Symposium MT03—Automated and Data-Driven Approaches to Materials Development—Bridging the Gap Between Theory and Industry

This presentation will summarize the results of our “Autonomous Research for Materials Development Workshop,” where a multidisciplinary group of materials researchers, computer scientists and AI/ML experts explored the opportunities, barriers and future investments. Closed-loop autonomous research systems are disrupting the research process.
The current materials research process is slow and expensive; taking decades from invention to commercialization. Researchers are now exploiting advances in artificial intelligence (AI), autonomy & robotics, along with modeling and simulation to create research robots capable of doing iterative experimentation orders of magnitude faster than today.
We propose a “Moore’s Law for the Speed of Research,” where the rate of advancement increases exponentially, and the cost of research drops exponentially. We consider a renaissance in “Citizen Science” where access to online research robots makes science widely available. This presentation will highlight advances in autonomous research and consider the implications of AI-driven experimentation on the materials landscape.

This presentation will focus on our self-driving laboratory for thin film materials discovery and optimization. Discovering high-performance, low-cost materials is an integral component of technology innovation cycles, particularly in the clean energy sector. The linear methodology currently used to develop optimal materials can take decades, which impedes the translation of innovative technologies from conception to market. Our interdisciplinary team is utilizing advanced robotics and machine learning to overcome this challenge. We are closing the feedback loop in thin film materials research by enabling our self-driving robotics platform named “Ada” to design, perform, and learn from its own experiments efficiently and in real time. As a proof-of-principle set of experiments, we will show how Ada discovers and optimizes high-performance, low-cost hole transport materials for use in advanced solar cells. I will also showcase how Ada’s modular design can enable the automated and autonomous discovery of materials for other clean energy technologies.

We present the results of an inter-laboratory high-throughput experimental (HTE) study which focused on measurement reproducibility and data exchange. Over the past 20 years, a great number of HTE techniques for synthesizing and characterizing thin-film oxides have been developed and reported. To date, however, there has not been a comprehensive study of how values measured for a series of properties (e.g. conductivity or optical band gap) on the same library compare across labs. Nor has there been a study that has attempted to normalize the hand-off of HTE samples and data. Here we report on the first such study using the Sn-Zn-Ti-O transparent conducting oxide system.
A series of Sn-Zn-Ti-O samples were deposited via Pulsed Laser Deposition and magnetron co-sputtering. At each institution a set of HTE measurements were made for typical properties including structure, thickness, conductivity, and optical bandgap. The samples were then exchanged between the two labs and the same set of properties were measured at the other lab and the data exchanged via an agreed upon uniform format.
A few lessons learned and several scientific observations regarding the reproducibility of HTE results gathered during this process will be discussed. An important lesson was the importance of deciding upon, and using, consistent measurement grids within a lab (and during exchanges) for all measurements, as this will impact future data archiving and retrieval. It was observed that qualitative trends are well reproduced even when two labs use very different methods for measuring a property, for instance ellipsometry versus transmission-reflection UV-VIS spectroscopy. However, quantitative comparisons were found to be measurement specific and spanned from excellent (bandgaps measured within a mean absolute error < 0.1 eV) to relatively poor (log resistivity measurements within a mean absolute error of 2). In the latter case, we believe that differences in sample probe geometries coupled to large changes in the properties of small composition regions were the most likely source of the poor correlation. The lessons learned and best practices obtained will be discussed.

The presence of microcracks in silicon wafers significantly reduces wafer strength, yielding wafer breakage during the manufacturing process, transportation and field operation. With the trend of decreasing wafer thickness for cost reduction purposes, thinner wafers are more prone to breakage in the presence of microcracks [1]. To enable a smooth transition to thin wafers for even cheaper photovoltaic modules, we recently developed a high-throughput prototype for in-line crack detection for silicon wafers [2]. This tool scans silicon wafer in the near-edge regions for micro-cracks and outputs linescan signals from a linescan camera, where no crack shows a smooth, undisrupted profile. As an in-line detection tool, it also requires a rapid and reliable algorithm that automatically identifies the presence of a micro-crack within a second after wafer scanning. In this work, we adopted an unsupervised machine learning method for anomaly detection, because the presence of microcracks above the critical length is typically a statistically rare event in the current PV production line (typically less than 5%). Specifically, a generative machine learning algorithm variational auto-encoder (VAE) is used to identify scans with microcracks [3]. The working principle of this algorithm is that: (1) VAE encodes the linescan profiles into lower-dimension vectors of latent variables, and then the latent variables are reconstructed back to linescan profile with the goal of minimized error; (2) because of most linescan profiles are very similar smooth curves without any cracks, the VAE model is trained to be biased toward linescan without cracks; (3) whenever a linescan profile for a crack appears, the trained VAE model generates a vastly different profile with a significant reconstruction error; (4) the crack is then detected by monitoring anomalous reconstruction error. The advantage of this unsupervised VAE method over the previous neural network method [4] is that it does not require a large amount of labelled crack data with different crack shapes (which can be very difficult to obtain). We demonstrate successful crack detections with several different wafer types (e.g., multi, mono, as-cut, and textured) and crack shapes (e.g., line-shape, cross-star, L-shape). We show that, with statistical analysis, this VAE-based anomaly detection could be a reliable and versatile method to enable the rapid detection of microcracks in silicon wafers.

Reference:
[1] S. Wieghold, Z. Liu, S. J. Raymond, L. T. Meyer, J. R. Williams, T. Buonassisi, and E. M. Sachs, “Detection of sub-500-μm cracks in multicrystalline silicon wafer using edge-illuminated dark-field imaging to enable thin solar cell manufacturing,” Solar Energy Materials and Solar Cells, vol. 196, pp. 70–77, Jul. 2019.
[2] Z. Liu, S. Wieghold, L. T. Meyer, L. K. Cavill, T. Buonassisi, and E. M. Sachs, “Design of a Submillimeter Crack-Detection Tool for Si Photovoltaic Wafers Using Vicinal Illumination and Dark-Field Scattering,” IEEE Journal of Photovoltaics, vol. 8, no. 6, pp. 1449–1456, Nov. 2018.
[3] H. Xu, W. Chen, N. Zhao, Z. Li, J. Bu, Z. Li, Y. Liu, Y. Zhao, D. Pei, Y. Feng, J. Chen, Z. Wang, and H. Qiao, “Unsupervised Anomaly Detection via Variational Auto-Encoder for Seasonal KPIs in Web Applications,” Feb. 2018.
[4] M. Demant, T. Welschehold, M. Oswald, S. Bartsch, T. Brox, S. Schoenfelder, and S. Rein, “Microcracks in Silicon Wafers I: Inline Detection and Implications of Crack Morphology on Wafer Strength,” IEEE Journal of Photovoltaics, vol. 6, no. 1, pp. 126–135, Jan. 2016.

The oxygen storage material (OSM), such as CeO₂ or pyrochlore type CeO₂-ZrO₂ (p-CZ), is used as a catalyst support for a three-way catalyst in automotive emission control systems. It has oxygen storage capacity (OSC) that is an ability to release and store oxygen reversibly by the fluctuation of cation valence depending on the reducing and oxidizing atmosphere. In this study, we explore high-capacity OSMs by using materials informatics (MI), which combines material science with inference algorithms in machine learning.
The OSC of 60 metal oxides supported Pd were experimentally estimated by the amount of produced CO₂ while switching between oxidizing (O₂) and reducing (CO) atmosphere every 2 minutes at temperature of 973, 773, and 573K. These experimentally measured OSC data were used as supervised data in our MI scheme. The support vector machine regression model was trained for the prediction of the OSC at each measured temperature. This model uses descriptors in which physical properties are considered to represent the features of the OSC. These features were automatically extracted using grid search to achieve each model with the highest accuracy. It is found that the features related to the stability of the oxygen atoms in the crystal and the crystal structure itself such as cohesive energy, which is obtained from the first-principles calculation, are highly correlated with the OSC. The present model predicts the OSC of 1,300 existing oxides registered in the in-house electronic structure calculation database. Several dozen materials with promising high OSC were proposed through this virtual screening. We synthesized one of the screened materials and experimentally confirmed that it indicates higher OSC than the conventional OSM, p-CZ.

Discovery in chemistry falls mainly into one of four types of areas with the discovery of new molecules, new reactions, new reactivity, and finally new physical properties of the resulting compounds or materials. Establishing new reactivity leads to new reactions which also leads to new molecules. This is therefore the order of impact for discoveries in terms of the amount of chemical knowledge that they contribute. Such findings must, by definition, belong outside the known or predictable; and they are outliers and as such can oppose conventions, assumptions and biases. By developing the meta-physics of chemistry and chemical reactivity we should be able to establish a new set of ontologies in chemistry that relate back to the practical core operations, but also can be translated into molecular structures and the discovery of function. The truth of chemistry lies with finding the intrinsic reactivity of the input chemicals, and then encouraging or enabling reactivity by process control. Whilst the new discovery and reaction should be translatable to chemical bonding theory, chemists need to grapple with the fact that the application of the current rules will not allow discovery, instead they will act to restrict it to the known rules. So chemical discovery requires that the current rules are updated, broken, or new ones are made where before there were none. The discovery of Diels-Alder or cross-coupling reactions are excellent examples of new rules that were just discovered without any prior warning.
Without a deeper development of a meta-physics of chemistry the use of big data and artificial intelligence will just tell us what we already know we know, and maybe predictable extensions, rather than enabling discovery. The challenge for the chemist is not the use of artificial intelligence, but the intelligent use of algorithms and automation for novel discoveries. In this talk I will explain how this might be possible.

References
[1] J. Granda, L Donina, V. Dragone, D. -L. Long, L. Cronin, Nature, 2018, 559, 377-381
[2] L. J. Points, J. W. Taylor, J. Grizou, K. Donkers, L. Cronin, Proc. Natl. Acad. Sci. USA, 2018, 115, 885-890.
[4] V. Duros, J. Grizou, W. Xuan, Z. Hosni, D. -L. Long, H. N. Miras, L. Cronin, Angew. Chem. Int. Ed., 2017, 56, 10815-10820.
[5] S. Steiner, J. Wolf, S. Glatzel, A. Andreou, J. Granda, G. Keenan, T. Hinkley, G. Aragon-Camarasa, P. J. Kitson, D. Angelone, L. Cronin, Science, 2019, 363, eaav2211.
[6] A. Henson, P. S. Gromski, L. Cronin, ACS Cent. Sci, 2018, 4, 793-804

Metal halide perovskites have emerged as promising materials for next-generation photovoltaic and optoelectronic devices. The discovery and full characterization of new metal halide perovskite-derived materials have been limited by the difficulty of growing high quality single crystals needed for single crystal X-ray diffraction studies. The formation of large single crystals is non-trivial, owing to the vastness of the chemical search space with enormous compositional degrees of freedom. We present the first automated, high-throughput approach for metal halide perovskite single crystal discovery based on inverse temperature crystallization (ITC) as a means to rapidly identify and optimize synthesis conditions for the formation of high quality single crystals. Using our automated approach, we have carried out a total of over 5000 metal halide perovskite synthesis reactions spanning six chemical systems. Through this unbiased search of the experimental space, we have more than doubled the number of metal halide perovskite materials accessible by ITC method and discovered a new perovskite structure. Combining machine learning and other statistical methods, we quantify the total experimental space and the likelihood of large single crystal formation. Moreover, machine learning models have been constructed for each chemical system, in which single crystal formation is modeled. This work is a proof of concept that a combined approach of high throughput experimentation and machine learning can be effective in the study of metal halide perovskite crystallization. The approach presented here is designed to be generalizable to different synthetic routes for the acceleration of materials discovery.

Self-driving laboratories combine algorithmic data analysis and experiment planning with robotic workflows to autonomously optimize one or more experimental figures of merit. This approach is applicable to challenging multi-parameter and multi-objective optimization problems such as the optimization of thin film materials within the vast design space of composition, deposition, and processing conditions. Here we describe a self-driving laboratory capable of formulating inks, depositing films via spin-coating, characterizing the resulting thin films using multiple techniques, and planning new experiments based on previous experimental data using the ChemOS experiment orchestration software1. The utility of this self-driving laboratory is demonstrated by autonomously optimizing optical and electronic properties of hole transport materials, which are crucial to the operation of a variety of thin-film-based optoelectronic devices. The autonomous optimization manipulates the film composition and annealing protocol to maximize a hole-mobility surrogate obtained by fusing data from transmission-reflection UV-Vis-NIR spectroscopy and 4-point probe conductivity measurements.

1. MacLeod, B. P. et al. Self-driving laboratory for accelerated discovery of thin-film materials. arXiv [physics.app-ph] (2019)

In the development of next-generation photovoltaics and light-emitting diodes, colloidal inorganic perovskite quantum dots (PQDs) have drawn notable attention for their highly tunable bandgap properties, high-charge carrier mobility and defect tolerance, and adaptability towards solution phase processing. However, studies of this material group and other colloidal semiconductor nanocrystals requires extensive exploration of their massive reaction parameter space within highly controlled reaction environments. Conventional flask-based, trial-and-error approaches are, therefore, highly unlikely to effectively capture the full potential and optimal synthesis conditions of these high-priority materials. Further complicating this process, across the accessible bandgap range, optimal synthesis parameters will vary significantly. Flow synthesis platforms have recently been demonstrated as a time- and material-efficient reaction monitoring strategy for synthesis, screening, and optimization of colloidal nanomaterials. The high sampling rate, low chemical consumption, and precise process control (automation) of flow reactors greatly reduces the challenges in exploring complex reaction spaces; however, high-throughput reaction screening technologies alone are likely not able to make significant breakthroughs, due to the massive scope of relevant colloidal synthesis conditions.
In this work, we present a modular microfluidic platform integrated with a machine learning (ML)-enhanced reaction optimization algorithm for on-demand synthesis of high-quality inorganic perovskite QDs with desired optical properties using a homogeneous anion exchange reaction. The intelligent QD synthesis platform consists of multiple computer-controlled pumps for on-demand reagent delivery/dosing, a flow path selector valve for automated selection of the halide salt source, and an in-line flow cell for automatic UV-Vis absorption and photoluminescence spectroscopy. Utilizing a utility function, an array of trained neural networks, and a global optimization algorithm, the intelligent QD manufacturing platform, approaches a target emission bandgap, while minimizing emission linewidth and maximizing quantum yield by tuning the concentrations of the precursors. Halide salt precursors are mixed within a highly efficient inline micromixer before combining with the perovskite QDs and gas-liquid segmentation. Monitoring each halide exchange condition requires less than 180 uL of total halide salt precursor and 170 uL of starting perovskite QDs per sample.
Integration with a ML-enhanced optimization algorithm enables the system to reach optimized synthesis conditions, across all six variables, for a target emission energy of 2.2 eV in 238 samples and 83 mL of chemicals without any prior training. More advanced optimization methods and pre-training with archived experimental data will further reduce this optimization time and cost. The versatility and modularity of the developed intelligent QD synthesis platform make it readily adaptable for on-demand synthesis of other colloidal nanomaterials.

The morphologies of solution-deposited thin films are frequently governed by complex combinations of processes from domains including multi-phase fluid flow, heat transfer, nucleation, solid mechanics, and interfacial phenomena. This complexity can frustrate both theoretical and empirical attempts to understand and control the morphologies of such films. Here we report an autonomous robotic system that uses machine vision feedback to determine optimal experimental parameters to achieve homogeneous, high-quality films via spin coating. An ink-formulating and spin-coating robot equipped with an imaging system¹ provides images of thin films to a computer vision algorithm which grades the quality of the thin films. This grading assessment provides input to an optimization algorithm that chooses the next ink formulation with the goal of identifying regions in the parameter space of ink formulation and spin-coating conditions that result in high-quality films.

[1] MacLeod, B. P. et al. Self-driving laboratory for accelerated discovery of thin-film materials. arXiv [physics.app-ph] (2019).

Many of the most relevant chemical properties of matter depend explicitly on atomistic and electronic details, rendering a first principles approach to chemistry mandatory. Alas, even when using high-performance computers, brute force high-throughput screening of compounds is beyond any capacity for all but the simplest systems and properties due to the combinatorial nature of chemical space, i.e. all compositional, constitutional, and conformational isomers. Consequently, efficient exploration algorithms need to exploit all implicit redundancies present in chemical space. I will discuss recently developed alchemical perturbation theory and quantum machine learning based approaches for interpolating quantum mechanical observables in compositional and constitutional space. Numerical results of our models indicate controlled accuracy and favourable computational efficiency.

Artificial Intelligence (AI) is a rapidly advancing field. Novel machine learning methods combined with reasoning and search techniques have led us to reach new milestones with increasing frequency, from self-driving cars to computer vision, machine translation, computer Go trained on human play, to Go and Chess world-champion level play using pure self-training strategies. These ever-expanding AI capabilities open up exciting new avenues for automating scientific discovery. I will discuss our work on using AI for accelerating and automating materials discovery. In particular, we have focused on high-throughput structure determination for combinatorial materials discovery and on solving the phase map diagram problem for composition libraries. While standard statistical and machine learning methods are important to address this challenge, they fail to incorporate relationships arising from the physics of the underlying materials. I will introduce an effective approach based on a tight integration of machine learning methods, to deal with noise and uncertainty in the measurement data, with optimization and inference techniques, to incorporate the rich set of constraints arising from the underlying physics. Finally, I will describe our vision and progress concerning Scientific Autonomous Reasoning Agent (SARA), a multi-Agent system to accelerate materials discovery integrating in a synergistic and complementary way, first principles quantum physics, experimental materials synthesis, processing, and characterization, and AI based algorithms for reasoning and scientific discovery, including the representation, planning, optimization, and learning of materials knowledge.

Functional groups link analytical, physical, organic, and materials chemistry and are therefore central to the chemical sciences. In both analytical and organic chemistry functional groups are used to predict the reactivity of molecules, select a solvent for a given reaction, and validate a reaction using measurable changes in the properties of a molecule. Current approaches to incorporate functional groups in the prediction, planning and verification of reaction conditions rely on human intervention and input. For example, the solvent used for a given transformation is chosen by a skilled organic chemist using intuition gained from the study of how the functional groups in a molecule dictate its solubility in a solvent. To verify if the reaction took place resulting in an unknown chemical entity, the state-of-the-art method is to accurately identify all functional groups of the reactants and products. This process is time-consuming, involves manual or database dependent analysis and interpretation of a Fourier Transform Infra-Red (FTIR) spectrum or Mass Spectroscopy (MS) data using previously established rules and experience of a skilled spectroscopist. These processes are subject to trial and errorfor compounds with multiple functional groups and for compounds that not well characterized in the literature. Such issues hinder the automated development and characterization of truly new materials with minimal human intervention. We present fast deep learning methods to select the optimal solvent for a given reaction in a transformation-free manner and identify all the functional groups for both the products and reactants for any given reaction. Our methods do not use any database, pre-established rules or procedures to perform either task and use the general definition of functional groups as a ‘collection of atoms’ instead of simple chemical groups traditionally assigned by chemists. We use Artificial Neural Networks (ANNs) to derive patterns and correlations between these collections of atoms and the solvents used to carryout a given chemical reaction using 2.3 million patented reactions available from the United States Patent and Trademark Office. Our methodology is the first to differentiate solvents by their precise chemical structure instead of simply choosing a solvent class and yields a 5-fold cross-validated average F1-score greater than 0.9. Solvent predictions obtained from this model have been validated both in silico using Density Functional Theory and using experimental in situ techniques. To verify that a reaction has occurred, we trained separate ANNs on 7393 publicly available FTIR and MS combined spectra obtained from the NIST Webbook. Instead of using multiple binary classifiers used in previous works to assign functional groups, our approach treats the classification problem in a multi-class, multi-label fashion. The model has a cross-validated F1 score higherthan 0.82 for 14 out of 17 defined functional groups. To showcase the practical utility of our method, we introduce two new metrics (Molecular F1 score andMolecular Perfection rate) to measure the performance of identifying all functional groups on molecules. The optimized model has a Molecular F1 score of 0.92 and a Molecular Perfection rate of 72%. Additionally, backpropagation of our model reveals IR patterns typically used by human chemists to identify standard groups, suggesting a convergence of the model on known spectral features that are diagnostic of particular functional groups. We further show that the introduction of additional functional groups does not decrease the performance of our model. Finally, we show redundancy in FTIR and MS data by encoding all our features in a latent space that retains the accuracy of the original model. These results reveal the importance of using machine learning for automated identification of new reaction conditions and functional groups to achieve autonomous processes in the future.

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. Machine learning has been tasked to aid in converting the collected materials property data into actionable knowledge, and more recently it has been used to assist in experiment design. In this talk we present the next step in machine learning for materials research - autonomous materials research systems. We first demonstrate autonomous measurement systems for phase mapping, followed by a discussion of ongoing work in building fully autonomous systems. For the autonomous measurement systems, machine learning controls X-ray diffraction measurement equipment both in the lab and at the beamline to identify phase maps from composition spreads with a minimum number of measurements. The algorithm also capitalizes on prior knowledge in the form of physics theory and external databases, both theory-based and experiment-based, to more rapidly hone in on the optimal phase map. The phase map is then exploited for functional material optimization.

The overwhelming majority of scientific knowledge is published as text, which is difficult to analyze by either traditional statistical analysis or modern machine learning methods. In contrast, the main source of machine-interpretable data for the materials research community has come from structured property databases which encompass only a small fraction of the knowledge present in the research literature. Beyond property values, publications contain valuable knowledge regarding the connections and relationships between the data items as interpreted by the authors. I will show multiple ways to extract useful information from scientific text in both supervised and unsupervised ways. I will show that materials science knowledge present in the published literature can be efficiently encoded as information-dense word embeddings (vector representations of words) without human labelling or supervision. These embeddings capture complex materials science concepts such as the underlying structure of the periodic table and structure-property relationships in materials. Furthermore, we demonstrate that an unsupervised method can recommend materials for functional applications several years before their discovery. This suggests that latent knowledge regarding future discoveries is to a large extent embedded in past publications.
In a more supervised way, we have also demonstrated the extraction of codified synthesis recipes from text. Extraction the details of synthesis, including precursor compounds, synthesis operations and their numerical details, requires a very high precision of information extraction, and a tolerance to deal with imprecise and non-standard language. I will show how a large data set of codified solid-state synthesis reactions has been obtained and be queried to obtain interesting information on choice of synthesis operations and precursors is related to the target material.

We are developing an autonomous scanning droplet cell (ASDC) capable of on-demand electrodeposition and real-time electrochemical characterization for investigating multicomponent alloy systems for favorable corrosion-resistance properties. The ASDC consists of a millimeter-scale electrochemical cell and an array of programmable pumps that can be used to electrodeposit an alloy film and immediately acquire polarization curves to obtain electrochemical quantities of interest, such as the passive current density and oxide breakdown potential. We model these quantities using Gaussian Process regression to select the most informative series of alloys to synthesize and characterize, continuously updating the model as new electrochemical data is acquired. Our initial studies focus on systems that are likely to form corrosion-resistant metallic glasses (MGs) and single-phase multi-principle element alloys (MPEAs).

The ASDC is an open exemplar autonomous system that provides insight into both technical and methodological aspects of building and deploying robust closed-loop synthesis and characterization platforms. Our approach is to build loosely-coupled modular experimental, automation, and communications systems to 1. support rapid prototyping, debugging, and verification while producing meaningful scientific output and 2. enable integration into future large-scale multi-user and multi-instrument distributed laboratory systems. We address both of the main challenges that autonomous science systems face: learning to reliably synthesize materials and mapping material specification and processing to structure and properties. We will discuss the incorporation of prior knowledge in the form of theoretical and data-driven predictive models, as well as the integration of online and offline multi-modal experimental data streams. Ultimately, closed-loop automated materials synthesis and characterization platforms offer much more than a means of engineering materials properties and performance through black-box optimization algorithms: they offer the potential to develop and deploy new algorithms for generating and testing scientific hypotheses.

An autonomous battery electrolyte experimental platform capable of mixing multi-component electrolyte systems and characterizing the transport and electrochemical properties in a high-throughput manner is disclosed. A Bayesian optimization software package found novel electrolyte compositions through optimization of the high-dimensional electrolyte design space over key design objectives like electrochemical stability and conductivity.
Electrolyte optimization is difficult because 1) electrolyte evaluation is expensive and takes time, and 2) the space of possible electrolytes is expansive, formed by many possible choices for solvents (often in ternary blends), salts (often in binary mixtures), and trace additives. Bayesian optimization methods are well suited for the optimization of high-dimensional functions with costly evaluations, often producing an efficient design-of-experiments to converge on multi-objective optimal formulations in few experiments. To expedite the optimization over the expansive design space, theoretical predictions of electrolyte properties via the Advanced Electrolyte Model were utilized as “priors” in the statistical model.
Implementation of the novel experimental platform was carried out by two novel test stands developed to automate the mixing and characterization of electrolytes: Otto (for aqueous systems) and Clio (for aprotic/organic systems). The test stands characterized the ionic conductivity and electrochemical stability of electrolyte systems, featuring a four-electrode conductivity probe, pH meter, and a flow-through three-electrode cell and potentiostat. Clio also integrated electrochemically active electrodes for optimization of electrolyte/electrode systems. The active electrode systems used were common functional ceramic metal oxides.
A software orchestration and data layer linked the test-stands to human experimenters and machine-learning packages through a web-services architecture; all experiment data and meta-data is saved in a database. Additional out-of-the-loop characterization was conducted on cathodic systems to validate composition, structure, and oxidation state.
The aqueous design space consisted of aqueous blends of lithium and sodium salts, including nitrates, sulfates, and other commonly-used battery salts. High-concentration aqueous electrolyte candidates were discovered by optimizing of electrochemical voltage stability and conductivity, including low-cost, high-performing alternatives to known but costly aqueous electrolytes (e.g. LiTFSI). Clio’s design space includes blends of both aprotic organic solvents and solutes in additional to various compositions of electrochemically active electrodes. The test-stands are demonstrated to be significantly faster than common human experimentation techniques, converging on novel, optimized electrolyte mixtures in mere hours or days of experimentation.

Iron-based superconductors are promising materials for high magnetic field applications because of their high transition temperature, high upper critical filed and small electromagnetic anisotropy [1]. Grain-boundaries (GBs) play a dominant role on the transport current performance of superconducting materials. Katase et al. reported that the critical current density through artificial bicrystal grain boundary is strongly dependent on grain boundary misorientation GB angles (θ_GB) and nearly constant up to a critical angle θ_c of ~9 degrees, which is substantially larger than the θ_c of ~5 degrees for YBa₂Cu₃O_7–δ. [2]. Foreseeing the future applications such as long-length wires and large bulk materials, it is important to understand the mechanisms of the grain-boundary formation during the solid-state sintering of polycrystalline iron-based superconductors. However, the mechanism of grain-boundary formation during the solid-state sintering has not been elucidated because in situ observation of grain-boundary at high temperature is technically not easy. In this sense, the numerical modeling of grain-boundary formation is a promising approach for understanding and controlling the grain-boundary structure in the sintered materials. In particular, phase-filed modeling [3,4] is a powerful tool to solve free boundary problems, such as the grain-boundary migration in polycrystalline materials. In this study, we performed the comparative study of experiments and phase-field modeling on grain-boundary formation of polycrystalline K-doped Ba122 iron-based superconductors. K-doped Ba122 polycrystalline bulk samples at different stages of sintering were fabricated by systematically changing heating conditions of mechanically alloyed powder. Their microstructure was analyzed by SEM and STEM to understand the mechanisms of grain-boundary formation during sintering. These experimental results were compared with the extended multi-phase-field modeling of solid-state sintering based on the model proposed by Wang [5].

[1] H. Hosono et al., Materials Today 21, 278-302 (2018).
[2] T. Katase et al., Nature Communications 2, 409 (2011).
[3] R. Kobayashi, Physica D, 63, 410-423 (1993).
[4] A. A. Wheeler et al., Physical Review A, 45, 7424-7439 (1992).
[5] Y. U. Wang, Acta Materialia 54, 953-961 (2006).

Recent publications analysing some commercially available graphene and GNP (graphite nanoplatelets) samples have highlighted the need for a robust metrology if graphene and related technologies are to develop.^1,2 Raman spectroscopy is a widely used tool for probing carbon due to the wealth of information that can be gained,³ however it is a highly localised probe and recording a single spectrum over a wide area often produces a distorted result dominated by minor features. This can be problematic when measuring inhomogeneous materials that contain a mixture of graphite flake sizes, thicknesses and various degrees of functionalization and defect concentrations.

To overcome these limitations, spectra can be recorded from a number of different points on a sample surface; however this poses a new question: how many points are needed? To answer this we have measured a selection of common graphitic materials including commercial GNPs and liquid exfoliated graphene to generate large data sets of Raman spectra from many points. A bespoke computer program has been developed in this study and was used to automatically fit the Raman signals for each independent spectrum; these parameters can then be used to infer some of the properties of the material. For example the heights of the D peak and G peak are often compared to estimate the concentration of defects in the sp² network.⁴

The convergence of these parameters is used to identify the approximate sample size required for robust characterization by plotting summary statistics (mean, interquartile range and P₁₀ & P₉₀) against the size of the data set; followed by a Monte Carlo based approach to resample data in smaller sets and probe the potential error when using an insufficient number of measurements. Whilst the focus of this work was on Raman spectroscopy, the methodology is applicable to other analysis techniques that require multiple measurements. Whilst different materials behave differently, in most cases it is clear that an average value can be derived after approximately one hundred data points however a full distribution describing the variety of materials present is only obtained after many hundreds of spectra are analysed. We also find many of these distributions are not normal, and simple metrics like mean and standard deviation fail for graphene materials, as reported previously.⁵

It is hoped this statistical approach can be applied in a variety of settings, both academic and industrial to inform the careful characterization and reporting of graphene materials. Raman spectroscopy is a non-destructive technique that is already in routine use and the spectral fitting program developed to accurately process these data sets will be freely available.

References
(1) Kauling, A. P.; Seefeldt, A. T.; Pisoni, D. P.; Pradeep, R. C.; Bentini, R.; Oliveira, R. V. B.; Novoselov, K. S.; Castro Neto, A. H. The Worldwide Graphene Flake Production. Adv. Mater. 2018, 30 (44), 1803784.
(2) Kovtun, A.; Treossi, E.; Mirotta, N.; Scidà, A.; Liscio, A.; Christian, M.; Valorosi, F.; Boschi, A.; Young, R. J.; Galiotis, C.; et al. Benchmarking of Graphene-Based Materials: Real Commercial Products versus Ideal Graphene. 2D Mater. 2019, 6 (2), 025006.
(3) Ferrari, A. C.; Basko, D. M. Raman Spectroscopy as a Versatile Tool for Studying the Properties of Graphene. Nat. Nanotechnol. 2013, 8 (4), 235–246.
(4) Ferrari, A. C. Raman Spectroscopy of Graphene and Graphite: Disorder, Electron–Phonon Coupling, Doping and Nonadiabatic Effects. Solid State Commun. 2007, 143 (1), 47–57.
(5) Kouroupis-Agalou, K.; Liscio, A.; Treossi, E.; Ortolani, L.; Morandi, V.; Pugno, N. M.; Palermo, V. Fragmentation and Exfoliation of 2-Dimensional Materials: A Statistical Approach. Nanoscale 2014, 6 (11), 5926–5933.

The transparent devices require complementary pairs of n-type and p-type transparent semiconductors to fabricate high efficient electronic devices. However, only a few p-type transparent oxide semicondutors were identified while there are a lot of n-type transparent oxide semiconductors like ZnO and In₂O₃, and InGaZnO with good device performances. Recently, several studies have been performed to search for p-type oxides using high-throughput screening with the band gap and effective mass (or branch point energy) as descriptors. However, these descriptors do not well distinguish the n-type and p-type oxide groups. Furthermore, none of suggested materials was verified experimentally, which means we need a better theoretical predictor. In this presentation, we propose a reliable descriptor for the p-type dopability based on the formation energy of hydrogen impurity (FEH). The predictive power of FEH is validated by that it can distinguish well known p-type and n-type oxides. Using FEH, we screen binary oxides and selected ternary compounds by considering the known design principles of p-type oxides (containing Sn²⁺ and Cu¹⁺ as well as oxychalcogenides). Considering FEH, band gap, and hole effective mass, we suggest La₂O₂Te and CuLiO as promising p-type oxides, which are validated through the full intrinsic defect calculations. [1] Furthermore, to enable high-throughput screening, we identify two coarse but simple descriptors - the band gap and oxygen partial weights at valence band maximum which correlate with FEH. Using the simplified descriptors, we screen over 17,700 oxygen-containing compounds in AFLOW database and select 426 compounds for the calculation of FEH. Finally, we identify 156 oxide semiconductors with the band gap larger than 1.1 eV and good p-type dopability. Furthermore, we classify the identified p-type oxides according to their valence band character and reveal chemical principles underlying the p-type dopability. [2]

References
[1] Yim, K. et al. npj Comput. Mater. 4, 17 (2018)
[2] Youn, Y. et al. Large-scale computational identification of p-type oxide semiconductors using hydrogen descriptor, submitted

Recently, various kinds of high resolution and multi-parameter spectral analysis is available by using high brilliant quantum beams like synchrotron radiation X-rays. Such advanced spectroscopy measurements potentially produce huge number of datasets. Developing the efficient method for the spectral data analysis is an urgent issue in the multi-dimensional measurements.
For example, operando SR X-ray scanning photoelectron microscopy system, called “3D nano-ESCA” (three-dimensional nanoscale electron spectroscopy for chemical analysis)[1], provides spatial, time and electric field dependence of photoemission spectra.This equipment has bias voltage application circuit for “operando analysis”, i.e. spectroscopy analysis during device operation, so we perform electric potential mapping by detecting the spatial distribution of XPS core-level peak shift in microscopic device structures under operation with high spatial resolution (~ 70 nm). However, the 3D nano-ESCA has been performed only for the pin-point or line-scan analysis that deals with few tens of spectral datasets by the inefficiency of peak fitting procedure, although spatial and time resolved measurement potentially provides over thousands of the datasets.
The method of spectral data analysis using machine learning technique has been studied to improve the resorting to the manual trial and error, e.g. applying a Bayesian peak separation with the exchange Monte Carlo method[2]. However, the computational cost was paid little attention because the number of datasets was not so large in conventional spectroscopy measurements.
Here we introduce the spectrum-adapted expectation-maximization (EM) algorithm for the high-throughput peak shift detection analysis of the large number of spectral datasets by considering the weight of the intensity corresponding to the measurement energy steps[3].
The EM algorithm is one of the machine learning techniques for estimating the parameters of the mixture model, including latent parameters, based on maximum likelihood estimation with iterative calculation between the expectation (E) step and the maximization (M) step. When ordinally EM algorithm is applied to the peak separation using a linear superpositions of distributions such as Gaussian distributions, the analyzed data are required to be one-dimensional, but the spectral data consists of N measurement steps of energy (x = {x₁, …, x_n, …, x_N}) corresponding to the intensity (w = {w₁, …, w_n, …, w_N}). So we solved this disadvantage by using the intensity (w) as a weight for each measurement step (x).
We applied the proposed method to the experimental datasets taken by the 3D nano-ESCA. Spectral datasets were collected from FETs using atomically thin films such as graphene and MoS₂ monolayer sheets. We confirmed drastic acceleration of peak fitting in comparison to the manual approach.In the presentation, we show concrete applications to the synthetic data of graphene field effect transistors (FETs) and other semiconductor fine devices.
This paper is based on results from a project (P16010) commissioned NEDO, JST-CREST(JPMJCR1761), JST PRESTO(JPMJPR17NB), and CORE lab (2016002) of "Five-star Alliance" in "NJRC Mater. & Dev."
Reference: [1] K. Horiba, K. Nakamura, N. Nagamura, et al., Rev. Sci. Instrum. 82(11), 113701 (2011). [2] N. Nagata, S. Sugita, Okada, Neural. Netw. 28, 82-89 (2012). [3] T. Matsumura, N. Nagamura, S. Akaho, K. Nagata, Y. Ando, Sci. Technol. Adv. Mat. in press (2019).

Materials informatics has been a relevant focus for accelerated materials development in an industry where traditional research methodologies demand a considerable investment in time and resources. In order to implement these informatics-enabled approaches, curation of well-structured datasets can be facilitated via high-throughput methods¹. An effective combinatorial screening procedure can be constructed with the fabrication of composition gradient polymer thin-film libraries via solution processing. High-throughput property characterization can then be applied to these gradient libraries to rapidly map the composition space for the response variable of interest. Here, we present work toward the high-throughput discovery of process-structure-property relationships in polymer composites. Many such polymer systems present challenges in solution processing due to low ambient solvent interactions and high viscosities. An automated flow system capable of high temperature operation was constructed for the solution-casting of polymer films. This system includes a custom-designed serpentine grooved-channel mixing component suitable for flow of high viscosity solutions and filler particles, allowing for fabrication of gradient films with spatial variations in composition. We discuss high-throughput characterization on these gradient libraries, which demonstrates the feasibility of our solution processing apparatus and gradient film approach.

[1] M. McBride et al., Processes (2018), 6, 79

Despite a rapid ramp-up of record efficiencies of lab-scale solar cells, perovskite as an absorber for solar cells has a great challenge to achieve large-area spatial uniformity [1], and long-term operation stability [2]. Both spatial uniformity and operation stability are important metrics to scale up the perovskite solar cells toward industrial commercialization. Particularly, during the degradation tests, the non-uniform film changes (e.g., local color change) are observed. We hypothesize this non-uniformity of degradation is likely due to process-induced defects acting as the degradation seed point (for example, water ingression), which could be reduced by a capping/buffer layer (e.g., ammonia salt 2D thin layer) on perovskite absorber. We examine this hypothesis by correlating multiple spatially resolved characterization tools along with the degradation tests. First, photoluminescence (PL) imaging is used to reveal the location of structural and electrical defects in the perovskite films. Second, calibrated red-green-blue (RGB) camera imaging is used to monitor spatial differences of degradation under various harsh environmental conditions (e.g., humid, hot and illuminated). Third, X-ray diffraction (XRD) mapping is used to identify the material phase change over time at different locations in the film. As a comparison, the experiments are repeated for the perovskite samples with a thin capping layer. With the correlative dataset of spatially resolved characterization results, we develop a machine learning method (i.e., convolutional neural network) that predict the material phase change (i.e., the results of XRD maps) from the fast and easy-to-implement PL and RGB images, because XRD mapping is a relatively slow process. Hence, we will be able to gain more understanding of the mechanism of non-uniform degradation by this correlative analysis.

Reference:
[1] Z. Li, T. R. Klein, D. H. Kim, M. Yang, J. J. Berry, M. F. A. M. Van Hest, and K. Zhu, “Scalable fabrication of perovskite solar cells,” Nature Reviews Materials, vol. 3, 2018.
[2] C. C. Boyd, R. Cheacharoen, T. Leijtens, and M. D. McGehee, “Understanding Degradation Mechanisms and Improving Stability of Perovskite Photovoltaics,” Chemical Reviews, vol. 119, no. 5, pp. 3418–3451, Mar. 2019.

Pairs of conjugate donor-acceptor fluorescent probes have proven themselves useful in stimulated emission depletion (STED) microscopy in recent years. For instance, it has been shown that the lifetime of said probes directly correlates to the resolution of the microscope [1]. However, once the lifetimes of the probes have been optimized, it is desirable to control their fluorescence in order to improve the resolution further. Here, we propose combining model-free control with sliding mode control to track nanosecond pulses of blue-shifted acceptor fluorescence in order to inhibit visible light emitted from the image plane, shrink the point spread function, and subsequently improve the resolution of the microscope. This is achieved by automatic adjustment of the STED laser beam pump power. This controller is numerically simulated against a generic model created from Förster resonance energy transfer (FRET) theory. However, since it is innately model-free and data-driven, it can be easily applied to varyious physical systems with drastically different dynamics. It is the goal of this paper to demonstrate that common fluorescent dyes can be used to increase the resolution in modern super resolution microscopy—paving way for biological imaging previously unachievable.
References
[1]
S. Deng, J. Chen, Z. Gao, C. Fan, Q. Yan and Y. Wang, "Effects of donor and acceptor's fluorescence lifetimes on the method of applying Forster resonance energy transfer in STED microscopy," Journal of Microscopy, vol. 269, no. 1, pp. 59-65, 2018.

The work function is a fundamental property defined as the minimum energy to remove an electron from the material, and depends sensitively on the bulk electronic structure and surface chemistry, such as surface orientation, surface termination, and presence of surface adsorbates. The work function is a crucial parameter to understand and control in numerous applications, including electron emitters, thermionic energy converters, oxide electronics, solar photovoltaics, (photo)electrochemical reactions and memristors.[1] In particular, there has been tremendous interest in the engineering of low work function materials, such as metals with adsorbed alkali species, semiconductors with functionalized organic species, and 2D materials such as graphene and MXene compounds which are both doped and functionalized with surface species. However, all of these solutions to create low work function materials rely on utilizing surface functional species or monolayers of electropositive elements which are unstable, particularly at elevated temperatures above room temperature. Therefore, a fundamental materials science research question is the following: How does one engineer a material with low work function (e.g. less than 2 eV), and have this low work function realized in a bulk, stable monolithic material?

To answer this question we turn our focus to perovskite oxides, a class of materials with applications ranging from catalysis to piezoelectrics to light absorbing layers in solar cells, which possess broad compositional flexibility enabling the tunability of many properties, including the work function.[2, 3] In particular, engineering of perovskite oxides presents the unique opportunity for realizing low work functions in a bulk, monolithic material by virtue of the intrinsic polar and stable {001} surfaces of the perovskite structure. Our recent work used Density Functional Theory (DFT) calculations to explore a representative set of 20 perovskite oxides for the dual purpose of (1) searching for perovskites which possess a low work function and (2) gaining an improved fundamental understanding of the factors governing the work function in perovskite oxides. From this work, we discovered that pure and Ba-doped SrVO₃ could exhibit low work functions of about 1.9 and 1.1 eV, respectively.[2]

Following this path, we have performed detailed experiments on pure, bulk polycrystalline samples of SrVO₃. The sol-gel method is used to produce high purity SrVO₃ powder, which is pressed into pellets and sintered in a reducing atmosphere. X-ray photoelectron spectroscopy (XPS) measurements indicate that bulk SrVO₃ exhibits a low work function of 1.8-2 eV, and similar values are obtained across the entire bulk sample surface to within +/- 0.2 eV, and is consistent with our DFT predictions. These results suggest that SrVO₃ is the first example of a bulk, monolithic material with an intrinsically low work function, potentially impacting a large array of technological applications and suggesting that engineered perovskite materials are a new class of low work function compounds.

References:
[1] Greiner, M. T., Helander, M. G., Tang, W. M., Wang, Z. Bin, Qiu, J. & Lu, Z. H. Nat. Mater. 11, 76–81 (2012).
[2] Jacobs, R., Booske, J. & Morgan, D. Adv. Funct. Mater. 26, 5471–5482 (2016).
[3] Pena, M. A., Fierro, J. L. G. Chem. Rev. 101 (2001).

Multiscale and Atomistic simulations of materials using empirical interatomic potentials (also called "force-fields", or more generically "models") play a key role in realistic scientific and industrial applications. The Open Knowledgebase of Interatomic Models project (https://OpenKIM.org) includes an automated user-extendable framework for testing the predictions of models for a host of material properties. Visualization tools have been developed to compare model predictions to help select the most appropriate one for a given application. Verification checks ensure the integrity of the models. Models in OpenKIM that conform to the KIM application programming interface (KIM API) can be seamlessly used with several major molecular simulation codes. Although KIM's objective is to have all models conform to the KIM API, this is not always immediately possible or practicable. So, OpenKIM also supports "Simulator Models" (SMs), i.e., models that are currently only available within a single molecular simulation code. SMs are treated just like any other model in OpenKIM so that results for verification checks, material properties predictions, and visualizations are available for SMs as well. This talk will describe the OpenKIM project and how the testing framework can assist materials researchers.

Thanks to the consistent effort in making DFT simulation more reliable and bringing this powerful tool to the engineer, first-principles quantum physics calculations can reliably be utilized to design materials and further to give insights into the material characterization results. In this work, we perform DFT simulation and experimental characterization on 2D material layers, such as InSe and graphene, and doped semiconductor ZnO, to explore and design material properties for different applications. We combine the DFT-atomic force microscopy (AFM)[1], DFT-Optical/Raman spectroscopy[2], DFT-predicted macroscopic surface wettability[3] and DFT device simulation[4] as an integration tool to directly compare the experimental results, such as AFM characterization, contact angle measurement, reflectance/Raman/FTIR spectroscopy, and solar cell IV curve. These full set of comparisons offer a great opportunity to allow the bottom-up realization of material structures that are tailored to deliver the desired device performances.
References
[1] J.-Y.Lu, C.-Y.Lai, I.Almansouri, and M.Chiesa, “Evolution in Graphitic Surface Wettability with First-Principles Quantum Simulations: The Counterintuitive Role of Water,” Phys. Chem. Chem. Phys., 2018.
[2] S. R.Tamalampudi et al., “Thickness-Dependent Resonant Raman and E’Photoluminescence Spectra of Indium Selenide and Indium Selenide/Graphene Heterostructures,” J. Phys. Chem. C, 2019.
[3] J. Y.Lu, Q.Ge, H.Li, A.Raza, andT.Zhang, “Direct Prediction of Calcite Surface Wettability with First-Principles Quantum Simulation,” J. Phys. Chem. Lett., vol. 8, no. 21, pp. 5309–5316, Nov.2017.
[4] B.Alfakes et al., “Optoelectronic Tunability of Hf Doped ZnO for Photovoltaic Applications,” J. Phys. Chem. C, 2019.

Interest in the application of big data and artificial intelligence to the manufacturing industry is increasing. However, this is feasible when all the structures of facilities/environment - sensor - communication - data acquisition - analysis are balanced. It is difficult to expect remarkable achievement in a short time considering the situation of small and medium-sized manufacturing companies. Especially in the foundry industry, there are many difficulties in introducing big data and artificial intelligence technology because it traditionally tends to depend on worker's know-how in controlling processing parameters. Nevertheless, there is still a need to build a smart factory, so it is necessary to study IoT sensing system for collecting big data of processing parameters and machine learning technique for prediction.

In this presentation, the difference of concepts between artificial intelligence, machine learning and deep learning is introduced. In connection with data acquisition using IoT sensors, an example of application to the centrifugal casting field using an Arduino-based module is introduced. In addition, a case study of machine learning model for prediction of hydrogen concentration in molten aluminum is introduced. The development of machine learning technology is very fast and it is necessary to grasp the flow while applying it in the manufacturing field even though it is quite simple case.

Current piezoelectric materials rely on a unique resilience of the polar response to disorder which produces a stable structure near a morphotropic phase boundary (MPB). However, many piezoelectric materials are not amenable to disorder, and are thus unusable in MPB systems. In this work, we apply a sensitivity analysis approach to ab initio calculated piezoelectric properties to determine the effect of disorder on the piezoelectric response. Vibrational properties are found to drive degradation in the piezoelectric effect with disorder, rather than the dielectric, or internal strain components. In well-known piezoelectric systems, (perovskites and tungsten bronze structures) multiple stable optical phonon modes are found to contribute to the piezoelectric response, providing a fingerprint for disorder tolerance. Hence, a multiple-phonon mode criteria is used to evaluate candidate materials suggested by the Materials Project and to screen the database for novel, disorder-tolerant piezoelectrics. Eight prototype structures are altered through chemical substitution to generate novel systems with high piezoelectric effect. These structural families may be explored to replace PZT as MPB systems beyond perovskites.

Laser Aided Additive Manufacturing (LAAM) is a kind of metal additive manufacturing process using the high-powered laser beam to melt the powdered or wire feedstock to build the near-net three-dimensional components. Studying thermal field induced by different laser parameters is important to evaluate and optimize the resultant residual stress and distortion distribution. However, it is very computationally expensive to simulate multi-bead deposition process using existing numerical model to analyze and select appropriate process parameter. This work proposed a laser process parameter planning frame to design an appropriate laser process parameter sequence for the LAAM process. The frame utilizes a recently developed finite element (FE) model to predict the thermal behaviour during the deposition process, and a reinforcement neural network to modify the laser process parameter based on the temperature field in process. The optimized process parameter is able to improve the temperature distribution homogeneity of the deposited part.

Hydrocarbons include a wide range of chemical substances that can be found in many products ranging from oil extracted directly from the reservoir to drug products and occur in a range of mixtures, degrees of purity and molecular size. Electron Affinity (EA) has been identified as an important factor for predicting the chemical properties of hydrocarbons. For instance, previous investigations found that the EA is a principle descriptor for modeling the gas phase oxidation rates of polyaromatic hydrocarbons (PAH) in diesel combustion which indicates that the initial step in the destruction of PAHs is capturing an electron. Additionally, previous studies could find a non-linear correlation between the phototoxicity and the EA of the PAHs. There are experimental and Quantum Mechanics (QM) techniques to calculate the EA of species. However, due to high experimental and computational costs, it is beneficial to develop machine learning (ML) methods based on reasonable amount of available data that can accurately and quickly estimate the EA of hydrocarbons. Supervised ML can be implemented on available data sets to predict properties based on individual structural features. In this study, we used Density Functional Theory (DFT) to calculate the EA of more than 1500 hydrocarbons including PAH, alkanes, cycloalkanes and alkenes. Then a ML technique was developed to explore the relationship between the EA and the physical structure of the species. Using this method, we could reveal the role of different geometrical features and functional groups on the EA of the hydrocarbons.

For the past several decades, the standard method of materials discovery has involved a combination of research intuition and comprehensive experimental investigation. Recently, a variety of new computational and experimental tools have been introduced to go beyond this standard paradigm. In this talk, I will describe our group's effort in creating an automated pipeline for suggesting and discovering new materials in silico. We have recently coupled our capability to perform high-throughput density functional theory calculations for a variety of materials properties (software implementation: https://atomate.org) with inverse optimization techniques that allow machine learning surrogate models to suggest the most promising compounds for exploration. This allows a computational optimization to be performed on supercomputing resources in a closed loop without any human intervention (software implementation: https://github.com/hackingmaterials/rocketsled). The machine learning model used for the optimization can be provided by the user, be chosen from a set of standard optimization routines, or use a new machine learning surrogate that we have developed that can be trained automatically and without any human intervention on any materials composition-property or structure-property data set (software implementation: https://github.com/hackingmaterials/automatminer). Finally, I will outline a collaborative effort at LBNL in which we analyze text data from materials abstracts to predict "gaps" in the research literature, and use this information to power an algorithm to suggest promising new compounds for functional applications. These tools make it possible to leverage large computing resources and the past scientific literature to focus and accelerate the experimental phases of materials discovery.

Scalable synthesis of two dimensional (2D) materials is a major bottleneck to more widespread adoption of layered material-based devices. Chemical vapor deposition (CVD) has emerged as a viable method for large-scale synthesis of 2D materials. However, neither experiment nor theory has been able to decipher mechanisms and selection rules for different growth scenarios or make predictions of optimized growth parameters. Experimental challenges stem from the use of atomic-resolution probes like TEM to characterize CVD growth reactions in situ under elevated temperatures and pressures. Computational synthesis, which simulates CVD processes using reactive molecular dynamics simulations provides the atomistic resolution necessary for the identification of reaction mechanisms and synthesis pathways. Here we use recurrent neural networks and reinforcement learning methods trained on trajectories from several hundred simulations of computational synthesis of a prototypical two-dimensional semiconductor, monolayer MoS₂, to uncover the dependence of product stoichiometry, crystallinity and phase distribution on reaction parameters like temperature, sulfur and hydrogen partial pressures, thus paving the way for rational design of CVD synthesis techniques for MoS₂ and other layered materials.

This work was supported as part of the Computational Materials Sciences Program funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award Number DE-SC0014607. Simulations were performed at the Argonne Leadership Computing Facility under the DOE INCITE program and at the Center for High Performance Computing of the University of Southern California.

Computational modeling and high-throughput techniques have the promise to enable the atom-by-atom design of nanoscale properties that give rise to essential changes in macroscale properties. In the quest for increasing energy efficiency and resource utilization, energy losses remain an outstanding challenge that can be solved in part through computational materials design. For example, during the operation of a car engine, only 15% of the fuel energy is used for car movement, whereas the remaining is lost. Friction reducers (FRs) additives can minimize the friction loss in these engines by reducing friction between moving parts at the contact area. Traditional FRs contain metals, sulfur, and phosphorus, which can poison exhaust system catalysts and diesel particulate filters. Thus, if suitably designed, organic friction reducers (OFRs) present a promising alternative solution. Here, we apply non-equilibrium molecular dynamics simulations together with density functional theory methods under a high-throughput workflow to enable the atom-by-atom design of OFRs. We directly compute the coefficients of friction (COFs) of OFRs on model-engine iron oxide surfaces. This atomistic simulation allows us to develop direct physical insight into the nanoscale properties of OFRs that give rise to the criteria of friction reducing characteristics. By carrying out coverage- and temperature-dependent simulations, we can explore several conditions not easily probed during experiments. These studies allow us to build a quantitative-structural-property relationship for predicting good OFR characteristics, enabling an iteratively improving materials design workflow.

Ferromagnetic semiconductors are valued for their potential applications in spintronics. For spin-polarized transport, combining strong ferromagnetism and attractive semiconducting properties in one material is highly desirable, but yet this remains an open problem. Here we conduct a search for concentrated ferromagnetic semiconductors through high-throughput computational screening. Our screening reveals the limited availability of semiconductors combining ferromagnetism and a low effective mass. Among the identified ferromagnetic semiconductors are Eu chalcogenides, Cr spinel chalcogenides, Bi manganites, Mn pyrochlore oxides, and Mn double perovskites. In particular, we show that the manganese pyrochlore oxide In₂Mn₂O₇, hitherto unknown to spintronic applications, is a promising candidate for spin transport as it combines low electron effective mass (0.29 m₀), a large exchange splitting of the conduction band (1.1 eV), good stability in air, and a Curie temperature (about 130 K) which is among the highest of concentrated ferromagnetic semiconductors. We rationalize the high performance of In₂Mn₂O₇ by the unique combination of a pyrochlore lattice favoring ferromagnetism with an adequate alignment of O-2p, Mn-3d, and In-5s, forming a dispersive conduction band and enhancing the Curie temperature. We further find that Sn and Mo can be incorporated on the In site while acting as shallow donors, suggesting that In₂Mn₂O₇ can be effectively n-doped.

The discovery of novel materials can generate immense technological, economic, and social benefits. However, these are slow, challenging, expert-intensive efforts. Our thesis is that new capabilities of cognitive computing – particularly natural language processing, knowledge representation, and automated reasoning – are poised to transform the process of materials discovery and take us from our current discovery paradigm driven by data science and machine learning to the next stage where cognitive systems seamlessly integrate information from human experts, experimental data, physics-based models, and data-driven models to speed discovery. We discuss the key bottlenecks to discovery that need to be removed to enable this new approach and illustrate progress towards this cognitive future with examples from IBM research efforts in Accelerated Materials Discovery.

The ever-increasing power of modern supercomputers, along with the availability of highly scalable atomistic simulation codes, has begun to revolutionize predictive modeling of materials. Molecular dynamics (MD), in particular, has led to breakthrough advances in diverse fields, including tribology, energy storage, catalysis, sensing. Furthermore, recent integration of MD simulations with X-ray characterization has demonstrated promise in real-time 3-D atomistic characterization of materials. The popularity of MD is driven by its applicability at disparate length/time-scales, ranging from ab initio MD (hundreds of atoms and tens of picoseconds) to all-atom classical MD (millions of atoms and tens of nanoseconds), and coarse-grained (CG) models (microns and tens of micro-seconds). Nevertheless, a substantial gap persists between AIMD, which is highly accurate but restricted to extremely small sizes, and those based on classical force fields (atomistic and CG) with limited accuracy but access to larger length/time scales. The accuracy and predictive power of classical MD is dictated by the empirical force fields, and their capability to capture the relevant physics.

In this talk, I will present some of our recent work on the use of machine learning (ML) to seamlessly bridge the electronic, atomistic and mesoscopic scales for materials modeling. Our automated ML framework aims to bridge the significant gulf that exists between the handful of research groups that develop new interatomic potential models (often requiring several years of effort) and the increasingly large user community from academia and industry that applies these models. Our ML approach showed marked success in developing force fields for a wide range of materials from metals, oxides, nitrides, hetero-interfaces to two-dimensional (2-D) materials and even water (arguably the most difficult system to capture from a molecular perspective). This talk will also briefly discuss our ongoing efforts to integrate such cheap yet accurate atomistic models with (a) AI techniques to perform inverse design and construct metastable phase diagrams of materials (b) Deep learning to improve spatiotemporal resolutions of ultrafast X-ray imaging.

The development of advanced materials via high-throughput experimentation at Intermolecular is accelerated using guidance from modeling and machine learning (ML). Rapid materials development at a reasonable cost implies that a large number of parameters describing the deposition outcome and materials performance (including composition, phase, texture, defects, stability/degradation) and the knobs that directly and indirectly affect those outcomes (including deposition method, precursors and ambients, temperature and other conditions, film thickness, and the identity of other materials included in the stack) form a vast space that needs to be efficiently navigated. This talk primarily focuses on methods, with a few examples of specific findings from our internal R&D.

Accessing external DFT databases (such as MaterialsProject) together with high-performance computing (HPC) resources for new DFT simulations and pymatgen-enabled workflows permits rapid on-demand theoretical modeling and thermodynamic analysis. This can be used to screen materials in a stack based on compatibility, allowed chemical potentials, possibility of templating, and wetting, so as to identify most promising combinations of material layers that make up the device stack, together with desired deposition conditions. For instance, additional material layers can be identified, deposition of which ensures high quality (e.g. low vacancy concentration) of preceding layers despite the use of aggressive chemistries in depositing subsequent layers. One can further theoretically pre-screen “dopants” that help stabilizing a given phase of a given material. While relatively straightforward in the case of stabilization of TiO₂ phases, such a theoretical pre-screening needs be more involved in the case of HfO₂. We compare different approaches to identifying regions of thermodynamic stability in multicomponent systems, including CALPHAD, theoretical models of disordered alloys used together with DFT database queries, phenomenological selection rules and recursive partitioning based on experimental data. Unexpected phases such as alloys between ordered compounds can be predicted using thermodynamic modeling with effective Hamiltonians such as cluster expansions (CEs).

HPC-enabled, DFT-driven search for novel material phases can help building a better reference set for phase identification in thin films; e.g. our analysis during development of ferroelectric HfO₂ films accounts for the novel rhombohedral phase, predicted by our DFT study[1] and independently demonstrated in recent experiments[2]. Combining DFT simulations with phenomenological modeling and data-driven search allows us to identify materials with better intrinsic properties, such as materials combining high band gap E_g with high dielectric constant κ. Some materials identified by our workflows combine attractive (κ,E_g) values with possibility of synthesis at ambient conditions, even though showing smaller improvement over the common high-κ materials than the exotic materials such as the high-pressure BeO phase predicted in literature[3] (that is yet to be stabilized at the ambient conditions). We discuss practical considerations that limit materials development based on such predictive workflows.

Finally, we illustrate co-optimization of multiple target material properties using ML based on experimental data. We discuss issues that may arise due to human pre-selection of promising phase space regions: ML algorithms readily pick up pre-detected trends, but may draw unjustified conclusions, particularly in the presence of noise/measurement errors. Use of physically meaningful descriptors, including those obtained using DFT simulations and data mining, leads to detecting important trends, improving predictions, and identifying optimal compositions.

[1] S.V. Barabash, J.Comput. Electron. 16, 1227 (2017).
[2] Y. Wei et al., Nature Mater. doi:10.1038/s41563-018-0196-0.
[3] K. Yim et al., NPG Asia Mater. 7, e190 (2015).

Time-of-flight secondary ion mass spectrometry (ToF-SIMS) is continuously advancing. The data sets now being generated are growing dramatically in complexity and size. More sophisticated data analytical tools are required urgently for the efficient and effective analysis of these large, rich data sets. Standard approaches to multivariate analysis are being customised to decrease the human and computational resources required and provide a user-friendly identification of trends and features in large ToF-SIMS datasets.

We demonstrate the generation of very large ToF-SIMS data matrices using mass segmentation of spectral data in the range 0 – 500 m/z in intervals ranging from 0.01 m/z to 1 m/z. No peaks are selected and no peak overlaps are resolved. Sets of spectra are calibrated and normalized then segmented and assembled into data matrices. Manual processing is greatly reduced and the segmentation process is universal, avoiding the need to refine peak lists for different sample types or variants.

ToF-SIMS data for standard polymers (PET, PTFE, PMMA and LDPE) and for a group of polyamides are used to demonstrate the efficacy of this approach. The polymer types of differing composition are discriminated to a moderate extent using PCA. PCA fails for polymers of similar composition and for data sets incorporating significant random variance.

In contrast, artificial neural networks, in the form of self organising maps (SOMs), deliver an excellent outcome in classifying and clustering different and similar polymer types and for spectra from a single polymer type generated using different primary ions. This method offers great promise for the investigation of more complex bio-oriented systems.

We compare the analysis of large scale mass segmented matrices with those formed using conventional selection of ToF-SIMS peak lists. SOMs are used to cluster and discriminate antibody fragments bound at surfaces and to demonstrate antibody orientation in optimised ELISA format assays.

The evolution of materials microstructure driven by radiation in nuclear applications poses an imminent need to describe the behavior for system reliability, safety and economics. Radiation effects on materials such as void swelling and embrittlement have been long standing complex phenomena yet to be well characterized under different conditions. Specifically, radiation-induced void swelling is a unique mode of degradation in nuclear materials. Serious efforts have been spent to increase radiation tolerance, which is to postpone the consequence of radiation damage. However, the behavioral sensitivity to a number of parameters such as chemical compositions and the irradiation environment makes the conventional trial-and-error experimental approach highly inefficient. Meanwhile, the physics-based computational models are also subject to inefficiency and critique of fidelity due to the complex underlying defect dynamics on multi-scales. In the era of increasing data resources from resource-consuming nuclear materials experiments, machine learning (ML) as an alternative could prove useful in more efficiently and accurately predicting macroscale materials response and guiding experiment testing. It is then applied to perform prediction of the incubation period of void swelling based on manually collected experimental data. By training decision tree based ensemble models, it delivers very good performance in prediction, and the identified main contributing factors such as temperature, Fe and Cr content, and dose rate are fairly consistent with established understandings. Ultimately, establishing a well-performed ML model is a promising route to radiation resistant material design. It is thus our plan to construct a public web platform to crowdsource more data for nuclear materials exploration. And from the methodology point of view, ML can be utilized as a subroutine in a physics-based modeling technique for acceleration. Such types of reduced-order model can constitute a future research direction, by exploring a hybrid of ML methodology and physical framework to alleviate current computational constraints in characterizing radiation-induced microstructural evolution.

Transmission electron microscopy (TEM) is the characterization method of choice to observe the atomic-scale and microstructural local features within materials that play a critical role in material performance. Current TEM capabilities include a wide variety of imaging modalities to probe a material’s structure. It is possible to resolve structures with 0.39 angstrom resolution [1], to perform rapid nanodiffraction experiments to characterize the microstructure of hard and soft materials [2], to resolve local chemistry and bonding [3], and much more. Furthermore, recent advances in fast electron detection enable imaging at nearly 100 kHz rates, enabling extremely rapid data acquisition. Directly incorporating these TEM capabilities for structural validation into a high throughput materials prediction, design and synthesis routine would profoundly speed up the materials discovery process.

However, a bottleneck exists between image acquisition and the extraction of relevant information that can be used in a materials design feedback loop. While image analysis of individual images can easily identify regions of interest and determine whether they contain defects, it is prohibitively time-consuming to manually perform this analysis on large numbers of images. This means that, for example, in a given nanoparticle synthetic optimization process, only a small number of successful end products are studied in detail. Failure cases are typically not characterized, and population heterogeneity statistics are not captured. Therefore, to automatically identify and quantify defects, size and shape statistics, and other key structural features which can dominate a material’s mechanical, electronic, and catalytic properties, a new approach is required.

Advances in machine learning and computer vision have made high accuracy automated image interpretation possible. While widely applied in the life sciences, this approach is only recently being applied to atomic resolution TEM images. Here, we present application of machine learning and other high-throughput methods to TEM images for nanoparticle identification and microstructural characterization. When combined with existing automatic image acquisition protocols, this approach is a viable option to close the materials design loop and incorporate TEM into high-throughput materials design and synthesis in a way not currently possible.

[1] Jiang, Yi, et al. "Electron ptychography of 2D materials to deep sub-ångström resolution." Nature 559.7714 (2018): 343.
[2] Ophus, Colin. "Four-Dimensional Scanning Transmission Electron Microscopy (4D-STEM): From Scanning Nanodiffraction to Ptychography and Beyond." Microscopy and Microanalysis (2019): 1-20.
[3] Bosman, M., et al. "Two-dimensional mapping of chemical information at atomic resolution." Physical Review Letters 99.8 (2007): 086102.

Visualization techniques, such as scanning and transmission electron microscopy, electron diffraction, X-ray computed tomography and magnetic resonance imaging, among others, are widely used providing high spatial resolution images of atomic arrangements, crystallographic information, material’s shapes, sizes and other microstructure information including defects and voids within materials. Significant advancements in materials characterization methods are providing higher resolutions images, complex information and faster data collection capabilities. With these breakthroughs in visualization techniques, the bottleneck in advancements in materials characterization will no longer be the capability limitations of the characterization tools themselves, but rather, the ability to rapidly analyze and interpret the large amount of complicated (high dimensional), high-quality data. To address this challenge, we are implementing machine learning approaches to accelerate image analysis process, extracting images features to correlate to mechanical performance.

Breakthroughs in machine learning have shown us that deep learning (DL) has significant advantages over traditional machine learning (ML) and computer vision (CV) techniques for a variety of applications, most notably: improved predictive performance and automated learning of feature representations with minimal human guidance. However, important limitations remain. In particular, DL typically requires more labeled training examples than traditional ML approaches, and it is often difficult to explain model performance. In order to assess application of computational tools for materials science, we chose to compare the two approaches: (1) a traditional ML approach (random forest) using state-of-the-art computer vision features and (2) an end-to-end deep learning approach. In this presentation, approaches taken to extract image features from SEM images of molecular solid crystals and using these features to predict materials performance will be discussed. We show that our image-based ML approach reduces root mean square error (RMSE) by an average of 51% over a non-image-based baseline.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 and was supported by the LLNL-LDRD Program under Project No. 19-SI-001.

The past decade has seen a rise in the concept of automated materials discovery. A large majority of the work in this area has focused around computation of intrinsic material characteristics and the creation of material property libraries¹. This has been paired with methods to search for suitable candidates from these libraries. However, these predictions still require experimental validation. To this end, high throughput synthetic methods have followed both in the area of organic and inorganic materials². The missing piece to automated materials discovery is then in high throughput characterization of materials. High throughput materials characterization represents the final feedback step to both computational predictions and high throughput syntheses.

High resolution transmission electron microscopy (HR-TEM) is particularly suited to the materials discovery pipeline because of its high resolution and local information. Most computational libraries are based on DFT calculations which are directly tied to atomic structure. By providing atomic structural information, TEM could be used to verify structural information in computation. This information is also critical to the synthesis of materials to verify the intended atomic structure is being created.

Since nanomaterials are particularly suited to analysis by HR-TEM characterization, in this talk we will address the current state of high throughput HR-TEM characterization of nanomaterials. The first fundamental problem which must be solved to allow for automation is segmentation of nanoparticle regions from the micrograph background. This is particularly challenging since the signal to noise of the particle above the background is extremely low in HRTEM micrographs. However, we have developed a convolutional neural network (CNN) which is able to segment nanoparticles with a 94% pixel-wise accuracy. The neural network is shown to be robust to changes in contrast and is sensitive to the presence of lattice fringes. We will also discuss automated detection of crystal structure and defects.

1. Green, M. L. et al. Fulfilling the promise of the materials genome initiative with high-throughput experimental methodologies. Appl. Phys. Rev. 4, 011105 (2017).

2. Chan, E. M. et al. Reproducible, High-Throughput Synthesis of Colloidal Nanocrystals for Optimization in Multidimensional Parameter Space. Nano Lett. 10, 1874–1885 (2010).

Laboratory data collection should be comprehensive but adaptable. An ideal system should provide a mechanism for specifying unambiguous machine-readable experiment that enable remote operation and replicability, presenting these as instructions to human operators and machines, capturing comprehensive data and metadata during experiment, and performing extraction, transformation, and loading (ETL) to facilitate machine learning. Yet, existing tools require significant development time that is incompatible with rapidly evolving scientific needs.

In this talk, I will describe ESCALATE (Experiment Specification, Capture and Laboratory Automation Technology) an adaptable open source package for experiment description and data collection. As a specific example, I will describe its application to Robotic-Accelerated Perovskite Investigation and Discovery (RAPID). The first generation of RAPID uses inverse temperature crystallization (ITC) to grow halide perovskite single crystals for x-ray structure determination and bulk characterization using commercial liquid handling robots. All experiment plans for the syntheses are contributed remotely, by both human scientists and algorithms trained on the reaction data. More >25 compounds have been produced by the RAPID:ITC (compared to only 4 known ITC-grown halide perovskites prior to our work), including >10 new compounds or polymorphs. Incoming data collected by ESCALATE is used to automatically train machine learning models, evaluate model performance and feature influence, and quantify reproducibility. A live web dashboard communicates these insights to the scientist and management in visual form. I will conclude by describing case studies about new scientific insights extracted from the comprehensive RAPID dataset, and discuss ongoing deployments of ESCALATE to perovskite thin film and vapor diffusion experiments.

With over a hundred elements in the periodic table, there exists a large number of potential new materials. Some of these new materials and rational optimization of devices based on them can provide solution to many of the growing and daunting challenges we face today, from climate change to water shortage and from faster internet to longer-lasting orthopedic implants. However, in spite of major investments in the last two decades resulting in dramatic advancement in performance of the material science tools, rate of discovery of new materials is still sluggish; it still takes over 20 years for development of a new technological material. The rate of new material discovery lags the performance enhancement of material synthesis and characterization tools because the traditional material discovery approaches based on human curation and perception are unable to cope with the revolution in data generation. They are unable to fully harness accelerating rate of data (both computational and experimental) into discovery of new materials.
Another way of looking at the problem is that the three stages of the discovery cycle, starting with predicting new experiments, performing measurements efficiently, and ending with finding new insights and discoveries in the measurements, are performed sub-optimally. Furthermore, these three stages are isolated from each other halting the discovery cycle. Suboptimal performance and broken links between the stages delay discoveries.
The first stage is suboptimal because human curators are finding it increasingly difficult to find trends hidden in rapidly growing and increasingly complex materials data. Overwhelmed researchers are unable to make accurate predictions of experimental conditions (and or compositions) where new desired material may be located. Absence of reliable prior hypotheses, combined with the inability of a human curator to keep pace with information contained in high throughput measurements in real-time, forces a strategy of rapid, but blind searches of vast experimental space, resulting in increasingly large fraction of data of poor quality and of little scientific value. In a blindly cast search, even when the grid catches evidence for a new material, it often takes human researcher weeks of frustrating work wading through large number of poor quality measurements to find it. It is like searching for a needle in a haystack in a lightless barn. The performance of the final stage of the discovery cycling is becoming worse as materials become more complex and role of processing becomes more significant, making experimental search space too complex (multi-dimensional) for a human curators, using traditional approaches, to find subtle but significant signatures of new discoveries even after prolonged data analysis.
Here, we use the search for new compositional complex alloys, to illustrate a new material discovery approach that leverages recent advances in big-data analytics, machine-learning, and artificial intelligence to link and accelerate performance of each of the three stages of the discovery cycle. The approach is based on two key learning machines: the first makes extraction of new knowledge from large and complex dataset fast, sophisticated, and independent of human curation; and the second engine then uses rapidly extracted information in a Bayesian approach to make decision-making smarter and automated. Data-driven decision-making can not only accurately predict the next experiment to perform, but optimize the experimental data collection measrement-by- measurement so only high quality measurements of with higher scientific content are collected. Close coupling of these two learning machines will bridge gaps that stall discovery cycle, and rapid iteration of one to the other will accelerate discoveries of new technological materials and devices.

Homogeneous catalysis using transition metal complexes is ubiquitously used for organic synthesis, as well as technologically relevant in applications such as water splitting and CO2 reduction. The key steps underlying homogeneous catalysis require a specific combination of electronic and steric effects from the ligands bound to the transition metal atom. Finding the optimal combination is a challenging task due to the exceedingly large number of possibilities and the non-trivial ligand-ligand interactions. In this work, we show how high-throughput density functional theory (DFT) and machine learning methods can be combined to accurately predict activation energies for H2 splitting within large chemical spaces containing thousands of derivatives of the Vaska’s complex (trans-[Ir(PPh3)2(CO)(Cl)]). A hybrid approach combining Gradient Boosting Regression with Gaussian Processes allows reach high accuracies (MAE = 0.59 kcal/mol). In contrast with DFT calculations requiring several days to be completed, the machine learning models were trained and used on a laptop on a time-scale of minutes. By using interpretable representations and ranking the feature importances, our approach allows us to interpret the machine learning models and extract design rules to systematically control activation energies. We will give an outlook on how this model can be generalized to other chemical reactions and to derive non-trivial application specific design rules in a fully automated way.

Next-generation photovoltaics require high reliability to scale up to market applications. Although a number of promising novel solar cell materials, such as organic solar cells or perovskites, have shown promising performance, improving reliability remains a difficult and time–consuming challenge due to complex and lengthy characterization, and challenging decoupling of dominant degradation mechanisms [1]. Furthermore, common degradation mechanisms among similar samples are often difficult to identify, due to the dynamic nature of degradation and the use of sparse experimental datasets. In this context, we propose a methodology to accelerate 10X–100X the reliability characterization and improvement of solar cell materials and devices by combining a semi-supervised machine learning approach, used for prediction of degradation, and an inference approach based on a device numerical simulator, used to identify root-causes of degradation. First, we adapt the Latent Dirichlet Allocation (LDA) [2] algorithm to identify common degradation trends and regimes in a set of current–voltage (JV) measurements of various unlabeled and weakly-labeled solar cells. Second, using the common degradation regimes extracted by LDA and the initial few hours of degradation, we develop and train a recurrent machine learning model [2] to predict the JV characteristics of the sample of interest at the 1000^th hour of degradation. Finally, we use the predicted JV degradation as function of time to infer, via a device model in a Bayesian setting, the degradation time–series of the intrinsic material and device properties, such as electron and hole mobility of the active layer, or the interfacial properties. By comparing the change in time of various material and device properties, the knowledge of these time–series allows us to quickly identify the main drivers of degradation and suggest improvements for solar cell reliability. To test our methodology, we synthesize, degrade and measure 300 samples of organic photovoltaic materials: established P3HT–PCBM, and next-generation PBDBT–ITIC and P3HT–IDTBR. Each sample in the dataset presents various degradation trends due to varying active layer compositions, hole and electron transport layer combinations, process variation and interfacial properties. Given, the high number of sources of variables, the dataset is inherently sparse. In this context, our methodology successfully determines the final degraded JV characteristics using less than 10% of the degradation measurements, on average, and successfully decouples the root-causes of degradation across the sparse dataset. Based in our results, we propose and test potential reliability improvements for PBDBT–ITIC and P3HT–DTBR solar cell materials and architectures.

Reference:

[1] Asghar, M. I., et al. "Device stability of perovskite solar cells–A review." Renewable and Sustainable Energy Reviews 77 (2017): 131-146.

[2] Blei, David M., Andrew Y. Ng, and Michael I. Jordan. "Latent dirichlet allocation." Journal of machine Learning research 3.Jan (2003): 993-1022.

[3] Harvey, Andrew C. Forecasting, structural time series models and the Kalman filter. Cambridge university press, 1990.

Automating critical steps of synthesis and screening experiments enables a variety of modes of materials exploration. High throughput experimentation comprises a family of techniques wherein materials systems can be comprehensively explored, and the resulting data relationships, e.g. composition-property and composition-structure-property relationships, are emblematic of the knowledge obtained from the experiments. Application of high throughput experimentation for solar fuels technology, in particular (photo)electrocatalysis of the oxygen evolution reaction, has led to a breadth of discoveries, many of which are based on high throughput computational screening. The resulting database of experiments, which is publicly released as the JCAP Materials Experiment and Analysis Database (MEAD) containing 6.5 million measurement files collected on 1.5 million materials samples, is a key resource for developing and evaluating algorithms that automate data interpretation. Successes to date include application of machine learning techniques to learn, identify, and communicate hidden data relationships. At a high level, these algorithms automatically generate answers to human-identified research questions, moving the frontier of artificial intelligence in materials discovery to the automatic identification of the interesting research questions.

Recently, materials scientists have started to utilize machine learning to accelerate experimental research. Active learning – an AI field dedicated to optimal experimental design – is a particularly promising tool; it provides systematic means to identify the shortest path toward a material with some desired properties or the experiments that maximize knowledge of the explored space. In many materials science tasks, however, the goal is to obtain a mapping between two or more experimentally measured quantities. Standard active learning algorithms may not be optimal for such complex scientific problems. Reducing the experimental effort to obtain such mappings can be optimized not by independently running several active learning tasks but rather by a strategy coordinating different experiments performed simultaneously. In this talk I will present the idea of cooperative learning, and illustrate it with examples from several different high-throughput studies.

We have significantly expanded the knowledge-base of metal catalysts through a unique combination of manifold learning, Gaussian process regression and clustering approaches. Given the complexity in performing catalytic measurements, the amount of data available for selecting ‘optimal’ catalysts for specific reactions is limited. The work described here develops an analysis framework suitable to the small number of measurements available, while also developing a large relevant descriptor-base. We have performed the foundational work needed to develop a catalyst discovery toolkit. Using volcano plots as a platform, we have fused manifold learning methods and graph network methods from which one can rapidly explore new chemistries for single atom alloy (SAA) catalysts. We use single atom systems for testing our robustness, with the added benefit that prior work on single atom systems has not utilized machine learning. Using SAAs allows for a rapid screening of the combinatorial design space. We developed a machine learning logic for screening chemistries to define necessary detailed DFT calculations and have identified 28 alloys which are most promising for further exploration.

Zeolites are inorganic nanoporous materials with broad industrial applications as catalysts, ion exchangers, and separators. Despite sustained research, controlling polymorphism is still a critical challenge in their design, relying mostly on trial-and-error. First-principles calculations could aid the search for new frameworks, but the number of theoretically accessible topologies and the complexity of their kinetic mechanisms render this approach computationally prohibitive. Here, we employ a suite of computational tools such as big-data, graph theory, structural kernels, density functional theory (DFT), and machine learning to explain and predict zeolite transformations. We first relate solid-state transformations to materials descriptors by combining crystallography with a graph-theoretical metric. Supported by exhaustive literature, we then show that interzeolite diffusionless transformations occur only between graph-similar pairs. Moreover, all known instances of intergrowth take place between either structurally- or topologically-similar structures. Our metric suggests hundreds of low-distance pairs between known frameworks and thousands of hypothetical frameworks for realizing novel transformations and intergrown crystals. Such insights are further refined by atomistic simulations. Building on millions of DFT calculations, we parameterize interatomic interactions in pure-silica zeolites using neural network models and active learning. The method enables accurate structural optimizations and off-equilibrium energy sampling with low computational cost, allowing the selection of favorable graph-driven phase transitions between frameworks and uncovering new synthetic pathways for zeolites.

A significant challenge in utilizing machine learning approaches to accelerate materials development is the lack of large and structured data sets. While there are ongoing community efforts to create collaborative materials databases and repositories for this purpose, the diversity and breadth of data types, length scales, and applications still makes it challenging to create such all-encompassing materials databases that are of broad practical use. However, if tools are developed to automatically process the vast scientific literature and extract and structure information of interest to a given user, then such tools can enable the easy creation of personalized databases with user-specified relevant information to which data mining approaches can then be applied.

We developed such tools for the automated extraction of a suite of information from the text of articles pertaining to nanomaterials synthesis and demonstrate their utility for nanomaterials synthesis. Attaining nanomaterials of desired composition, dimension, and morphology is critical for end-use applications but challenging to do, often requiring time-consuming iterations of synthesis and characterization. Using a corpus of 35k nanomaterials-related articles, we first use a simple unsupervised classification algorithm based on the frequency of occurring terms to identify the primary nanomaterial composition and morphology in each article. Classifying and analyzing articles based on their targeted nanomaterial composition and morphology by itself provides a bird’s eye view and can help identify “hot topics” in the field or alternatively under-studied or challenge-to-synthesize nanomaterials. Next, we apply a supervised machine learning approach to our corpus to identify and extract from articles’ text the sentences related to the nanomaterials’ synthesis protocols, thereby yielding a useful synthesis reference library. Interesting, we find that function words (i.e., to, in, for, of, at) commonly omitted in natural language processing of non-scientific text are in fact a characteristic trait in discriminating between synthesis- and non-synthesis-related sentences in scientific text. With synthesis protocols in-hand, we further process these via chemical entity recognition (CER) to identify and extract the chemicals used in various nanomaterials’ syntheses. We evaluate a variety of open-access CER tools, as well as our own in-house developed CER tool, that each utilize different tokenizers for parsing the text and techniques for identifying chemicals, and we find that, despite the variety of approaches undertaken, most tools have comparable performance with a peak f1 score of 87%. Normalizing the chemicals names extracted from articles, we then have the opportunity to compare the frequency of use of chemicals for various nanomaterial morphologies. We demonstrate how such analysis provides useful insights as to the importance of chemicals in directing the growth of nanoparticles during synthesis to desired morphologies, like for example nanowires versus nanospheres versus nanocubes. We package this database created entirely by extracting information from existing nanomaterials literature into a browser-based visualization tool we developed that enables easy exploration of the data, thereby helping guide hypothesis generation and reduce the potential parameter space during experimental design.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS-777678

Bridging the gap between computational predictions and industrial applications requires acceleration and automation of experimental synthesis, characterization and data analysis. High Throughput Experimentation (HTE). also known as combinatorial experiments, is one possible approach to accelerate materials research. Thus, HTE combinatorial methods have been regarded as a promising approach to fulfill the promise of the materials genome initiative (MGI)[], complementary to high throughput computations and industrial research and development.
This presentation will focus on the state of high throughput experimental materials research methods at National Renewable Energy Laboratory (NREL). First, I will discuss methods for creating later gradients of thin film sample properties, in particular going beyond chemical composition of metals [1]. Then, I will talk about spatially-resolved characterization methods, including interlaboratory exchange of samples [2]. Finally, I will highlight our efforts in streamlining data analysis in combinatorial materials science, including a recently published COMBIgor software package [3]. These methods will be illustrated by a broad range of materials examples, including oxides, chalcogenides, and nitrides.
[1] ACS Combinatorial Science (2018) 20 436
[2] ACS Combinatorial Science (2019) 21 350
[3] ACS Combinatorial Science (2019), DOI: 10.1021/acscombsci.9b00077

The so-called 4^th revolution in science began with the advent of machine learning (ML), as well as high-throughput (HT) experimentation and robotization. Herein, we describe a workflow that enables rapid screening of a parameter space of hybrid composites, comprised by carbon nanotubes and poly-3-hexylthiophene (CNTs:P3HT) for thermoelectric (TE) applications, with Bayesian optimization embedded in the feedback loop in order to explore the space in a more efficient fashion.
The parameter space under scrutiny includes 4 types of nanotubes (both single- and multiwall), 16 CNTs:P3HT ratios, 2 solvents (o-DCB, and chloroform) and 3 doping conditions, which result in 384 unique synthetic parameters. The setup used involves a robotic pipettor and the microfluidic flow reactor with XY stage for automatic drop casting. We synthesize more than 150 samples per hour, as compared to 4 samples per hour with traditional manual procedures. Characterization was done with means of hyperspectral imaging and 4-point-probe measurements to ascertain optical properties and electrical conductivity.
The conductivity data of the initial experiments was used to train the ML algorithm. After training, the algorithm inferred new experimental conditions for achieving highest possible conductivity, closing the feedback loop. In the end, 3 iterations of the experiments yielded the value of conductivity higher than 150 S cm^-1.
Further work includes the optimization of the experimental setup as well as the ongoing effort to use hyperspectral imaging to bypass the bottleneck step i.e. profilometry, as it is the main source of error.

o meet increasing global energy demand, it is critical yet challenging to explore new methods to accelerate the development of novel energy materials. In recent years high-throughput experimentation (HTE) and machine-learning techniques have become increasingly accessible to scientific researchers. We herein demonstrate a case study on the data-driven design of perovskite-inspired materials for photovoltaic applications, where we employed machine-learning techniques to guide the synthesis of new halide perovskites for photovoltaic applications. Halide perovskites (ABX₃, where A = Cs, methylammonium (MA), formamidinium (FA); B = Pb, Sn; and X = Cl, Br, I) have shown great promise as light absorbers. Solar cells based on perovskites have surprised the energy community as an emerging low-cost photovoltaic technology with a record power conversion efficiency (24.2%) now exceeding polycrystalline Si cells (22.3%).[1] In this study, we developed a high-throughput experimental platform for thin-film synthesis and characterization, and investigated 75 unique perovskite compositions interest for energy-harvesting applications in a two-month period. To achieve desired optoelectronic properties, we established a set of selection criteria for screening. A deep neural network is employed to classify compounds into 0D, 2D, and 3D perovskite structures via X-ray diffraction patterns analysis. [2] The combination of fast synthesis and machine-learning assisted data diagnostics achieves an acceleration of over an order of magnitude per experimental learning cycle over our laboratory baseline. Among the 41 Pb-free perovskite compositions studied, we identified the optimised doping level in a multi-site alloy series, Cs₃(Bi_1-xSb_x)₂(I_1-xBr_x)₉, where a desired structural (2D) and optical properties (< 2 eV). [3] Our work contributes to the prospect of automated materials discovery and we envision an accelerated development in functional materials in the next decade aiming to provide new energy solutions.

[1] NREL. National Renewable Energy Laboratory, Best Research Cell Efficiencies http://www.nrel.gov/ncpv/images/efficiency_chart.jpg. (accessed June 14, 2019).
[2] Oviedo, F.; Ren, Z.; Sun, S.; Settens, C.; Liu, Z.; Hartono, N. T. P.; Ramasamy, S.; DeCost, B. L.; Tian, S. I. P.; Romano, G.; et al. Fast and Interpretable Classification of Small X-Ray Diffraction Datasets Using Data Augmentation and Deep Neural Networks. npj Comput. Mater. 2019, 5 (1), 60.
[3] S. Sun, N. T. P. Hartono, Z. D. Ren, F. Oviedo, A. M. Buscemi, M. Layurova, D. X. Chen, T. Ogunfunmi, J. Thapa, S. Ramasamy, C. Settens, B. L. DeCost, A. G. Kusne, Z. Liu, S. I. P. Tian, I. M. Peters, J. P. Correa-Baena and T. Buonassisi, Joule, , DOI:10.1016/j.joule.2019.05.014.

Structure zone diagrams (SZD) are frequently used to estimate thin film microstructures based on a few chosen synthesis parameters. Despite their usefulness, the predictive power of classical SZD is very limited due to the complexity of the synthesis-microstructure relationship of thin films. Furthermore, the complicated interplay of many synthesis parameters and compositional complexity hinders a generalisation. Classical SZD have in common that they are based on a small number of observations. Underlying trends were extracted by the scientists‘ expertise and in a creative process abstracted into a diagram representation of microstructural features. Several refined SZD were proposed, which implemented more physical parameters. With emerging developments in combinatorial thin film synthesis and high-throughput characterization a fast, high-quality acquisition of microstructure data is now possible. This and progress in machine learning of images now provides tools to handle complex image data and improve SZD: We present a dataset containing > 100 samples of SEM surface images from Cr-Al-O-N material libraries, each featuring a different chemical composition and synthesis condition such as deposition temperature, ion energy and sputter frequency of high power impulse magnetron sputtering (HiPIMS). We train convolutional variational autoencoders (VAE) on this dataset of augmented SEM surface image data. Results show that VAEs can cluster microstructure data through latent space representations. The performance of different neural network architectures is discussed. The VAEs generative capabilities to predict SEM surface images from chemical composition and synthesis parameters are investigated. By sampling of the latent representation, we are able to generate SZDs for different variations and combinations of input parameters like temperature, ion energy and chemical composition. The qualitative trends which we observe demonstrate the prediction of microstructure by generative deep learning models.

Generative models, a recent paradigm in machine learning has revolutionized the industry by generating ‘natural looking’ data. While such models have found limited applications in the domain sciences, they display untapped potential in generating materials or molecular structures commensurate with target properties and desired functionalities. While the protein folding problem has been addressed previously by multilevel computational methods and various deep convolutional neural networks, unfortunately, the key step of encoding atomic structures for computational treatment is a challenge. Historical efforts have focused on pre-process featurization that relies upon traditional string representation without any structural information, expert-designed heuristics-based inputs, or on volumetric modeling that presumes a specific predetermined conformation without associated functionality. Here we demonstrate the first implementation of generative models, more precisely, generative adversarial networks with graph encoding of atomic connectivity within the biomolecules, for data-driven prediction of peptide and protein conformations associated with particular functionalities such as binding to atomically flat surfaces and biomineralization. In the graph input, atoms are considered as nodes and the bonds are considered edges, while angles in the molecule are encoded as a third order tensor between any three nodes. The generator tries to output secondary structures in terms of the bond edges and angle tensors while the discriminator network learns from existing sequences and their secondary structures from pdb files and MD simulations. We test the predicted results with MD simulations as well as circular dichroism experiments. Results show that the generative model developed herein is generalizable to any functionality and more accurate than existing methodologies for predicting functionality-associated peptide conformations for practical implementations in disease diagnostics, drug screening, biosensing and bioelectronic devices.As part of the Materials Genome Initiative, the research is supported by NSF-DMREF program through the grant DMR-1629071.

Sequential learning is a data-driven workflow for accelerating materials development. In this iterative workflow, machine learning models are used to explore a “design space,” the set of possible experiments that could be performed, to surface promising candidate materials. Experimental data for those candidate materials are used to retrain the models so that they can provide successively better-informed suggestions.

For sequential learning to be effective, a relevant design space of candidate materials must first be constructed. These design spaces often include complex constraints, as well as a mix of continuous and categorical variables. The machine learning model can be used to sift through the design space to surface the most promising candidates. For these top candidates, it is valuable to have insights into how the model made its predictions and why it predicts high performance. Interpretability analysis can increase confidence in the model predictions, uncover sample bias in the underlying training data, and provide information on the robustness of the predicted performance. This talk will discuss approaches for constructing relevant design spaces and for interpreting model predictions, and show how these approaches fit into the overall sequential learning workflow.

One of the holy grails of materials science, unlocking structure-property relationships, has largely been pursued via bottom-up investigations of how the arrangement of atoms and interatomic bonding in a material determine its macroscopic behavior. Here we consider a complementary approach, a top-down study of the organizational structure of networks of materials, based on the interaction between materials themselves. We demonstrate the utility of applying network theory to materials science in two applications: First, we unravel the complete “phase stability network of all inorganic materials” as a densely-connected complex network of 21,000 thermodynamically stable compounds (nodes) interlinked by 41 million tie-lines (edges) defining their two-phase equilibria, as computed by high-throughput density functional theory. We find that the node connectivity in the materials network has a lognormal distribution, and the connectivity decreases with the number of elemental constituents in a material. Analyzing the topology of this network of materials has the potential to uncover new knowledge inaccessible from traditional atoms-to-materials paradigms. Using the connectivity of nodes in this phase stability network, we derive a rational, data-driven metric for material reactivity, the “nobility index”, and quantitatively identify the noblest materials in nature. Second, we apply network theory to the problem of synthesizability of inorganic materials, a grand challenge for accelerating their discovery using computations. We combine the above phase stability network with timelines for the first experimental synthesis of each compound from literature citations. This allows us to create a time-dependent network, and from the time-evolution of the underlying network properties, we use machine-learning to predict the likelihood that hypothetical, computer generated materials will be amenable to successful experimental synthesis. ** In collaboration with S. Kirklin, L. Hung, S. Suram, P. Herring, and J. Hummelshoj

Atomically precise nanoclusters can be used for numerous applications due to the unique properties they possess. Despite their wide range of applications, the structures and properties of many small elemental clusters remain unknown. We present the “The Quantum Cluster Database,” an open-access source containing the structures and properties of tens of thousands of cluster structures of up to 55 atoms for 55 elements. The structures are compared against previous computational and experimental data where available. We discuss the methods that are being used to accelerate the construction of the database and describe how the database can be accessed for cost-effective, data-driven materials design.

New materials enable new technologies, so discovery of new materials is one of the most important directions in materials research. Oxides and some other materials chemistries, which have been extensively explored in the past, yielded many spectacular properties. Other chemistries, such as nitrides, have been barely touched: for every 14 documented oxides there is only 1 known nitride.
We will present on data-driven experimental discovery of new nitride materials, focusing on experimental synthesis and characterization, while also featuring computational predictions and machine learning. The data mining efforts followed by first-principles calculations and machine learning analysis indicated that there are 93 unexplored ternary metal nitride chemical spaces, with 244 new predicted stable ternary materials, and explained the stability trends among these and other nitrides [1].
Experimental synthesis using high-throughput combinatorial methods realized 7 of these compounds, including Zn-M-N (M= Sb, Mo, W) with wurtzite-derived crystal structures and Mg-TM-N (TM = Nb, Ti, Zr, Hf) with rocksalt-derived crystal structures. Physical property characterization results of the ternary rocksalts indicate that they are semiconductors with 1.8-2.1 eV optical absorption onsets and large dielectric constants [2]. The Zn-Sb-N wurtzite is the first ever reported Sb-containing nitride, with Sb in unusually high 5+ valence state, and measured room-temperature photoluminescence near 1.6-1.7 eV solar matched band gap [3]
Overall, these results both demonstrate the power of data-driven materials discovery, and suggest that many new previously unreported nitride materials remain to be synthesized.
[1] W. Sun et al, Nature Materials (2019), DOI: 10.1038/s41563-019-0396-2
[2] S. Bauers, et al Proc. Nat. Acad. Science (2019), DOI: 10.1073/pnas.1904926116.
[3] E. Arca et a;, Materials Horizons (2019), DOI: 10.1039/c9mh00369j

We have explored the design of nanophotonic structures, such as subwavelength-scale spectral filters, using an advanced active machine learning algorithm that efficiently explores multiple physical models with different approximation fidelities and costs. Our method, which is applicable to a variety of nanophotonics optimization problems, employs a novel strategy consisting of a mutual information based multi-fidelity Gaussian process optimization algorithm (MF-MI-Greedy). It consists of two components: an exploratory procedure to gather information about the target (i.e., the highest fidelity) function via querying lower fidelity functions, followed by an exploitative procedure to optimize the target level fidelity with the previously gathered information. Our results on several pre-collected nanophotonics datasets demonstrate the compelling performance of the multiple-fidelity Bayesian optimization approach. These experiments suggest that there is a significant potential in utilizing cheap, multi-fidelity simulations to aid the discovery of optimal photonic nanostructures.

Exhaustive experimental mapping of non-equilibrium processing phase diagrams demands a prohibitively huge allocation of resources for even a single realistic system with more than two compositional degrees of freedom, even with current state-of-the-art high-throughput techniques. Advanced data analysis methods are thus called for to accelerate such explorative efforts, focusing on multimode analysis of critical boundary points in the phase diagram where transitions are observed. Here, we present a hierarchical, prioritized data analysis structure to optimize usage of costly experimental resources. By combining data analysis methods based on optical characterization and x-ray diffraction with sophisticated active learning algorithms, we can efficiently map phase boundaries in composition-time-temperature processing phase diagrams. We demonstrate the utility of our approach by constructing processing phase diagrams for spike annealed multicomponent oxide materials

Hybrid organic-inorganic perovskites solar cells have recently increased scientific interest for their manufacturing simplicity and high performance, challenging the best thin-film photovoltaic devices. Perovskites solar cells have broad and strong light absorption along with excellent transport properties that partly explain their record power conversion efficiency above 24% [1, 2, 3]. Compositional engineering of perovskites is a time-consuming effort. Reaching high efficiency in compositions with complex and diverse dopants that usually requires significant trial and error and hundreds of measurements of full solar cells [4]. In addition to the abundance of applied research on this subject, there is great interest in understanding the fundamentals of transport and photophysical properties of various perovskites compositions [5]. High-throughput methods for screening compositions at the film level could be a potentially powerful alternative to investigate the complex perovskite composition space efficiently. Nevertheless, screening for high-efficiency perovskite compositions with high-throughput methods is not yet firmly established, in part due to the complexity of photophysical characterization experiments at the film level. In this work, we report for the first time a combination of high-throughput conventional steady-state photoconductivity method, (SS-PC) to determine diffusion lengths and Bayesian optimization methodology. This approach allows us to investigate the complex compositional space efficiently by just making films and not full solar cells. By using QSS-PC as a proxy for efficiency, we use Bayesian optimization to guide compositional changes and obtain the best solar cell efficiency for a given material, accelerating material screening by 10X.


[1] Silver-Hamill Turren-Cruz et al. Energy Environ. Sci., 2018, 11, 78, DOI: 10.1039/c7ee02901b
[2] Best Research-Cell Efficiencies (NREL, accessed 02 January 2019); https://www.nrel.gov/pv/assets/pdfs/pv-efficiency -chart.20181221.pdf
[3] Jiajun Peng et al. Chem. Soc. Rev., 2017, 46, 5714, DOI: 10.1039/c6cs00942e
[4] Ian L. Braly at al. J Phys. Chem Lett. 2018, 9, 3779-3792, DOI: 10.11021/acs.jpclett.8b11520
[5] Y. Chen et al. NATURE COMMUNICATIONS | 7:12253 | DOI: 10.1038/ ncomms12253 |

Machine learning is often used in materials design nowadays. They are called as Materials Informatics [1] and Material Genome [2]. Recent progress of the technology to predict physical properties of materials is remarkable, especially in metal material [3] and small organic molecule consisting of a single raw material [4].
By contrast, in the case of thermosetting resins, mixed raw materials are often used to synthesize. In such a case, it is hard to predict the material properties due to the existence of missing data derived by the difference in the number of raw materials in each resin. Thus, we suggest a method of machine learning using explanatory variables incorporating raw material information concerning chemical structure.
First, we predicted the classification of raw materials by the random forest, where the extended circular fingerprint (ECFP) [5] was used as the explanatory variable. Then, we aggregated ECFP for each classification predicted by the random forest. After that, we constructed the prediction model by using the aggregated ECFP, feature quantities of reaction intermediates, and curing conditions of resin as explanatory variables. As a result, the model was able to predict in high accuracy (R^2 = 0.8), for example, the elastic modulus of thermosetting resins. Furthermore, we show the result of verification of prediction accuracy in first step, such as using the one-hot-encording.
Therefore, we confirmed that the properties of thermosetting resins could be predicted using mixed raw materials by the proposed method.

[1] K. Rajan. Materials Today, 8, 38 (2005)
[2] A. Jain, et. al., APL Materials, 1, 011002 (2013)
[3] A. Agrawal, et. al., Integr. Mater. Innovation 3, 1-19 (2014)
[4] R. Gomez-Bombarelli. et. al., Nature Materials, 15, 1120-1127 (2016)
[5] D. Rogers, M. Hahn, J. Chem. Inf. Model. 50, 742 (2010)

We proposed a method that can accurately predict the vapor-liquid equilibrium (VLE) of a mixture through the deep learning (DL) based on thermodynamic theory. VLE data of binary hydrocarbons over 1,000 points were studied by various DL methods, followed by compared with the experimental data to predict the VLE outside the learned range. Whereas The simple DL method exhibited good correlative ability in the learned domain, the predictive ability outside the learned domain showed a definite limitation. On the other hand, the thermodynamic-based DL method proposed in this study has the excellent predictive ability even outside the learning range. When we compare our results with conventional thermodynamic model UNIFAC, similar or better prediction results have been obtained. From this study, it is believed that our proposed DL system would be applicable to the prediction of the various thermodynamic parameters for the design of new materials.

The discovery and design of novel structures for reactive membranes which purify or enrich contaminated air, either without or with limited use of toxic chemicals, still have a significant environmental impact. Airborne byproducts of manufacturing and fuel combustion such as particulate matter (PM_2.5; particulate diameter < 2.5 µm) have proven to be a global health risk and while current filtration membrane materials such as polyester and fiberglass benefit from tunable pore areas and fine fiber diameters, they are non-recyclable and must be replaced regularly due to fouling caused by the accumulation of pollutants. In collaboration with Ecovative Design, a bio-fabrication company working with mycelium (the root structure of mushroom), we studied the application of mycelium films as air-based filtration membranes. Surface proteins on mycelium hyphae are bio-adsorbants of several heavy metals and air contaminants and are therefore ideal candidates for membrane development. Like other naturally growing materials, mycelium has a heterogeneous structure, and in its optimization for membrane design with high filtration efficiency, a comprehensive quantification of its structure is necessary. In this study, through a combination of high-resolution and high-throughput imaging of the membrane, across several location and depths of the membrane, we quantitatively estimated several physical parameters such as pore area, fiber diameter, network topology, and fiber orientation of these heterogeneous membranes. Scanning Electron Microscope and atomic force microscope images were acquired to provide micron-level details of the mycelium network. These images were sampled across a range of magnifications, and image processing techniques such as statistical region mapping and axial thinning were employed for feature extraction. By obtaining a distribution of the fibers radii, Gaussian mixed models were used to identify three unique fibers indicating bifurcation as the main network growth mechanism. Additionally, unsupervised learning tools were employed to appropriately identify pores from the processed images, which showed a positively skewed data with an average pore area of 4 µm² and a mode of 0.5 µm² across the growth. These pore areas put mycelium in the magnitude for PM_2.5 filtration, verifying mycelium’s potential as an air filtration membrane. The results accelerate the development of mycelium-based biofiltration products by establishing a feedback loop with Ecovative Design to optimize their growth conditions and species selection to generate optimized microstructures for filtration.

Conventional methods to predict 3D structures of the molecule are based on iterative stochastic optimization techniques by moving each atom based on energy calculation using physics-based electronic structure modeling such as DFT or MD. Therefore, computing cost of physics-based modeling of 3D molecular structure is significantly increasing with the number of iterations to calculate energy until the total energy of structure is converged, and so with the number of atoms in the target molecule. Therefore, the conventional iterative stochastic optimization may not be optimal for the molecules with a large number of atoms but relatively simple structure. Specifically, for the molecules containing multiple same substructures, such as aromatic hydrocarbon family which has derivative structures of benzene rings, the number of atoms tends to be very large compared with the simple structural pattern of aromatic hydrocarbon family.
As the cost-efficient alternatives, we propose a novel reinforcement learning algorithm to optimize 3D structure of molecules in aromatic hydrocarbon family based on DDPG (Deep Deterministic Policy Gradient) method. There are three main contributions of our proposed approach: (1) To fasten the computation, we applied novel technique to reduce the degree of freedom to move atoms by grouping repetitive substructures in aromatic hydrocarbon family as a unique entity and move them as a unit, (2) We developed the general strategy to build action space applicable to any small aromatic hydrocarbon molecules, (3) We tested whether the knowledge obtained from several aromatic hydrocarbon molecules can be transferred to new aromatic hydrocarbon molecules. To demonstrate the efficiency of our model, we predicted the 3D structure of 13 aromatic hydrocarbon molecules and compared with results from the conventional DFT calculation. Our experiments show that our model succeed to predict 3D structure of our 10 target molecules with the faster convergence than DFT calculation.

Prediction of stoichiometric properties of small molecules is an important process in developing and designing new materials having the desired properties. Specifically, fast screening to identify a promising density of the molecules is significant to develop high performance explosive and high energy molecule. To calculate the density property of molecules, sophisticated quantum level electronic structure calculations, such as DFT or MD simulation, have been employed in optimizing 3D molecular structure and find the energy minimum. However, there are two problems in employing quantum level calculation. First, it requires a significant amount of computing resources which polynomially increases as the size of the molecule. Therefore, it is computationally expensive and not scalable. Second, there are many cases the results from calculation is different with experimental value. Therefore, finding an efficient computational method to accurately predict experiment density of large number of small molecules is of great value. Recent advances in deep learning lead to rise in numerous neural networks based architectures, and they have recently been applied to various tasks in material science domain, such as small molecular design, predicting physical attribute of molecules, etc.
Despite the impressive performances of previous works, there have been only a few work to predict structural properties of molecules which requires through featurization of 3D structure of the target molecule. Specifically, density prediction requires careful selection of input data and sophisticated featurization method to represent both 3D structure and physical property accordingly. In this work, we propose methods to predict experimental density value of small molecules employing various combinations of molecular featurization techniques and neural net based architectures. We used published experimental density values and 3D structure of molecules from cambridge structural database. Comparison study between different neural net architectures and different molecular fingerprint is presented and accuracy of density prediction is reported. Several preprocessing techniques including various fingerprints are also compared and discussed with the results. To the best of our knowledge, this is the first work to predict experimental density using deep learning.

Finding the “ideal” catalyst is a matter of great interest in the communities of chemists and material scientists, partly because of its wide spectrum of industrial applications. Information regarding a physical parameter termed “adsorption energy”, which dictates the degrees of adhesion of an adsorbate on a substrate is a primary requirement in se- lecting the catalyst for catalytic reactions. Both experiments and in-silico modeling are extensively being used in estimating the adsorption energies, both of which are Edisonian approach and demands plenty of resources and are time consuming. In this work, em- ploying a data-mining approach, we predict the adsorption energies of mono-atomic and di-atomic gases on the surfaces of many transition metals (TMs) in no times. With less than set of 10 simple atomic features, our predictions of the adsorption energies are within a root-mean-squared-error (RMSE) of 0.4 eV with the quantum many-body perturba- tion theory estimates, a computationally expensive with good experimental agreement. We minimized the RMSE further up to 0.11 eV by using the pre-computed adsorption
energies obtained with conventional exchange and correlation (XC) functional as one component of the feature vector. Based on our results, we developed a set of scaling relation between the adsorption energies computed with many-body perturbation theory and conventional DFT XC-functionals.

Understanding and developing maps of chemical synthesis using machine learning has the potential to revolutionize efficient synthesis. The design of a chemical synthesis often relies on a combination of chemical intuition and Edisonian trial-and-error methods which are not just inefficient but inherently limited in their ability to quantitatively predict synthetic outcomes, easily defeated by complex interplays between variables, and oftentimes based on suppositions that are limited in validity. The synthesis of nanomaterials has been especially prone to empiricism given the combination of complex chemical reactivity as well as mesoscopic nucleation and growth phenomena spanning multiple temporal and spatial dimensions. Here, utilizing the synthesis of two-dimensional CsPbBr3 nanoplatelets as a model system, we demonstrate an efficient machine learning navigation of reaction space that allows for predictive control of layer thickness down to sub-monolayer dimensions. Support vector machine (SVM) classification and regression models are used to initially separate regions of the design space that yield quantum-confined nanoplatelets from regions yielding bulk particles and subsequently to predict the thickness of quantum-confined CsPbBr3 nanoplatelets that can be accessed under specific reaction conditions. The SVM models are not only just predictive and efficient in sampling the available design space but also provide fundamental insight into the influence of molecular ligands in constraining the dimensions of nanocrystals. The results illustrate a quantitative approach for efficient navigation of reaction design space and pave the way to navigation of more elaborate landscapes beyond dimensional control spanning polymorphs, compositional variants, and surface chemistry.

Solid-state magnetic refrigeration employing the magnetocaloric effect (MCE) is a field of very active research. Compared with conventional gas compression-expansion refrigeration, magnetic refrigeration based on MCE offers improved energy efficiency, reduced environmental impact, and noise-free operation. In this study, we performed data-driven predictions of complex iron-based amorphous alloys laying in the 18-dimensional composition range: Fe_100–ΣNi_0–38Co_0–40Cu_0–1Zr_0–11B_0–29Cr_0–10Nb_0–10Gd_0–4Y_0–22Mo_0–10Si_0–10Nd_0–15Ce_0–13Sm_0–3Mn_0–24Ti_0–8V_0–14.
To design alloys that would achieve higher-than-ever reported performance in refrigeration at a desired operating temperature, we employed a combination of Machine Learning and Search Algorithms. The framework aims to maximize the magnitude of peak magnetic entropy change |-ΔS_M^peak| and tune Curie temperature T_C to value the granting of the highest performance. Based on the contents of the database we collected, our predictive tool is expected to be useful when designing alloys for applications requiring any T_C between 200K and 500K. This temperature range includes room temperature (298K), which is considered to be the most important, as it would allow the creation of MCE-based household refrigerators.
Since April, we have been experimentally validating our predictions, first by testing a state-of-the-art alloy reported in the literature and then testing novel alloys designed with our predictive framework based on that alloy. Results obtained so far show excellent agreement between our predictions and experimental results.
In our MRS presentation, we pay particular attention to the design of Search Algorithms. Based on our experience, this step is by far the most challenging due to the considerable dimensionality of composition and vast search space, which in our case included up to 10³⁴ possible alloys. It is also the most crucial step to the prediction of alloys that are reasonable and are feasible to be made in the lab. We not only report our final method of search but also elaborate on challenges we overcame along the way. This knowledge should be useful to any group working on the design of high-entropy amorphous metals.

Transport current is one of the critical parameters for practical applications of superconducting materials, such as transmission cables and strong magnets. On the other hand, early studies have shown that there is a large gap between the intrinsic physical properties and the macroscopic transport properties obtained as a material due to “weak-link” effects at grain-boundaries, vortex pinning performance and microstructural defects [1,2]. Establishing the prediction model of transport properties is an important issue for the utilization of automated and/or data-driven approaches to superconducting materials research and development. Connectivity is a useful parameter that represents the effective cross-sectional area of the transport current [3]. Matsushita et al. has proposed a mean-field model that quantitatively links connectivity and microstructural defects [4]. In this study, we studied transport property prediction method based on numerical modeling. The transport mechanism of polycrystalline materials composed of crystals with electrical anisotropy and containing voids as microstructural defects was investigated by means of a random resistance network based on a three-dimensional cubic site percolation model. The local current distribution and the macroscopic connectivity of the system were obtained by numerical simulation using the finite element method. In addition, the simulated result was compared with the measured electrical resistivity of the synthesized samples.
[1] T. Katase et al., Nature Communications 2, 409 (2011).
[2] J. H. Durrell et al., Reports on Progress in Physics 74, 124511 (2011).
[3] J. M. Rowell, Superconductor Science & Technology 16, R17 (2003).
[4] A. Yamamoto, T. Matsushita et al., Superconductor Science & Technology 20, 658 (2007); 21, 015008 (2008).

Raman spectroscopy is used ubiquitously in the characterization of condensed materials, spanning from minerals, polymers, to biomaterials, as it provides a unique fingerprint of local bonding and environment. In this work, we demonstrate a robust and automatic computational workflow for Raman spectra that employs first-principle calculations based on density functional perturbation theory. Sets of calculated Raman spectra peaks was compared to those obtained from established experimental databases to estimate the accuracy of the calculated properties across chemical systems and structures. Our work, consisting 55 mineral compounds and 205 pairs of matched peaks, shows that the mean wavenumber deviation is -9.66 cm^-1 with a standard deviation of 18.58 cm^-1, indicating our calculations tend to slightly underestimate the Raman peak locations. With a mean absolute relative error (MARE) of -2 % for wavenumbers and the low bias exhibited by the intensities, we conclude that our computational data set is in reasonably good agreement with the experimental data and have explored potential reasons for the deviations.

With integrated circuit manufacturers aiming to produce sub-10nm feature sizes, extreme ultraviolet lithography (EUV) is the next developing technology for the job. The challenges with using polymer-based photoresists for EUV lithography can be eliminated by using oxohydroxo metal nanoparticle photoresists instead. Oxohydroxo metal clusters have the potential to provide faster writing speeds, higher resolutions, and better etch resistance than chemically amplified polymer resists. In a joint computational-experimental effort we examine the stability of organotin Keggin clusters for use as sensitive high-resolution photolithographic resists. A one-step synthesis obtained the β-isomer (β-NaSn12), the γ-isomer (γ-NaSn12), and a γ-isomer capped with a butyltin (γ-NaSn13). The Sn Keggin ions crystallized readily without counterions which increased the simplicity of the synthesis as well as improved the yield, purity, and reproducibility. Solution characterization (SAXS, NMR, ESI-MS) verified that solutions contained only the Na-centered dodecamers. Computational modeling was used to determine the ground state electronic structure of the three butyltin Keggin structures, as well as the capped β-isomer (β-NaSn13), and the hypothetical α-isomers (α-NaSn12 and α-NaSn13). The computed hydrolysis Gibbs free energy and HOMO-LUMO gap provided the stability ranking: β-NaSn12 > γ-NaSn12 > α-NaSn12 which was consistent which experimental observations. The uncapped isomers were computationally evaluated to be more stable than their respective capped analogues. In-depth structure-energy analysis reveals a balance between corner-linking to minimize cation-cation repulsion, and edge-lining to maximize stability via bond formation. Therefore, this sodium centered tin Keggin ion represents the only Keggin ion family so far, that favors the isomers of lower symmetry. Finally, the system’s neutral charge makes it a valuable model system for understanding the fundamental patterning mechanisms at play.

Recently, the breakthrough developments of artificial neural network (ANN) accelerate the in the field of materials science. In the last few years, J. Behler and M. Parrinello introduced a method, which is a new kind of neural-network representation of density functional theory potential-energy surfaces, called "symmetry function (SF)". Here, we calculate the total energies and atomic forces of variety Si clusters, which is fundamental and significant material in modern semiconductor industry, by using ANN. More than 11,000 training set data are generated by carrying out trajectories of ab initio molecular dynamic simulations with canonical ensembles at 300 and 700 K, and the static calculations of randomly positioned coordination of Si_n clusters. To describe chemical atomic environment properly, we carefully determined number of radial and angular SFs with different parameters. The hyper-parameters in ANN have been tuned by random search and grid search method. To evaluate accuracy of ANN models, we addressed root mean square errors between value of the DFT calculations and ANN calculations through decent-gradient algorithm. The calculated error of total energy for training and test data-sets are less than 0.08, 0.18 eV/atom, and error of atomic force for training and test-sets are less than 0.10 and 0.74 eV/A, respectively. For applications of machine learning based potential, we compared the mechanical and crystallographic properties with density functional theory calculations.

A major challenge towards efficient materials discovery is the reliance on human intuition in the experimental design. Currently, the navigation in the search space for targeted material properties using high-throughput experiments and computations often necessitates large resources. Here, we investigate how machine learning accelerates the search and discovery of new materials using sparse datasets. In particular, model-based search explores iteratively key parameters to guide the experiments and computations. The approach utilizes uncertainties and makes predictions from a surrogate model together with a utility function that prioritizes the decision making process on unexplored data in multidimensional space. We demonstrate this approach on the fabrication of laser-reduced graphene circuits on flexible substrates. The surrogate model is fed with surface characterizations with a view to automate the manufacturing process in a closed-loop. We demonstrate effective explorations and rapid optimizations even with sparse data and discuss the integration of machine learning in advanced manufacturing.

Computational resources for high-throughput data generation offer incredible potential for accelerating scientific discovery, especially if used in conjunction with well-managed computational workflows. The open-source signac data management framework enables researchers to maintain well-formed, sharable, and reusable data spaces from early exploration through production runs on supercomputers. This is achieved through a transparent, application-agnostic data and workflow model, as well as a simple and unobtrusive programmatic interface. The framework has been applied to molecular simulations, quantum chemistry, photonics, computational fluid dynamics, machine learning, graph mining, and even the management of experimental data. Implemented in Python, signac interoperates organically with other well-established libraries within the scientific software ecosystem such as NumPy and pandas, and it integrates naturally into Jupyter notebooks for data presentation and exploration. The framework readily supports the use of HDF5 files for the storage of large hierarchical numerical data. With special emphasis on enabling rapid prototyping, workflows implemented with signac are easily executed both on local resources and powerful supercomputing clusters like OLCF’s Summit. We will present representative examples of scientific applications that demonstrate the efficacy and versatility of signac across a range of materials research domains.

Autonomous Research Systems like ARES™ are disrupting the research process by using AI and Machine Learning to drive closed-loop iterative research. ARES™ is our autonomous research robot capable of designing, executing and evaluating its own experiments at a rate of up to 100 iterations per day. Previously ARES taught itself to grow carbon nanotubes at controlled rates (NPJ Comp Mat 2016). Here we discuss recent research campaigns on maximizing carbon nanotube growth rates using a Bayesian optimization planner. We also use HOLMES and knowledge gradient descent to introduce advanced decision policies with local parametric models to control nanotube diameter. Implications for nanotube materials development will be discussed. Finally, we have developed a new research robot for additive manufacturing, AM ARES™, which is at the early stages of teaching itself to print structures with unknown inks. We plan to make the AM ARES™ Robot available online so that a broad community of researchers can test concepts and approaches for AI/ML and experimental design as applied to 3D printing, thus building to the larger goal of enhancing citizen science.

Accurate cure kinetic model of polymeric sealants is essential to determine processing conditions of industrial processes, e.g., an automotive painting process involving a body sealant because their mechanical behaviour is highly dependent on their degree of cure during processes. Sourour-Kamal model with seven kinetic parameters is the most widely used cure kinetic model, but it cannot describe the degree of cure in both isothermal and dynamic process conditions. In this study, additional parameters related with the temperature rate were introduced and its validity was proved by carrying out a series of cure experiments of an epoxy-base automotive sealant using a differential scanning calorimetry, finally demonstrating that the degree of cure can be calculated under arbitrary time and temperature profile. Later, the thermal expansion, the cure shrinkage, and the rheological and viscoelastic properties of the sealant were characterized with a function of the degree of cure using a thermomechanical analysis, rheometry and dynamic mechanical analysis, respectively. The developed cure kinetic model and the four properties were incorporated into ABAQUS through a user material subroutine. Finally, the mechanical behavior of the sealant under various time-varying temperatures in automotive heating and coating processes was simulated to determine whether the sealant fails or not during and after the process.

Manipulating physicochemical and structural properties of materials by customizing their microstructures has been the pivotal element in materials science. However, where purely artificial efforts, e.g., heat-and-beat, usually involve complicated materials processing, nature’s design process, evolution, achieves similar functionality for materials at ambient conditions. While enantioselective directed evolution of enzymes has become recently popular in chemistry, the potential for directed-evolution in manipulating materials’ microstructure has been largely untapped. Here we demonstrate how millions of peptide sequences collected from combinatorial libraries and next-generation-sequencing paradigms, can not only be used to de-novo predict materials science relevant peptide sequences (for example, biomineralization), but also be used to scan naturally occurring pan-functional proteins for deriving peptides with targeted functionalities (e.g., target binding). As examples to PDpep approach, we derive hydroxyapatite biomineralizing peptides from Amelogenin and use these Amelogenin-Derived-Peptides for developing formulations towards a utility in repairing defective dental tissues; ice-inhibiting peptides, AFpep (antifreeze peptides), from antifreeze proteins for controlling ice nucleation and growth; and magneto-receptive peptides, MRpep, from cryptochrome-4 for biosensing applications, we demonstrate here that the methodology can be used to mineralize tailored microstructures with predictable materials phases as well as interface these structures with biology and materials science at the molecular level towards genetic design of functional materials in practical technological systems. As part of the Materials Genome Initiative, the research is supported by NSF-DMREF program through the grant DMR-1629071.

The growing research efforts have been put into developing solid-state electrolytes (SSEs) that can alleviate many of the issues of Li-ion batteries arising from the utilization of the organic liquid electrolytes. Up to date, several SSEs such as Li₁₀GeP₂S₁₂, Li₇La₃Zr₂O₁₂, and Li_1.3Al_0.3Ti_1.7(PO₄)₃ have been proposed as candidates in previous experiments. However, none of those SSEs are fully satifactory due to the inssuficient ionic conductivity or chemical and mechanical stability or high electronic conductivity. Nevertheless, the capability of numerous materials for SSE has been unexplored, and thus we might be able to discover new interesting candidates by a systematic investigation of relevant properties of thousands of materials. In particular, owing to the high accuracy of first-princpiles calculations as well as recent advances in computing power, computational screening approaches have been applied to searching publicly available materials databases for a new type of SSEs.[1,2]
In this work, we take one step further toward discovering potential SSEs; in addition to simply exploring a pre-existing materials database, we try to design new materials with aliovalent substitution of cations. This aliovalent substituion is known to facilitate kinetics of Li diffusion.[3,4] We first screen potential materials, which maybe be useful in themselves or become mother materials for generating new ones by aliovalent substituion, considering fundamental properties such as the presence of transition metals, thermodynamic stability, and band gap. Afterward, we crudely examine the potential energy surface (PES) around the Li ion at interstitial sites. The materials with the most smooth PES are then chosen for the further studies. Starting from 42,337 structures from the Materials Project databae, we find and design a number of new[U1] Li ionic conductors, some of them are predicted to exceed the Li ionic conductivity of the state-of-the-art Li₁₀GeP₂S₁₂.

[1] Chem. Mater., 29, 281 (2017)
[2] Energy Environ. Sci., 10, 306 (2017)
[3] Chem. Rev., 116, 140 (2016)
[4] Nat. Comm., 8, 1 (2017)

First-principles crystal structure prediction (CSP) is a well established technique which is routinely used to predict crystal structure in a diverse range of systems such as periodic solids, interfaces, encapsulated nanowires etc. However, one downside of CSP is that the number of minima in the potential energy surface scales exponentially with system size. As such, it rapidly becomes computationally unfeasible to search for more complex structures with larger unit cells. Here we discuss how ligand symmetries can be exploited to accelerate the prediction of Metal Organic Frameworks (MOFs).

Symmetry is often imposed during CSP to reduce the dimension of the search space and expedite energy evaluations. When atoms are used as the constituent units of the structure it is simple to apply symmetry constraints. However, if, as with MOFs, we start with extended units then imposing symmetry constraints is not trivial. Our method, Wyckoff Allowed Molecules (WAM), makes use of molecule symmetries to generate trial structures with molecules on special Wyckoff sites, as well as on the general site.

To test our approach we searched for a variety of Zinc based MOFs [1] using the ab initio random structure searching method with WAM and the CASTEP plane wave DFT code. These results will be presented and used to highlight strengths and weaknesses of the approach.

1. Darby, James P.; Arhangelskis, Mihails; Katsenis, Athanassios D.; Marrett, Joseph; Friscic, Tomislav; Morris, Andrew J. (2019): Ab Initio Prediction of Metal-Organic Framework Structures. ChemRxiv. Preprint
doi.org/10.26434/chemrxiv.8204159.v2

We discuss a strategy and present a simple tool to facilitate scientific data reproducibility by making available, in a distributed manner, all data and procedures presented in scientific papers, together with metadata to render them searchable and discoverable [1]. We also discuss accessibility to data presented in scientific papers and in general material simulation data generated by diverse groups and communities.

[1] Qresp, A Tool for Curating, Discovering, and Exploring Reproducible Scientific Papers, Marco Govoni, Milson Munakami, Aditya Tanikanti, Jonathan H. Skone, Hakizumwami B. Runesha, Federico Giberti, Juan de Pablo, Giulia Galli
Scientific Data, 6, 190002 (2019).

Two-dimensional (2D) ferroelectric (FE) materials displaying spontaneous polarizations are promising candidates for miniaturized electronic and memory devices. However, stable FE orderings are only found in a small number of 2D materials by experiment so far. In the current work, based on high-throughput screening of a 2D van der Waals layered materials database and first-principles calculations, we demonstrate niobium oxide dihalides NbOX₂ (X = Cl, Br and I), a group of experimentally synthesized yet underexplored van der Waals layered compounds, as a new family of 2D materials that simultaneously exhibit intrinsic in-plane ferroelectricity and antiferroelectricity. Similar to FE perovskite oxides, polar displacement of Nb cations relative to the center of the anion octahedral cage can lead to experimentally measurable FE polarizations up to 27 μC cm⁻² in layered NbOX₂. The presence of low-lying antiferroelectric (AFE) phases can effectively reduce the energy barrier associated with polarization switching, suggesting switchable ferroelectricity is experimentally achievable. In addition, the mechanism driving FE phase transitions in NbOX₂ monolayers around Curie temperature T_C is clearly revealed by our finite-temperature simulations. NbOCl₂ monolayer is predicted to be a stable ferroelectric with T_C above room temperature. Moreover, application of NbOBr₂ and NbOI₂ monolayers as 2D dielectric capacitors is further developed, where electrostatic energy storage of nearly 100% efficiency can be achieved in the 2D single-layer regime.

The discovery of earth abundant, functional materials is critical for sustainable technological advancement. There is a concerted global effort to reduce the time it takes to realize such materials via databases, high-throughput screening and informatics,^[1] but forming a four-component compound from the first 103 elements results in excess of 10¹² potential compositions.^[2] Such a search space is intractable to high-throughput experiment or first principles calculations.

In this work, we present a low-cost, virtual high-throughput materials design workflow and use it to identify earth-abundant materials for solar energy applications from the quaternary oxide chemical space. A statistical model that predicts bandgap from chemical composition is built using supervised machine learning and the trained model forms the first in a hierarchy of screening steps. Further data-driven algorithms are used to assign crystal structures and to discard unlikely chemistries based on oxidation state information.^[3]

We demonstrate the utility of this process by screening over 1 million oxide compositions generated using the open-source SMACT package.^[4] We find that, despite the difficulties inherent to identifying stable multi-component oxides, several compounds produced by our workflow are calculated to be thermodynamically stable or metastable and have desirable optoelectronic properties according to first-principles calculations. The predicted phases are Li₂MnSiO₅, MnAg(SeO₃)₂ and two polymorphs of MnCdGe₂O₆, all four of which are found to have direct electronic bandgaps in the visible range of the solar spectrum.

[1] K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, A. Walsh, Machine learning for molecular and materials science, Nature, 2018; DOI: 10.1038/s41586-018-0337-2

[2] D. W. Davies et al., Computational screening of all stoichiometric inorganic materials, Chem, 2016; DOI: 10.1016/j.chempr.2016.09.010

[3] D. W. Davies et al., Materials discovery by chemical analogy: role of oxidation states in structure prediction, Faraday Discussions, 2018; DOI: 10.1039/c8fd00032h

[4] D. W. Davies et al., SMACT: Semiconducting materials by analogy and chemical theory, JOSS, 2019; DOI: 10.21105/joss.01361

One of the conventional approaches of materials informatics (MI) is on the basis of supervised learning, which expects a property of unknown materials from known materials. In other words, the learning process is appropriate for accurate and broad prediction within a search space where a supervised dataset covers. In this study, to discover truly new materials by a MI approach, we present a search system to explore outside of what we have known so far.
Recently, Yang and Tsuda group proposed an automatic-design framework for organic materials, by using Monte-Carlo tree search (MCTS) to select a prospective molecular fragment. This search process is driven by iterative evaluations in optimization of a target property. However, long-simulation time to evaluate the property is a major bottleneck in MCTS. In the case of engine oils, for example, a molecular dynamics (MD) method using Green-Kubo formalism is a proper evaluator for viscosity property. This MD-based evaluation of viscosity requires a large number of MD time steps, which is the bottleneck of the search system.
We present an ultra- fast MD evaluation method of viscosity, which realizes the automatic oil-molecular design. This acceleration is achieved by an idea from a theoretical model of glass transition, in which viscosity can be represented by Arrhenius-type equation. This model states that viscosity can be estimated by a small number of MD steps compared to a direct evaluation methodology. We combine this fast MD evaluation with the MCTS, and examine this search system, by setting viscosity index (VI) as a target property. The VI is an indicator of viscosity susceptible to temperature; typically, high VI oil is regarded to high quality in machinery equipment. In this closed-loop search, the total number of the evaluated molecules is 55,000, which is done in the “K” super computer. We select one prospective molecule as a motif, and make it modified so as to synthesis. This MI-designed oil molecule is experimentally confirmed that it indicates high VI essentially comparable to high-VI base oils in the commercial market.

Metal oxide semiconductors, which support equilibrium populations of electron and hole charge carriers, have widespread applications including batteries, solar cells, and display technologies. It is often difficult to predict in advance if a redox process will result in localized or delocalized charge carriers; the former is associated with polaron transport and the latter with band transport. Determining polaron ground states directlty within the density functional theory (DFT) framework is a formidable task and invariably requires careful parameter control of multiple expensive calculation runs^[1].

Thanks to the existence of extensive databases covering calculated properties of known and hypothetical systems^[2], we are able to utilize data from routine DFT calculations to predict the energetic driving force for carrier localization. By considering the competition between the loss in kinetic energy (from wavefunction localization) and gain in potential energy (from dielectric polarization), the net polaron binding energy can be determined in line with the early work of Pekar^[2] and Frölich^[3]. We demonstrate how this data can parameterise a simple screening metric for predicting the nature of charge carriers, circumventing the need for explicit polaron modelling in high-throughput studies. Our results are consistent with observations regarding carrier dynamics and lifetimes. We apply our screening metric to identifying p-type metal oxides (targeting low hole polaron binding energy) yielding a number of promising candidates including LiAg₃O₂ and Ca₄Bi₂O, in addition to recovering known p-type oxides such as the Delafossite structure CuRhO₂.

This study highlights the utility of modern materials datasets to facilitate rapid screening of hard-to-access properties. By comparison with new data generated from hybrid DFT calculations, we explore current limitations in terms of data quality and availability of necessary properties.

[1] M. Reticcioli et al., Small polarons in transition metal oxides, in: Handbook of materials modelling, Springer Nature, 2019; DOI: 10.1007/978-94-009-5107-5_34
[2] K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, A. Walsh, Machine learning for molecular and materials science, Nature, 2018; DOI: 10.1038/s41586-018-0337-2
[3] S. I. Pekar, Local quantum states of electrons in an ideal ion crystal J. Exp. Theor. Phys., 1946
[4] H. Fröhlich, Electrons in lattice fields, Adv. Phys, 1954

In the recent decade, materials science has seen a huge growth in available data from combinatorial experiments as well as high-throughput first principles calculations. With this data explosion, we now stand at the cusp where machine learning techniques can make meaningful predictions of many properties of materials almost instantaneously. In this talk, I will discuss the potentially transformative impact that this “instant” materials property prediction can have on materials research, from providing new chemistry insights that will greatly improve our ability to “guess” new materials with superior properties to accessing large length / time scales at near DFT accuracy. I will highlight some of the most promising machine learning approaches thus far, focusing, in particular, on techniques to address fundamental data size and diversity limitations in materials science. Finally, I will outline some of the key obstacles that still remain to ML-enabled materials science.

Discovery of a new material with desired properties is the ultimate goal of materials research. To date, a generally successful strategy has been to use chemical intuition and empirical rules to design new materials, but these conventional approaches require a significant amount of time and cost due to almost unlimited combinatorial possibilities of materials to explore in chemical space. A promising way to significantly accelerate the latter process is to incorporate all available knowledge and data to plan the synthesis of the next material. In this talk, I will present a few initial frameworks we have developed along this line to perform machine-learned density functional calculations, to predict the properties of a material using simple representations, and to generate new materials for a given property using materials deep generative model.

Post-processing tools for PAW wavefunctions are important for the automation of DFT studies. For example, calculating the overlap operator between wavefunctions from different structures is critical for electron-phonon coupling calculations as well as new formation energy correction methods being developed for point defect calculations. PAW augments plane-wave pseudo wavefunctions with atom-centered orbitals on a radial grid to minimize the size of the plane-wave basis set needed for accurate calculations. This also makes calculating these overlap operators nontrivial. To overcome this difficulty and enable arbitrary overlap calculations, we developed a general formalism for doing so and implemented it in an open-source code called Pawpyseed. Pawpyseed provides a powerful, parallelized C backend for handling large systems and a flexible Python front-end that takes advantage of cutting-edge research tools like the Python Materials Genomics (pymatgen) package. This performance-usability balance is ideal for automated post-processing of DFT calculations. We demonstrate one potential use of pawpyseed by implementing a perturbation theory-inspired formation energy correction for point defect calculations that improves upon previously proposed band edge shifting techniques.

From metal contacts to batteries, interfaces dominate the physics of devices. Whilst modelling of the individual components of such systems is well understood, the modelling of interfaces is limited due to the structure of the interface being relatively unknown. Prediction of these interfaces is driven by human intuition and computational ease, which is fraught with problems. We have developed a structural prediction approach for exploring interface structures with the goal of eliminating human bias. Here, using two test cases; 1) metal contacts of 2D systems, and 2) oxide-graphite batteries, we show how we can produce interface structures for these complicated systems in order to find the lowest energy interface.

Here, we present our software package, ARTEMIS (Ab initio Restructuring Tool Enabling the Modelling of Interface Structures), developed to ease the generation and modelling of interfaces in order to explore the energy space of interfaces and find the lowest energy interface. We model various metal contacts on a set of transition metal dichalcogenides, investigating different termination planes and surface stoichiometries. The software also enables the exploration of intermixing and plane shifting in order to find the lowest energy interface between any two materials. Hence, by using ARTEMIS, we show how a variety of metal contacts for 2D systems could be explored, as well as understanding electronic properties of graphitic/oxide mixtures for applications such as batteries.

Demand for functional materials targeted for specific applications is ever increasing as societal needs and demands mount with advancing technology. One class of inorganic-organic hybrid materials, metal-organic frameworks (MOFs), has burgeoned in recent years due, in part, to effective design strategies (i.e. reticular chemistry) for their synthesis and their inherent [and readily interchangeable] hybrid, highly functional character. Metal-organic materials, specifically metal-organic frameworks (MOFs), have emerged as a unique class of materials amenable to design and manipulation for desired function and application. Several design strategies have been utilized and developed to target viable MOF platforms, from the single-metal-ion molecular building block (MBB) approach to the hierarchical supermolecular building block and supermolecular building layer approaches (SBB and SBL, respectively) to our newly introduced merged net approach allowing for the constrcution of intricate MOFs with multiple ligands, . This inherent built-in information allows access to highly stabile and made-to-order porous materials toward applications pertaining to energy and environmental sustainability.

Designing energy conversion materials with simultaneous satisfying activity and stability properties is a challenge especially for electrochemical reaction such as the water splitting reaction. Often, high electrocatalytic activity of a catalyst material is accompanied by structural instability. In this presentation, I will discuss a core-shell oxide catalyst material design, where an active core catalyst is covered with a stable shell, as an attractive solution to the problem of catalyst instability. So far, the influence of the core during electrocatalytic reactions on the surface of the shell of oxide systems is not well understood. In this contribution, we demonstrate how ultrathin heterostructures composed of unstable active and stable inactive layers can be used for studies of model core–shell oxide architectures. In particular, we show that as little as a one unit cell subsurface layer of an active perovskite SrRuO₃‘‘core’’ can activate a SrTiO₃‘‘shell’’ surface layer towards oxygen evolution reaction (OER). While, as little as a two unit cell shell layer is enough to completely protect the inherently unstable catalyst against corrosion during OER. Using density functional theory calculations and a simplified electronic structure model, we unravel the mechanism for the subsurface activation for a class of oxide heterostructures and rationalize the choice of the SrTiO₃-SrRuO₃system, which is realized experimentally.¹

1. A. Akbashev, L. Zhang, J.T. Mefford, J. Park, B. Butz, H. Luftman, W. Chueh, and A. Vojvodic, "Activation of SrTiO₃with Subsurface SrRuO₃for Oxygen Evolution Reaction", Energy & Environmental Science,11, 1762-1769 (2018).

In materials development, a structural analysis of materials is indispensable to understand the structure property relationships. Especially, with increasing demand for nanoscale devices, in which peculiar atomic arrangements influence the material properties more than those in bulk, the importance of the local structure analysis is rapidly increasing. Among a variety of analytical techniques, core-loss spectroscopy, that is, electron energy loss near edge structure (ELNES) and X-ray absorption near edge structure (XANES), is strongly authentic with nano- or sub nano-scaled spatial resolution and nano- or femto-scaled time resolution, enabling to analyze the lattice defects and chemical reaction. Due to the their superior spatial- and time-resolutions, huge number of spectra, several thousands and million, can be observed in one experiment.
On the other hand, interpreting the core-loss spectra is not straightforward. Although theoretical simulations of ELNES/XANES have been performed to interpret the spectra, calculating just single spectrum requires huge computational time, such as tens hours and days. The one-by-one simulation is quite difficult for the large spectrum data-set.
In this situation, we need an alternative approach to analyze a thousand of spectra. Here, we developed new data-driven approaches to predict the core-loss spectra and quantify the material properties by the machine learning [1-3].
First, we have used the machine learning to predict the core-loss spectra. A spectral database of O-K edges of silicon dioxides, including 1,171 spectra, was constructed by the theoretical calculation. Then, we constructed a neural network model, whose input layer is the partial density of states (PDOS) at the ground state and output layer is the core-loss spectra. After training, the prediction model can produce the core-loss spectra only by the PDOS. Different from the core-loss spectrum, the PDOS at the ground state can be obtained within a few minutes.
Furthermore, the prediction model constructed by the "crystalline" silicon dioxides was applied to predict the O-K edge spectra of "amorphous" silicon dioxides. Based on this investigation, we revealed the critical difference between the crystalline and amorphous silicon dioxides [2].
In addition to the spectrum prediction, we used the machine learning to find the connections between the ELNES/XANES and the materials structures and properties. The same database as above was used and considered mean bond length, bond angle and Voronoi volume as geometrical features, and bond overlap population, Mulliken charge and excitation energy as chemical bonding features. As the results, we achieved to predict the geometrical properties and the chemical bonding properties correctly. Furthermore, our prediction model constructed by the simulation was applied to an experimental spectrum, and we found that our model could predict the properties correctly from the experimental spectrum even though it had large noise.
The details of the two studies will be discussed in my presentation.

References
[1] S. Kiyohara et al., Data-driven approach for the prediction and interpretation of core-electron loss spectroscopy. Sci. Rep. 8, 13548 (2018).
[2] S. Kiyohara et al., in preparation
[3] S. Kiyohara et al., Quantitative estimation of properties from core-loss spectrum via neural network. J. Phys. Mater. 2, 024003 (2019).

The use of machine learning (ML) and automation has recently emerged to accelerate material development and optimization. This automated closed-loop cycle of material development requires accurate and rapid characterization to provide feedback for the optimization of subsequent syntheses. In this study, we propose an ML-based method to rapidly extract optical constants and thickness of thin films from transmittance and reflectance (TR) spectra. The TR spectra of a thin film capture convoluted information of its optical constants (i.e., refractive index n and extinction coefficient k) and thickness t. Today, a trained expert fits measured TR spectra with a material-specific parametric model and optical model to indirectly extract optical constants and thickness, and validates with low-throughput measurements (e.g., profilometry or scanning electron microscopy). Hence, the extraction of n, k, and t from TR spectra becomes a bottleneck for a high-throughput experiment loop.

This study demonstrates semi-automatic extraction of n, k, and t from TR spectra. First, we generate our training dataset using optics simulations. Values for n and k are simulated by combining several optical oscillator models (e.g., Tauc-Lorentz model); then, transfer matrix formulation of Fresnel equations are used to simulate TR spectra. The simulated data are augmented to reflect the noise and imperfection of experimental measurements, following the strategy in [1]. Second, we use a variational autoencoder (VAE) to reduce the dimension of the dataset, and then we feed the dimensionally-reduced latent variables into a deep neural network (NN) to estimate the optical parameters and thickness of the film. We found a 95% confidence interval is within ±30 nm for the whole range of 100-nm- and 2000-nm-thick films. Third, after training of the machine learning model is complete, we apply the pre-trained ML model to simulated and experimental TR spectra of organic and inorganic metal-oxides. The NN-predicted parameters are further used as the inputs to a forward-fitting Bayesian inference model to accurately extract n, k, and t. Lastly, we discuss how this approach might generalize to other material systems and characterization methods.

[1] https://www.nature.com/articles/s41524-019-0196-x

Recently, machine learning (ML) has emerged as a useful tool for photovoltaics (PV) system design and performance evaluation [1]. However, many PV materials have been demonstrated, with different sensitivities to the solar spectrum and operating temperatures. In this contribution, we seek to demonstrate how ML can accelerate the adoption of PV technology, by customizing it to the local environment, leveraging a technology-dependent energy-yield model that uses open-source satellite information [2]. First, we apply ML to classify PV operating conditions (spectra + temperature) globally, using an unsupervised clustering approach with embedded physics domain knowledge. Then, on the basis of this classification, we envision a way to more accurately determine power ratings and predict energy yield for PV modules made from different materials in each climate zone, identifying opportunities for R&D innovation with strong future market pull.

[1] https://www.cell.com/joule/pdf/S2542-4351(18)30570-1.pdf
[2] https://www.cell.com/joule/pdf/S2542-4351(18)30098-9.pdf

Predicting the performance of functional materials systems such as rechargeable batteries in real-time is of great importance to both scientific research and industry. Due to the large capacity and power needed for batteries in electric vehicles and the relatively closed environment of cars, the safety of Li-ion batteries has become the paramount concern, since failure due to short circuit or leakage of chemicals can easily catch fire. Also, a typical battery test in battery research and development can be quite time-consuming. For example, up to thousands of charge/discharge cycles are needed for a cycling performance test. Thus, there is strong interest across both the academic and industrial community in predicting the cycling performance, life and critical failure event based only on the first cycle, the initial cycles, or the few cycles before battery failure, which holds great potential of accelerating the research and development of battery materials and improving the safety control procedures of electric vehicles. Previously, sophisticated apparatus is required in order to measure indicator properties of performance, while machine learning approaches based on feature engineering procedures require a priori expertise, which impedes the application in the complicated real-world environment. Here we show a novel end-to-end machine learning approach, free of feature engineering, toward effective real-time prediction of the battery life and failure, using just the raw images of the charge-discharge voltage profiles. Our method can make unsupervised real-time automatic extraction of latent physical factors that control the battery performance beyond human expertise. We demonstrated the real-time classification of cyclability using just the first cycle, and the real-time prediction of battery malfunction up to 20 cycles prior to the catastrophic failure. Our results show significant effectiveness even for just a proof-of-concept level dataset, which paves the way toward the real-world application.

Increasing computational sophistication and resources can enable a larger and more integrated role of theory in the discovery and understanding of new materials. Most materials studies start in a data-poor regime where the material of interest is unrelated to previous to studies (new structure, composition etc) or the computational methods are incompatible with previous studies (different exchange-correlation functionals, methods, etc). Efficient methods to quickly define, schedule, and organize necessary simulations are thus important and enable the application of online design of experiments approaches. I will discuss on-going work and software development to enable data science methods in catalysis including open datasets for the community. I will describe applications of our approach to ordered bimetallic alloy catalysts, with applications to several electrochemical catalyst discovery efforts including CO2 reduction, oxygen reduction, and water splitting chemistry. I will also discuss the methods and approaches we use to share data among group members and educate new students with the necessary skills to pursue these research directions (including statistics, machine learning, computer science, etc). Finally, I will discuss the transition from data-poor to data-rich regimes and our experiences when data-intensive deep-learning methods become more appropriate than simpler models based on chemical intuition.

The long term electric insulating behavior of polymers critically depends on the electronic structure at polymer-metal interfaces. Particularly, injection of charged particles (electrons/holes), from the electrodes (metals) to the dielectric (polymer) at these interfaces are known precursors to dielectric aging which culminates in the breakdown of the dielectric. In the absence of defect states on the interface, the probability of this charge injection is decided by the energy difference between the Fermi level of the electrode and the edges of valence and conduction bands of the dielectric. Conventionally, to bypass the need for laborious experimental measurements, computational schemes based on density functional theory (DFT) are employed to calculate these properties. But, computing these properties to match experimental observations requires calculations to be performed on large metal-polymer interfacial structures containing hundreds of atoms. This is computationally expensive and limits the chemical space one can explore to find a suitable combination for a given application. Here, we try to overcome this limitation by using a Multi-Fidelity Information-Fusion approach [1] to create a model that can predict the charge injection barrier, given a metal-polymer pair. To develop this model, we use charge injection barriers calculated at three different levels of fidelity. In the lowest level of fidelity, band edges positions of isolated single-chain of polymers and Fermi level of isolated metal slabs are computed to obtain the charge injection barriers. These calculations are cheap and hence a database of these values are created from quick calculations performed on a large number of metal-polymer systems. Further, property values obtained from expensive computations, considering large polymer-metal interfaces, are performed on a subset of these systems and are used as data for the second level of fidelity. Finally, the experimentally observed values of the property are used as the highest fidelity data. A co-kriging method is then used to model the relationship between a numerical representation of the metal-polymer system and values of the property at different fidelities. This approach provides an accurate, inexpensive and flexible alternative to expensive DFT calculations to obtain reliable values of charge injection barriers. These models can be readily used for screening metal-polymer systems for insulating applications.

Perovskite is a classic class of materials that have been studied for a long time. Their interesting properties, such as ferroelectricity, have been enabled a wide range of applications. Recently, the Ruddlesden-Popper (RP) phase of perovskites is becoming an exciting research topic, owing to the interesting properties like superconductivity and optoelectronics. The RP perovskite has layered crystalline structure, consisting of alternated ABC₃ and AC layers with a compound formula of A_n+1B_nC_3n+1 where n = 1, 2, … (Note that the conventional ABC₃ can be seen as n = infinity). However, our understanding of RP perovskite is limited. It is highly desirable to gain an overall picture of the structure, energetic, and electronic structures of RP perovskites, particularly how the size confinement effect (i.e., the finite n) could differentiate their physical properties from the conventional ABC₃. In this work, we did data-mining for the RP perovskites based on the Materials Project database (including over 100 000 inorganic compounds). It is well recognised that octahedra rotation plays a critical role in the electronic, ferroelectric, and magnetic properties of perovskites. We, therefore, carried out in-depth structure and energy analysis of octahedra rotation of RP perovskite and compared them with the conventional ABC₃. Some physical insights are obtained to understand the physical origins of octahedra rotation. The detailed observations and conclusions are listed below.
(1) 254 RP perovskite or inverse perovskite crystal structures are identified. (2) The RP perovskites cover diverse chemical species, such as oxide, halide, hydride, and have a wide composition range. (3) The majority of RP perovskite compounds with n=1, i.e., 38 out of 56, have a larger band gap than ABC₃ counterparts. There are 12 semiconductor RP compounds whose ABC₃ counterparts are metallic. (4) Most of the RP perovskites inherit the octahedra rotation patterns of the ABC₃ counterparts. (5) Five RP perovskites have multiple stable octahedra rotation patterns. (6) For most RP perovskites, there is a clear trend that the total energy difference between the rotated and non-rotated configurations are reduced. For some RP perovskites, the non-rotated configurations even become the ground states. (7) Our in-depth analysis suggests that the pseudo-Jahn-Teller effect could be the driving force of octahedra rotation and the relative A-site ion size (concerning B-site as expressed by tolerance factor) could serve as a resistance force for the octahedra rotation. (8) As for RP perovskites, the existence of interface moves the d-orbital of B-site ion to lower energy levels, weakening the pseudo-Jahn-Teller effect and thus the tendency of octahedra rotation.
Our work provides an overview of the structure, energetic, and electronic properties of RP perovskites, which could lay the ground for further investigation of properties of novel two-dimensional perovskites exfoliated from RP phase.

Molecular dynamics is an essential computational tool for the development and discovery of new materials. However, large length and time scale systems can be computationally intractable. In particular, integration times must be small enough to capture the most rapid motions in the system. In addition, computational expense increases with system size, making it impossible to observe the dynamics of protein folding events, for example. In light of this, coarse graining procedures to explicitly replace atoms with coarser “beads” are often needed to simulate dynamics even at millisecond timescales.

While predominantly guided by chemical and physical intuition, recent works have been aimed at designing automated procedures for determining coarse grained units [1,2,3]. To this end, we demonstrate the efficacy of spectral clustering algorithms in coarse graining a variety of systems that have not been previously explored with these methods. Alternative metrics for comparing atomic trajectories are considered that can preserve rare-events at the coarse grained level, and we explore the physical intuition of the latent space that is afforded by these procedures as opposed to other machine learning methods.

In addition, we consider the coupled problem of finding a force model for the coarse grained units along with that of “fine graining” - given a latent space representation of the atomic system, how do the units interact, and how can we go back to the atomic representation despite the dramatic loss of information? Recent work using Gaussian process regression for ab initio molecular dynamics [4] has demonstrated the advantage of Bayesian methods for force field construction, and we extend these ideas to the dynamics of latent space force fields. The variance of force predictions provides a natural indicator of when the coarse grained system is insufficient for capturing the system dynamics, from which we consider a method for going between the atomic and coarse grained representations.

[1] Luca Ponzoni, Guido Polles, Vincenzo Carnevale, and Cristian Micheletti, “Spectrus: A dimensionality reduction approach for identifying dynamical domains in protein complexes from limited structural datasets,” Structure 23 (2015).
[2] Michael A. Webb, Jean-Yves Delannoy, and Juan J. de Pablo, “Graph-based approach to systematic molecular coarse-graining,” J. Chem. Theory Comput. 15 (2019).
[3] Wujie Wang and Rafael Gómez-Bombarelli, “Coarse-Graining Auto-Encoders for Molecular Dynamics,” arXiv preprints arXiv:1812.02706.
[4] Jonathan Vandermause, Steven B Torrisi, Simon Batzner, Alexie M Kolpak, and Boris Kozinsky, “On-the-Fly Bayesian Active Learning of Interpretable Force-Fields for Atomistic Rare Events,” arXiv preprints arXiv:1904.02042.

Organic light emitting diodes (OLEDs) are one of the leading technologies used as active optoelectronic device in displays. Currently, in OLED displays the most challenging quest is to synthesize and design efficient and new stable blue emitters or improve already existing emitters. To achieve this goal it is of great interest to evaluate a large number of possible molecules prior to synthesis and to optimize their molecular design.

In this study a large amount, i.e. 151 derivatives, of phosphorous fluorescent compounds were evaluated with first principle simulations regarding their light emitting properties and molecular orbital alignment. As basic structure the central phosphorous atom is bound to one derivative of 8-Quinolinol as well as 4 fluorine atoms. 8-Quinolinol was systematically varied by introduction of substituents from strong electron withdrawing groups (EWG) towards strong electron donating groups (EDG). Besides the mesomeric and inductive electron withdrawing and donating effects also the position of the substituent has a crucial influence on the calculated emitting wavelength of these fluorophores. To determine the emitting wavelength of an isolated molecule time-dependent density functional theory (B3LYP/6-31+G(d,p)) was employed. EWGs tend to increase the bandgap if placed on the benzene ring of the ligand, where EDGs show the same effect when substituted on the pyridine ring of the ligand. This allows to tune the calculated bandgap between 3.2 - 4.1 eV and to choose configurations with energetically aligned molecular orbitals.