Session K22: Predicting and Classifying Materials via High-Throughput Databases and Machine Learning II

An Automated Application Framework to Model Disordered Materials Based on a High Throughput First Principles Approach

Predicting material properties of disordered systems remains a long-standing and formidable challenge in rational materials design. To address this issue, we introduce an automated software framework capable of modeling partial occupation within disordered materials using a high-throughput (HT) first principles approach. At the heart of the approach is the construction of supercells containing a virtually equivalent stoichiometry to the disordered material. All unique supercell permutations are enumerated and material properties of each are determined via HT electronic structure calculations. In accordance with a canonical ensemble of supercell states, the framework evaluates ensemble average properties of the system as a function of temperature. As proof of concept, we examine the framework's final calculated properties of a zinc chalcogenide (ZnS$_{1-x}$Se$_x$), a wide-gap oxide semiconductor (Mg$_{x}$Zn$_{1-x}$O), and an iron alloy (Fe$_{1-x}$Cu$_{x}$) at various stoichiometries.   

An Extensive Database of Electronic Structure Calculations between Transition Metals

Density Functional Theory and its derived application methods, such as the Augmented Plane Wave (APW) method, have shown great success in predicting the fundamental properties of materials. In this work, we apply the APW method to explore the properties of diatomic pairs of transition metals in the CsCl structure, for all possible combinations. A total of 435 compounds have been studied. The predicted Density of States, and Band Structures are presented, along with predicted electron-phonon coupling and Stoner Criterion, in order to identify potential new superconducting or ferromagnetic materials. This work is performed to demonstrate the concept of “high-throughput” calculations at the crossing-point of “Big Data” and materials science.   

High throughput ab-intio modeling of proton transport in solid electrolytes

Solid oxide materials that can selectively transport protons have great potential for fuel cell applications. However several fundamental questions remain unanswered such as (a) How do the dopants organize at various dopant concentrations, (b) How spatial organization of dopants influence proton migration energy, (c) How disorder and strain in a material influence its ionic transport. $\backslash $In this work have developed an integrated high throughput framework to calculate proton transport properties by integrating open source packages (such as pymatgen, fireworks) The high throughput framework scales well on supercomputing clusters. We have used this framework to analyze over 100 perovskites compounds with over 12 different dopant atoms. These computational models enable us to obtain insights how the proton transport properties depend on host and dopant atoms. Further, we also perform ab-initio modeling to understand how dopants spatially organize at different dopant concentrations, and how this spatial organization affects proton conductivity. This analysis enabled us to obtain fundamental insights on why proton conductivity decreases in Y doped BaZrO3 at high dopant concentrations.   

From organized high throughput data to phenomenological theory: The example of dielectric breakdown

Understanding the behavior (and failure) of dielectric insulators experiencing extreme electric fields is critical to the operation of present and emerging electrical and electronic devices. Despite its importance, the development of a predictive theory of dielectric breakdown has remained a challenge, owing to the complex multiscale nature of this process. Here, we focus on the \emph{intrinsic} dielectric breakdown field of insulators---the theoretical limit of breakdown determined purely by the chemistry of the material, \emph{i.e.}, the elements the material is composed of, the atomic-level structure, and the bonding. Starting from a benchmark dataset (generated from laborious first principles computations) of the intrinsic dielectric breakdown field of a variety of model insulators, simple predictive phenomenological models of dielectric breakdown are distilled using advanced statistical or machine learning schemes, revealing key correlations and analytical relationships between the breakdown field and easily accessible material properties. The models are shown to be general, and can hence guide the screening and systematic identification of high electric field tolerant materials.   

The Robustness of Cluster Expansion: Assessing the Role of Relaxation

Cluster expansion (CE) has been used widely in combination with first-principles calculations to predict stable structures of metal alloys. CE treats alloys as a purely configuration problem, i.e., a problem in the distribution of the alloying elements on a fixed lattice. CE models are usually built from data taken from ``relaxed'' first-principles calculations where the individual atoms assume positions that minimize the total energy. A perennial question in the cluster expansion community is how the accuracy of the CE is affected by relaxations---technically, the formalism of CE breaks down when the underlying lattice is not preserved---but practitioners often argue that there is a one-to-one correspondence between relaxed and unrelaxed structures so that the formalism holds. We quantify the effect of relaxation on the robustness of cluster expansions by comparing CE fits for relaxed and unrelaxed data sets. Our results give a heuristic for when CE models can be trusted.   

The New NRL Crystallographic Database

For many years the Naval Research Laboratory maintained an online graphical database of crystal structures for a wide variety of materials. This database has now been redesigned, updated and integrated with the AFLOW framework for high throughput computational materials discovery (http://materials.duke.edu/aflow.html). For each structure we provide an image showing the atomic positions; the primitive vectors of the lattice and the basis vectors of every atom in the unit cell; the space group and Wyckoff positions; Pearson symbols; common names; and Strukturbericht designations, where available. References for each structure are provided, as well as a Crystallographic Information File (CIF). The database currently includes almost 300 entries and will be continuously updated and expanded. It enables easy search of the various structures based on their underlying symmetries, either by Bravais lattice, Pearson symbol, Strukturbericht designation or commonly used prototypes. The talk will describe the features of the database, and highlight its utility for high throughput computational materials design.   

First-principles determination of low-energy structures in epitaxially-strained perovskite SrMnO$_{3}$

Using a physically-motivated form for the energy as a function of magnetic ordering and lattice distortions around the high symmetry reference structure, we present a systematic method for determining the ground state and low-energy structures of transition-metal ABO$_{3}$ compounds from first principles. The structural information obtained through this method forms the foundation for the first-principles structural determination of the structure of perovskite oxide superlattices. The method is demonstrated for SrMnO$_{3}$, which has a nontrivial phase sequence with varying epitaxial strain that has been of recent interest both in first-principles and experimental investigations.   

Finding new ternary transition metal selenides and sulphides

The transition metal oxides exhibit many interesting physical properties, and have been explored in detail over time. Recently, the transition metal chalchogenides including selenium and sulfur have been of interest because of their correlated electron properties, as seen in the iron based superconductors and the layered transition metal dichalchogenides. However, the chalchogenides are much less explored than the oxides, and there is an open question of whether there may be new materials heretofore undiscovered. We perform a systematic combined theoretical and experimental search over ternary phase diagrams that are empty in the Inorganic Crystal Structure Database containing cations, transition metals, and one of selenium or sulfur. In these 27 ternary systems, we use a probabilistic model to reduce the likelihood of false negative predictions, which results in a list of 24 candidate materials. We then conduct a variety of synthesis experiments to check the candidate materials for stability. While the prediction method did obtain compositions that are stable, none of the candidate materials formed in our experiments. We come to the conclusion that these phase diagrams are either truly empty or have unusual structures or synthesis requirements.   

High-Throughput Identification of Unique Structure Prototypes in the Inorganic Crystal Structure Database

High-throughput computational assessment of materials properties is currently a major component of the effort to develop new useful materials by uncovering trends and correlations between structures, compositions, and functionalities. Efficient implementation of this approach thus requires a systematic identification of distinct material structure prototypes. We have developed a robust algorithm that calculates the level of similarity between crystal structures independent of the unit cell representation, using the comparison method proposed by Burzlaff [1]. This algorithm has been implemented in the high-throughput framework, Automatic Flow (AFLOW) [2], and applied to the Inorganic Crystal Structure Database (ICSD) [3] entries in the AFLOWLIB.org [4] online repository. We have determined the uniqueness statistics for the ICSD and have created a comprehensive set of the unique structural prototypes represented in it. [1] H. Burzlaff and Y. Malinovsky Acta Cryst. A53, 217-224 (1997). [2] S. Curtarolo et al. Comp. Mater. Sci. 58, 218-226 (2012). [3] FIZ Karlsruhe and NIST, Inorganic Crystal Structure Database, http://icsd.fiz-t karlsruhe.de/icsd/ [4] S. Curtarolo et al. Comp. Mater. Sci. 58, 227-235 (2012).   

Prediction of the first stable compound with flat hexagonal tin layers

An analysis of stability trends in a large family of metal stannides has directed our attention towards a previously unknown compound featuring a backbone of flat hexagonal tin layers. Ab initio calculations show that this compound is at least metastable under ambient conditions and is furthermore stabilized under pressure. Compounds with such layered frameworks may possess exotic electronic properties and also serve as precursors for the synthesis of 2D derivatives.   

Can k-point integration for metals be dramatically improved?

Our group has spent hundreds of millions of cpu hours calculating the energies of different materials and their competing structures. The energy of the occupied electron states is a small part of the total energy of a given material, but electron energy accounts for almost all of the numerical error in these calculations (at least in metals). Current methods of integrating electron band energies are simple (usually rectangle rule + smearing) but converge very slowly, requiring many, many k-points, even for simple metals. But integration approaches with better convergence rates, such Gauss quadrature, are hard to utilize. Because of the multivalued nature of electron bands (leading to crossings, kissings, etc.) standard interpolation methods are ineffective. We will discuss a number of improvements we have made and discuss a possible solution to the interpolation problem.   

The Quantum Monte Carlo Database: towards high-accuracy and high-throughput calculation of material properties~

Quantum Monte Carlo (QMC) techniques comprise a class of promising methods that offer a path towards higher accuracy for materials property prediction. However, their application in bulk materials has historically been limited to one-at-a-time evaluation of a given material. While these results often provide benchmark-level accuracy for quantities of interest, they do not allow for high-throughput analysis of the data since each calculation is done slightly differently. We present a combined data format and automatic generation platform based on the QWalk code for QMC data: QMCDB. This platform collects QMC results and provenance information automatically and stores the information in a database. We will report on the construction of this database and what lessons can be learned about using QMC for high-throughput applications.   

\textbf{Predictive modeling of surface morphology of multicomponent catalysts for their optimum performance}

Multi-component microstructures of artificially engineered catalysts are promising for the best ever performance in alternative fuel production. We have designed and implemented a set of intelligent algorithms capable of predicting the surface morphology of multicomponent catalysts for their optimum performance. For example we come up with three kinds of different catalysts. Based on a database obtained from the density functional theory based kinetic Monte Carlo simulations, the first kind of single component catalytic surface promotes and helps dissociative adsorption of chemical species, but it hinders the diffusion of intermediate species. On the other hand, the second kind of single component catalytic surface promotes the diffusion of intermediate species, but suppresses the reactions and desorption processes. However the third kind of single component catalytic surfaces can significantly enhance reactions among intermediate species. Therefore no single component material surface would be a suitable candidate for becoming a good catalyst. However a combination of all above mentioned kind of materials may exhibit the maximum ever performance. Our algorithm models the surface morphology of these multicomponent catalysts by varying the surface area of each component and also by changing the shape of each component in such a way that the catalyst gives the highest rate of chemical formation. Our results confirm the best ever performance of our artificially engineered catalysts.   

\textbf{Doping Li and K into Na}$_{\mathrm{\mathbf{2}}}$\textbf{ZrO}$_{\mathrm{\mathbf{3}}}$\textbf{ Sorbent to Improve Its CO}$_{\mathrm{\mathbf{2}}}$\textbf{ Capture Capability}

Carbon dioxide is one of the major combustion products which once released into the air can contribute to global climate change. Solid sorbents have been reported in several previous studies to be promising candidates for CO$_{\mathrm{2}}$ sorbent applications due to their high CO$_{\mathrm{2}}$ absorption capacities at moderate working temperatures. However, at a given CO$_{\mathrm{2}}$ pressure, the turnover temperature (T$_{\mathrm{t}})$ of an individual solid capture CO$_{\mathrm{2}}$ reaction is fixed and may be outside the operating temperature range ($\Delta $T$_{\mathrm{o}})$ for a particularly capture technology. In order to shift such T$_{\mathrm{t}}$ for a solid into the range of $\Delta $T$_{\mathrm{o}}$, its corresponding thermodynamic property must be changed by changing its structure by reacting (mixing) with other materials or doping with other elements. As an example, by combining thermodynamic database searching with \textit{ab initio} thermodynamics calculations, in this work, we explored the Li- and K-doping effects on the T$_{\mathrm{t}}$ shifts of Na$_{\mathrm{2}}$ZrO$_{\mathrm{3}}$ at different doping levels. The obtained results showed that compared to pure Na$_{\mathrm{2}}$ZrO$_{\mathrm{3}}$, the Li- and K-doped mixtures Na$_{\mathrm{2-\alpha }}$M$_{\mathrm{\alpha }}$ZrO$_{\mathrm{3}}$ (M$=$Li, K) have lower T$_{\mathrm{t}}$ and higher CO$_{\mathrm{2}}$ capture capacities.