. Rising, S. Vander Wiel

Abstract: Nuclear data tabulate physics reaction mechanisms for a large number of isotope and materials. They are key input quantities for application calculations in various fields, for instance, in nuclear energy, Stockpile Stewardship, global security, criticality and safety, astrophysics and nuclear medicine. The evaluated nuclear data of a particular observable (e.g., a fission cross-section) are obtained by combining statistically all experimental and theoretical information pertinent to this observable. A combination of nuclear data is validated by using them in simulations of small-scale experiments that are representative of specific applications. Usually, several thousand nuclear data observables are used to simulate one value that can then be compared to the corresponding experimental value of the small-scale experiment. One key problem of nuclear data validation is that it is impossible to asses all these complex inter-dependencies between nuclear data and simulated small-scale experiments with a human brain. This problem inhibits our ability to pin-point the specific nuclear data values that lead to a difference between simulation and experiment, and, hence, to pin-point what future nuclear data research is needed to improve the predictive capability of nuclear data for applications.

Here, we tackle this problem by applying Machine Learning techniques (Random Forest, Elastic Net) to identify the most important nuclear data related to a high difference between simulated and experimental values of criticality small-scale experiments. It will be shown that the machine learning algorithms correctly identify known issues in evaluated nuclear data. But the machine learning algorithm also highlighted problems in nuclear data previously overlooked by expert judgment validation.


28


Materials Informatics for Novel Scintillator Discovery Ghanshyam Pilania

Abstract: Applications of inorganic scintillators—activated with lanthanide dopants, such as Ce—are found in diverse fields, such as, medical imaging, radiation detection and nuclear nonproliferation. From a physics-based perspective, electronic structure of a viable scintillator must exhibit certain necessary conditions. For instance, as a strict requirement to exhibit scintillation, the 4f ground state and 5d lowest excited state levels induced by the activator must be appropriately placed within the host bandgap. This talk will discuss a new machine learning based screening strategy that relies on a high throughput prediction of the lanthanide dopants’ ground and excited state energy levels with respect to the host valance and conduction band edges for efficient chemical space explorations to discover novel inorganic scintillators. Using specific examples, it will be demonstrated that the developed approach is able to (i) capture systematic chemical trends across host chemistries and (ii) effectively screen promising compounds in a high throughput manner. While a number of other application-specific performance requirements need to be considered for a viable scintillator, the present scheme can be a practically useful tool to systematically down-select the most promising candidate materials in a first line of screening for a subsequent in-depth investigation.

Data analysis framework for enabling real-time feedback during microstructure evolution Reeju Pokharel

Abstract: Experimental techniques based on electron and high-energy X-ray diffraction for mesoscale microstructure characterization have advanced significantly in the last three decades. Large amount of data can be gathered at very high-rates enabling microstructure evolution measurements at relevant length and time scales. However, the major limitation faced by these measurement techniques is less than optimal raw data inversion algorithms that are unable to provide real-time feedback during microstructure evolution experiments. In this talk, we report on the supervised learning-based approaches developed for microstructure reconstruction from electron backscatter diffraction (EBSD) measurements. The results show that these methods can significantly speed up reconstruction process in comparison to the current state-of-the-art methods. The pre-trained models are able to produce reconstructions as fast as the practical scanning rates during data acquisition. The method developed here


29


will be extended for improving reconstruction speed for high-energy X-ray diffraction microscopy (HEDM) data.

Data-Informed Turbulence Models Juan A. Saenz

In Collaboration with: Sidharth GS, Oliver Hennigh, Susan Kurien, Gavin Portwood, Gyrya Vitaliy

Abstract: Turbulence involves complex interactions between processes at a large range of physical and temporal scales. For simplified and/or idealized turbulent flows, we are able to perform direct numerical simulations (DNS) that resolve all the important scales. However, simulation of complex turbulent flows still relies on the solution of averaged Navier-Stokes (NS) equations, which only partially, at best, resolve the range relevant scales while modeling the interactions between resolved and un-resolved scales in the form of unclosed correlations that result from non-linearities in the NS equations. Models, in this situation, mathematically relate the unclosed terms to resolved, prognostic quantities, are developed based on our understanding of idealized flows, and are thus limited by our understanding. On the other hand, availability of high resolution DNS of several canonical turbulent flows, along with progress in data science and machine learning tools have led to encouraging tools to relate data where we lack theoretical understanding to do so. We discuss ongoing efforts towards developing methodologies to combine our current theoretical understanding of turbulence with data from DNS to inform the development and enhancement of turbulence models, using recent progress in physics-informed machine learning. In particular, we describe methodologies to use data to 1) develop models for Reynolds-Averaged NS in variable density turbulence, 2) develop sub-grid scale models for large eddy simulations, and 3) to extract functional relations between statistical quantities.

Theoretical Division at Los Alamos National Laboratory: An Overview Focusing on Computational Data Science and Materials Science Jack Shlachter

Abstract: Theoretical Division at Los Alamos National Laboratory comprises a unique assemblage of physicists, chemists, biologists, mathematicians, materials scientists, and engineers. These talented researchers perform theory, modeling and simulations to solve problems with national security implications in the


30


broadest sense. In this presentation, I’ll share a little history about this Division which dates back to the Manhattan Project in 1943. There will also be some statistics about the personnel in the Division and the array of sponsors who support this research. The talk will showcase some of the recent scientific highlights from the Division and illustrate the wide-ranging topics addressed by our staff. An effort will be made to emphasize areas of importance to the computational data and materials science communities. In conclusion, I’ll suggest potential growth areas for the Division.

A Modular, Multi-Level Data Scheme for Molecular Opacity Calculations Eddy Timmermans

Authors: Eddy Timmermans (XCP-5), Cristiano Nisoli (T-4), Dima Mozyrsky (T-4), Leanne Duffy (AOT-AE) and Mark Zammit (T-1).

Abstract: At low temperatures (kBT<eV), atomic particles in a plasma combine into molecules. Low energy photons (Ephoton<2-3 eV) then interact mostly with the mechanical vibrations and rotations of these molecules. The long-lasting vibrations, with nuclei moving at speeds much slower than those of the electrons that govern the atomic transitions in the higher temperature atomic opacities, give sharply peaked spectral features in the molecular opacities (cross-section per gram) that vary on a scale as small as 10-6 eV (given here are typical numbers for pressures near ambient atmospheric conditions). In contrast, the energy-resolution of time-resolved spectroscopes, typically Einstr~10-3 eV, and the band-widths for multi-group radiative transfer calculations Egroup> 10-1 eV are much larger. The divide that separates the natural from relevant energy scales is as large as the separation between the quantum-governed atomic and continuum-modeled mesocopic scales in materials. Tables of band-averaged opacities, the radiative-transfer equivalent of coarse-graining, are useful but different averages (straight, Plank and Rosalind means) are appropriate for use in different optical regimes. Comparing different legacy tables can then be challenging and many of these tables are not well documented. In spite of the exponential increase in computational resources and a recent resurgence of interest in molecular opacities fueled by the astrophysical need for modeling exoplanet atmospheres, the efforts to fill in the ‘1 eV-gap’ in opacity data have only met with partial success. We propose a novel scheme for systematizing the bridging of the natural-to-relevant scale divide with a multi-level scheme of molecular data-handling. The proposed method combines recently developed radiative transfer methods, the modern, computer-based ease of data-transfer and the old insights


31


of statistical line models. The systematic approach that we propose allows comparing and using differently band-averaged data, as well as comparing high-resolution data for low resolution use. The method can store, check and replace partial narrow band data at any time of the calculations while allowing for filling in unknown data from tables or experiments.

Machine Learning for Molecular Properties Sergei Tretiak

Abstract: Modern theoretical chemistry relies on computer simulations typically involving numerical solution of Schrodinger equation. Such quantum mechanical (QM) methods tend to be more accurate than their classical counterparts, however, their computational scaling is frequently prohibitively expensive to treat realistic systems. Machine learning-based (ML) QM property predictors are capable of fitting directly to reference QM data with low error while remaining computationally as fast as classical techniques. We apply various ML models for QM property prediction. These techniques are trained to large QM datasets and then shown to generalize well outside of the training set. The targeted properties include interatomic potentials, various atomic charge partitioning schemes, dipoles and quadruples, infrared spectra, and reduced Hamiltonians. Our results show the applicability of these accurate ML property predictors to systems many times larger than those in the training set with a several magnitude speedup over reference QM methods, an exciting prospect for computational sciences.

Flow and Fracture in Microstructure Accelerated by Machine Learning Hari Viswanathan

Abstract: Microstructural information (fracture size, orientation, etc.) plays a key role in governing the dominant physics for two timely applications of interest to LANL: dynamic fracture processes like spall and fragmentation in metals (weapons performance) and detection of gas flow in static fractures in rock due to underground explosions (nuclear nonproliferation). Micro-fracture information is only known in a statistical sense, so representing millions of micro-fractures in 1000s of model runs to bound the uncertainty requires petabytes of information. We currently either ignore or idealize microscale information, since we lack a framework that efficiently utilizes it in its entirety to predict macroscale behavior in brittle materials. Our goal is to exploit the underlying discrete structure of fracture networks common to both applications and discover a compact graph representation that requires far fewer degrees of freedom (dof) to


32


capture micro-fracture information. Our critical IS&T advance is to integrate computational physics, machine learning and graph theory to make a paradigm shift from computationally intensive grid-based models to efficient graphs. Including the microstructural fracture physics in a parallel graph implementation will improve accuracy of predictions with at least 3 orders of magnitude speedup. The use of graph theory and machine learning in a parallel implementation to identify and isolate the key regions of a static DFN has never been attempted and has allowed us to resolve flow through DFNs for a range of scales that are currently regarded as impossible. We use dynamic graphs and machine learning to evolve fracture growth, capture crack interaction and coalescence making it possible to finally include the sub-micron scale physics of fracture propagation to predict macroscale brittle material failure. Our UQ framework will use distributions on the topological and geometric characteristics on the graph to efficiently quantify uncertainties in the original DFN.

High-Throughput Data Analysis and Machine Learning for f-Electron Materials Jian-Xin Zhu

Abstract: Actinides and lanthanides are important materials for their exotic properties like unconventional superconductivity, magnetism, heavy fermion, and anomalous structural properties. However, modeling these f-electron systems is challenging due to the complex interplay between strong electron-electron interaction of f-electrons and their hybridization with itinerant conduction electrons, and spin-orbit coupling. In this talk, I present our recent implementation of a data-driven approach to aid the materials discovery process in these materials. Following a discussion on the implemented f-electron database, the application of this database is illustrated by considering the f-electron localization-delocalization trend with the lattice constants. In addition, by constructing new datasets from high-throughput density functional theory simulations for magnetic actinide compounds, a use of state-of-the-art machine learning tools to predict magnetic moment size is also discussed.

Posters Analysis of the X-Ray Diffraction Experimental Parameter Space through Simulation Comparison using Geometric Hashing


33


Divya Banesh

Abstract: X-ray diffraction is used as a structural characterization technique, and in particular, has been used for studying the shock-response of single crystals of α-cyclotrimethylene trinitramine (𝛂𝛂-RDX). However, these experiments can contain significant uncertainty in the crystal’s orientation, since the sample is mounted, transported, and fastened into the gas gun experiment. Uncertainty in the orientation increases the difficulty indexing Bragg spots, which may prevent further analysis. Therefore, we present a robust methodology to index X-ray diffraction patterns based on geometric hashing.

Parameter Tuning and Physics Model Comparison Using Statistical Learning

Ayan Biswas

Abstract: TBD

FragData—High-fidelity Data on Dynamic Fragmentation of Materials

Nitin Daphalapurkar

Abstract: Problems involving dynamic failure, such as disruption of asteroids, protection materials under impact, and debris formation of construction materials under blast loading, all have one thing in common—these materials fail by activation and propagation of a massive number of cracks and their coalescence leads to formation of fragments. The fragment distributions (size, shape, velocity) are strongly affected by the loading history and the materials microstructure. We constructed a massive dynamic fragmentation database (FragData) for materials undergoing catastrophic failure. In the early stage of this project, the development of establishing a database and python-based data analytic tools has been accomplished using SciServer—an NSF-DIBBs funded program (http://www.sciserver.org). The fragmentation data was derived from a high-fidelity, mesoscale simulations of dynamic fracture using the finite element methods. FragData assists in elucidating the formation of fragments, evolution of damage, statistics of fragment distributions as a function of input energy and materials microstructure with application to materials undergoing dynamic damage and failure. The design of the database and the database itself, along with tools to carry out in situ analysis, serve as a central platform for other researchers to study the data from state-of-the-art simulation techniques on fragmentation.


34

http://www.sciserver.org


Network Intrusion Detection in Smart Grid for Imbalanced Attack Types Using Machine Learning Models

Dipanjan Das Roy

Abstract: Smart grid has emerged as next generation power grid which enables the transfer of real time information between the power grid and the end users via smart meters. But at the same time making it more vulnerable to different types of attacks. These vulnerabilities allow an attacker to breakdown integrity and confidentiality and deduce personal habits from the detailed consumption data. Thus, the research of anomaly-based intrusion detection within smart grids is investigated by many researchers. Problem emerges when we try to use common approaches of pattern recognition in imbalance distributed data. Which means that there are much more data instances belonging to normal behavior than to attack data, the common approaches cause a low detection rate for the minority class. Therefore, various machine learning models to overcome this drawback will be investigated by using two different datasets CSE-CIC-IDS2018 and AWID.

Towards Random Generation of Microstructures of Spatially Varying Materials from Orthogonal Sections.

Robert C. Foster

Abstract: New additive manufacturing techniques create materials with microstructures that are not easily represented by traditional simulation techniques, such as ellipsoid packing algorithms. Building upon previous techniques, a new framework for geometric simulation is introduced which allows for more flexible simulation and estimation of microstructure inf materials with spatial variation, such as additively manufactured materials. The framework recreates growth of grains, allowing for a per-grain specification of growth properties, which is unavailable to techniques that rely on optimal packing procedures. An estimation technique is also introduced that allows for estimation of the hyperparameters of the growth model based only on a partial sample of orthogonal slices.

Assessment of machine learning approaches for predicting the crystallization propensity of active pharmaceutical ingredients

Ayana Ghosh


35


Abstract: Crystallization is an important procedure for the purification and isolation of specialty chemicals. For new molecular entities, understanding their inherent tendency to crystallize is a key step for developing efficient industrial processes. To date, understanding the crystallization propensity of molecular solids has been primarily driven by empirical approaches. These studies show that crystallization may be influenced by a wide variety of processing parameters, including those of a structural, thermal, chemical and kinetic nature. Unfortunately, detailed trial-and-error studies that evaluate crystallization propensities for a diverse range of compounds have been lacking because they are expensive, time-consuming and inefficient, but more importantly failed experiments are rarely reported in the literature. Computational data-driven approaches based on a complete set of historical data coupled with machine learning methods would be invaluable to a number of industries that depend on successful crystallizations of new molecular entities. This work primarily focuses on assessing the capabilities of different machine learning algorithms and identifying important molecular descriptors to predict crystallization propensities of a set of small organic compounds (<709 Da). The influences of varied training set sizes used and experimental factors such as the solvents used, presence of impurities and/or degradants, influence of potential seeded crystallizations and implied supersaturation levels are also explicitly investigated. The best performing API only model has an RMSE of 30% whereas for the API + solvent models the RMSE is found to be 20%. Beyond inclusion of the solvent, it is found that the presence of impurities and/or degradants has the greatest influence on model accuracy. When these experiments are excluded, an additional improvement of up to 10% RMSE is observed in some cases.

Chemical Modifications of Halide Perovskites for Enhanced Optoelectronics and Photodetection

Dibyajyoti Ghosh

Abstract: Inorganic and hybrid lead halide perovskites such as CsPbX3, CH3NH3PbX3 (MAPbX3) and CH(NH2)2PbX3 (FAPbX3) (X= I, Br) have great potential as solar cell materials. Due to their high effective atomic number, large carrier diffusion lengths, low charge trap density, these materials are also promising for high energy radiation detection. In this work, combining with experimental studies, we investigate various chemical approaches to enhance the optoelectronic and photodetection capability of these hybrid perovskites. We find that controlled Cl incorporation in MAPbBr3 can tune the band gap as well as can affect the charge extraction by


36


realigning the band edges. On another work, we demonstrate that Cs coating on CsPbX3 is beneficial for photodetection as it includes low-energy absorption peaks.

Design of Materials for Energy Applications Guided by Computational Chemistry

Ivana Gonzales Matanovic

Abstract: Rational design of new cost-effective materials for clean energy demands a deeper understanding at the atomic level of the origins of relevant properties of materials and associated processes. Computational chemistry plays a key role in this effort by predicting various properties of materials for energy storage and conversion, including the activity, selectivity, and durability of the (electro-)catalysts. This work will illustrate our efforts to combine computational and experimental approaches in order to explore various aspects in the design of both platinum group metal and platinum group metal-free materials for energy applications.

The first example will include our work on the development of electrocatalysts with high yield and Faradaic efficiency for nitrogen reduction to ammonia. We will present DFT calculated kinetic volcano plots and free energy diagrams for hydrogen evolution and nitrogen reduction reactions aimed at understanding the activity and selectivity of different surfaces of molybdenum nitrides and carbides. Secondly, we will illustrate how thermodynamic data can be used to map out possible stable phases of an electrochemical system as a function of pH, potential, and temperature, which are important for determining the stability of the considered system in aqueous media.

The second example will include our work aimed at understanding of the catalyst-ionomer interface and identifying ionomer features crucial for its efficient implementation in the anion alkaline membrane fuel cells. Namely, our results shown that the benzene adsorption on the Pt catalyst is a major factor determining the low performance of the hydrogen oxidation reaction on the anode side of the alkaline membrane fuel cells. Understanding the correlation between the ionomer structure and the electrocatalysts performance can be applied to other electrochemical energy devices leading to more advanced energy technologies.

Reconstructing Grain Growth During Rapid Solidification with Spatial Statistical Methods

Nathan Johnson

Abstract: Rapid solidification of metal alloys in small volumes has proven useful for a wide range of en-gineering applications including welding, cladding, melt spinning, and


37


additive manufacturing. Large thermal gradients and fast cooling rates during rapid solidification dominate the growth behavior of grain structures in alloys. Orientation and size of grains are related to the direc-tion of thermal fluxes and the rate of solidification, respectively. Both of these microstructural characteristics have an impact on the mechanical properties of a weld join, cladded surface, additively manufactured part, etc.

The nature of grain growth can be revealed through high energy transmission X-ray diffraction (HEXRD). Droplets of titanium and stainless steel (SS) alloys from a cold metal transfer welder were analyzed for crystallographic grain information, including lattice parameter, strain, and orientation using HEXRD. Measurements were taken at 100 × 100 µm spot sizes at various locations throughout 2.5 mm diameter droplets. All measurements were taken from the moment of melting through the solidification process.

Spatial statistical modeling methods were applied – particularly Gaussian Process Modeling –in order to reconstruct crystallographic strain fields throughout full solidifying volumes. Un-certainty and standard error are also computed for every point in the reconstructed volume. The dependence of grain growth on the direction of thermal flux is revealed, as well as insights into temperature field inside solidifying volumes. Temperature fields for SS were also computed us-ing a polynomial fit of lattice strains. The reliability of temperature measurements from lattice strain measurements is discussed.

Multitasking neural networks for first-principles based statistical mechanics of alloys and magnetic systems

Massimiliano Lupo Pasini

Abstract: Computing the thermodynamics of alloys and magnetic systems via first principles calculations can become computationally expensive as the system size increases. Reducing the computational cost thus becomes essential. We tackle this problem by constructing highly accurate surrogate models to estimate physical quantities using neural networks for nonlinear regression. In particular, we perform joint training of neural networks to predict multiple physical quantities such as total energy, charge density, and magnetization simultaneously. Our numerical experiments show that this practice improves the predictive power of the regression model. In addition, there is a strong correlation between the total energy of a solid crystal structure and the other physical properties like charge density and magnetization. This strong correlation


38


acts as a physical constraint during the training process, which helps eliminate numerical artifacts like overfitting and leads to more reliable prediction

Experiment Design Frameworks for Materials Discovery

Anjana Talapatra

Abstract: Over the last decade, there has been a paradigm shift away from labor-intensive and timeconsuming materials discovery methods, and materials exploration through informatics approaches is gaining traction at present. Current approaches are however typically centered around the idea of achieving this exploration through high-throughput experimentation/computation. Such approaches, however, do not account for the practicalities of resource constraints which eventually result in bottlenecks at various stage of the workflow. Regardless of how many bottlenecks are eliminated, the fact that ultimately, a human must make decisions about what to do with the acquired information implies that HT frameworks face hard limits that will be extremely difficult to overcome. In this work, we present frameworks capable of optimally exploring the materials design space in order to attain an optimal materials response. Specifically, we use variants of the Efficient Global Optimization algorithm to deploy an autonomous computational materials discovery platform capable of performing optimal sequential computational experiments in order to find optimal materials. We demonstrate single and multi-objective optimization and we also show how this framework can be made robust against selection of non-informative features by using Bayesian Model Averaging approaches. The complete framework thus demonstrates the possibility of attaining a robust and autonomous platform for computer-driven materials discovery.

List of Participants

Speakers External Speakers Sorelle Friedler Brian Giera Olexandr Isayev Surya Kalidindi Ying Wai Li Nikolay Makarov Andrew Rappe Matthew Rever

Haverford College, PA [email protected] Lawrence Livermore National Laboratory [email protected] University of North Carolina – Chapel Hill [email protected] Georgia Tech [email protected] Oak Ridge National Laboratory [email protected] UbiQD, Inc. nikolay@ubiqd. Com University of Pennsylvania [email protected] Lawrence Livermore National Laboratory [email protected]


39

mailto:[email protected]




mailto:nikolay@ubiqd.%20Com


Antia Sanchez-Botana Arizona State University [email protected] Subramanian Sankaranarayanan

Argonne National Laboratory [email protected]

Arunima Singh Arizona State University [email protected] Bobby Sumpter Oak Ridge National Laboratory [email protected] Steve Waiching Sun Columbia University [email protected] Congwang Ye Lawrence Livermore National Laboratory [email protected]

LANL Speakers James Ahrens CCS-7: Applied Computer Science [email protected] Christine Anderson-Cook CCS-6: Statistical Sciences [email protected] Kipton Barros T-1: Physics and Chemistry of Materials [email protected] Alexandrov Boian T-1: Physics and Chemistry of Materials [email protected] Chris Biwer CCS-7: Applied Computer Science [email protected] Marc Cawkwell T-1: Physics and Chemistry of Materials [email protected] Edward Kober T-1: Physics and Chemistry of Materials [email protected] Ricardo Lebensohn T-3: Fluid Dynamics and Solid Mechanics [email protected] Turab Lookman T-4: Phys of Condensed Matter &

Complex Sys [email protected]

David Mascarenas NSEC: National Security Education Center [email protected] Michael McKerns CCS-3: Information Sciences [email protected] Arvind Mohan T-4: Phys of Condensed Matter &


Denise Neudecker XCP-5: Materials and Physical Data [email protected] Ghanshyam Pilania MST-8: Materials Science in Radiation &

Extremes [email protected]

Reeju Pokharel MST-8: Materials Science in Radiation & Extremes

[email protected]

Juan Saenz XCP-4: Methods and Algorithms [email protected] Jack Shlachter T-DO: Theoretical Division [email protected] Eddy Timmermans XCP-5: Materials and Physical Data [email protected]

Sergei Tretiak T-1: Physics and Chemistry of Materials [email protected]

Hari Viswanathan EES-16: Computational Earth Science [email protected] Jianxin Zhu T-4: Phys of Condensed Matter &



40









Poster Presenters Divya Banesh Ayan Biswas Nitin Daphalapurkar Dipanjan Das Roy Robert C. Foster Ayana Ghosh Dibyajyoti Ghosh Ivana Gonzales Matanovic Nathan Johnson Massimiliano Lupo Pasini Anjana Talapatra


41

wisdom of crowds business intelligence market study tm ... · capability development for the...

Documents