The discovery of structures with desired functionality is inextricably linked with our ability to enumerate molecules with unique chemotypes and novel three-dimensional architectures. The size of the “small molecule universe” is astronomically vast, making exhaustive searches impossible. I will describe how we are constructing “representative" universal libraries that span the small molecule universe and that samples the full extent of feasible small molecule chemistries. The SMU is found to have much more structural diversity than found previously for molecular scaffolds and, as such, represents an untapped resource. This project is being conducted in collaboration with Dr. Aaron Virshup, Dr. Julia Contreras-García, Prof. Peter Wipf, and Prof. Weitao Yang.

The quest for new materials demands the reliable knowledge of their stable and meta(!)stable structures under operational conditions. For instance, materials may exhibit the desired function at finite temperature and exposed to an atmosphere of reactive molecules in the gas phase. In this sense, not only an accurate account of the total (electronic) energy should be aimed at, but it should be also always considered that the degeneracy of states (related to the configurational entropy) may play an important role in indicating the lowest free-energy structures. We show how massive parallel replica exchange (RE) based on ab initio molecular dynamics (MD) is a powerful tool for, not only identifying a set of (meta)stable structures, but also for estimating in an unbiased fashion the relative free-energy of the structures and therefore their relative population under environmental conditions. As an illustrative example of our approach, we apply it to gold clusters in an atmosphere of oxygen and carbon monoxide, aiming at the description of the gold-cluster-mediated oxidation of CO. The electronic structure of the Au_N clusters (2<N<40) and the ligands is described by means of van der Waals corrected density functional theory. We find that a) intra-cluster van der Waals interactions play an important role in describing the relative energetics of the clusters, b) at many sizes the clusters present themselves as an ensemble of often fluctional isomers already at relatively low temperature (e.g. T = 100 K) c) structures that are stable at T = 0 K according to harmonic analysis may be unstable at finite temperature (e.g. T = 300 K). A way to incorporate in REMD a variable number of adsorbed species, via a grand canonical scheme, is also described.

The utilization of first prinicples density functional theory (DFT) is gradually shifting towards predictive materials design. In the present study, we use DFT calculations in combination with cluster expansion and quantitative structure property relationship (QSPR) techniques to identify new dielectric materials for high energy density capacitor applications. The current state-of-the-art in high energy density capacitors is mainly dominated by polyethylene (PE) and polypropylene (PP) based polymers. However, these polymer dielectrics suffer from a relatively low dielectric permittivity, which eventually leads to a lower energy storage density. In our rational approach to this problem, we investigate and identify various possible local chemical modifications to PE, which would result in much larger dielectric permittivity than PE, while still preserving its attractive insulating properties (such as large band gap). More specifically, we allow the -CH2- unit in the PE backbone segment to be replaced by -SiF2-, -SiCl2-, -GeF2-, -GeCl2-, -SnF2-, or -SnCl2- units in a systematic, progressive, and exhaustive manner. The motivation behind substituting C with the larger group 14 elements (i.e, Si, Ge and Sn) in the polymer backbone is to ensure chemical compatibility by preserving local chemical environment and bonding as well as to insert more polarizable units than C in the backbone that would result in an enhanced dielectric response. Furthermore, constraining the side chain pendent groups to small groups with high electronegativity such as F and Cl would provide large dipole moments that would not only contribute to the orientational part of the dielectric permittivity but are also expected to lead to large band gap. We first use our recently developed method that employs DFT calculations of isolated single polymer chains in combination with the effective medium theory to accurately estimate the dielectric permittivity of the chemically modified PE chains for a set of limited compositions and configurations. QSPR based techniques are then used to mine knowledge from the data and identify quantitative chemical trends between electronic, dielectric and structural properties. Furthermore, by successfully parameterizing a cluster expansion from the DFT computed data set we are able to predict the dielectric permittivity of systems spanning a much larger part of the configurational and compositional space. A set of most promising PE modifications (with simultaneously large dielectric constant and band gap) is identified using the cluster expansion parameterization, and the predictions made are further validated through DFT computations.

Polymer Nanocomposite (PNC) literature has demonstrated a myriad of potential structures and self assembled morphologies that could significantly impact commercial and military applications. It has been recognized, however, that PNCs are far from optimized and for them to achieve their promised potential in advanced applications, it is important to develop the ability to perform a priori virtual design by circumventing the dominant trial and error approach. A priori prediction could be realized rigorously by multiscale modeling that starts with ab initio models at the atomic and molecular level and progresses upto continuum scale models that predict bulk properties. However, developing and implementing exact scale-bridging methods is a challenge fraught with difficulty and limited by both computational power and scale linking methods. An alternative approach is to use heuristic techniques, which are based on first principles, to bridge length scales. Here a heuristic informatics-based approach (Materials Quantitative Structure Property Relationships - MQSPR) is used, coupled with physics-based continuum models and experimental validation to predict the thermomechanical properties of spherical nanoparticle filled polymers. The hypothesis driving this strategy is that the dispersion and distribution of the nanofillers and the mobility of the polymer chains near the nanoparticles can be predicted from the polar and dispersive components of the surface energy. Materials Quantitative Structure-Property Relationship (MQSPR), using models trained on molecular-level features of the polymers and particles from literature has been shown to predict the polar and dispersive components of polymer and nanoparticle surface energies. Experiments were performed on an array of PNCs with four different matrices of known surface energies ranging from polar to non polar (Poly(2-VinylPyridine), Polymethylmethacrylate, Polyethylmethacrylate and Polystyrene), filled with colloidal silica nanoparticles modified with three different monofunctional silanes to obtain quantitative measures of dispersion and Tg deviations from neat polymers. The effect on energetics on PNC microstructure, was analyzed by relating derived energetic parameters (contact angle (theta;), relative work of adhesion (ΔWa)) to quantitative measures of dispersion (Cluster size and intercluster distance), obtained by the characterization of digital micrographs. Measured Tg deviations were related to a compound dispersion and a mobility related term (theta;,ΔWa , Work of Spreading (Ws)). These relations were used to generate a 3D continuum model from the knowledge of the MQSPR predicted surface energies. Simulations run on these 3D models are used to predict the thermomechanical response of the nanocomposites. This effort represents the first time that MQSPR methods have been used within a fully connected set of scales, and presents a tremendous opportunity to predict bulk nanocomposite properties.

Electronic structure methods based on density functional theory have reached a level of speed and accuracy where they can be used to describe complete catalytic reactions on transition metal surfaces. Simulations now complement experiments to give unprecedented insight into surface processes, allowing us to pinpoint the origin of the catalytic activity of e.g. a metal in terms of its electronic structure. However, the use of electronic structure theory as a tool for directly designing or searching for new materials is a rather expensive process [1] and has until now been extremely limited. The complexity arises to a smaller extent from the computationally expensive electronic structure calculations and to a larger extent from the enormous number of atomic configurations that might be considered when ordering even just a few atoms in a molecule or a few atoms in the unit cell of a crystalline system. In a few cases the underlying reactivity trends for a reaction have been understood well and were subsequently successfully used to computationally seek out new catalysts [2]. I here discuss the search for especially new hydrogenation catalysts [3,4]. The focus will be to show the generality of this trend- and descriptor-based approach by linking a few specific case stories to more fundamental underlying linear-energy relations for the adsorbate-surface bonds [5,6,7,8]. The general applicability of linear energy relations in the complexity reduction for finding new catalyst makes it natural to develop standardized tools for analyzing catalytic reactions [9]. References: [1] G.H. Johannesson, T. Bligaard, A.V. Ruban, H.L. Skriver, K.W. Jacobsen, and J.K. Noslash;rskov, Phys. Rev. Lett. 88, 255506 (2002) [2] J. Greeley, T.F. Jaramillo, J. Bonde, I. Chorkendorff, and J.K. Noslash;rskov, Nature Materials 5, 909 (2006) [3] M. P. Anderson, T. Bligaard, A. Kustov, K. E. Larsen, J. Greeley, T. Johannessen, C. H. Christensen, J. K. Noslash;rskov, J. Catal. 239, 501 (2006) [4] F. Studt, F. Abild-Pedersen, T. Bligaard, R.Z. Soslash;rensen, C.H. Christensen, and J.K. Noslash;rskov, Science 320, 1320 (2008) [5] J.K. Noslash;rskov, T. Bligaard, A. Logadottir, S. Bahn, L.B. Hansen, M. Bollinger, H. Bengaard, B. Hammer, Z. Sljivancanin, M. Mavrikakis, Y. Xu, S. Dahl, and C.J.H. Jacobsen, J. Catal. 210, 275 (2002) [6] F. Abild-Pedersen, J. Greeley, F. Studt, J. Rossmeisl, T.R. Munter, P.G. Moses, E. Skulason, T. Bligaard, and J.K. Noslash;rskov, Phys. Rev. Lett. 99, 016105 (2007) [7] E.M. Fernández, P.G. Moses, A. Toftelund, H.A. Hansen, J.I. Martínez, F. Abild-Pedersen, J. Kleis, B. Hinnemann, J. Rossmeisl, T. Bligaard, and J.K. Noslash;rskov, Angew. Chem. Int. Ed. 47, 4683 (2008) [8] J.K. Noslash;rskov, T. Bligaard, J. Rossmeisl, C.H. Christensen, Nature Chemistry 1, 37 (2009) [9] J.S. Hummelshoslash;j, F. Abild-Pedersen, F. Studt, T. Bligaard, J.K. Noslash;rskov, Angew. Chem. Int. Ed. 51, 272 (2012)

Better oxygen reduction catalysts are needed to improve the efficiency and lower the cost of fuel cells. Metal nanoparticles are good candidates for new catalysts because their catalytic properties are different from bulk metals, and are sensitive to particle size, shape and composition. The electronic structure can be determined for small particles, making it possible to optimize particles for a desired reaction. Here, we calculate the electronic structure of 1 nm core/shell particles and show how the energy of electrons in the shell can tune the binding of oxygen by varying the core metal. Transition state calculations for O2 dissociation on the nanoparticle surface show that the d-band center is a good measure of the activation and reaction energies. Two factors are found to be significant for determining the catalytic activity of small nanoparticles; charge transfer in core/shell particles and the rigidity of alloy particles.

Over the past fifty years, metallic glasses have evolved significantly from their origins as a scientific novelty but still are not commonly used engineering materials. They are excellent candidates for many applications and combine the unique properties of amorphous materials and conventional metals, such as excellent mechanical and soft magnetic properties. One of the key limitations preventing metallic glasses from being used in more applications is the difficultly in designing new metallic glass alloys, which often involves a time-consuming experimental search guided by empirical rules. Consequently, there are only a handful of highly stable amorphous alloys known, many of which contain expensive and toxic elements. To help accelerate the design of new alloys, we have developed a set of computational tools to rapidly evaluate possible candidates. Effectively designing a metallic glass using computational methods requires the ability to generate interatomic potentials rapidly, calculate properties on a large scale, and assess glass-forming ability without the need for experimental input. We have developed a novel technique to create accurate embedded atom method potentials with a small computational expense based on combining existing elemental potentials. In combination with the large number of elements available in interatomic potential libraries, this simple and powerful technique enables classical molecular dynamics as a tool for alloy design. In order to search rapidly through a large set of possible alloy compositions, we have developed software to manage the optimization of alloy properties using genetic algorithms. In this way, it was possible to run an automated, targeted search in a space with millions of possible alloy compositions within a matter of weeks. A similarly broad search would be infeasible using conventional experimental methods. After a final alloy is selected based on favorable properties, we are able evaluate the ability of the alloy to be cast into the amorphous state using a newly-developed approach based on evaluating the driving force for crystallization and kinetics in the supercooled liquids. Combining these techniques, we have identified an amorphous alloy with an optimized stiffness to weight ratio, which we are currently working to validate experimentally.

Perovskites with the chemical formula ABO3 (with A and B being cations) offer material solutions for diverse technological applications such as electronics, sensors and catalysis. In order to tune their chemical and physical properties for a specific application, chemical doping in perovskites is considered to be an important and promising route. For instance, pure BaTiO3 absorbs only a small fraction of the visible light in the solar spectrum. However, for efficient photo-electrochemical activity, one needs to extend its photo response to the visible part of the spectrum by tailoring its electronic properties through lattice defect chemistry or via oxygen stoichiometry. Furthermore, dopants are introduced to preserve the insulating behavior of multilayer ceramic capacitors in the presence of lattice oxygen vacancies. Owing to the intrinsic capability of the perovskite structure to host ions of various sizes across the periodic table, a wide range of dopants can be, and have been, successfully accommodated in the ABO3 lattice. Although chemical doping in BaTiO3 has been targeted by many experimental studies in the literature, the underlying strategies have largely been based on Edisonian approaches. There is thus a need to explore and understand the dopant chemical space of perovskites using rational approaches such as first principles methods to achieve deeper insights. Here we present a comprehensive high-throughput first principles based study of chemical doping in BaTiO3. As possible dopants, we have considered the 3d, 4d and 5d transition metals along with the neighboring IA through VA group elements in the periodic table. For a clear understanding of dopant&’s behavior at different sites, doping at both Ba and Ti sites has been studied. We first evaluate the electronic structure and energetics of dopant formation in BaTiO3 lattice. We have attempted to correlate dopant properties such as its formation energy and changes to the electronic structure to dopant atom properties such as electron affinity, electronegativity, ionic radius, oxidation state, ionization potential, polarizability etc. The regularities governing the dopant properties are investigated by using statistical regression methods. It is found that the observed behavior mainly occurs due to the interplay of three factors, viz., ionic radius, oxidation state and electro-negativity. We also find that the monovalent and divalent dopants have a tendency to substitute the host atom of similar size while dopants with higher oxidation states always favor the Ti site. On the basis of the correlations unearthed from our statistical analysis of first principles data, we have developed a thorough and comprehensive theoretical understanding of the behavior of dopants within the perovskites lattice.

In order to develop a robust and general method to calculate melting points via ab initio calculations, we propose a scheme which drastically improves the efficiency of Widom's original method to calculate chemical potentials based on the insertion of a test particle. Our scheme can be easily implemented into computer code to automatically find, explore and study large cavities which can accommodate the test-particle at low energy cost and whose contribution to the chemical potential thus dominates. This selective sampling helps reduce the computational cost by orders of magnitude, so that it is, for the first time (to our knowledge), practical to calculate chemical potentials of liquids entirely from first-principles, without the help of any high-quality empirical potentials as reference systems. This distinct advantage makes our method easily generalizable to a wide variety of materials, compared to the commonly used thermodynamic integration method (which depends heavily on the quality of reference potentials, in particular, their ability to stabilize the correct solid-state crystal structure). This is especially useful in automated materials screening effort in which melting point is a design parameter, since one does not need to develop empirical potentials for each of the chemical system explored. We benchmark our method using liquid copper, comparing our results to both experiments and other calculations.

Data mining is a broad discipline that comprises a variety of techniques for extracting meaningful information and patterns from data. It draws on knowledge and “know-how” from various scientific areas such as statistics, graph theory, linear algebra, databases, mathematics, and computer science. Recently, materials scientists have begun to explore data mining ideas for discovery in materials. We will present our recent study of data mining applications in materials. Three techniques: unsupervised clustering, supervised classification, and regression, will be illustrated with materials examples. By mining properties of the constituent atoms, materials research relevant tasks such as the separation of a number of crystals into subsets in terms of their crystal structure, grouping of an unknown crystal into the most characteristically similar peers, and a specific property prediction can be achieved.

This presentation will provide a critical assessment of the role of informatics in materials science. We will show that informatics provides a mathematical foundation that integrates the many facets of informatics including databases, modeling, high throughput experiments and data sharing. The role of informatics in actually making the “materials genome” a reality and not just a conceptual framework for materials science is discussed. It is argued that a “systems biology” paradigm which integrates different length scales (e.g. genomics, proteomics, metabolomics, etc) is more appropriate for materials science and a better guide for advancing materials informatics. Finally we show how by exploiting the full definition of “BigData”, which goes far beyond the issue of simply large data volumes and accounts for other characteristics of data, provides a means for materials informatics to be key for the rational design and discovery of materials.

First-principles high-throughput screening of novel materials requires simultaneously inexpensive and accurate predictive computations of key properties. The first and most difficult challenge is the selection of the appropriate descriptors that are relevant to the material performance, and formulating the computational strategy. We will present two examples of this process in the fields of materials screening in Li-air batteries and new ferroelectric perovskites. The second challenge is the need to establish a materials&’ informatics infrastructure able to automatically prepare and execute calculations on large classes of materials, to monitor calculations, and to store, retrieve and analyze complex data. We accomplish this by integrating storage databases with grid-enabled computational workflow into a powerful flexible environment adaptable to diverse purposes. We are developing and making available this open-source software platform named AIDA (“Automated Infrastructure and Database for Ab-initio design”) to make computational design efforts faster, easier, and fully integrated with automatic data collection and community sharing.

Recently, a technique from the field of signal processing, compressive sensing, has emerged as an efficient and robust way to construct models for describing materials' properties. Compressive sensing exploits the widely-held intuition that the properties of materials can be expressed using a small number of variables. Using this assumption to restrict the solution search results in an efficient way for building very robust models. One way to restrict the model space is through the use of Bayesian inference. In a natural way, Bayesian methods provide error bars on predictions made, a systematic approach for adding data, and noise quantification. We demonstrate Bayesian compressive sensing applied to a cluster expansion model, but the approach is general and could be used in many other model building approaches. This new technique for building materials models, combined with high-throughput {\emph ab-initio} databases, will allow the fast construction of alloy models for hundreds of systems, representing a major step forward in the endeavor to discover the underlying ``genome'' of alloy physics.

During this talk, I will describe our group's efforts in the Clean Energy Project (http://cleanenergy.harvard.edu) , a collaboration with the IBM World Community Grid to search for novel materials for organic photovoltaics and organic electronics using computational resources from volunteer donors around the world. Our project aims to find new materials using techniques from ab initio quantum chemistry combined with cheminformatics tools that are usually employed for the discovery of novel pharmaceutical compounds. To date, we have computed more than eight million structures using first-principles methods, and have analyzed three million using cheminformatics. I will describe our progress so far, and describe immediate goals. A computationally-predicted material with an unusally high hole mobilty of 13 cm2/Vs was synthesized by Zhenan Bao's group at Stanford. I will describe this experimental collaboration as well.

First-principles codes can nowadays provide hundreds of high-fidelity enthalpies on thousands of alloy systems with a modest investment of a few tens of millions of CPU hours. But a mere database of enthalpies provides only the starting point for uncovering the "alloy genome." What one needs to fundamentally change alloy discovery and design are complete searches over candidate structures (not just hundreds of known experimental phases) and models that can be used to simulate both kinetics and thermodynamics. Despite more than a decade of effort by many groups, developing robust models for these simulations is still a human-time-intensive endeavor. We have developed an automatic framework for extracting cluster expansion-based models from a large database (www.aflowlib.org) of alloy enthalpies. Such a framework will uncover, in a general way across the periodic table, the important components of such models and reveal the underlying "genome" of alloy physics.

The replacement of expensive or environmentally detrimental material building blocks with abundantly available building blocks is one of the pressing questions to be addressed by the materials science community. We present a mathematically precise formulation to solve the so-called building block replacement problem, developed for general systems and applied here to hierarchical materials. We use material ologs, category-theoretic descriptions of hierarchical materials, to express structure-function relationships of arbitrary complexity. We exemplify this approach for the case of a laminated composite, chosen for its analytical simplicity. With a rigorous mathematical description in hand, we show how to solve the building block replacement problem in this system. That is, we define concrete conditions under which we can exchange certain components of the structure without compromising the system&’s overall functionality. Our work shows that material ologs offer a powerful strategy to capture the complex interactions and relationships that constitute hierarchical materials and represent them in a formal and unified structure, also allowing for the rigorous translation of material concepts from one domain to another.

A significant advantage of computations is that they can be automated, thereby enabling the computational screening of thousands of compounds. Such high-throughput computing has recently led to the Materials Genome idea of trying to calculate properties of all known inorganic materials. One problem that faces such high-throughput calculations is the fact that no single density functional theory approximation is valid across all of chemical space. The inconsistency of errors in materials with very different bonding can lead to large and systematic errors in phase stability prediction. To circumvent this problem we have developed a general formalism to combine the results from different functionals each applied in a distinct chemical space. This approach is general and allows for example to mix GGA with GGA+U calculations, or experimental data and computed data. Until functionals are available that are accurate across diverse chemical spaces, such an approach will be critical to advance high-throughput phase stability calculations. I will demonstrate this approach by showing results from a large-scale accuracy test in predicting the formation energy of ternary oxides. The high-throughput infrastructure has now been applied to the discovery of Li-battery cathodes, novel photocatalyst materials, and Na battery materials, and I will show some examples from these areas. A. Jain, et al, Physical Review B, 84 (4), 045115 (2011) G. Hautier, et al, Physical Review B, 85 (15), 155208 (2012).

High-throughput density functional theory calculations enable the computational screening of materials properties over a wide range of compounds and compositions. However, the data generated in high-throughput simulations is limited. For example, only a small fraction of multi-component alloys can be covered because of the combinatorial explosion of the number of possible structures and compositions. Furthermore, a direct interpretation of chemical and structural trends from the density functional data is not possible, such that often simplified models of the electronic structure help to gain insight into the bond chemistry of the material. In this talk I will discuss two focus points of our current activities. The formation of topologically close-packed (TCP) phases in Ni-based superalloys may be understood from a systematic coarse-graining of the electronic structure and explained by a few physically intuitive parameters. The structural stability of multi-component thermoelectric materials may be estimated by making use of similarities in the bond chemistry of the binary subsystems.

Metal organic frameworks (MOFs) have attracted significant attention as nanoporous solid sorbants for cost-effective CO₂ capture from fossil fuel combustion. MOFs are composed of metal ions and organic linker groups that form crystalline porous networks. Design of these materials, however, remains one of the greatest challenges in the field due to an almost infinite combination of linking organic groups and the functionalization of the pores. We have carried out high throughput screening with both DFT and parameterised models to assess hundreds of thousands of combinatorially generated hypothetical MOFs for their suitability for CO₂ capture. Several promising synthetic targets have been identified with higher capacities and affinities for CO₂ than high performance materials previously reported. In addition to identifying synthetic targets from this large scale screening, we are also developing effective chemoinformatic tools to discover general structure-performance relationships and analyze the local chemical environments of the CO₂ binding sites in these enormous data sets. Computational machine learning techniques combined with chemoinformatic tools adapted to mine relevant knowledge from large data sets of nanoporous materials will be presented.

Predictive modeling of strength from first-principles electronic structure methods offers great promise to inform Mg alloy design. Simulating the mechanical behavior for new alloys requires an understanding of mechanisms for deformation at atomic-length scales, with accurate chemistry, extended to larger length- and time-scales. Modern computational approaches can now investigate dislocations from first-principles, and compute interactions with solutes across the periodic table. We can predict metallurgical trends with changes in size and chemical misfits, and connect those to predictions of mechanical behavior through predictive models, including solute strengthening and thermally-activated cross-slip. Comparing alkali, alkali earth, and transition metals with rare earth solutes provides new connections between electronic structure and mechanical behavior. Moreover, the computational approach provides a blueprint for attacking new challenges in deformation behavior beyond solute strengthening and softening.

The Materials Project (www.materialsproject.org) aims to leverage the information age for materials using the only tool that can efficiently scan multiple materials properties in a reasonable amount of time: computations. While ab initio computations have already started to show promise for accelerating the traditionally slow development process for new materials, integration with web-based free dissemination and a user-dynamic workspace will lead to a new paradigm for how materials science is performed. Today, the Materials Project contains over 20,000 inorganic compounds, which are freely available for searching and modifications. Through our project, we hope to give both experimentalists and theorists access to materials properties of all known inorganic compounds and beyond to scan, analyze and provide inspiration for novel materials development. In this talk we will give examples of high-throughput materials design projects within the energy storage space. Furthermore, we will also present the software infrastructure and high-throughput workflow algorithms behind the Materials Project as well as demonstrate some of its existing materials design capabilities, such as phase diagram generator, reaction calculator, structure predictor, Li-ion battery analytics, etc. Future directions for a collaborative efforts and data generation and analysis through crowd sourcing will be given.

First-principles, high-throughput screening of novel materials relies on our ability to tackle three challenges of very different origin. In order of increasing complexity, we need to establish a materials&’ informatics infrastructure able to deal automatically with calculations done on entire classes of materials, to monitor calculations as appropriate, and to store, retrieve and analyze very heterogeneous sets of microscopic and macroscopic materials&’ data. Second, we need to be able to calculate materials properties that may have a very complex microscopic origin, and for which there is not an algorithmic or even a theoretical framework available or practical, so we need to nurture the human and theoretical skills required to accomplish this while establishing effective protocols for verification. Last, and most critically, we need to achieve meaningful, predictive accuracy in first-principles calculations for realistic systems, and be able to validate it. Depending on the problem at hand one might argue that all, some, or even none of these challenges are currently met. I will focus here on thermal properties of materials - from thermodynamic stability of perovksite oxides, to thermal conductivity of bulk and nanostructured thermoelectrics, showing how second- and third-order density-functional perturbation theory can provide reliable estimators of stability and performance. This effort is embedded into open-source projects available and used by the community at large, for quantum simulations (www.quantum-espresso.org) and transport calculations (www.wannier.org), and organized through AIDA (“Automated Infrastructure and Database for Ab-initio”).

For many materials and molecules, accurate property predictions are feasible from quantum-mechanical first principles alone. However, there is a key prerequisite: Knowledge of the realizable structure or structural ensembles. For instance, the flexible, environment-dependent conformations of peptide molecules and proteins govern essentially all their function. Predicting the relevant conformations out of the combinatorial explosion of possible ones from first principles is a challenge that can only be addressed by high-throughput, largely automated search and screening techniques. We here use a van der Waals corrected density functional theory approach for peptide molecules that are one to twenty aminoacids long (up to 220 atoms): (i) a series of polyalanine-based peptides Ac-Ala_n-LysH⁺ (n=4-8) in direct comparison to gas-phase experiments, (ii) a "glass-forming" peptide Ac-LysH⁺-Ala₁₉, (iii) an exhaustive benchmark database of the conformations of all 20 biological aminoacids in the presence of divalent cations (Ca²⁺, Sr²⁺, Ba²⁺ Pb²⁺, Cd²⁺, Hg²⁺). We follow a two-step approach: First, collect a large number of force field based structure candidates (10⁵ or more for peptides of 6 or more amino acids); second, follow up with first-principles post-processing for thousands of candidates. For cases (ii) and (iii), we show that an additional, first-principles only refinement by replica exchange molecular dynamics runs is both effective and essential, leading to trends in excellent agreement with available experiments.

The Human Genome Project in Biotechnology has received increasing attention in the materials community through programs such as the Materials Genome Initiative by the White House and the Scientific Discovery through Advanced Computing with the Department of Energy. These programs seek to enhance the processes of discovery and development of less expensive materials by using materials data and developing computational infrastructures. In this talk, we will discuss a specific informatics approach for developing advanced devices for solar energy. In particular, we will demonstrate the value of the multi-way statistical technique, N-way Partial Least Squares, that generates a multi-linear model using all of the reliability data simultaneously for study on the photovoltaic loss mechanisms of n-CdS/CdTe heterojunction solar cells. Here, our goal is to rapidly construct analytical framework for identifying key factors related to cell degradation throughout accelerated lifetime testing. With the multi-way approach, we are able model variables of interest such as cell efficiency while representing the data in a lower-dimensional space in which salient features are more easily identified. We track the multiple associations between multi-way capacitance-voltage parameters and a performance metric, fill factor, for the study of photovoltaic degradations. Even with the inclusion of a noisy parameter, and with a relatively small number of cell devices, we will show how to effectively identify key factors that are highly related to performance degradation. As one of the materials informatics tools, the multi-way approach addresses critical needs in Materials Genomics. It includes integration of heterogeneous and multi-scale data, extraction of knowledge from it, and incorporation of the comprehension into the design of enhanced processes, enabling data-intensive PV materials/devices-by-design.

An automated high-throughput code has been applied to tackling combinatorial diffusion-based problems in two very different systems. This code is both versatile and general, using a keyword based input file in order to specify both the system information and the VASP run parameters. The code gives complete control over the run parameters at each stage of the calculation and provides automatic status checking, submission, and resubmission of each stage, and convenient user access to the active runs and their statuses. Data is organized into SQL tables, which can be queried to produce useful data subsets. Coordinate combinations can be plotted and regressed using a tool developed for MATLAB, producing a visual and numerical means of discovering correlations between coordinates. Results have been generated for vacancy formation and migration energies in simple elements and for the oxygen sublattice in potentially promising oxygen conducting perovskites. The perovskite data is analyzed with partial least squares to show how a few simple geometric descriptors can provide a surprisingly good representation of the migration energetics, providing both insight into oxygen migration in these materials and enhancing rates of ab initio screening for fast oxygen conductors.

Recently there have been significant advancements in developing novel high-energy molecules (HEMs). A range of new HEMs and relevant additives have been synthesized and applied to civil and military research. The high power of those molecules can be attributed to the intriguing molecular structure and high nitrogen content. On the other hand, the highly sensitive nature of HEMs requires great caution when charging and handling. Researchers who design and synthesize new HEMs have pushed themselves to derive more powerful, yet safe HEMs. However, it is difficult to find promising HEM candidates because more powerful HEMs are generally sensitive, and vice versa. Owing to this inversely proportional tendency between performance and insensitivity, the search a promising novel HEM requires good strategies, not just a trial and error approach. It is important to understand the various molecular aspects and pinpoint the molecular aspects to enhance either the explosive performance or safety, for example, like the molecular design in drug discovery. So it is extremely important to select good candidate molecules in the early stages of development. One good way of identifying promising candidates is to have a good DB, which archives all previous information including the failures and possible virtual candidates, such as ICT thermochemical database which contains more than 12,000 compounds, although it only archives experimental values of known molecules. We developed the new on-line management system for HEMs, called MS-HEMs. It is designed to collect and store virtual HEMs as well as known HEMs to derive a new successful HEM, contrary to other previous HEM DBs including ICT thermochemical DB. With this feature, the on-line MS-HEMs can manage virtual HEMs designed by theoretical scientists as well as existing HEMs synthesized previously. Owing to the virtual molecules dealt with this DB, a large proportion of input records consist of molecular descriptors that can be obtained from quantum mechanical calculations. The user-friendly designed input screen and the search routine, where the users can find the exact HEM and its properties either by text/value or by 2D-chemical drawing, allow users easy to access or handle it. The DB has own built-in calculation routine, where a range of compositional and topological descriptors, heat of formation, and crystal density can be computed automatically when a new HEM is inputted into the MS-HEMs. In addition, the impact sensitivity can be calculated using in-house ANN codes. The on-line DB for HEMs is a potential molecule pool that combines real and virtual HEMs, and provides important knowledge in locating good HEM candidates with a sufficiently safe nature. This study was financially supported by ADD and the Defense Acquisition Program Administration(DAPA).

Choosing a promising candidate is of great importance in getting good high energy density materials (HEDMs), since the synthesis of new HEDMs requires a great deal of effort and takes much longer time. Rapid virtual screening is probably one of the best ways in selecting promising candidates. In order to predict explosive performance and safety character of new explosive molecules starting from molecular structures alone, Agency for Defense development (ADD) in South Korea developed a new systematic procedure which we called ADD Method-1 [1]. ADD Method-1 includes three theoretical steps, i.e. (1) ab initio calculation of molecular structure and energy, (2) computation of molecular descriptors, and (3) estimation of explosive performance and sensitivity. However, ADD Method-1 required laborious computational procedures to find the global minimum in a potential energy surface, and relevant molecular descriptors. Our new approach, ADD Method-1(2D), excluded conformational search procedures, and obtained heat of formation, density, and other molecular descriptors from simple compositional, fragmental, topological descriptors, which were automatically calculated in our HEDM database, MS-HEMs [2]. ADD Method-1(2D) will serve as a useful tool in deriving novel HEDM candidates in an efficient way. 1. Cho, S. G. A Systematic Procedure to Predict Explosive Performance and Sensitivity of Novel High-Energy Molecules in ADD, ADD Method-1, In Handbook of Material Science Research, Charles René and Eugene Turcotte, Editors, Nova Science Publishers, Inc., New York, 2010, Chapter 11, pp.417-431. 2. Lee, S. K.; Cho, S. G.; Park, J. S.; In, Y. Y.; No, K. T. MS-HEMs: An On-line Management System for High-Energy Molecules at ADD and BMDRC in Korea. Bull. Korean Chem. Soc. 2012, 33, 855-861.

We performed a virtual screening of nitroimidazole derivatives with various functional groups in 1-position with ADD Method-1, a theoretical procedure to evaluate explosive performance and sensitivity of new explosive molecules developed in ADD, South Korea [1]. Imidazole moiety is considered to be an excellent skeleton for high energy density molecules due to high nitrogen content and aromatic nature [2]. Molecular structures and chemical properties of nitroimidazole derivatives have been investigated at high levels of density functional theories. Explosive performances and impact sensitivities have been estimated at the global minimum of potential energy surface. As more nitro groups are introduced, the explosive performances of nitroimidazole derivatives are enhanced, while the impact sensitivity becomes more sensitive. A two-dimensional plot between explosive performance and impact sensitivity has been utilized to comprehend the technical status of new explosive candidates. Based on locations in the two-dimensional plot, the upper limit of nitroimidazole derivatives in explosive performance is close to that of HMX, which is one of the most powerful explosive molecules widely used in current military application. However, tetranitroimiadzole is predicted to be too sensitive to use as the main ingredient of the secondary explosive formulations. 1-Aminodinitroimidzole isomers appears to have a potential to be good candidates for insensitive explosive formulations. 1. Cho, S. G. A Systematic Procedure to Predict Explosive Performance and Sensitivity of Novel High-Energy Molecules in ADD, ADD Method-1, In Handbook of Material Science Research, Charles René and Eugene Turcotte, Editors, Nova Science Publishers, Inc., New York, 2010, Chapter 11, pp.417-431. 2. Su, X.; Cheng, X.; Meng, C.; Yuan. X. Quantum Chemical Study on Nitroimidazole, Polynitroimidazole and Their Methyl Derivatives. J. Hazard. Mat. 2009, 161, 551-558.

Up to now, various kinds of combinatorial high-throughput materials exploration process have established for film, powder, polymer preparation and so on. Our group have hitherto developed the high-throughput apparatus "M-ist Combi" for finding functional powder/film using the solution process based on the electrostatic spray deposition (ESD) method. The "M-ist Combi" system has function of not only materials preparation by ESD method but also evaluation system of electric property by changing the triaxial robot hand from the spray nozzle to the four point probe. And, we can evaluate various physical property by development of new equipment which can attach to the "M-ist Combi" system. As an example, we prepared VO2 film library (16 samples) and evaluated thermochromic property using the above system. And, it enables us to evaluate 16 samples on the library within 2 hours. In the conventional measuring method such as two-terminal method, the measurement of only one sample had taken about 4 hours. However, the big difference arose in the resistance value of 16 samples. In relation to the measuring of resistance value by the four-point resistance method, the value of each sample varies by surface area, thickness, and the position to measure. From perspective of informatics of high throughput materials exploration, only the critical temperature from semiconductor to metal condition is important information for finding the candidate thermochromic materials. In this study, I will introduce some kinds of exchangeable evaluation probe and high throughput evaluation system, and treatment method how to take available information for finding candidate functional materials from enormous quantity data.

Metal-organic frameworks (MOFs) are a class of porous solids formed by the self-assembly of structural building units - metal ions (or clusters) and polydentate organic linkers. The limitless combinations of metals and ligands can yield diverse structural and functional properties that can in principle be tuned by modification of the building units. MOFs have been targeted as potential enabling materials for vehicular storage of gaseous fuels and for energy efficient carbon capture and sequestration from fossil fuel combustion. However, the large-scale discovery of MOFs with enhanced performance brings about a combinatorial design challenge that demands efficient experimental and computational screening tools. While, high throughput (HT) screening and combinatorial chemistry techniques are extensively used in drug design, the implementation of HT screening technologies for nanoporous material discovery is in an early stage. In contrast, advances in the computational simulation of gas adsorption properties of porous materials have recently enabled the in-silico HT screening of gas uptake parameters by grand canonical Monte-Carlo simulations. In this context, we are developing sophisticate chemoinformatic tools to retrieve relevant knowledge from large-scale virtual screening data on gas uptake parameters of MOFs. Here, we reported two webserver applications that serve as predictor of the gas adsorption capacity of MOFs (GADMOF) and gas binding site locator (BISILO) in MOFs. Both webservers use structural information and Support Vector Machines (SVMs) learning to build quantitative-structure-property relationships (QSPR) models of the gas adsorption capacity and analyze the local chemical environments of the gas binding sites in the data sets. To build the QSPR models, ~130,000 and ~30,000 structures of the hypothetical MOFs in the Northwestern University database were in silico screened for methane and CO₂ uptake, respectively. The structural features of the hypothetical MOFs were mapped to their theoretical gas adsorption capacities at different pressures using SVMs. While the methane storage capacity correlates very well with pore size, surface area and void fraction of the MOFs with R² values > 0.85, the CO₂ uptake capacities exhibited very poor correlations with these structural features. However, when training the SVMs with the radial distribution functions computed from the spatial distribution of atoms in the MOF, the QSPR models yielded the most accurate predictions with R² values of 0.70. The machine learning approach also exhibited great potential to recognize CO₂ binding sites in MOFs with preliminary accuracy of 75%. Our results indicate that the methane storage capacity in MOF is mainly control by the shape and size of the pores while the CO₂ adsorption is mainly driven by the chemistry of the pores. The webservers are freely available online at http://titan.chem.uottawa.ca/gadmof and http://titan.chem.uottawa.ca/gadmof/BISILO.

It is a timely goal in the biological and materials sciences to computationally design novel compounds that exhibit specific chemical properties and are straightforward to synthesize. Some of the most relevant and promising materials properties depend explicitly on atomistic details, rendering an atomistic resolution of any employed simulation model mandatory. Alas, even when using high-performance computing, brute force high-throughput screening of all the possible compounds is beyond any capacity for all but the simplest systems and properties due to the combinatorial nature of chemical compound space (compositional, constitutional, and conformational isomers). Consequently, when it comes to properties or systems that require first principles calculations, a successful optimization algorithms must not only make a trade-off between sufficient accuracy of applied models and computational speed, but must also aim for rapid convergence in terms of number of compounds "visited". I will discuss recent contributions related to the former aspect. Molecular graph based quantitative structure property relationships for charge hopping rates of poly-aromatic hydrocarbons will be presented [1]. Thereafter, kernel ridge regression and neural network models will be introduced that permit rapid prediction of electronic structure properties (atomization energies, HOMO/LUMO, polarizability, excitation energies) with an accuracy similar to density functional theory [2]. Finally, I will discuss necessary and desired requirements of molecular descriptors, as well as novel approaches, that are particularly well suited for machine learning in chemical space [3]. [1] M. Misra, D. Andrienko, B. Baumeier, J-L. Faulon, OAvL, J Chem Theory Comput (2011); [2] M. Rupp, A. Tkatchenko, K.-R. Muller, OAvL, Phys Rev Lett (2012); G. Montavon, M. Rupp, V. Gobre, A. Vazquez-Mayagoitia, K. Hansen, A. Tkatchenko, K-R. Mueller, OAvL, submitted (2012); [3] OAvL, A. Knoll, to be submitted (2012).

We live in an era of data explosion. This is especially true in the domain of materials. Voluminous heterogeneous datasets are being continuously generated. Extracting relevant information from these datasets is a formidable challenge. To process voluminous datasets we need computational algorithms that are both time and memory efficient. In particular, out-of-core algorithms that employ parallel computers are called for. In this talk we give a brief introduction to computational techniques that are vital for Material Genome. In particular we will talk about out-of-core algorithms, parallel computing, data mining, and learning techniques. The analyses involving materials data are not only memory intensive but also time intensive necessitating the employment of parallel architectures. To make a parallel program efficient, it is essential to optimize both computation and communication. We have invented a technique called LessTalk that can be used to reduce communication costs. The idea of LessTalk is to perform some redundant computations in order to reduce the number of communication steps. We have employed this technique in the context of simulation of coastal waves, fuel cells, and biological cells to obtain superlinear speedups on CPU clusters. We will briefly introduce LessTalk in this talk. Given the volume of data involved, the core memory of a typical desktop, a server, or even a parallel computer may not be enough to hold all the data to be analyzed. Therefore, efficient out-of-core computing techniques are necessary. With advances in storage technology, core memory and secondary memory sizes on computers are ever-increasing. Unfortunately, the sizes of datasets are increasing at a much faster pace. A memory hierarchy can be expected to always exist on computers. As a result, gaps will remain across the different levels of the hierarchy. Performance can be vastly improved if we learn to wisely manage data across the different levels. This is precisely what we mean by out-of-core computing. In this talk we will also introduce some basic out-of-core computing ideas. One way of extracting useful information from large datasets is using data mining techniques. There are several forms of data mining that are currently being employed. We introduce clustering and association rules mining in this talk. We also will say something about text mining. Text mining refers to the problem of automatically processing millions of articles (i.e., text files) and identifying a short list of articles that are likely to contain information of a desired kind. Some learning and data reduction techniques will also be touched upon in this talk.

Intuition suggests that important properties of materials are primarily determined by just a few key variables. Examples of such cases are numerous. For instance, most magnets can be described using a Heisenberg model with only a few non-zero parameters, alloy thermodynamics can be treated using truncated cluster expansions with a small number of short-ranged interactions, and models of bulk dynamics require interatomic potentials with a few short-ranged terms. Traditionally, physical intuition has been used to aid the construction of these models. However, this intuition often does not exist or is difficult to develop, and often there is no clear path to achieve systematic improvement. We show that a recently developed technique in the field of signal processing, compressed sensing (CS), provides a simple, general, and efficient way of constructing physics-based models. CS is a new paradigm for model building - its models are just as robust or better than those built by current state-of-the-art approaches, but can be constructed at a fraction of the computational cost and user effort.

Over the last few decades, the tools of materials research have become significantly more sophisticated, allowing for the rapid synthesis and characterization of large numbers of samples. As a result, materials researchers can now collect sample characterization data faster than they can analyze it. This disparity in data collection and analysis time is fueling interest in new machine learning algorithms, also known as data-mining techniques, for accelerating data processing. In this talk we will discuss two algorithms that can be utilized to quickly sort data from combinatorial libraries spanning large composition phase spaces. We will also show how crystal structure data from crystallographic databases can be used to improve the performance of these algorithms. This talk will focus on X-ray diffraction data obtained from Fe-Ga-Pd and Fe-Co-Ni thin-film ternary composition spreads.

After a brief and gentle introduction into kernel based machine learning (ML), I will review several recent successful applications of the ML framework to atomistic simulations [1,2,3]. Subsequently I discuss what lessons can be learned from these different applications; what are methodological limits and open ends with respect to scaling, accuracy and further application fields in quantum chemistry. [1] Rupp, M., Tkatchenko, A., Muller, K.-R., von Lilienfeld, O.A., Fast and Accurate Modeling of Molecular Energies with Machine Learning, Physical Review Letters, 108, 058301 (2012) [2] Pozun, Z.D., Hansen, K., Sheppard, D., Rupp, M., Muller, K.-R., Henkelman, G., Optimizing transition states via kernel-based machine learning, Journal of Chemical Physics, 136, 174101 (2012) [2] Snyder, J., Rupp, M., Hansen, K., Muller, K.-R., Burke, K., Finding density functionals with machine learning, Physical Review Letters, (2012), to appear This is joint work with O. Anatole von Lilienfeld, Alexandre Tkachenko, Gregoire Montavon, Katja Hansen, Matthias Rupp, Kieron Burke, John Snyder, Graeme Henkelman, Zachary Pozun, Dan Sheppard, Andreas Ziehe, Siamac Fazli, Franziska Biegler and many others. Support from DFG, EU, BMBF and by the World Class University Program through the National Research Foundation of Korea funded by the Ministry of Education, Science, and Technology, under Grant R31-10008 is gratefully acknowledged.

Predictions of structure formation by computational methods have the potential to accelerate materials discovery and design. We present two computational approaches for the prediction of crystal structures and the morphology of nanoparticles. First, many materials properties are controlled by composition and crystal structure. Finding the ground state composition and crystal structure of a material is a difficult problem because of the large number of minima in the potential energy surface. We developed an approach based on evolutionary algorithms coupled to ab-initio relaxations that accurately predicts the arrangements of atoms into crystal structures and the composition of the ground state phases without any prior information about the system. We will describe our genetic algorithm for structure prediction (GASP, http://gasp.mse.cornell.edu) and present applications of the approach to energy storage materials, such as the the battery material Li-Si. Second, the self-assembly of nanocrystals into mesoscale superlattices provides a path to the design of materials with tunable electronic, physical and chemical properties for various applications. The self-assembly is controlled by the nanocrystal shape and ligand-mediated interactions between them. To understand this, it is necessary to know the effect of the ligands on the surface energies (which tune the nanocrystal shape), as well as the relative coverage of the different facets (which control the interactions). We will discuss how calculations of ab-initio surface and ligand-binding energies for PbSe nanocrystals predict the equilibrium shape of the nanocrystals and a transition from octahedral to cubic when increasing the ligand concentration during synthesis. Our results furthermore suggest that the experimentally observed transformation of the nanocrystal superlattice structure from fcc to bcc is caused by the preferential detachment of ligands from particular facets, leading to anisotropic ligand coverage.

Organic solar cells have the potential for widespread usage due to their low cost-per-watt and mechanical flexibility. Their wide spread use, however, is bottlenecked primarily by their low solar efficiencies. Experimental evidence suggests that a key property determining the solar efficiency of such devices is the final morphological distribution of the electron-donor and electron-acceptor constituents. By carefully designing the morphology of the device, one could potentially significantly enhance their performance. This is an area of intense experimental effort that is mostly trial-and-error based, and serves as a fertile area for introducing mechanics and computational thinking. We use a data-driven approach that uses ideas from data-mining, homology and graph theory to establish process-structure-property relationships with ultimate goal to design more efficient organic solar cells. We design a two stage approach to this problem. In the first stage, we develop a purely mechanistic based description of the performance of the device by characterizing performance using carefully chosen morphology descriptors. This mechanics based ‘surrogate model&’ that links morphology with photovoltaic performance is built using (a) concepts from graph theory, (b) and the underlying equivalence between a two phase morphology and an undirected, weighted graph. In this stage, we correlate morphological features with properties of the device (determined using full-scale analysis) and rank them according to their relative importance to the efficiency. To interrogate large set of morphologies we develop framework that is based on a MapReduce algorithm (used by Google, Amazon). The inherent parallelism of MapReduce and its support for cloud computing make it a very attractive platform to enable efficient and fault tolerant analysis. The MapReduce strategy allows us to automatically characterize and extract morphology-performance trends. It also allows to identify set of morphology descriptors that the best encode the performance of the devices. In the second stage, we utilize the surrogate modes along with the set of the most-physically meaningful morphology descriptors to define the cost function of the inverse problem. We develop a massively parallel genetic algorithm that enables the inverse design of morphologies that maximize certain performance criterion. We illustrate the differences in optimized morphologies resulting from 2D and 3D assumptions.

Modern computational modeling enables the study and development of novel organic, inorganic and hybrid photovoltaic structures to achieve the desired performance and cost objectives. The power of this tool is however limited by the availability of information on materials, their combination, and their behavior in different conditions. As materials science research publishes more and more information on these materials, computation modeling tools are not able to explore them directly for the development of new structures. Therefore, data mining and information retrieval techniques are being used for the extraction of this information from scientific papers and archiving it in dedicated databases [1]. In this situation, the problem is that such techniques retrieve data from the manuscripts published, and more specifically from the text, while information presented graphically remains not retrieved. To address this issue, we have developed an electronic device which enables extracting graphical data [2]. In parallel with that, because of the lack of computation packages that use data mining and information retrieval techniques in this area, we have developed a software package which is able to extract information from peer-reviewed papers and incorporate it directly into the software data base. By combining these tools, we are now having access to more information about various materials, and are able to use it for the study and development of novel photovoltaic devices. For the demonstration of this solution, we have made a prototype which enables the modeling of several compositions of II-VI and III-V semiconductor materials for the development of novel single-junction, multi-junction and spectral beam-splitting solar photovoltaic devices. [1] S. Liao, P. Chu and P. Hsiao, Expert Syst. Appl. 39 (2012) 11303. [2] A. Mokri, M. Emziane, US Patent Application 13/459,996.

Evaluation of biological effects of materials and nanomaterials (NMs) using in silico approaches is of critical importance to enable rational design of materials with the desired properties. We review the potential of modern cheminformatics methods such as Quantitative Structure - Property Relationship (QSPR) modeling to develop statistically significant and externally predictive models that can accurately forecast biological effects of materials from their physical, chemical, and geometrical parameters. We discuss major approaches for model building and validation using both experimental and computed properties of materials. We illustrate these concepts with case studies for which we successfully built and validated predictive models. We argue that similar to conventional applications of QSPR modelling for the analysis of bioactive organic molecules, this approach can be successfully used for (i) predicting activity profiles of novel materials solely from their representative descriptors and (ii) designing and manufacturing safer materials with the desired properties. We will discuss (i) the need for integrated databases compiling all published information concerning NMs, especially their measured physical/chemical properties and associated biological profiles; (ii) the challenges of developing new descriptors of materials and NM; and (iii) the development of statistically significant and externally predictive Quantitative Nanostructure-Property Relationship (QNAR) models. We will also present the first study reporting the computer-aided design of new surface-modified carbon nanotubes with the desired bioactivity and safety profiles: (1) we have successfully developed robust QNPR models for 84 carbon nanotubes decorated with organic surface modifiers; (2) these models were further applied for screening a library of 240,000 ligands potentially attachable to carbon nanotubes and identifying the ones giving desired bioactivity and toxicity properties; (3) selected nanotubes decorated by the ligands prioritized by our QNAR models were experimentally synthesized and validated by our collaborator, Dr. Bing Yan at St. Jude Children Research Hospital. Our study demonstrates how QNPR models can be used to predict activity profiles of NMs and bias the design and manufacturing towards better and safer products.

For the computational screening of novel materials we need methods that are faster than density-functional theory (DFT) while remaining sufficiently accurate for quantitative predictions. The semi-empirical tight-binding (TB) method makes it possible to simulate structures with tens of thousands of atoms while still providing insight into the electronic structure. TB models for practical calculations are usually derived by fitting band structures and total energies to reproduce experimental data or results of DFT calculations. We have developed an alternative approach in which the TB parameters are determined directly from DFT wave functions of arbitrary reference configurations without extensive fitting. Our method [1] is conceptually different from previous approaches [2] as it is based on a projection [3] of basis-set-converged wave functions from mixed-basis DFT computations onto a minimal basis of atomic orbitals. The radial shape of the atomic orbitals is optimized by minimizing the loss (spillage) in the projection procedure. The Slater-Koster tables are then calculated with the optimized minimal basis using the self-consistent DFT Hamiltonian. In conventional TB models the intra-atomic (on-site) matrix elements of the Hamiltonian are approximated by constant atomic values and crystal field splittings are neglected. Using our projection technique it is, however, straightforward to analyze also the on-site integrals. We find that the values of these matrix elements may depend strongly on the atomic environment, and neglecting the crystal field splittings leads to significant errors in the electronic structure. We therefore introduce a new crystal field TB (CF-TB) scheme which allows to take the environmental dependence of the on-site terms into account.[4] Comparisons of the CF-TB band structures and densities of states for a representative set of benchmark cases with DFT reference results demonstrate the accuracy and an enhanced transferability of the CF-TB parametrization. [1] A. Urban, M. Reese, M. Mrovec, C. Elsaesser, and B. Meyer, Phys. Rev. B 84 (2011) 155119. E.R. Margine, A.N. Kolmogorov, M. Reese, M. Mrovec, C. Elsaesser, B. Meyer, R. Drautz, and D. Pettifor, Phys. Rev. B 84 (2011) 155120. [2] D. Porezag, Th. Frauenheim, Th. Koehler, G. Seifert, R. Kaschner, Phys. Rev. B 51 (1995) 12947. [3] D. Sanchez-Portal, E. Artacho, J.M. Soler, Sol. State Comm. 95 (1995) 685. [4] A. Urban, C. Elsaesser, and B. Meyer, in preparation.

Materials informatics offers the brave new world of functionally and materially optimized systems. However the jounrney to that world, involves the complex generation of systems and platforms that can support the interaction with, the generation of, the storage, the analysi of and the data. These applications can derive and integrate vast fields of diverse data, so an agnostic platform is needed as well as the cabolity to present this in a simple unified way to consumers of the data, derive value added compuitation as well as experimental analyses and to use evolutionary or hpping approaches to find new possible combinations of componenets, levels or materials that satisfy a tradeoff non singular design space. Further this capability then needs to link or connect inot the quality management and hte product lifecycle that occurs in many commercial organizations

Molecular dynamics simulations of large systems critically depend on the accurate description of the underlying potential energy surface (PES). While first-principles methods such as density-functional theory (DFT) can provide very accurate energies and forces, they are computationally too demanding to address many interesting systems. High-dimensional Neural Networks (NN) trained to first-principles data have been shown to provide accurate PESs for systems containing a single atomic species, while being many orders of magnitude faster than conventional DFT.[1-3] We have generalized this method to multicomponent systems with arbitrary chemical composition including long-range interactions and charge transfer.[4] Here we demonstrate the capabilities of the NN method for three different material classes: a transition metal (copper), a semi-conducting metal oxide (zinc oxide), and a small molecule (methanol). We report a number of properties like structural energy differences, vacancy formation energies, and surface energies for different copper and zinc oxide surfaces. We find that the predicted geometries, energies, forces, and atomic charges are in excellent agreement with reference DFT calculations, making NN potentials promising candidates for large-scale simulations. [1] J. Behler, and M. Parrinello, Phys. Rev. Lett. 98 (2007) 146401. [2] J. Behler, J. Chem. Phys. 134 (2011) 074106. [3] N. Artrith, and J. Behler, Phys. Rev. B 85 (2012) 045439. [4] N. Artrith, T. Morawietz, and J. Behler, Phys. Rev. B 83 (2011) 153101.

We present a method for leveraging density functional theory (DFT) calculations to carry out large-scale molecular dynamics (MD) simulations without the use of empirical interatomic potentials. The atomic system is decomposed in to smaller clusters of atoms centered on a particular atom of interest. A database of these cluster configurations and the resulting forces generated by fully ab-initio calculation is generated dynamically. During the large scale MD simulation, the database is continually queried and updated to find the most appropriate forces on each atom in its local environment and store the forces resulting from new configurations. A metric for comparing cluster geometries is proposed and is shown to uniformly converge with database size. Using this metric, an adaptive multilevel sorting and indexing technique that runs in parallel with the MD simulation allows for significantly faster force evaluation. We compare our method against fully ab-initio MD as well as MD using a reactive bond order interatomic potential for a (10,0) carbon nanotube. The method can be applied to treat complex systems of thousands of atoms composed of elements anywhere on the periodic table with near DFT-level accuracy.

The easy access to cheap fossil fuels is a necessary condition to keep up the high living standard created in the world during the last century. The development of sustainable energy forms is one of the most important problems of our time because of the ever increasing energy consumption together with the CO₂ related climate problems. The conversion of solar light into electrons and holes and their subsequent collection at spatially separated regions is one of the possible ways to address the world's pressing energy supply and storage problem. The properties determining the usefulness of a material to be used as light harvester in a photochemical cell include (i) a narrow band gap allowing the utilization of a significant fraction of solar spectrum, (ii) good mobility allowing electrons and holes to reach the surface and reduce/oxidize the targets before recombining, and (iii) chemical/structural stability under irradiation [1]. In addition, low cost and nontoxicity are necessary properties for an eco-friendly material. We focus on three main applications: (i) single- and (ii) two-photon water splitting [2, 3] and (iii) protection against photocorrosion [3] using a recently implemented DFT-functional, called GLLB, that gives good results for the evaluation of the band gap [4]. We find 20 and 12 promising materials for visible light harvesting in the single- and two-photon schemes and 15 for the transparent shielding in the space of 20000 cubic perovskites obtained by combining 52 metals together with oxygen, nitrogen, sulfur and fluorine as anions. We also apply the screening procedure for more complex structures, like double and layered perovskites. We thus investigate the structural and electronic properties of the new combinations with special attention to the possible applications related to light harvesting. A database, called Computational Materials Repository [5], has been developed for easy access and analysis of the data. References [1] A. Kudo, and Y. Miseki, Chem. Soc Rev. 38, 253 (2009). [2] I.E. Castelli, T. Olsen, S. Datta, D.D. Landis, S. Dahl, K.S. Thygesen, and K.W. Jacobsen, Energy Environ. Sci., 5, 5814 (2012). [3] I.E. Castelli, D.D. Landis, K.S. Thygesen, S. Dahl, I. Chorkendorff, T.F. Jaramillo, and K.W. Jacobsen, PrePrint. [4] M. Kuisma, J. Ojanen, J. Enkovaara, and T.T. Rantala, Phys. Rev. B 82, 115106 (2010). [5] Computational Materials Repository, https://cmr.fysik.dtu.dk/.

The operation of mechanical assemblies under extreme conditions requires that the frictional behaviors of the materials in sliding contact be rapidly determined. However, it is frequently the case that during the design phase of these assemblies the friction coefficients of the materials being considered for use are not known. This presentation will present an approach by which data mining and materials informatics are used to generate a predictive model that enables efficient high-throughput screening of metal oxide materials. In particular, a combination of principal component analysis and recursive partitioning using a small dataset of intrinsic material properties is used to develop a decision tree based model that consists of if-then rules. This model is then used to estimate the friction coefficients of a wide range of materials derived from the interrelationships between the intrinsic material properties. This work paves the way for new studies in predictive modeling and for the tailoring of materials with specific tribological characteristics. This work is supported by the Office of Naval Research.

Designing new materials and formulations with desired properties is an important and difficult problem, encompassing a wide variety of products in the specialty chemicals and pharmaceuticals industries. Traditional trial-and-error design approaches are laborious and expensive, and cause delays time-to-market as well as miss some potential solutions. Furthermore, the growing avalanche of high throughput experimentation data has created both an opportunity, and a major modeling and informatics challenge, for material design and discovery. A systematic way to convert the raw data from analytical technologies to information and first principles knowledge that can be used for online decision decision making is lacking. A new paradigm is needed that increases the idea flow, broadens the search horizon, and archives the knowledge from today&’s successes to accelerate those of tomorrow. All these present considerable challenges, as well as great opportunities, in the area of molecular products design and engineering. Cyberinfrastructure will play a crucial role in molecular products design, process development and commercial scale manufacturing by streamlining information gathering, data integration, model development, and managing all these for easy and timely access and reuse. In this talk, I will discuss a novel cyberinfrastructure framework called Ontological Discovery Informatics. The foundation of such an infrastructure is the explicitly and formally modeled information, called an Ontology. This framework enables the management of information complexity, accumulation of knowledge, systematic model development, and efficient search for new materials with desired performance characteristics. I will discuss the application of this paradigm for industrial molecular products design problems in the specialty chemical and pharmaceutical industries.

Porous materials such as zeolites and metal organic frameworks have been of growing importance as materials for energy-related applications such as CO2 capture, hydrogen and methane storage, and as catalysts. Very large databases of virtual materials are being developed with a promise that they can be screened to discover optimal materials for these applications. The current state-of-the-art molecular simulations allow for accurate in silico prediction of materials properties but the computational cost of such calculations prohibits their application in the characterization of the entire database of structures, which would be required to perform brute-force screening. Our work focuses on the development of efficient screening approaches. Our techniques relay on the similarity principle exploited in a combination of diversity selection, similarity searching and (inverse) docking. The resulting screening approach requires expensive characterization only for carefully selected and statistically relevant subset of a database, therefore enabling discoveries at a minimal computational cost.

It would be appealing to exploit the notion of similarity in computational design of nanomaterials since it is often possible to prescribe a reference system with the targeted material properties. The search in the design space could then be expedited by dismissing the configurations that are dissimilar to the reference system. Putting this idea in practice requires a generally-valid mathematical expression of similarity in terms of a set of adequate material descriptors, and encourages one to adopt a design approach based on the notion of similarity. To this end and in line with the recent advances in the conceptual density functional theory, atom-partitioned quantum-chemical descriptors have been employed to define a number of similarity indices. First-principles calculations have been carried out to compute these indices as well as to obtain the size-, shape- and composition-dependent nanocrystal energetics. A database of atomic clusters, nanocrystals, and bulk solids have been constructed in order to explore correlations between the local similarity indices and the atomic chemical potentials. The variation of the chemical potential with the similarity indices has been investigated, leading to the development of quantum similarity-based informatics in nanocrystal energetics. Within the framework of the latter it is demonstrated that the energy differences utilized in the atomistic materials design and/or in elucidating the observed physicochemical phenomena could be obtained from the quantum similarity measures.This work is supported by TUBITAK under Grant No. 109T677.

I will present a detailed study of the energetics of water clusters (H$_2$O)$_n$ with $n \le 9$, comparing diffusion Monte Carlo (DMC) and approximate density functional theory (DFT) with well converged coupled-cluster benchmarks. We used the many-body decomposition of the total energy to classify the errors of DMC and DFT into 1-body, 2-body and beyond-2-body components. Using both equilibrium cluster configurations and thermal ensembles of configurations, we find DMC to be uniformly much more accurate than DFT, partly because some of the approximate functionals give poor 1-body distortion energies. Even when these are corrected, DFT remains considerably less accurate than DMC. When both 1- and 2-body errors of DFT are corrected, some functionals compete in accuracy with DMC; however, other functionals remain worse, showing that they suffer from significant beyond-2-body errors. Combining the evidence presented here with the recently demonstrated high accuracy of DMC for ice structures, we suggest how DMC can now be used to provide benchmarks for larger clusters and for bulk liquid water.

Crystal structure solution from diffraction experiments is one of the most fundamental tasks in materials science, chemistry, physics, and geology. Unfortunately, numerous factors render this process labor-intensive and error-prone. Experimental conditions, such as high pressure or structural metastability, often complicate characterization. Further, many materials of great modern interest, such as batteries and hydrogen storage media, contain light elements like Li and H that only weakly scatter X-rays. Finally, structural refinements generally require significant human input and intuition, as they rely on good initial guesses for the target structure. To address these many challenges, we demonstrate a new hybrid approach, called First Principles-Assisted Structure Solution (FPASS), which combines experimental diffraction data, statistical symmetry information, and first-principles-based algorithmic optimization to automatically solve crystal structures. We use FPASS to clarify five important crystal structure debates: the hydrogen storage candidates MgNH and NH3BH3; Li2O2, relevant to Li-air batteries; the shape-memory alloy NiTi; and high-pressure silane, SiH4.

SISYPHUS (Stochastic iterations to strengthen yield of path hopping over upper states) is a hybrid stochastic and deterministic algorithm. While maintaining fully atomistic resolution, SISYPHUS allows one to achieve milliseconds and longer time scales for several thousands of atoms. The method exploits rare event nature of the dynamics like other such methods but goes beyond them by (i) not having to pick a scheme for biasing the energy landscape, (ii) providing control on accuracy of the boosted time-scale, (iii) not assuming any Harmonic Transition State Theory (HTST), and (iv) not having to identify collective co-ordinates. We validate SISYPHUS by calculating diffusion constants for vacancy mediated diffusion in Iron and Tantalum metal at low temperatures, and comparing against brute-force high temperature Molecular Dynamics and low temperature HTST as well. The method is then applied to perform tensile tests on Gold nanopillars on strain rates as low as 10/s, bringing out the perils of high strain-rate Molecular Dynamics calculations. SISYPHUS thus promises to be a tool to bridge the time-scale discrepancy between simulations and experiments.

Computational Simulation is getting more important to develop new materials and nano-scale processing technologies such as semiconductor devices, Lithium-ion batteries and organic EL materials. In those research and development, both structural and electronic studies must be considered. To realize this purpose, TB based first-principles molecular-dynamics simulations are frequently used from the point of the total computational cost instead of a conventional technique such as LDA calculations, because long-time dynamics of large-scale systems should be investigated. Although the TB technique has a various potentiality, the parameters should be transferable in any circumstances. Our TB method is based on Extended Hückel Molecular Orbital Atom Super-Position and Electron Delocalization (EHMO-ASPED) Tight Binding theory. The charge self-consistency corresponding to various atom configurations in molecules or materials are also taken into account. The s-, p-, d-atomic orbitals are involved and all Slater-type orbitals are described in double zeta; functions.[1] In addition to the effective calculation, the appropriate TB parameters are required. [2] Based on the TB parameterization scheme developed by Nishino and Fujiwara [3], we have been developing a program package to determine TB parameters, by which the band structure can be reproduced so as to agree with that of first-principles molecular-dynamics calculation. As a benchmark, we have applied this to Si bulk system, and obtained our twelve TB parameters in our TB method to reproduce the band structure excellently. The energy differences of energy levels are typically less than a few hundred meV from lowest energy level up to tenth level. We will report the results applied to other bulk systems widely. References: [1] J. Cedra and F. Soria, Phys. Rev. B 41, 5652, (1990). [2] S. Nishino, T. Fujiwara, H. Yamasaki, S. Yamamoto, and T. Hoshi, Solid State Ionics, to be published. [3] S.Nishino and T.Fujiwara, in preparation.