Using first-principles electronic-structure codes, a huge number of materials is being studied since some years. The amount of already created data is immense. Thus, the field is facing the challenges of “Big Data”, which are often characterized in terms of the “four V”: Volume (amount of information), Variety (heterogeneity of the form and meaning of the data), Veracity (uncertainty of the data quality), and Velocity at which data may change or new data arrive.
Obviously, the computed data may be used as is: query and read out what was stored. However, for achieving deeper and novel scientific insight, the four V need to be complemented by an “A”, the Big-Data Analysis. On this branch, big data studies will identify correlations between putative causes and the properties of interest. However, the science starts where the correlations reflect causal inference.
From the above-mentioned issues, the 4V & A, and for first-principles computational materials science and engineering, the two key challenges concern big-data veracity and analysis. These are at the focus of this talk.
(*) Work performed in collaboration with Luca M. Ghiringhelli, Jan Vybiral, Sergey V. Levchenko, and Claudia Draxl

Organic electronic materials have significantly large and complex design space, providing great opportunity to achieve diverse physical and chemical properties. There is a pressing need for the development of low-cost, informatics-based modeling tools to rapidly explore the vast chemical space and advise experimental efforts. In this work, we present fast and user-friendly cheminformatics tools wrapped-up in a graphical user interface to make it ideal for the discovery and design of novel organic electronic compounds. With the help of Schrodinger&’s high-throughput quantum chemistry and virtual screening technology, cheminformatics analysis methods such as scaffold decomposition and R-group analysis are validated against high-level density functional predictions for a wide range of organic light-emitting diode (OLED) and organic photovoltaic (OPV) materials. Once identified with the key structural motifs that determine the properties of these materials, 2D-based quantitative structure-property relationships (QSPR) models are constructed for an extensive list of properties such as redox potentials, reorganization energies, and triplet exited-state energies. We confirm that kernel-based partial least square (KPLS) regression with binary fingerprint provides good predictions for the optoelectronic properties, enabling an effective in silico design scheme for a wide selection of organic electronic materials.

Gold nano-clusters are considered for applications such as optoelectronics, bio-recognition, and catalysis. In spite of their enticing potential, the atomic structures of very small (number of atoms <15) Au clusters are still elusive. For instance, there is ongoing debate on the critical cluster size, reportedly between 7 to 15 atoms, beyond which globular -3D- structures become more stable compared to planar -2D- ones. Furthermore, we found that some isomers at a given cluster size are energetically very close (<10 meV/atom), the reasons of which should be sought in finer details of electronic structures rather than mere differences in the number of dangling bonds at the edges or surfaces. Here, we present high throughput density functional theory calculations (DFT) coupled with a global structural optimization scheme using genetic algorithm (GA) to identify the ground state structures of Au nano-clusters near the critical cluster size. By investigating the electronic and vibrational properties of the low energy structures, we explore fundamental changes that drive the structural transformations.

This work identifies pathways to new chemical substitutions for rare earth elements in high temperature superalloys via an informatics based design strategy that involves coupling graph theoretic mapping with dimensionality reduction. Multiple non-linear pathways are traced through a weighted graph constructed in high dimensional data space, built by mapping discrete elemental data onto existing theories of alloy design to account for mechanical properties and thermodynamic stability, in the course of elemental substitution. This provides a manageable and reasonable target space to look at when substituting for elements linked by specific properties, a feature not captured in the visualization of the periodic table, and the weighted graph, thus, provides a new visual representation of the periodic table.

Materials informatics-based tools and workflows have matured under the auspices of the Materials Genome Initiative (MGI) and the Integrated Computational Materials Engineering (ICME) design approach. One ICME vision is co-evolved in-silico materials and process design with experimental verification. This requires the integration of computational and experimental workflows, often with different goals, maturity levels, and timescales. Automation and optimization of this combined workflow further requires a feedback loop between computation and experimental characterization.
Here we present our efforts to “close the loop” between modeling and experimental efforts. In our combined workflow, process controls are automated and informatics-based Materials Quantitative Structure-Property Relationships (MQSPRs) are built on in-situ materials characterization data. A genetic algorithm suggests new experiments based upon prior knowledge and a targeted goal. Thus, the workflow enables us to perform experiments, collect and process in-situ characterization data, update models, and perform new experiments based upon the success or failure of previous experiments in an autonomous fashion. We use this approach to study the fundamental mechanism by which carbon nanotubes nucleate and grow.

To date, trial and error strategies guided by intuition have dominated the identification of materials suitable for a specific application. We are entering a data-rich, modeling-driven era where such Edisonian approaches are gradually being replaced by rational strategies which couple predictions from advanced computational screening with targeted experimental synthesis and validation. Consistent with this emerging paradigm, we propose a strategy of hierarchical modeling with successive down-selection stages to accelerate the identification of polymer dielectrics that have the potential to surpass `standard' materials for a given application. Specifically, quantum mechanics based combinatorial searches of chemical space are used to identify polymer repeat units that could lead to desirable dielectric properties, followed by configurational space searches to determine the 3-dimensional arrangement of polymers (and their properties) built from the desirable repeat units. Successful synthesis and testing of some of the most promising identified polymers and the measured attractive dielectric properties (which are in quantitative agreement with predictions) strongly supports the proposed approach to material selection.

Heterogeneous interfaces between solids or between solids and liquids play a fundamental role in determining materials properties. For example, the understanding and control of the microscopic structure of solid/water and solid/electrolyte interfaces is of key importance in order to successfully predict the properties of photocathodes for water catalysis and the production of clean fuels, or optimal materials for energy storage. Likewise in solar cells, e.g. nanoparticle based or multi-junction cells, solid/solid interfaces are a critical part of the device, greatly influencing the energy conversion process. We present the results of ab initio vibrational and opto-electronic spectroscopies integrated with large scale ab initio molecular dynamics simulations, used to study realistic interfaces, directly comparable with experimental conditions. In particular, we focused on interfaces between semiconductors and oxides, and aqueous solutions of interest to water photo-catalysis, and on solid/solid interfaces present in nanostructured materials for third generation solar cells.

Refractory materials are of clear importance in a wide range of applications (gas turbines, rocket thrusters, shields or armors, leading edge of hypersonic vehicles, etc.). Employing a recently-developed small size coexistence approach for efficient ab initio melting point calculations, we study the melting temperatures and phase diagrams of the most refractory carbides known to date, i.e., hafnium carbide (HfC), tantalum carbide (TaC) and their mixtures. Very few measurements of the melting points of tantalum and hafnium carbides have been reported, because of the obvious experimental difficulties at extreme temperatures. We report ab initio calculations of these melting points and identify several key factors that contribute to their high values. Inspired by these factors, we identify a new class of materials which should exhibit even higher melting temperatures than Ta-Hf-C alloys. We explore this class through large-scale ab initio calculations and identify a likely candidate with the highest known melting point.

Recently, challenges for efficient materials exploration with the aid of information technology have been well demonstrated. One of the approaches uses high-throughput screening of materials database that is generated by first principles density functional theory (DFT) calculations. Thanks to recent progress of computational power and technique, a large number of DFT calculations can be made with the accuracy comparable to experiments, which can be used for the efficient materials exploration. Another approach is based upon state-of-the-art machine-learning algorithm to search the optimum. A combined approach of them should also be useful. A few examples of accelerated discovery of materials for energy applications will be given in my talk. Discovery of cathode materials with prolonged cycle life for lithium-ion battery is the first example. A wide chemical compositional space for co-substituted LiFePO₄ was explored using DFT calculations. Based upon the results of the screening, synthesis of selected materials was targeted. The optimized materials were actually synthesized and confirmed to show excellent cycle-life performance. Second example is on the discovery of solid state electrolytes for lithium ion battery. A systematic set of DFT calculations were combined to the cluster expansion method. Many first-principles molecular dynamics (FPMD) simulations were carried out for a diverse range of chemical compositions. A machine-learning technique was used to combine theoretical and experimental datasets to predict the conductivity of each composition. The third example used our own dataset of thermal conductivity computed by the first principles anharmonic force constant method as implemented in our phonopy code. The information was found to be very useful for discovery of new thermo-electric crystals.

New inexpensive and efficient catalyst materials for energy transformations are necessary for future sustainable energy solutions. Theoretical heterogeneous catalysis based on atomistic simulations provides a fast way to screen materials and predict their relative performance. Current increases in computing power and algorithms will soon enable us to generate date on millions of catalysis-relevant materials properties. Our ability to efficiently handle large amounts of computational data, and our ability to integrate the efforts and data from many delocalized research groups will become key to high productivity and accelerated new discoveries in the field of catalysis. We are presenting the current status of our efforts to develop the software infrastructure, in the form of a computational catalysis data warehouse, in order to carry out this task. Ultimately, the Data Warehouse can also be utilized in other research fields, but initially we shall focus entirely on catalysis relevant properties. A central part of this infrastructure is to store not just simulation results but also already at run-time to store an adequate amount of "meta-data" to be able to redo the simulations automatically in different electronic structure codes and at different levels of accuracy. The Data Warehouse will be directly accessible through a restful web-based interface, thus establishing an open repository of high quality electronic structure simulations of relevance for materials (and especially catalyst) design.

First-principles high-throughput screening of novel materials requires a materials informatics infrastructure able to automatically prepare and execute HPC calculations on large classes of materials, to monitor calculation progress, and to store, retrieve and analyze complex data. In order to speed up computational discovery and optimization, we have developed a powerful flexible environment that integrates these capabilities and is adaptable to diverse applications and use cases. The new platform called AiiDA ("Automated Interactive Infrastructure and Database for Atomistic simulations") combines database storage with user-defined grid-enabled computational workflows. We demonstrate how automated AiiDA workflows make computational design efforts faster, easier, and fully integrated with data collection and community sharing.

The discovery and the development of new materials is a formidable challenge due to the unlimitedly large chemical search space. New scientific methodologies aimed to find the high performance materials by reducing the search space are needed. The focus of our research for the accelerated materials discovery is based on utilizing the concepts of Materials Genomics coupled with the tools of Materials Informatics. In this talk, we will present a successful high-throughput materials discovery process. Our computational studies aim to find promising small electroactive molecules for aqueous flow batteries that would provide grid-scale energy storage. We will address the following main components of a Materials Genomic approach in the search for suitable candidate molecules: (i) first principles high-throughput calculations for massive computational data generation, (ii) a robust and quick theoretical method for virtual screening, (iii) informatics-based data analysis for rapid extraction of quantitative structure property relationships, and (iv) validation of models and results through experiments.

High-throughput experimental measurements are essential for effective validation of computational predictions and for serving as input to materials databases. Compositional gradients and libraries of different phases created in combinatorial thin films and bulk diffusion multiples together with high-spatial resolution localized property measurements can provide reliable materials property data as a function of composition and crystal structure. This talk will provide a snapshot of the state-of-the-art development of such high-throughput experimentation as well as its role in establishing materials databases for accelerated design/development of new materials.

Incorporation of small molecules in inorganic single crystals remains relatively untouched upon and poorly understood. While existing studies provide initial data, the range of molecular species and inorganic crystal systems investigated is currently very narrow due to a lack of a consistent, general experimental approach. Here, a general rapid-screening protocol for crystal preparation and characterization is described, introducing a novel application of high-throughput apparatus. This includes utilizing well plates as micro-volume crystallizers, accurate liquid-handling pipetting workstations for sample preparation; plate readers for colori-, turbidi- or fluoro-metric analysis; SEM with sample mapping technology for shape and size analysis; in situ pXRD for phase identification and Raman spectroscopy. For demonstrative purposes, characterization of amino acid incorporation into calcite has been characterized for comparison to previous studies; however the technique has been developed further for the analysis of other small molecules with a variety of inorganic crystal systems and a wide range of other environmental and chemical variables, such as crystal growth protocol, ionic strength and co-dopants. It is foreseen that this powerful and generic technique could be expanded across a gamut of different additive/mineral combinations and chemical environments for compiling vast datasets to uncover theoretical rationale behind effects of small molecules on kinetics, shape, structure and incorporation.

The Materials Project (www.materialsproject.org) is an effort to compute the properties of all known inorganic materials and beyond.¹ The current release contains data derived from density functional theory (DFT) calculations for over 50,000 materials, representing over 20 million CPU-hours of computing at the NERSC supercomputing center. This talk discusses the uses of this database by the authors towards the design of new Li ion battery cathodes materials. In addition to such data-driven discovery, we discuss how this data was used to mine physical trends in cathode materials, such as the nature of the inverse relationship between cathode voltage and safety. We also present statistical data on voltage found from this data.
The talk will cover some of the tools offered by the Materials Project that can be used by researchers for materials informatics, including interactive web-based tools like the Phase Diagram App as well as open-source codebases and data access tools such as the pymatgen materials analysis library,² FireWorks workflow software,³ and Materials API. We will discuss up-and-coming features and developments, such as integration of elastic constant data, additional band structure-derived properties (e.g., towards thermoelectrics), user-submitted structures for computations, and collaborations with outside projects such as MaterialsHub.⁴ All these tools can be used by any researcher for data-driven materials discovery and analysis.
References
¹ A. Jain, S.P. Ong, G. Hautier, W. Chen, W.D. Richards, S. Dacek, S. Cholia, D. Gunter, D.
Skinner, G. Ceder, and K.A. Persson. APL Materials 1 (2013) 011002
² S.P. Ong, W.D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V.L.
Chevrier, K.A. Persson, and G. Ceder. Comp. Mat. Sci 68 (2013) 314-319
³ http://pythonhosted.org/FireWorks/
⁴ http://www.materialshub.org

The formation of dendrites on the anode surface of metal air batteries, such as the lithium air (Li-air) battery, causes significant decreases in performance over the lifetime of the battery and poses safety concerns due to short circuiting. This research focuses on improving the performance and durability of advanced Li-air battery technologies through computational modeling and materials informatics. We use predictive computational methods to model the physical battery process and materials informatics to identify novel electrolyte materials which will reduce dendrite growth. Dendrite growth depends strongly on the reactive transport near the anode-electrolyte interface and heterogeneities on the anode surface. We are investigating increasing the mixing near the anode-electrolyte interface through novel electrolyte designs that directionally affect the transport properties of the dendritic components.
Materials informatic approaches are used in a discovery mode to identify promising material combinations for the electrolyte. Our informatics techniques have predictive capabilities based upon electrolyte materials knowledge base, allowing us to evaluate materials based on their thermodynamic properties and the desired properties needed for the specific application to Li-air batteries. Focusing on the more promising materials, physical models are being constructed to simulate the physical processes at the atomic and mesoscopic scales to evaluate their stability and performance in a Li-air battery.
This research is supported by Samsung Electronics Company through the Samsung GRO Program.

The design of inorganic scintillators is accelerated by mathematically defining the relationship between crystal chemistry, defect structure, and scintillation properties with material informatics. The host lattices analyzed includes garnet and perovskite crystal structures, with both structures able to tolerate a prohibitively large and complex chemical space. Defect structure is modified through co-doping schemes which we assess through mapping of oxygen vacancy concentrations and solution energies. Based on both these design thrusts, we identify new scintillator materials with predicted high light yield and fast decay time. These predictions supplement the use of manifold mappings for screening the potential compounds and to focus the regions requiring modeling. Density functional theory (DFT) and thermodynamic modeling is used to assess the scintillation mechanism, provide a comparison with predicted results, and link to the material database. This new design strategy reduces the scintillator search space, and accelerates the discovery of new scintillator materials.

Density functional theory (DFT) calculations suggest that polymers having tin atoms in their backbones may have high dielectric constants ^[1,2]. A class of organotin polymers, called tin esters, were then synthesized, realizing this prediction ^[3,4]. Motivated by this result, we developed a high-throughput screening scheme to predict the chemical environments which would correspond to the desired performances for an energy storing dielectric, i.e high band gaps and dielectric constants. Our scheme is based on four essential components. The first component is a database of crystalline organotin compounds/polymers, built up from the low-energy structures which are either predicted by the minima-hopping method ^[5,6] or taken from the Crystallography Open Database ^[7]. The properties of these compounds, i.e., energies, band gaps, and dielectric constants, were calculated at the DFT level. Second, a fingerprint is carefully developed to capture the bonding environment of the crystalline compounds. Third, Kernel Ridge Regression model is used to quickly predict the desired properties of a compound/polymer solely from its fingerprint, given that the model is trained with the database of related materials. Finally, an inverse design mechanism is developed to identify the chemical environment (in terms of a fingerprint) corresponding to the desired performances (band gap and dielectric constant). We describe the scheme and show that it works very well for organotin polymers as well as other related crystalline compounds.
References
[1] C.C. Wang, G. Pilania, & R. Ramprasad, Phys. Rev. B 87, 035103 (2013).
[2] G. Pilania, C.C. Wang, K. Wu, N. Sukumar, C. Breneman, G. Sotzing, & R. Ramprasad, J. Chem. Inf. Model 53, 879 (2013).
[3] A. F. Baldwin, T.D. Huan, Rui Ma, A. Mannodi-Kanakkithodi, N. Katz, R. Ramprasad & G. Sotzing, (submitted)
[4] A.F. Baldwin, Rui Ma, A. Mannodi-Kanakkithodi, T.D. Huan, C.C. Wang, J.E. Marszalek, M. Cakmak, R. Ramprasad, & G. Sotzing, (submitted)
[5] S. Goedecker, J. Chem. Phys. 120, 9911 (2004).
[6] M. Amsler & S. Goedecker, J. Chem. Phys. 133, 224104 (2010).
[7] Grazulis, S., Chateigner, D., Downs, R. T., Yokochi, A. T., Quiros, M., Lutterotti, L., Manakova, E., Butkus, J., Moeck, P. & Le Bail, A. (2009) "Crystallography Open Database - an open-access collection of crystal structures". J. Appl. Cryst.42, 726-729.

We develop first-principles data driven discovery approach to screen for experimentally-known transition metal oxide compounds with low band gaps and optimal band edges for water splitting applications, and that are stable in aqueous environments. Our screening process initially focuses on the transition metal oxides (TMOs) in the Inorganic Crystal Structure Database, using a new, broadly-applicable high-throughput workflow automating density functional theory (DFT) and hybrid functional calculations of bulk and surface electronic and magnetic structure implemented within the Materials Project (www.materialsproject.gov) framework. With approximate band gaps from the Heyd-Scuseria-Ernzerhof (HSE) hybrid functional for ~2000 TMOs and machine learning techniques, our workflow can predict HSE-quality gaps for TMOs with significantly more chemically and structurally complex unit cells with a mean error of ~0.4 eV. Our analysis also reveals interesting correlations between the band gap and other physical quantities including the valence electron density and bonding strength. Additionally, we discuss trends in absolute band edge energies with surface orientations and specific facets, and uncover useful correlations between band edges and surface formation energies. Finally, we assess the stability of candidate TMOs in acidic or alkaline aqueous environments from calculated Pourbaix diagrams. Applying our new approach, we successfully identify several new TMOs with promising band gaps and edges that are predicted to be stable under aqueous conditions relevant to water splitting. We discuss the recent successful synthesis of several predicted candidates, where trends in optical spectroscopy are found to be in good overall agreement with our theoretical predictions. We also further comment on best practices for analysis of these data, and the efficacy of hybrid functionals for trend-wise and quantitative prediction of optical and electronic properties. This work is supported by the Materials Project Predictive Modeling Center and the Joint Center for Artificial Photosynthesis through the U.S. Department of Energy, Office of Basic Energy Sciences, Materials Sciences and Engineering Division, under Contract No. DE-AC02-05CH11231. Computational resources also provided by the Department of Energy through the National Energy Supercomputing Center.

A systematic information theory based approach to deriving ab initio Hamiltonians for lattice dynamics of crystalline materials is proposed. Compressive sensing lattice dynamics (CSLD) allows to include all anharmonic terms up to a certain order and within a certain maximum distance. The relevant terms that are necessary to reproduce the interatomic forces calculated from the density-functional theory are selected by minimizing the L₁ norm of the scaled force constants. The accuracy and efficiency of the method are demonstrated for cubic Si, rocksalt NaCl, and Cu₁₂Sb₄S₁₃, a recently proposed earth-abundant thermoelectric based on the natural mineral tetrahedrite.

Directed self-assembly has enabled the development of a wealth of advanced functional materials, with applications in microelectronics, optoelectronics, drug delivery, and antimicrobials. Molecular simulations implementing physics-based force fields can predict the structural rearrangements driving assembly, but it remains a challenge to identify the underlying mechanisms and pathways driving assembly. The complex nature of self-assembly has hindered the rational design of building blocks to self-assemble into aggregates with desired properties. A full accounting of the assembly process requires an integrated kinetic and thermodynamic description of both what structures can be formed and how they assemble.
By combining tools from statistical mechanics, molecular simulation, and nonlinear machine learning, we have developed an approach to systematically identify assembly mechanisms from molecular simulations, and define good order parameters with which to describe the assembly pathways. These order parameters define a manifold containing the fundamental collective dynamical modes governing the long-time assembly of the system, providing a kinetically meaningful low-dimensional embedding of the high-dimensional configurational space. We show how to perform accelerated sampling directly in these variables to recover free energy landscapes that encapsulate both the thermodynamics and kinetics of self-assembly, providing a quantitative bridge between building block properties and assembly behavior. These landscapes may be subsequently used to “reverse engineer” building block structure and chemistry to control the morphology, stability, and kinetic accessibility of self-assembled materials with desired structure and function.
We have applied our methodology to compute the free energy landscapes for self-assembly of experimentally realizable patchy colloids into icosahedral aggregates as a coarse-grained model for viral capsid self-assembly or prototypical vehicles for small molecule encapsulation. Our approach systematically identifies the low energy pathways driving assembly, and the energetic and entropic barriers, and kinetic traps impeding cluster formation. We then discuss how these landscapes inform an iterative design process to rationally manipulate building block architecture and chemistry to elevate the stability and accessibility of the desired icosahedral aggregates.

We present a demonstration of Bayesian inference applied to a representative problem in materials research. Our goal is to infer the length scale of a spatially varying field on a 2-D substrate from the behavior of a phase-separating film deposited on the substrate and modeled by the Cahn Hilliard equation. Inference was performed based on a combination of parametric and non-parametric model approximations followed by the use of Bayes&’ rule to determine the probability distribution of the underlying length scale. The techniques demonstrated are general and may be extended to other problems involving the assessment of reduced order models in materials science.
This work was supported by the US Department of Energy (DOE), O#64259;ce of Basic Energy Sciences under Award No. DE-SC0008926.

The density of states (DOS) has been shown to be a key statistical descriptor for predicting material properties. The elastic moduli and DOS data have been calculated using first-principles, density functional theory (DFT) for dopants in Ni₃Al to form Ni₃AlX, where X = Cr, Zr, Ce, or B. These properties were then used to develop a data informatics model based on principal component analysis (PCA) to predict the lattice site preference (Ni vacancy, Al vacancy, or either one) of generic individual dopants. The model can predict the site-substitution preference of Cr without doing further DFT calculations. This has been confirmed by Energy Dispersive Spectroscopy within the group. These findings are being used to guide the experimental development of new Ni₃Al alloys. This work is supported by Air Force Office of Scientific Research grant # FA9550-12-0496 and the National Science Foundation (DMR-1307840).

While the current polymer dielectric standard for high-energy density capacitors is biaxially oriented polypropylene (BOPP), efforts are underway to identify other organic polymers^[1,2] with similar chemical connectivity to BOPP but with much improved dielectric properties. We have attempted to do this by considering the crystals of a number of polymers formed by certain organic building blocks in the repeat unit, namely, CH₂, CO, CS, O, NH, C₄H₂S and C₆H₄. Density functional theory calculations combined with the minima hopping method^[3,4] was used to determine the stable crystal structures of these polymers, as well as their dielectric constants and band gaps.
A heuristic assessment of the data was done, leading to correlations between the atomic polarizabilities and the electronic component of the dielectric constants, and between the dipole moments and the ionic component of the dielectric constants^[5]. Additionally, machine learning methods^[6] were used to mine the data and to develop methods for the rapid and high-fidelity property predictions. The learning method used was kernel ridge regression (KRR) which maps a fingerprint based on crystal and chemical structure information to the dielectric constants and band gaps. It is shown that the model works extremely well in predicting the properties of polymers outside of the training sample set. Finally, based on a correlation analysis between the fingerprint components and the dielectric constants, a decision tree is prepared to identify the ideal organic chemical environment suitable for high energy density capacitor applications^[7].
References
[1] V. Sharma, C.C. Wang, R. G. Lorenzini, Rui Ma, Q. Zhu, D.W. Sinkovits, G. Pilania, A.R. Oganov, S. Kumar, G.A. Sotzing, S.A. Boggs, & R. Ramprasad, submitted.
[2] C.C. Wang, G. Pilania, S.A. Boggs, S. Kumar, C. Breneman, & R. Ramprasad, Polymer 55 (2014)
[3] S. Goedecker, J. Chem. Phys. 120, 9911 (2004).
[4] M. Amsler & S. Goedecker, J. Chem. Phys. 133, 224104 (2010).
[5] G. Pilania, C.C. Wang, K. Wu, N. Sukumar, C. Breneman, G. Sotzing, & R. Ramprasad, J. Chem. Inf. Model 53, 879 (2013).
[6] G. Pilania, C.C. Wang, X. Jiang, S. Rajasekaran, R. Ramprasad, Scientific Reports 3 (2013)
[7] A. Mannodi-Kanakkithodi, R. Ramprasad (manuscript under preparation)

The last few decades have seen significant advancements in materials research tools, allowing researchers to rapidly synthesis and characterize large numbers of samples - a major step toward high-throughput materials discovery. The main bottleneck has now shifted to data analysis, as researchers can collect data faster than they can be analyzed. This disparity in data collection and analysis time is fueling interest in new machine learning algorithms for accelerated data processing, with particular interest in algorithms that are capable of analyzing complex spectral data. In this talk we will discuss a set of hyperspectral algorithms that can quickly sort data from combinatorial libraries spanning large composition phase spaces. We will also show how incorporating crystal structure data from both experimental and DFT crystallographic databases can improve the performance of these algorithms. This talk will focus on X-ray diffraction data obtained from Fe-Ga-Pd, Fe-Co-X, and Al-Co-W thin-film ternary composition spreads. The databases of interest are the ICSD and the DFT database AFLOWLIB.

The use of machine learning algorithms like artificial neural networks are increasing in popularity as ways to capture complex relationships between a set of inputs and desired outputs. Of particular interest in materials science is the ability to learn the relationship between the atomic configuration and the resultant total energies and forces, i.e., of learning the ground-state solutions of the Schrodinger equation. Fast and accurate predictions of these quantities would allow for rapid MD simulation beyond the length and time scales accessible by conventional fully ab-initio techniques. We describe previous implementations of this neural network approach to potential energy surfaces, and present new results incorporating recent improvements to this approach inspired by similar advancements in the field of computer vision.

The discovery and design of new materials can be accelerated by the identification of simple descriptors of material properties. However the identification of the most relevant descriptors and how they relate to the properties of interest is not always obvious. We demonstrate how machine learning, in the form of genetic programming, can be used to identify relevant descriptors for predicting hole trap depths in hydrogenated nanocrystalline and amorphous silicon. Amorphous silicon is an inexpensive and flexible photovoltaic material, but its efficiency is limited by low hole mobility. To identify the structural descriptors that are most indicative of hole trap depth, we have used genetic programming to analyze a library of thousands of hydrogenated amorphous and nanocrystalline silicon structures. Of 243 structural descriptors considered, our calculations reveal three general classes of structural features that influence hole trap depth and predict that multiple interacting defects may result in deeper traps than isolated defects. These results suggest a possible mechanism for the Staebler-Wronski effect, in which exposure to light degrades the performance of amorphous silicon over time.

Materials selection is one of the most important steps in product design and manufacturing. Product designers are interested in selecting materials that satisfy the minimum functional requirements of the design, as well as optimize the design trade-offs such as cost, performance and environmental impact. Given the importance of choosing the optimal materials for each application and its associated trade-offs, there has been growing interest in the application of multi-criteria decision analysis (MCDA) methods to materials selection problems. One problem in applying MCDA methods to materials selection is that data regarding the properties of materials is not always necessarily available, and when available, is imprecise or involves a degree of uncertainty. In such cases, materials databases represent the properties of materials in the form of interval ranges. As a result, the application of MCDA to materials selection must account for imprecision and uncertainty in the data relating to the properties of materials. Another problem in applying popular MCDA methods to materials selection is that they require the designer to specify trade-offs in terms of a ranking over properties and their relative numeric weights, which is often unavailable a priori to the designer, or available only qualitatively during materials selection. The existing MCDA methods handle the above issues according to their specific assumptions, which has implications on the ranking of materials they produce. While many MCDA methods applied to materials selection are themselves mathematically well-founded, we find that the nature of the materials data, and the designer's trade-offs over them makes the applicability of these methods to materials selection questionable. To address this challenge, we introduce the application of a new formal method grounded in qualitative decision theory and logical reasoning to the problem of materials selection. The new method is specifically appropriate for materials selection problems where the data is imprecise, expressed qualitatively, as a range of values, or if the data contains missing values. An additional advantage of the method is that it guarantees the soundness of the decision made, and backs the selection of one material over another with a proof of correctness, in the form of a sequence of logical equations that certify the correctness (optimality) of the decision using the stated trade-offs and principles of rational choice. We demonstrate the application of the method to a case study of chemistry and processing design of aluminum alloys.

In crystalline solids, structural defects like dislocations control plastic flow in the material. In amorphous solids, it has been difficult to find structural defects that govern plastic flow. We recently showed that such defects can be identified in 2D systems under shear or compression using machine learning methods like the support vector machine, using only local geometric information. Here we will show that the predictive power of the inference algorithm can be improved by taking into account the stochastic nature of rearrangements (i.e. Regions that are "soft," or vulnerable to rearrangement, do not always rearrange immediately). Our method works in 3D and with thermal rearrangements. Moreover, the displacements of particles in rearrangement can be predicted from the spatial gradient of the field quantifying softness.

Utilizing dopants to optimize, enhance, or fundamentally change the behavior of a parent material has been exploited in many situations ranging from material strengthening to electronics to electrochemistry. The search and identification of suitable dopant candidates has been laborious though, and dominated either by lengthy trial-and-error strategies (guided by intuition) or plain serendipity. We are entering an era where such Edisonian approaches are gradually being augmented (and sometimes, replaced) by rational strategies based on advanced computational screening. In this contribution, we offer such a paradigm for the selection of suitable dopants in metal oxides for catalytic reactions involving oxygen. As an example, we explore dopants across the periodic table for enhanced thermochemical splitting of water on doped ceria surfaces, identify the best candidates, and determine the primary factors that make these dopants the most attractive. Our screening criteria are inspired by Sabatier&’s principle, and are based on requirements placed on the thermodynamics of the elementary steps. Among the 33 dopants across the periodic table considered, Sc, Cr, Y, Zr, Pd and La are identified to be the most promising ones. The surface O vacancy formation energy is revealed as the primary descriptor correlating with enhanced water splitting performance. Experimental evidence exists for the enhanced activity of ceria for water splitting when doped with Sc, Cr and Zr. The proposed strategy can be readily extended for dopant selection in other oxides for different chemical conversion processes (e.g., CO₂ splitting, chemical looping, etc.).

Over the last few decades, multi-scale modeling (MSM) has become a
dominant paradigm in materials modeling and simulation. The practical
impact of MSM depends, to a great extent, on its ability to utilize
modern computing platforms. However, since there are no general
numerical and computational frameworks for MSM, the vast majority of
multi-scale material models or simulations are developed on a
case-by-case basis. We seek to formulate an adaptive numerical and
computational framework for MSM. We do not plan to develop a specific
method for MSM simulations, but instead, aim to develop a broad and
flexible numerical framework for designing and developing such
simulations. Our focus is primarily on new scalable numerical
algorithms applicable to a wide range of MSM applications. These
algorithms fall into one of the three areas: i) scalable data
transfer between parallel applications, ii) adaptive strategies for
MSM, and iii) data analytics for MSM. We present a formulation of
our numerical and computational MSM framework. Subsequently, we
describe development of a two-scale multi-scale model of composite
materials, as well as, a high-throughput capability for battery
research, both utilizing our framework.

In the present work we propose a novel computational approach to explore the broad configurational space of possible interfaces formed from known crystal structures to find new heterostructure materials with potentially interesting properties. In the series of subsequent steps with increasing complexity and accuracy, the vast number of possible combinations is narrowed down to a limited set of the most promising and chemically compatible candidates. This systematic screening encompasses (i) establishing the geometrical compatibility along multiple crystallographic orientations of two (or more) materials, (ii) simple functions eliminating configurations with unfavorable interatomic steric conflicts, (iii) application of empirical and semi-empirical potentials estimating approximate energetics and structures, and (iv) use of DFT based quantum-chemical methods to ascertain the final optimal geometry and stability of the interface in question. We will show how the interplay between the four steps can be optimized for selected materials and will discuss representative results including semiconductors and oxides. For efficient high-throughput screening we have developed a new method to calculate surface energies, which allows for fast and systematic treatment of materials terminated with non-polar surfaces. We show that our approach leads to a maximum error of around 3%.

The microstructure mediated design approach has been proposed to discover the underlying rules governing material properties. In designing polymer nanocomposites, the morphology of the microstructure is represented by statistical descriptions to bridge the gap between processing parameters and materials properties, which are essential for predictive materials design. So far, the relationships between microstructure and properties of the nanocomposites have been widely studied using Finite Elements Analysis (FEA) modeling and statistical learning algorithms. However, the relationship between processing and microstructure remains a challenge. While the effects of individual processing parameters have been studied, no systematic investigation has been conducted to taken into account multiple processing parameters simultaneously. In this research, we propose to use two physics-based intermediaries, the mechanical energy consumed during processing and the interfacial energies of the constituents, as descriptors to characterize the processing. The microstructure of the polymer nanocomposites will be represented by morphological descriptors. A semi-empirical hybrid modeling approach is then developed to establish the relationship between processing descriptors and microstructure descriptors, by using energy parameters as intermediate variables throughout the mapping. A case study on nanodielectrics is carried out to validate the proposed methodology.
For determining the processing parameters, polymer extrusion is described with six main functional modules: solids conveying, plasticating or melting, melt conveying or pumping, devolatilization, mixing and die forming. Analytical models are employed to estimate the energy consumptions for each individual module. These energy estimations are then integrated with the surface energy parameters of the constituent phases of the polymer nanocomposites to generate system-level energy prediction. Next, image-based microstructure characterization is utilized to statistically quantify the microstructure using descriptors. In addition, key statistical descriptors are identified using statistical learning to reduce the dimensionality of microstructure representation. A data mining technique is employed to discover the latent governing rules of the processing-structure relationships. Validation of the proposed methodology is also examined through comparison between data mining predictions and experimental results, i.e., the microscopic images from a set of dielectric polymer composites at different stages of processing.
To summarize, we develop a novel energy-mediated approach to establish the processing-microstructure relationship, which facilitates the predictive design and processing of such materials. With this approach, the high cost of exploratory synthesis could be drastically reduced by predictive design. Application of the methodology can be extended to a wide variety of material systems.

Multimillion atom reactive molecular dynamics simulations are used to investigate structural and dynamical correlations under highly nonequilibrium conditions, and reactive processes in nanostructured materials under extreme conditions. In this talk I will discuss four simulations:
(1) Reactive molecular dynamics (MD) simulations of heated aluminum nanoparticles have been performed to study the fast oxidation reaction processes of the core (aluminum)-shell (alumina) nanoparticles and small complexes.
(2) Cavitation bubbles readily occur in fluids subjected to rapid changes in pressure. We use billion-atom reactive molecular dynamics simulations on a 163,840-processor BlueGene/P supercomputer to investigate chemical and mechanical damages caused by shock-induced collapse of nanobubbles in water near silica surface. Collapse of an empty nanobubble generates high-speed nanojet, resulting in the formation of a pit on the surface. The gas-filled bubbles undergo partial collapse and consequently the damage on the silica surface is mitigated.
(3) Quantum molecular dynamics (QMD) simulation reveals rapid hydrogen production from water by an Al superatom. We find a low activation-barrier mechanism, in which a pair of Lewis acid and base sites on the Al_n surface preferentially catalyzes hydrogen production.
(4) We introduce an extension of the divide-and-conquer (DC) algorithmic paradigm called divide-conquer-recombine (DCR) to perform large QMD simulations on massively parallel supercomputers, in which interatomic forces are computed quantum mechanically in the framework of density functional theory (DFT). A benchmark test on an IBM Blue Gene/Q computer exhibits an isogranular parallel efficiency of 0.984 on 786,432 cores for a 50.3 million-atom SiC system. As a test of production runs, LDC-DFT-based QMD simulation involving 16,661 atoms is performed on the Blue Gene/Q to study on-demand production of hydrogen gas from water using LiAl alloy particles.
The Simulation data have been rendered with Nanovol, a grid-based GPU volumetric visualization software based on a scalable ray casting technique developed by Argonne National Laboratory, and visualized on CAVE2, the 37 Megapixels 3D large-scale virtual reality environment at the Electronic Visualization Laboratory at the University of Illinois at Urbana Campaign.

Many device structures combine the functionality of materials with very different bonding types: metallic, ionic and covalent. Traditional empirical potentials have been designed to consider one type of bonding only. The Charge Optimized Many-Body (COMB) approach allows for the seamless simulation of structures composed of dissimilar materials. This is because COMB includes a charge equilibration method that allows each atom to autonomously and dynamically determine its charge, and a sophisticated description of bond order, by which the strength of an individual pair bond is modulated by the presence and strength of other local bonds. Simulations using COMB potentials are orders of magnitude faster than electronic-structure calculations, can consider much larger systems and can easily simulate dynamically behavior. The power of this approach is illustrated from problem of interest for various condensed phase systems including U/UO₂, Zr/ZrO₂ and Cu/SiO₂

Simulations of realistic catalyst particles 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 accurately interpolated PESs that allow to speed up simulations by many orders of magnitude compared to conventional DFT [1].
We demonstrate the capabilities of NN potentials, using a new open source NN code [2], to showcase how a combination of first-principles computations and large-scale Monte-Carlo simulations based on high-dimensional NN potentials can be employed to study the equilibrium surface structure and composition of bimetallic Au/Cu nanoparticles (NP) which have recently been of interest as stable and efficient CO₂ reduction catalysts.
We show that the inclusion of explicit water molecules at a first-principles level of accuracy is necessary to predict experimentally observed trends in Au/Cu NP surface composition; in particular, we find that Au-coated core-shell NPs are thermodynamically favored in vacuum, independent of Au/Cu chemical potential and NP size, while NPs with mixed Au-Cu surfaces are preferred in aqueous solution. [3]
[1] J. Behler and M. Parrinello, Phys. Rev. Lett.98 (2007) 146401.
[2] N. Artrith and A. Urban, An Open Source Implementation of Atomistic Interaction Potentials Based on Artificial Neural Networks; Department of Mechanical Engineering and Department of Materials Science and Engineering, Massachusetts Institute of Technology: Cambridge, MA, 2014.
[3] N. Artrith and A. M. Kolpak, Nano Lett.14 (2014), 2670.

Polyelectrolyte block copolymers, which combine structural features of polyelectrolyte, block copolymers and surfactants, may self-assemble in a variety of nanoaggregates in aqueous environment, such as micelles, vesicles, lamellar mesophases or micellar aggregates. The morphology and size of formed aggregates are determined by the characteristically complex equilibrium of noncovalent forces (electrostatic, steric, hydrogen bonding, Van der Waals, and hydrophobic interactions). The strength of repulsive Coulomb interactions between the polyelectrolyte segments can be efficiently tuned by variations in ionic strength or/and pH in the aqueous solution. In order to explore the self-assembly process of polyelectrolyte block copolymers, we developed implicit solvent ionic strength (ISIS) model for use with the Dissipative Particle Dynamics (DPD) method to simulate the behavior of polyelectrolyte block copolymers with incorporated electrostatic interactions to achieve a good balance between reasonable physical description and computational feasibility. We applied this coarse-grained model to explore the influence of block length, block architecture, and solvent quality on the properties of the assemblies formed in aqueous solutions. Our DPD model enables us to obtain the main characteristics of the micelles formed (the aggregation number, the corona and core sizes, and anisotropy) as a function of the block length and salt concentration. Based on a comprehensive set of data obtained we constructed a morphological diagram of polyelectrolyte block copolymers in aqueous solution. The coarse-grained modeling and simulation, which is demonstrated as a complimentary approach in addition to experimental and theoretical methods, can deliver insight into self-assembly processes of block copolymers and provide evaluation of the size of aggregates obtained along with their scaling relation representation. The simulation results suggest that this coarse-grained simulation scheme gives a route wherein one can effectively and efficiently capture the self-assembly behaviors of polyelectrolyte block copolymers, such as DNA, RNA and other natural and synthetic polycations and polyanions.