Bulletin of the American Physical Society
APS March Meeting 2015
Volume 60, Number 1
Monday–Friday, March 2–6, 2015; San Antonio, Texas
Session D16: Focus Session: Machine Learning For Materials Discovery |
Hide Abstracts |
Sponsoring Units: DMP Chair: Marco Nardelli, University of North Texas Room: 101AB |
Monday, March 2, 2015 2:30PM - 2:42PM |
D16.00001: Machine Learning methods in fitting first-principles total energies for substitutionally disordered solid Qin Gao, Sanxi Yao, Michael Widom Density functional theory (DFT) provides an accurate and first-principles description of solid structures and total energies. However, it is highly time-consuming to calculate structures with hundreds of atoms in the unit cell and almost not possible to calculate thousands of atoms. We apply and adapt machine learning algorithms, including compressive sensing, support vector regression and artificial neural networks to fit the DFT total energies of substitutionally disordered boron carbide. The nonparametric kernel method is also included in our models. Our fitted total energy model reproduces the DFT energies with prediction error of around 1 meV/atom. The assumptions of these machine learning models and applications of the fitted total energies will also be discussed. [Preview Abstract] |
Monday, March 2, 2015 2:42PM - 2:54PM |
D16.00002: Phase Transitions of Boron Carbide: Pair Interaction Model of High Carbon Limit Sanxi Yao, Michael Widom, William Huhn, Qin Gao Boron carbide is a structure that exhibits a broad composition range, implying a degree of intrinsic substitutional disorder. While the observed symmetry is rhombohedral, the enthalpy minimizing structure has lower, monoclinic, symmetry. With high melting temperature, it is difficult to experimentally study its phase transition at low temperature and there is discrepancy among different research groups. Moreover, the widely-accepted phase diagram suggests substitutional disorder at low temperature, implying a non vanishing entropy. Here we use computational method to study its phase transition. We implement a pair interaction model and fit to a database of structural energies. Utilizing histogram methods to analyze Monte Carlo simulations of this model, we investigate the symmetry-restoring phase transition that explains the observed rhombohedral symmetry at high temperatures. [Preview Abstract] |
Monday, March 2, 2015 2:54PM - 3:06PM |
D16.00003: Unsupervised machine learning on atomistic configurations: examples on amorphous defects and energy landscapes Ekin Cubuk, Samuel Schoenholz, Andrea Liu, Efthimios Kaxiras Due to the recent availability of very large datasets, machine learning (ML) methods are gaining popularity as approximation and optimization tools in solid state physics. We have recently shown that supervised ML can also be used to identify and analyze soft particles, particles susceptible to rearrangement, in amorphous solids [1]. Our method can be used to understand what makes certain configurations of particles more prone to rearrangement, and design stronger materials. We use unsupervised ML and nonlinear dimensionality reduction methods, where we do not need a ``training set'' to train the algorithm, to explore better representations of atomic configurations. These representations are shown to provide important physical insights into the structure of soft spots and stable regions in several computational and experimental glassy systems, as well as the energy landscapes of quantum mechanical systems based on Density Functional Theory calculations. By discovering an improved representation and visualization of relevant energy landscapes, discovery and optimization efforts can be simplified.\\[4pt] [1] arXiv: 1409.6820 [Preview Abstract] |
Monday, March 2, 2015 3:06PM - 3:42PM |
D16.00004: Machine Learning Methods for the Sampling of Chemical Space From First Principles Invited Speaker: Raghu Ramakrishnan Computational brute force high-throughput screening of compounds is beyond any capacity for all but the most restricted systems due to the combinatorial nature of chemical space, i.e. all the compositional, constitutional, and conformational isomers. Efficient computational materials design algorithms must therefore make good trade-offs between the accuracy of the applied model and computational speed. Overall, rapid convergence in terms of number of compounds visited is highly desirable. In this talk, I will describe recent contributions in this field based on statistical approaches that can serve as inexpensive surrogate models to reduce the computational load of quantum mechanical calculations. Such surrogate machine learning (ML) models infer quantum mechanical observables of novel materials, rather than solving approximate variants of Schroedinger's equation. We developed accurate ML models for the rapid prediction of atomization energies and enthalpies, cohesive energies, and electronic properties that conventionally can only be predicted using quantum mechanics. All our ML models have been trained using large data bases containing properties of thousands of chemical compounds and materials. I will exemplify our approach for the prediction of properties from scratch for out-of-sample compounds. These predictions reach quantum chemical accuracy and are basically instantaneous, i.e. at a computational cost reduced by several orders of magnitude. [Preview Abstract] |
Monday, March 2, 2015 3:42PM - 3:54PM |
D16.00005: Informatics guided Search for Magnetic Apatites Prasanna V. Balachandran, Turab Lookman Materials with apatite crystal structure have applications ranging from biomaterials to electrolytes for solid oxide fuel cells. Their chemical flexibility and structural diversity provide a fertile ground to tune functionalities as potential candidates for many applications. However, magnetic apatites are rare. In this work, we use machine learning methods to rapidly screen a vast chemical space and identify novel apatite compositions with magnetic ions. We first construct a database of known materials from surveying the experimental literature. We then augment the database with features that capture the trends in geometry and bonding characteristics of apatites. Supervised classification learning form the basis of our machine learning approach through which we uncover design rules that enable prediction of potentially stable magnetic apatite compositions, prior to experimental synthesis. Finally, we validate our predictions using density functional theory calculations. [Preview Abstract] |
Monday, March 2, 2015 3:54PM - 4:06PM |
D16.00006: Using Data Mining Algorithms in Solid State Physics Troy Lyons, Nicholas Mecholsky, Stefano Curtarolo, Marco Buongiorno Nardelli, Marco Fornari We processed large materials databases with data mining methods such as clustering and classification in order to answer specific questions in the field of thermoelectric materials and transparent conductors. Our goal is to extract meaningful information from band structures repositories such as AFLOWLIB. Our implementation is validated using a toy database that mimics the complexity of AFLOWLIB, which has also been solved analytically. We found that even when the analytical solution is known, proper data analysis can help to understand physical phenomena. [Preview Abstract] |
Monday, March 2, 2015 4:06PM - 4:18PM |
D16.00007: Interplay of Atomic and Electronic Structure in Second Harmonic Generating Nonlinear Optical Materials Antonio Cammarata, James Rondinelli Group theoretical methods and \emph{ab initio} electronic structure calculations are combined to formulate a general Symmetry-Assisted Functional Optical Response (SAFOR) protocol to understand and predict the second harmonic generation (SHG) response in nonlinear optical crystals. We show that the SHG coefficients may be decomposed into atomic contributions from various inversion symmetry lifting distortions, which we parametrize as symmetry-adapted displacement patterns that transform as irreducible representations of a relevant centrosymmetric parent structure. The SAFOR protocol is then combined with an electronic descriptor for bond covalency to explain the origin of SHG in noncentrosymmetric-nonpolar $A$TeMoO$_6$ telluromolybdate compounds. We show that the SHG response has a complex dependence on the asymmetric geometry of the polyhedral units and the orbital character at the valence band edge. The atomic scale and electronic structure understanding of the macroscopic SHG behavior obtained with these descriptions is then used to identify hypothetical HgTeMoO$_6$ as a candidate telluromolybdate, which we predict should exhibit the largest SHG response in the $A$TeMoO$_6$ family. [Preview Abstract] |
Monday, March 2, 2015 4:18PM - 4:30PM |
D16.00008: Materials Cartography: Representing and Mining Material Space Using Structural and Electronic Fingerprints Corey Oses, Olexandr Isayev, Denis Fourches, Eugene Muratov, Kevin Rasch, Alexander Tropsha, Stefano Curtarolo As the proliferation of high-throughput approaches in materials science is increasing the wealth of data in the field, the gap between accumulated-information and derived-knowledge widens. We address the issue of scientific discovery in materials databases by introducing novel analytical approaches based on structural and electronic materials fingerprints. The framework is employed to (i) query large databases of materials using similarity concepts, (ii) map the connectivity of the materials space (i.e., as a materials cartogram) for rapidly identifying regions with unique organizations/properties, and (iii) develop predictive Quantitative Materials Structure-Property Relationships (QMSPR) models for guiding materials design. In this study, we test these fingerprints by seeking target material properties. As a quantitative example, we model the critical temperatures of known superconductors. Our novel materials fingerprinting and materials cartography approaches contribute to the emerging field of materials informatics by enabling effective computational tools to analyze, visualize, model, and design new materials. [Preview Abstract] |
Monday, March 2, 2015 4:30PM - 4:42PM |
D16.00009: Big Data of Materials Science -- Critical Role of the Descriptor Luca M. Ghiringhelli, Jan Vybiral, Sergey V. Levchenko, Claudia Draxl, Matthias Scheffler Statistical learning of materials properties or functions so far starts with a largely silent, non-challenged step: the introduction of a multidimensional descriptor. However, when the scientific relationship of the descriptor to the actuating mechanisms is unclear, causality of the trained (learned) descriptor-property relation is uncertain. Thus, scientific advancement, trustful prediction of new promising materials and identification of anomalies is doubtful. We discuss and analyse this issue and define requirements for a descriptor that is suited for statistical learning of materials properties and functions. We show how a meaningful descriptor can be found systematically, by means of compressed sensing techniques. These concepts are demonstrated for examples in materials science: prediction of the relative stability of zincblende/wurtzite vs rocksalt octet binary semiconductors, and prediction of their band gaps, by using simple atomic input for building the descriptor. [Preview Abstract] |
Monday, March 2, 2015 4:42PM - 4:54PM |
D16.00010: Structure classification of AB solids via machine learning J.E. Guberntis, G. Pilania, T. Lookman We explored the use of machine learning methods, specifically support vector machines and various forms of cross-validation, for the task of classifying the crystal structures of the octet AB solids. We partitioned a set of 75 solids into rocksalt and non-rocksalt structures and thus performed a binary classification task. We found that using the standard indices $(r_\sigma, r_\pi)$, suggested by St.\ John and Bloch several decades ago, enabled an average success in classification of $92\%$. Our main new result is our finding that using just $r_\sigma$ and the excess Born effective charge $\Delta Z_A$ of the A atom,computed by DFT, enabled an average success of $98\%$, prompting us to propose $(r_\sigma, \Delta Z_A)$ as a replacement for the St.\ John-Bloch pair. In general, we found that adding one or two other features to the St.\ John-Bloch pair, unless they include the excess Born effective charge, generally decreases the average success rate. [Preview Abstract] |
Monday, March 2, 2015 4:54PM - 5:06PM |
D16.00011: Study on Electro-polymerization Nano-micro Wiring System Imitating Axonal Growth of Artificial Neurons towards Machine Learning Nguyen Tuan Dang, Megumi Akai-kasada, Tetsuya Asai, Akira Saito, Yuji Kuwahara Machine learning using the artificial neuron network research is supposed to be the best way to understand how the human brain trains itself to process information. In this study, we have successfully developed the programs using supervised machine learning algorithm. However, these supervised learning processes for the neuron network required the very strong computing configuration. Derivation from the necessity of increasing in computing ability and in reduction of power consumption, accelerator circuits become critical. To develop such accelerator circuits using supervised machine learning algorithm, conducting polymer micro/nanowires growing process was realized and applied as a synaptic weigh controller. In this work, high conductivity Polypyrrole (PPy) and Poly (3, 4 - ethylenedioxythiophene) PEDOT wires were potentiostatically grown crosslinking the designated electrodes, which were prefabricated by lithography, when appropriate square wave AC voltage and appropriate frequency were applied. Micro/nanowire growing process emulated the neurotransmitter release process of synapses inside a biological neuron and wire's resistance variation during the growing process was preferred to as the variation of synaptic weigh in machine learning algorithm. [Preview Abstract] |
Monday, March 2, 2015 5:06PM - 5:18PM |
D16.00012: Force field development from first principles for materials design Maria Chan, Alper Kinaci, Badri Narayanan, Fatih Sen, Stephen Gray, Michael Davis, Subramanian Sankaranaryanan The ability to perform accurate calculations efficiently is crucial for computational materials design. In this talk, we will discuss a stream-lined approach to force field development using first principles density functional theory training data and machine learning algorithms. We will also discuss the validation of this approach on precious metal nanoparticles. [Preview Abstract] |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700