Bulletin of the American Physical Society
APS March Meeting 2016
Volume 61, Number 2
Monday–Friday, March 14–18, 2016; Baltimore, Maryland
Session E22: Predicting and Classifying Materials via High-Throughput Databases and Machine Learning IFocus
|
Hide Abstracts |
Sponsoring Units: DCOMP Chair: Gus Hart, Brigham Young University Room: 321 |
Tuesday, March 15, 2016 8:00AM - 8:12AM |
E22.00001: Novel energy materials through structural search Maximilian Amsler, Stefan Goedecker, Chris Wolverton Sophisticated structure prediction methods have been developed and become essential tools when designing new materials with desired properties. Their successful applications to many systems at various conditions and the increasing amount of available computational power have strongly contributed to their popularity. \\ The Minima Hopping Method (MHM) is a powerful tool to find low energy structures given only the chemical composition of a system and allows the prediction of structures at any boundary condition. Recently, not only the thermodynamic ground states, but also metastable phases accessible through various synthesis methods have drawn considerable interest for energy applications. We present the discovery of novel energy materials, ranging from low-density silicon allotropes with improved absorption in the visible to thermoelectric materials, by optimizing the MHM to imitate synthesis pathways. [Preview Abstract] |
Tuesday, March 15, 2016 8:12AM - 8:24AM |
E22.00002: High-throughput Screening and Statistical Learning for the Design of Transparent Conducting Oxides Christopher Sutton, Luca Ghiringhelli, Matthias Scheffler Transparent conducting oxides (TCOs) represent a class of well-developed and commercialized wide-bandgap semiconductors that are crucial for many electronic devices. Al, Ga, and In-based sesquioxides are investigated as new TCOs motivated by very intriguing recent experimental work that has demonstrated bandgap engineering in ternary (AlxGayIn1-x-y)2O3 ranging from 3.8 eV to 7.5 eV by adjusting the ratio of In/Ga[1] and Ga/Al.[2] We employed DFT-based cluster expansion (CE) models combined with fast stochastic optimization techniques (e.g., Wang-Landau and diffusive nested sampling) in order to efficiently search for stable and metastable configurations of (AlxGayIn1-x-y)2O3 at various lattice structures. The approach also allows for a consideration of the effect of entropy on the relative stability of ternary TCOs. Statistical learning/compressed sensing is being used to efficiently identify a structure-property relationship between the targeted properties (e.g., mobilities and optical transparency) and the fundamental chemical and physical parameters that control these properties. $\backslash $[1] Zhang et al., Solid State Commun, 186, 28 (2014). $\backslash $pard[2] Ito et al., Jpn. J. Appl. Phys., 51, 100207 (2012); Zhang et al., Appl. Phys. Lett., 105, 162107 (2014). [Preview Abstract] |
Tuesday, March 15, 2016 8:24AM - 8:36AM |
E22.00003: Many-body physics via machine learning Louis-Francois Arsenault, O. Anatole von Lilienfeld, Andrew J. Millis We demonstrate a method for the use of machine learning (ML) to solve the equations of many-body physics, which are functional equations linking a bare to an interacting Green’s function (or self-energy) offering transferable power of prediction for physical quantities for both the forward and the reverse engineering problem of materials. Functions are represented by coefficients in an orthogonal polynomial expansion and kernel ridge regression is used. The method is demonstrated using as an example a database built from Dynamical Mean Field theory (DMFT) calculations on the three dimensional Hubbard model. We discuss the extension to a database for real materials. We also discuss some new area of investigation concerning high throughput predictions for real materials by offering a perspective of how our scheme is general enough for applications to other problems involving the inversion of integral equations from the integrated knowledge such as the analytical continuation of the Green’s function and the reconstruction of lattice structures from X-ray spectra. [Preview Abstract] |
Tuesday, March 15, 2016 8:36AM - 9:12AM |
E22.00004: Deep Wavelet Scattering for Quantum Energy Regression Invited Speaker: Matthew Hirn Physical functionals are usually computed as solutions of variational problems or from solutions of partial differential equations, which may require huge computations for complex systems. Quantum chemistry calculations of ground state molecular energies is such an example. Indeed, if $x$ is a quantum molecular state, then the ground state energy $E_0(x)$ is the minimum eigenvalue solution of the time independent Schr\"{o}dinger Equation, which is computationally intensive for large systems. Machine learning algorithms do not simulate the physical system but estimate solutions by interpolating values provided by a training set of known examples $\{ (x_i, E_0(x_i) \}_{i \leq n}$. However, precise interpolations may require a number of examples that is exponential in the system dimension, and are thus intractable. This curse of dimensionality may be circumvented by computing interpolations in smaller approximation spaces, which take advantage of physical invariants. Linear regressions of $E_0$ over a dictionary $\Phi = \{ \phi_k \}_k$ compute an approximation $\widetilde{E}_0$ as: $ \widetilde{E}_0 (x) = \sum_k w_k \phi_k (x), $ where the weights $\{ w_k \}_k$ are selected to minimize the error between $E_0$ and $\widetilde{E}_0$ on the training set. The key to such a regression approach then lies in the design of the dictionary $\Phi$. It must be intricate enough to capture the essential variability of $E_0(x)$ over the molecular states $x$ of interest, while simple enough so that evaluation of $\Phi (x)$ is significantly less intensive than a direct quantum mechanical computation (or approximation) of $E_0 (x)$. In this talk we present a novel dictionary $\Phi$ for the regression of quantum mechanical energies based on the \textit{scattering transform} of an intermediate, approximate electron density representation $\rho_x$ of the state $x$. The scattering transform has the architecture of a deep convolutional network, composed of an alternating sequence of linear filters and nonlinear maps. Whereas in many deep learning tasks the linear filters are learned from the training data, here the physical properties of $E_0$ (invariance to isometric transformations of the state $x$, stable to deformations of $x$) are leveraged to design a collection of linear filters $\rho_x \ast \psi_{\lambda}$ for an appropriate wavelet $\psi$. These linear filters are composed with the nonlinear modulus operator, and the process is iterated upon so that at each layer stable, invariant features are extracted: $ \phi_k(x) = \| || \rho_x \ast \psi_{\lambda_1} | \ast \psi_{\lambda_2} | \ast \cdots \ast \psi_{\lambda_m} \|, \quad k = (\lambda_1, \ldots, \lambda_m), \enspace m = 1, 2, \ldots $ The scattering transform thus encodes not only interactions at multiple scales (in the first layer, $m = 1$), but also features that encode complex phenomena resulting from a cascade of interactions across scales (in subsequent layers, $m \geq 2$). Numerical experiments give state of the art accuracy over data bases of organic molecules, while theoretical results guarantee performance for the component of the ground state energy resulting from Coulombic interactions. [Preview Abstract] |
Tuesday, March 15, 2016 9:12AM - 9:24AM |
E22.00005: Machine learning bandgaps of double perovskites Ghanshyam Pilania, Arun Mannodi-Kanakkithodi, Blas Uberuaga, Rampi Ramprasad, James Gubernatis, Turab Lookman The ability to make rapid and accurate predictions of bandgaps for double perovskites is of much practical interest for a range of applications. While quantum mechanical computations for high-fidelity bandgaps are enormously computation-time intensive and thus impractical in high throughput studies, informatics-based statistical learning approaches can be a promising alternative. Here we demonstrate a systematic feature-engineering approach and a robust learning framework for efficient and accurate predictions of electronic bandgaps for double perovskites. After evaluating a set of nearly 1.2 million features, we identify several elemental features of the constituent atomic species as the most crucial and relevant predictors. The developed models are validated and tested using the best practices of data science (on a dataset of more than 1300 double perovskite bandgaps) and further analyzed to rationalize their prediction performance. [Preview Abstract] |
Tuesday, March 15, 2016 9:24AM - 9:36AM |
E22.00006: A Multi-Objective Optimization Technique to Model the Pareto Front of Organic Dielectric Polymers J. E. Gubernatis, A. Mannodi-Kanakkithodi, R. Ramprasad, G. Pilania, T. Lookman Multi-objective optimization is an area of decision making that is concerned with mathematical optimization problems involving more than one objective simultaneously. Here we describe two new Monte Carlo methods for this type of optimization in the context of their application to the problem of designing polymers with more desirable dielectric and optical properties. We present results of applying these Monte Carlo methods to a two-objective problem (maximizing the total static band dielectric constant and energy gap) and a three objective problem (maximizing the ionic and electronic contributions to the static band dielectric constant and energy gap) of a 6-block organic polymer. Our objective functions were constructed from high throughput DFT calculations of 4-block polymers, following the method of Sharma et al., Nature Communications 5, 4845 (2014) and Mannodi-Kanakkithodi et al., Scientific Reports, submitted. Our high throughput and Monte Carlo methods of analysis extend to general N-block organic polymers. [Preview Abstract] |
Tuesday, March 15, 2016 9:36AM - 9:48AM |
E22.00007: A New Silicon Allotrope with a Direct Band Gap for Optoelectronic Applications Yaguang Guo, Qian Wang, Yoshiyuki Kawazoe, Puru Jena Silicon structures with direct band gaps have been hotly pursued for solar cell applications. To effectively harvest the sunlight in the whole frequency region, it is a good strategy to use arrays consisting of Si structures with different direct band gaps. However, the structure with a direct band gap about 0.6 eV has been missing according to current progress made in the direction. Here we report our findings that the missing structure can be constructed by using Si triangles as the building blocks, which is stable dynamically and thermally, not only exhibiting the desirable band gap, but also showing high intrinsic mobility and low mass density. These advantages over the existing Si structures would motivate new experimental effort in this direction. [Preview Abstract] |
Tuesday, March 15, 2016 9:48AM - 10:00AM |
E22.00008: A comprehensive polymer dataset for accelerated property prediction and design Huan Tran, Arun Kumar Mannodi-Kanakkithodi, Chiho Kim, Vinit Sharma, Ghanshyam Oilania, Rampi Ramprasad Emerging computation- and data-driven approaches are particularly useful for rationally designing materials with targeted properties. In principle, these approaches rely on identifying structure-property relationships by learning from a dataset of sufficiently large number of relevant materials. The learned information can then be used to rapidly predict the properties of materials not already in the dataset, thus accelerating the design of materials with preferable properties. Here, we report the development of a dataset of 1,065 polymers and related materials, which is available at http://khazana.uconn.edu/. This dataset is uniformly prepared using first-principles calculations with structures obtained either from other sources or by using structure search methods. Because the immediate target of this work is to assist the design of high dielectric constant polymers, it is initially designed to include the optimized structures, atomization energies, band gaps, and dielectric constants. The dataset will be progressively expanded by accumulating new materials and including additional properties calculated for the optimized structures provided. We discuss some information ``learned`` from the dataset and suggest that it may be used as the playground for further data-mining work. [Preview Abstract] |
Tuesday, March 15, 2016 10:00AM - 10:12AM |
E22.00009: Application of Machine Learning tools to recognition of molecular patterns in STM images Artem Maksov, Maxim Ziatdinov, Shintaro Fujii, Manabu Kiguchi, Shuhei Higashibayashi, Hidehiro Sakurai, Sergei Kalinin, Bobby Sumpter The ability to utilize individual molecules and molecular assemblies as data storage elements has motivated scientist for years, concurrent with the continuous effort to shrink a size of data storage devices in microelectronics industry. One of the critical issues in this effort lies in being able to identify individual molecular assembly units (patterns), on a large scale in an automated fashion of complete information extraction. Here we present a novel method of applying machine learning techniques for extraction of positional and rotational information from scanning tunneling microscopy (STM) images of $\pi $-bowl sumanene molecules on gold. We use Markov Random Field (MRF) model to decode the polar rotational states for each molecule in a large scale STM image of molecular film. We further develop an algorithm that uses a convolutional Neural Network combined with MRF and input from density functional theory to classify molecules into different azimuthal rotational classes. Our results demonstrate that a molecular film is partitioned into distinctive azimuthal rotational domains consisting typically of 20-30 molecules. In each domain, the ``bowl-down'' molecules are generally surrounded by six nearest neighbor molecules in ``bowl-up'' configuration, and the resultant overall structure form a periodic lattice of rotational and polar states within each domain. [Preview Abstract] |
Tuesday, March 15, 2016 10:12AM - 10:24AM |
E22.00010: Accurate Models of Formation Enthalpy Created using Machine Learning and Voronoi Tessellations Logan Ward, Rosanne Liu, Amar Krishna, Vinay Hegde, Ankit Agrawal, Alok Choudhary, Chris Wolverton Several groups in the past decade have used high-throughput Density Functional Theory to predict the properties of hundreds of thousands of compounds. These databases provide the unique capability of being able to quickly query the properties of many compounds. Here, we explore how these datasets can also be used to create models that can predict the properties of compounds at rates several orders of magnitude faster than DFT. Our method relies on using Voronoi tessellations to derive attributes that quantitatively characterize the local environment around each atom, which then are used as input to a machine learning model. In this presentation, we will discuss the application of this technique to predicting the formation enthalpy of compounds using data from the Open Quantum Materials Database (OQMD). To date, we have found that this technique can be used to create models that are about twice as accurate as those created using the Coulomb Matrix and Partial Radial Distribution approaches and are equally as fast to evaluate. [Preview Abstract] |
Tuesday, March 15, 2016 10:24AM - 10:36AM |
E22.00011: Description of interatomic interactions with neural networks Samad Hajinazar, Junping Shao, Aleksey N. Kolmogorov Neural networks are a promising alternative to traditional classical potentials for describing interatomic interactions. Recent research in the field has demonstrated how arbitrary atomic environments can be represented with sets of general functions which serve as an input for the machine learning tool. We have implemented a neural network formalism in the MAISE package [1] and developed a protocol for automated generation of accurate models for multi-component systems. Our tests illustrate the performance of neural networks and known classical potentials for a range of chemical compositions and atomic configurations. [1] Module for Ab Initio Structure Evolution, http://maise-guide.org [Preview Abstract] |
Tuesday, March 15, 2016 10:36AM - 10:48AM |
E22.00012: Learning targeted materials properties from data Turab Lookman, Prasanna V Balachandran, Xue Dezhen, James Theiler, John Hogden We compare several strategies using a data set of 223 M$_2$AX family of compounds for which the elastic properties [bulk ($\mathrm{B}$), shear ($\mathrm{G}$), and Young's ($\mathrm{E}$) modulus] have been computed using density functional theory. The strategy is decomposed into two steps: a \emph{regressor} is trained to predict elastic properties in terms of elementary orbital radii of the individual components of the materials; and a \emph{selector} uses these predictions to choose the next material to investigate. The ultimate goal is to obtain a material with desired elastic properties. We examine how the choice of data set size, regressor and selector impact the results. [Preview Abstract] |
Tuesday, March 15, 2016 10:48AM - 11:00AM |
E22.00013: Neural-network-biased genetic algorithms for materials design Tarak Patra, Venkatesh Meenakshisundaram, David Simmons Machine learning tools have been progressively adopted by the materials science community to accelerate design of materials with targeted properties. However, in the search for new materials exhibiting properties and performance beyond that previously achieved, machine learning approaches are frequently limited by two major shortcomings. First, they are intrinsically interpolative. They are therefore better suited to the optimization of properties within the known range of accessible behavior than to the discovery of new materials with extremal behavior. Second, they require the availability of large datasets, which in some fields are not available and would be prohibitively expensive to produce. Here we describe a new strategy for combining genetic algorithms, neural networks and other machine learning tools, and molecular simulation to discover materials with extremal properties in the absence of pre-existing data. Predictions from progressively constructed machine learning tools are employed to bias the evolution of a genetic algorithm, with fitness evaluations performed via direct molecular dynamics simulation. We survey several initial materials design problems we have addressed with this framework and compare its performance to that of standard genetic algorithm approaches. [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