Symposium CB
Big Data and Machine Learning Methods for Materials Advancements
ABSTRACTS
CB-1:KL1 Network Theory Meets Materials Science
C. WOLVERTON1, V. Hegde1, M. Aykol2, 1Northwestern University, Department of Materials Science and Eng., Evanston, IL USA; 2Toyota Research Institute, Los Altos, CA USA
One of the holy grails of materials science, unlocking structure-property relationships, has largely been pursued via bottom-up investigations of how the arrangement of atoms and interatomic bonding in a material determine its macroscopic behavior. Here we consider a complementary approach, a top-down study of the organizational structure of networks of materials, based on the interaction between materials themselves. We demonstrate the utility of applying network theory to materials science by analysis of the complete “phase stability network of all inorganic materials” as a densely-connected complex network of 21,000 thermodynamically stable compounds (nodes) interlinked by 41 million tie-lines (edges) defining their two-phase equilibria, as computed by high-throughput density functional theory.
CB-1:IL01 Autonomous Combinatorial Experimentation
ICHIRO TAKEUCHI, University of Maryland, College Park, MD, USA
As a branch of machine learning, active learning has attracted much attention recently. It can effectively help navigate experimental sequences in materials research. We are actively incorporating active learning in screening of combinatorial libraries of functional materials. The array format with which samples with different compositions are laid out on combinatorial libraries is particularly conducive to active learning driven autonomous experimentation. For some physical properties, each characterization/measurement requires time/resources long/large enough that "high"-throughput measurement is not possible. Examples include detection of martensitic transformation and superconducting transitions in thin film libraries. By incorporating active learning into the protocol of combinatorial characterization, we can streamline the measurement and the analysis process substantially. We will present examples of mapping of phase change memory materials as well as others functional materials.
This work is performed in collaboration with A. Gilad Kusne, V. Stanev, H. Yu, and A. Mehta. This work is funded by SRC, ONR and AFOSR.
CB-1:IL04 Understanding General Grain Boundaries in 8+ Dimensions for Big Data Enhanced Materials Design
JIAN LUO, University of California San Diego, La Jolla, CA, USA
Understanding and controlling grain boundaries (GBs) in 8+ dimensions, including 5 GB macroscopic degrees of freedom (DOFs), 2+ thermodynamic DOFs (temperature and composition) and external (e.g., electric) fields, is a critical component for big data enhanced materials design. Here, our special focus is on the important, but less understood, general GBs. This talk will first review a series of our studies to compute GB counterparts to bulk phase diagrams as a useful tool for data-enabled materials design. A recent collaborative study used genetic algorithm-guided deep learning to predict GB properties as a function of five macroscopic DOFs plus temperature and composition in a 7D space [Materials Today 2020]. Highly asymmetric interfacial superstructures were discovered and modeled in WC based materials [Materials Horizons 2020; Science Advances 2021]. An electrochemically induced GB transition discovered in ZnO based ceramics showed yet another dimension to alter the GB structure and microstructural evolution with an applied electric field [Nature Communications 2021]. Here, we showed in multiple cases that successes can be achieved via combining thermodynamic models, DFT, atomistic simulations, AIMD, machine learning, advanced microscopy, and controlled experimentation.
CB-1:IL06 Active Learning of Bayesian Force Fields for Fast Molecular Dynamics Simulations of Rare Events
B. KOZINSKY, Harvard University, Cambridge, MA, USA
High-fidelity ab-initio simulations of atomistic dynamics are limited to small systems and short times, and development of surrogate machine learning models for force fields is an emerging promising direction to access long-time large-scale dynamics of complex materials systems. However, the main challenges are high accuracy, reliability, and computational efficiency of these models, which critically depend on the training data sets. We develop ML interatomic potential models that are interpretable and uncertainty-aware, and orders of magnitude faster than reference quantum methods. Principled uncertainty quantification built into these models enables the construction of autonomous data acquisition schemes using active learning. We demonstrate on-the-fly learning of machine learning force fields and use them to gain insights into previously inaccessible physical and chemical phenomena in ion conductors, catalytic surface reactions, 2D materials phase transformations, and shape memory alloys [1,2,3].
1. J. Vandermause, et al, NPJ Computational Materials, 6, 20 (2020)
2. J. S. Lim, et al, JACS. 2020, 142, 37, 15907–15916 (2020)
3. Y. Xie et al, https://arxiv.org/abs/2008.11796
CB-1:IL08 AiiDA and Materials Cloud: Coupling High-throughput Computational Automation with Data Management for the Creation of Materials Properties Databases
G. PIZZI, M. BERCX, EPFL and NCCR MARVEL, Lausanne, Switzerland
In recent years, "materials by design" has become a very powerful approach, but it requires running large numbers of simulations on supercomputers to build databases of computed properties. Key challenges are: automatically prepare, execute and monitor workflows; guarantee reproducibility; facilitate data querying and analysis; and share the results. In this talk, I describe our Open Science Platform developed by the Swiss MARVEL NCCR and the European MaX Centre of Excellence, based on: 1) Widely-used, community-based open-source simulation engines. 2) The AiiDA[1] materials informatics infrastructure, to manage, persist and reproduce computational workflows by automatically tracking the full provenance of data and calculations. 3) The Materials Cloud[2] web platform, to enable sharing and dissemination of research data with full provenance, and to run simulations on the cloud. I will describe our infrastructure and show concrete examples of how it enables the creation of materials databases and the discovery of novel materials and their properties, focusing also on the understanding of the microscopic nature of the paraelectric-ferroelectric transition in oxide perovskites.
[1] S.P. Huber et al., Sci. Data 7, 300 (2020) [2] L. Talirz et al., Sci. Data 7, 299 (2020)
CB-1:L09 Smart Data Analysis for Machining of Ceramic Matrix Composites
R. GOLLER, P. Leon-Perez, University of Applied Sciences, Augsburg, Germany
The use of process data become more interesting for industrial production and research. Machining of Ceramic Matrix Composites (CMCs) is known as a critical process step because damage can destroy a high value part. To evaluate process data during machining operation like force, power or vibration can help to detect non-conform or out of spec events. In the presented work machining data from CMC machining operations are first streamed from the process, filtered and analyzed. In this way tool wear and other changes in process can be detected. Correlations between operation parameters and surface quality can be observed.
CB-2:KL1 Monte-Carlo Random Walks for Solving Equations in Large Images. Application to Simulations of CMC Processing and Degradation
G.L. VIGNOLES, C. Charles, C. Heisel, C. Descamps, University of Bordeaux, CNRS, CEA, Safran : LCTS , Pessac, France
Ceramic-Matrix Composites (CMCs) are materials with complex architectures for which the assessment of an effective behavior, as far as heat and mass transfer are concerned, is a sometimes-difficult task. This occurs e.g. when trying to design and optimize processing methods, like e.g. Chemical Vapor Infiltration, or to interpret and understand the material behavior under ablation. Image-based modeling arises as a powerful tool for the prediction of morphological evolution of a CMC or a Carbon/Carbon (C/C) composite subjected to gas-solid reactions (either deposition or etching), possibly in presence of strong temperature gradients; however, the size of these images can be extremely large, turning numerical simulation cumbersome. Monte-Carlo Random Walks (MCRW) are an attractive method to address these questions, because they are intrinsically economical in terms of required computer memory and capable of massive parallelization. We will describe how equations describing mass and heat transfer coupled to morphological evolution can be solved by MCRW methods, focusing on the research of low memory requirements. Examples of Chemical Vapor Infiltration and of ablation simulations in massive images obtained by X-ray tomographic imaging will be presented and discussed.
CB-2:KL2 Theoretical and Machine Learning Studies of Grain Boundary Solute Segregation
F. Abdeljawad, Department of Mechanical Engineering, Department of Materials Science and Engineering, Clemson University, Clemson, SC, USA
Grain boundary (GB) solute segregation has been experimentally shown to mitigate grain growth and thermally stabilize the grain structures of nanocrystalline materials. However, most studies are focused on the thermodynamic aspect of GB segregation, and the role of dynamic solute drag remains poorly understood. Based on a recently developed sharp-interface model of GB segregation, we present machine learning studies that quantify GB solute drag and identify regimes where kinetic stabilization is dominant. The dynamical equation governing GB segregation is used to generate datasets to train and validate a neural network model that predicts the solute drag hypersurface—solute drag is a function of several parameters describing bulk and interface thermodynamics, interfacial transport coefficients, and segregation asymmetry. Analysis is presented using the neural network model to demonstrate the complex trends in solute drag as a function of alloy thermodynamics, temperature, and solute concentration. In broad terms, our proposed model provides a design tool to rapidly screen for new materials with enhanced microstructural stability under extreme environments.
CB-2:IL05 Data-driven Spectral Analysis for Materials Characterization
TERUYASU MIZOGUCHI, The University of Tokyo, Tokyo, Japan
Data driven approaches are now indispensable for modern materials characterization due to rapid increase of size and dimension of data observed in experiments and simulations. Based on this backgrounds, we are developing data-driven methods for the materials characterization. The present authors have reported the applications of machine learning for crystal interface analysis and XAFS/EELS spectrum. In the crystalline interface studies, the ML approaches, including virtual screening, kriging, and transfer learning, was used to accelerate structure determination of the interface. In this presentation, we are going to present about topics machine learning for the XAFS/EELS. XAFS/EELS is core-loss spectroscopy observed using X-ray/electron, and the spectral features around near edge are called XANES/ELNES which reflect partial density of states of conduction band. We applied a neural network to predict the excited states of the XANES/ELNES from their ground states. Consequently, our model correctly learned and predicted the excited states from their ground states, providing several thousand times computational efficiency. The machine learning was also applied to directly predict the materials structure and functions. The details will be presented in the presentation.
CB-2:IL09 Machine Learning Analysis of Multiphase Magnetic Microstructures
A. Kornell, A. Kovacs, M. Gusenbauer, T. Schrefl, Danube University Krems, Wiener Neustadt, Austria
Sustainable energy production and environmentally friendly transport require high-performance soft and hard magnetic materials. In addition to the intrinsic properties of magnetic materials, the microstructure of the magnet has a decisive influence on its hysteresis properties. We apply machine learning methods to understand the influence of microstructure on the resistance of permanent magnets to external fields. We identify the microstructural properties of weak grains that trigger the magnetization reversal in high-power NdFeB magnets. As a direct result of our findings machine learning analysis we show that edge hardening by Dy-Diffusion leads to higher coercivity fields. We also show how machine learning could be used to predict the demagnetization from microstructure images with data-driven machine learning. We combine model order reduction and neural network regression for estimating the hysteresis properties of nanocrystalline permanent magnets. The structure of the magnet is encoded into a low dimensional latent code which serves is input to a multi-target neural network regressor for the hysteresis properties. The financial support by the Austrian Federal Ministry for Digital and Economic Affairs, and the Christian Doppler Research Association is gratefully acknowledged.