Bulletin of the American Physical Society
APS March Meeting 2017
Volume 62, Number 4
Monday–Friday, March 13–17, 2017; New Orleans, Louisiana
Session A1: Computational Discovery and Design of Novel Materials IFocus
|
Hide Abstracts |
Sponsoring Units: DMP DCOMP Chair: Anatole von Lilienfeld, University of Basel Room: 260 |
Monday, March 13, 2017 8:00AM - 8:12AM |
A1.00001: Combining first principles modeling, experimental inputs, and machine learning for nanocatalysts design Fatih Sen, Spencer Hills, Alper Kinaci, Badri Narayanan, Michael Davis, Stephen Gray, Subramanian Sankaranarayanan, Maria Chan Nanocatalysts are of technological and scientific relevance for a large variety of catalytic processes. Due to the diverse geometries and complex structure-activity relationships, computational modeling and machine learning techniques are helpful in order to sample configuration space, incorporate experimental information, and account for co-variations in stability and catalytic activity. We will discuss structural determination of Au and IrO2 nanocatalysts from single and multi-objective global optimization algorithms, using as inputs density functional theory (DFT) calculations [1], a combination of energetic and simulated pair distribution function (PDF) data, and a combination of energetic and activity objectives. DFT data from thousands of Au nanostructures are fitted using a genetic algorithm to a hybrid bond-order potential (HyBOP)[2], which is able to predict structural and energetic properties of Au nanoclusters to bulk. Similarly, genetic algorithm is used to parametrize a variable charge potential for IrO2[3], which is instrumental in the combined multi-objective optimization of stability and activity. [1] A. Kinaci, et al, Sci. Rep. 6, 34974 (2016). [2] B. Narayanan, et al, J. Phys. Chem. C 120, 13787 (2016). [3] F. G. Sen, et al, J. Mater. Chem. A 3, 18970 (2015). [Preview Abstract] |
Monday, March 13, 2017 8:12AM - 8:24AM |
A1.00002: Automatic high-throughput screening of colloidal crystals using machine learning Matthew Spellings, Sharon C. Glotzer Recent improvements in hardware and software have united to pose an interesting problem for computational scientists studying self-assembly of particles into crystal structures: while studies covering large swathes of parameter space can be dispatched at once using modern supercomputers and parallel architectures, identifying the different regions of a phase diagram is often a serial task completed by hand. While analytic methods exist to distinguish some simple structures, they can be difficult to apply, and automatic identification of more complex structures is still lacking. In this talk we describe one method to create numerical ``fingerprints'' of local order and use them to analyze a study of complex ordered structures. We can use these methods as first steps toward automatic exploration of parameter space and, more broadly, the strategic design of new materials. [Preview Abstract] |
Monday, March 13, 2017 8:24AM - 8:36AM |
A1.00003: Comparative Analysis of Particle Swarm and Differential Evolution via Tuning on Ultrasmall Titanium Oxide Nanoclusters Eric Inclan, Jack Lassester, David Geohegan, Mina Yoon Optimization algorithms (OA) coupled with numerical methods enable researchers to identify and study (meta) stable nanoclusters without the control restrictions of empirical methods. An algorithm's performance is governed by two factors: (1) its compatibility with an objective function, (2) the dimension of a design space, which increases with cluster size. Although researchers often tune an algorithm's user-defined parameters (UDP), tuning is not guaranteed to improve performance. In this research, Particle Swarm (PSO) and Differential Evolution (DE), are compared by tuning their UDP in a multi-objective optimization environment (MOE). Combined with a Kolmogorov Smirnov test for statistical significance, the MOE enables the study of the Pareto Front (PF), made of the UDP settings that trade-off between best performance in energy minimization (``effectiveness'') based on force-field potential energy, and best convergence rate (``efficiency''). By studying the PF, this research finds that UDP values frequently suggested in the literature do not provide best effectiveness for these methods. Additionally, monotonic convergence is found to significantly improve efficiency without sacrificing effectiveness for very small systems, suggesting better compatibility. [Preview Abstract] |
Monday, March 13, 2017 8:36AM - 8:48AM |
A1.00004: Using Machine Learning to Improve Cluster Expansion Predictions Wiley Morgan, Kevin Francis, Gus Hart Cluster Expansion is used to predict the energies of different configurations and concentrations of a several elements arranged on a fixed lattice. Applications include ground state searches, modeling the energy of solid solutions, and precipitate formation. For materials with a small lattice mismatch, these predictions are generally reliable. However, when the lattice mismatch becomes large, the cluster expansion method often fails. In a number of ternary cases, we have found that the errors in the predictions appear to be connected to the concentration. Partitioning composition space and constructing a cluster expansion model for each partition allows us to make accurate predictions within most of the partitions. We report on atemtps to use machine learning to predict in the partitions where cluster expansion is inaccurate. [Preview Abstract] |
Monday, March 13, 2017 8:48AM - 9:00AM |
A1.00005: Why is Formation Enthalpy Bad for Cluster Expansion Fitting? Andrew Nguyen, Mark Transtrum, Gus Hart Cluster expansion (CE) allows one to map a relationship between energy and configuration. We often use total energy or formation enthalpy to build a CE model. Since the formation enthalpy is just the transformation of the total energy, one expects that total energy or formation enthalpy would yield similar CE models, i.e., similar errors and a similar number of coefficients. In this study, we are examining the effect of input energies (total energy vs. formation enthalpy) in the cluster expansion formalism. We show that the formation enthalpy generates worse fits, i.e., higher prediction errors and a higher number of coefficients in most cases. We show that correlated noise has little impact on CE models except when that noise is very large. However, the main problem is the transformation of total energy into formation enthalpy which amplifies noise. Thus, we find that it is best to always use the total energy in the fitting. One can always perform the transformation into formation enthalpy after constructing the CE model. [Preview Abstract] |
Monday, March 13, 2017 9:00AM - 9:12AM |
A1.00006: Invariant Representations for Robust Materials Prediction Gus Hart, Conrad Rosenbrock, Gabor Csanyi The high-throughput approach for computational materials science has led to the generation of huge databases of DFT-based calculations. Direct mining of this data has led to the discovery of new materials and is of considerable utility. But the real potential for these data to impact American competitiveness, as envisioned in the MGI, is in "interpolation"---using the data to discover materials not present in the databases. I will discuss an approach for materials interpolation that combines cluster expansion, the new SOAP (smooth overlap of atomic positions) representation, and machine learning. [Preview Abstract] |
Monday, March 13, 2017 9:12AM - 9:48AM |
A1.00007: TBD - Computational Discovery and Design of Novel Materials Invited Speaker: Risi Kondor |
Monday, March 13, 2017 9:48AM - 10:00AM |
A1.00008: Invariance to deformations: A new representation for materials space Chandramouli Nyshadham, Gus L. W. Hart Huge databases of known materials have been developed using computational and experimental methods over the last century. The existing databases cover a very small fraction of the complete materials space. The future of materials discovery lies in intelligently exploring the materials space (composition and structure space) using machine learning methods. Recently, it has been understood that details of the mathematical representation of materials are key to developing effective algorithms that leverage the machine learning models. One of the main challenges in representing materials space is to incorporate the ``deformation stability"--- that is small changes in the material imply small changes in representation---a kind of ``differentiability". The well-known Fourier based approaches for representing materials space cannot handle the invariance to deformations. In this talk, we will present a new, easy to understand representation based on scattering transforms. Scattering transforms are formally stable to deformations and more effective in interpolating the materials space than the Fourier based approaches. Machine learning models based on scattering transforms offer the potential of high accuracy at the speed of machine learning, thus accelerating materials discovery. [Preview Abstract] |
Monday, March 13, 2017 10:00AM - 10:12AM |
A1.00009: Geometric space - the extension of extremely dense unit cells Antony Bourdillon The Quasicrystal is a relatively new kind of solid, intermediate between crystals and compound glasses. It has many peculiar properties including non-Drude conductivity; geometric electronic band structures; peculiar mechanical and magnetic effects etc. However the greatest benefit they have taught us is the fact of geometric space with sharp coherence [1]. This provides opportunities for finite element simulations with fast convergence and avoidance of subsidiary maxima or minima. As Einstein's curved space is locally Euclidean; dense atomic space is locally icosahedral, and geometric in extension. Intermediate linear periodicity, in crystals, is constrained by unit cells that are less dense at short range. [1] Diffraction line width in quasicrystals -- sharper than crystals, A.J. Bourdillon, (2016) \textit{Journal of Modern Physic,}\textbf{7, }1558-1567 \quad (2016) DOI: 10.4236/jmp.2016.712142 [Preview Abstract] |
Monday, March 13, 2017 10:12AM - 10:24AM |
A1.00010: Making Sense of the Multitude of Brillouin Zone Integration Methods Jeremy Jorgensen, Gus Hart Over the past 50 years, a host of Brillouin zone (BZ) sampling and integration methods have been proposed. After outlining the principal difficulties associated with BZ integration, we explore the evolution of BZ sampling methods, starting with the mean-value point method, ending with the maximal packing fraction method [1]. We also examine the standard techniques that have been employed in performing BZ integrations, which include various projection methods. Finally, in order to illustrate the complications that arise during BZ integration, we employ an intuitive and realistic toy model, and also use it to investigate band energy convergence with increasing sampling point densities. \\ \\ [1] Wisesa, Pandu, Kyle A. McGill, and Tim Mueller. "Efficient generation of generalized Monkhorst-Pack grids through the use of informatics." Physical Review B 93.15 (2016): 155109. [Preview Abstract] |
Monday, March 13, 2017 10:24AM - 10:36AM |
A1.00011: Enhancements to the $k$-point grid server: generating highly efficient grids through the use of informatics Pandu Wisesa, Wan Wan, Tim Mueller Calculating material properties often involves using a grid of points, commonly known as $k-$points, to approximate an integral over the Brillouin zone in reciprocal space. The choice of grids directly affects the computational resources consumed and accuracy of the calculation. Finding a grid that minimizes computational cost for a desired level of accuracy can be computationally expensive, but we have facilitated the process by creating a publicly-available $k-$point grid server backed by a database of hundreds of thousands of efficient, pre-calculated $k$-point grids. We estimate that for well-converged calculations these grids on average reduce the resources consumed by approximately half while maintaining the same level of accuracy. We discuss recent updates to the server and how to make use of them, including new features and support for additional software packages. [Preview Abstract] |
Monday, March 13, 2017 10:36AM - 10:48AM |
A1.00012: A Computational Framework for Automation of Point Defect Calculations Anuj Goyal, Prashun Gorai, Haowei Peng, Stephan Lany, Vladan Stevanovic A complete and rigorously validated open-source Python framework to automate point defect calculations using density functional theory has been developed. The framework provides an effective and efficient method for defect structure generation, and creation of simple yet customizable workflows to analyze defect calculations. The package provides the capability to compute widely accepted correction schemes to overcome finite-size effects, including (1) potential alignment, (2) image-charge correction, and (3) band filling correction to shallow defects. Using Si, ZnO and In2O3$_{\mathrm{\thinspace }}$as test examples, we demonstrate the package capabilities and validate the methodology. We believe that a robust automated tool like this will enable the materials by design community to assess the impact of point defects on materials performance. [Preview Abstract] |
Monday, March 13, 2017 10:48AM - 11:00AM |
A1.00013: Accurate atomistic potentials and training sets for boron-nitride nanostructures Isaac Tamblyn Boron nitride nanotubes exhibit exceptional structural, mechanical, and thermal properties. They are optically transparent and have high thermal stability, suggesting a wide range of opportunities for structural reinforcement of materials. Modeling can play an important role in determining the optimal approach to integrating nanotubes into a supporting matrix. Developing accurate, atomistic scale models of such nanoscale interfaces embedded within composites is challenging, however, due to the mismatch of length scales involved. Typical nanotube diameters range from 5-50 nm, with a length as large as a micron (i.e. a relevant length-scale for structural reinforcement). Unlike their carbon-based counterparts, well tested and transferable interatomic force fields are not common for BNNT. In light of this, we have developed an extensive training database of BN rich materials, under conditions relevant for BNNT synthesis and composites based on extensive first principles molecular dynamics simulations. Using this data, we have produced an artificial neural network potential capable of reproducing the accuracy of first principles data at significantly reduced computational cost, allowing for accurate simulation at the much larger length scales needed for composite design. [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