Focus on Advanced Material Modelling, Machine Learning and Multiscale Simulation

The effective mass approximation (EMA) could be an efficient method for the computational study of semiconductor nanostructures with sizes too large to be handled by first-principles calculations, but the scheme to accurately and reliably introduce EMA parameters for given nanostructures remains to be devised. Herein, we report on an EMA approach based on first-principles-derived data, which enables accurate predictions of the optoelectronic properties of quantum nanostructures. For the CdS/ZnS core/shell quantum rods, for which we recently reported its experimental synthesis, we first carry out density functional theory (DFT) calculations for an infinite nanowire to obtain the nanoscopic dielectric constant, effective mass, and Kohn-Sham potential. The DFT-derived data are then transferred to the finite nanorod cases to set up the EMA equations, from which we estimate the photoluminescence (PL) characteristics. Compared with the corresponding method based on bulk EMA parameters and abrupt potential, we confirm that our EMA approach more accurately describes the PL properties of nanorods. We find that, in agreement with the experimentally observed trends, the optical gap of nanorods is roughly determined by the nanorod diameter and the PL intensity is reduced with increasing the nanorod length. The developed methodology is additionally applied to CdSe nanoplatelets, where reliable experimental data became recently available. Here, we again obtain excellent agreements between calculated and measured optical gap values, confirming the generality of our approach. It is finally shown that the abrupt confinement potential approximation most adversely affects the accuracy of EMA simulations.

The machine learning models developed on a dataset comprising particular class of materials show poor transferability across different classes. The problem can be partially solved by increasing the variability in the dataset at the cost of prediction accuracy. To develop a model on a highly variable database, we propose a localized regression based patchwork kriging approach for capturing most of the complex details in the data. In this approach, the data is partitioned into smaller regions with shared patches of few datapoints across the neighboring boundaries. Local regression functions are developed in each partition with a constrain to give similar performance at the boundary. Out of 17 different properties tried for partitioning the data, the decomposition with respect to target output κ_l gave local models with unprecedented accuracies. The partitioning with respect to κ_l, however, requires its estimate for any unknown compound beforehand. To address this, we developed a global model for the entire database. The global model accurately predicts the order of magnitude of κ_l for the compounds in the dataset and hence, directs them towards a particular partition for more accurate prediction. We define this stepwise approach as guided patchwork kriging, which can be applied to develop highly accurate transferable prediction models.

We propose a data-driven method to extract dissimilarity between materials, with respect to a given target physical property. The technique is based on an ensemble method with Kernel ridge regression as the predicting model; multiple random subset sampling of the materials is done to generate prediction models and the corresponding contributions of the reference training materials in detail. The distribution of the predicted values for each material can be approximated by a Gaussian mixture models. The reference training materials contributed to the prediction model that accurately predicts the physical property value of a specific material, are considered to be similar to that material, or vice versa. Evaluations using synthesized data demonstrate that the proposed method can effectively measure the dissimilarity between data instances. An application of the analysis method on the data of Curie temperature (${T}_{{\rm{C}}}$) of binary 3d transition metal- 4f rare-earth binary alloys also reveals meaningful results on the relations between the materials. The proposed method can be considered as a potential tool for obtaining a deeper understanding of the structure of data, with respect to a target property, in particular.

We use a machine learning approach to identify the importance of microstructure characteristics in causing magnetization reversal in ideally structured large-grained Nd₂Fe₁₄B permanent magnets. The embedded Stoner–Wohlfarth method is used as a reduced order model for determining local switching field maps which guide the data-driven learning procedure. The predictor model is a random forest classifier which we validate by comparing with full micromagnetic simulations in the case of small granular test structures. In the course of the machine learning microstructure analysis the most important features explaining magnetization reversal were found to be the misorientation and the position of the grain within the magnet. The lowest switching fields occur near the top and bottom edges of the magnet. While the dependence of the local switching field on the grain orientation is known from theory, the influence of the position of the grain on the local coercive field strength is less obvious. As a direct result of our findings of the machine learning analysis we show that edge hardening via Dy-diffusion leads to higher coercive fields.

Strongly anisotropic bands near the Fermi level via band structure engineering have been proposed to enhance thermoelectric performance in functional materials. Recent works exhibit the presence of flat-and-dispersive-band-like strong anisotropy in a class of ternary transition metal pnictides. Taking KZnP as a representative example, here we investigate the thermoelectric properties of this class of materials based on first-principles calculations and semiclassical Boltzmann transport theory. Strikingly, the calculated lattice thermal conductivity ${\kappa }_{L}$ shows a strong anisotropy with a small value of about 5.24 (2.58) W m⁻¹K⁻¹ along the in-plane (out-of-plane) lattice direction at room temperature. Based on the electron relaxation time calculated from intensive ab initio electron–phonon interactions, a thermoelectric figure of merit zT of 0.32 is predicted for n-type doped KZnP, about two times lower than the value estimated by the constant relaxation time approximation from deformation potential theory. Nanostructures with the characteristic length shorter than 13 nm can reduce the ${\kappa }_{L}$ by 40%, enhancing zT to 0.52 along the c axis direction. This work supports that KZnP is a potential candidate for thermoelectric applications.

The identification of descriptors of materials properties and functions that capture the underlying physical mechanisms is a critical goal in data-driven materials science. Only such descriptors will enable a trustful and efficient scanning of materials spaces and possibly the discovery of new materials. Recently, the sure-independence screening and sparsifying operator (SISSO) has been introduced and was successfully applied to a number of materials-science problems. SISSO is a compressed sensing based methodology yielding predictive models that are expressed in form of analytical formulas, built from simple physical properties. These formulas are systematically selected from an immense number (billions or more) of candidates. In this work, we describe a powerful extension of the methodology to a 'multi-task learning' approach, which identifies a single descriptor capturing multiple target materials properties at the same time. This approach is specifically suited for a heterogeneous materials database with scarce or partial data, e.g. in which not all properties are reported for all materials in the training set. As showcase examples, we address the construction of materials properties maps for the relative stability of octet-binary compounds, considering several crystal phases simultaneously, and the metal/insulator classification of binary materials distributed over many crystal prototypes.

Localized structures in nano- and sub-nano-scales strongly affect material properties. Thus, some spectroscopic techniques have been used to characterize local atomic and electronic structures. If material properties can be directly 'measured' via spectral observations, the atomic-scale understanding of the material properties would be dramatically facilitated. In this paper, we have attempted to unveil the hidden information about the material properties directly and quantitatively based on core-loss spectra. We predicted six properties, including three geometrical and three chemical bonding properties, by a simple feedforward neural network, and achieved considerably sufficient accuracy. Moreover, we applied the constructed model to the noisy experimental spectrum and could predict the six properties precisely. This successful prediction implies that this method can pave the way for local measurement of the material properties.

Faithful representations of atomic environments and general models for regression can be harnessed to learn electron densities that are close to the ground state. One of the applications of data-derived electron densities is orbital-free density functional theory (DFT). However, extrapolations of densities learned from a training set to dissimilar structures could result in inaccurate results, which would limit the applicability of the method. Here, we show that a non-Bayesian approach can produce estimates of uncertainty which can successfully distinguish accurate from inaccurate predictions of electron density. We apply our approach to DFT where we initialise calculations with data-derived densities only when we are confident about their quality. This results in a guaranteed acceleration to self-consistency for configurations that are similar to those seen during training and could be useful for sampling-based methods, where previous ground state densities cannot be used to initialise subsequent calculations.

A major goal of computation is to optimize an objective for which a forward calculation is possible, but no inverse solution exists. Examples include tuning parameters in a nuclear reactor design, optimizing structures in protein folding, or predicting an optimal materials composition for a functional application. In such instances, directing calculations in an optimal manner is important to obtaining the best possible solution within a fixed computational budget. Here, we introduce Rocketsled, an open-source Python-based software framework to help users optimize arbitrary objective functions. Rocketsled is built upon the existing FireWorks workflow software, which allows its computations to scale to supercomputing centers and for its objective functions to be complex, long-running, and error-prone workflows. Other unique features of Rocketsled include its ability to easily swap out the underlying optimizer, the ability to handle multiple competing objectives, the possibility to inject domain knowledge into the optimizer through feature engineering, incorporation of uncertainty estimates, and its parallelization scheme for running in high-throughput at massive scale. We demonstrate the generality of Rocketsled by applying it to optimize several common test functions (Branin-Hoo, Rosenbrock 2D, and Hartmann 6D). We highlight its potential impact through two example use cases for computational materials science. In a search for photocatalysts for hydrogen production among 18 928 perovskites previously calculated with density functional theory, the untuned Rocketsled Random Forest optimizer explores the search space with approximately 6–28 times fewer calculations than random search. In a search among 7394 materials for superhard candidates, Rocketsled requires approximately 61 times fewer calculations than random search to discover interesting candidates. Thus, Rocketsled provides a practical framework for establishing complex optimization schemes with minimal code infrastructure and enables the efficient exploration of otherwise prohibitively large search spaces.

Data visualisation is an important part of understanding the distributions, trends, correlations and relationships in materials data sets, as well as communicating results to others. Traditionally visualisation has been straightforward, particularly when studying single-structure/single-property relationships. It is not so straightforward when confronted with a materials data set represented by a large number of features, and containing multi-structure/multi-property relationships. Here we use Kohonen networks, or self-organising maps, to aid in the visualise sets of silver and platinum nanoparticles based on structural similarity and overlay functional properties to reveal hidden patterns and structure/property relationships. We compare these maps to a popular alternative dimension reduction method and find them superior for our cases where the structure/property relationships are highly nonlinear, and the data set is imbalanced, as they often are in materials science.

The Novel Materials Discovery (NOMAD) Laboratory is a user-driven platform for sharing and exploiting computational materials science data. It accounts for the various aspects of data being a crucial raw material and most relevant to accelerate materials research and engineering. NOMAD, with the NOMAD Repository, and its code-independent and normalized form, the NOMAD Archive, comprises the worldwide largest data collection of this field. Based on its findable accessible, interoperable, reusable data infrastructure, various services are offered, comprising advanced visualization, the NOMAD Encyclopedia, and artificial-intelligence tools. The latter are realized in the NOMAD Analytics Toolkit. Prerequisite for all this is the NOMAD metadata, a unique and thorough description of the data, that are produced by all important computer codes of the community. Uploaded data are tagged by a persistent identifier, and users can also request a digital object identifier to make data citable. Developments and advancements of parsers and metadata are organized jointly with users and code developers. In this work, we review the NOMAD concept and implementation, highlight its orthogonality to and synergistic interplay with other data collections, and provide an outlook regarding ongoing and future developments.

Interfaces, in which the atomic structures are greatly different from those in the bulk, play a crucial role in the material properties. Therefore, determination of a central structure that is involved with the interface properties is an important task in materials research. However, determination of the interface structure requires a huge number of calculations. We previously proposed a powerful machine learning technique based on virtual screening (VS) to determine interface structures (Kiyohara et al 2016 Sci. Adv. 2 e1600746). Here, we discuss the feasibility, versatility, and robustness of the prediction model for VS. Through this study, the prediction model constructed using only 5 types of grain boundaries determines the energies and structures of the 52 grain boundaries. Furthermore, based on the constructed prediction models, we investigated the geometrical differences between the grain boundaries of different rotation axes. We also investigated the structure-property relationship at the grain boundary (GB). We found that a short bond at the GB is the key factor for preferential vacancy formation at the GB.

Recent advances in experimental and computational methods are increasing the quantity and complexity of generated data. This massive amount of raw data needs to be stored and interpreted in order to advance the materials science field. Identifying correlations and patterns from large amounts of complex data is being performed by machine learning algorithms for decades. Recently, the materials science community started to invest in these methodologies to extract knowledge and insights from the accumulated data. This review follows a logical sequence starting from density functional theory as the representative instance of electronic structure methods, to the subsequent high-throughput approach, used to generate large amounts of data. Ultimately, data-driven strategies which include data mining, screening, and machine learning techniques, employ the data generated. We show how these approaches to modern computational materials science are being used to uncover complexities and design novel materials with enhanced properties. Finally, we point to the present research problems, challenges, and potential future perspectives of this new exciting field.

We propose a data-driven method to extract dissimilarity between materials, with respect to a given target physical property. The technique is based on an ensemble method with Kernel ridge regression as the predicting model; multiple random subset sampling of the materials is done to generate prediction models and the corresponding contributions of the reference training materials in detail. The distribution of the predicted values for each material can be approximated by a Gaussian mixture models. The reference training materials contributed to the prediction model that accurately predicts the physical property value of a specific material, are considered to be similar to that material, or vice versa. Evaluations using synthesized data demonstrate that the proposed method can effectively measure the dissimilarity between data instances. An application of the analysis method on the data of Curie temperature (${T}_{{\rm{C}}}$) of binary 3d transition metal- 4f rare-earth binary alloys also reveals meaningful results on the relations between the materials. The proposed method can be considered as a potential tool for obtaining a deeper understanding of the structure of data, with respect to a target property, in particular.

Despite recent achievements in the design and manufacture of advanced materials, the contributions from first-principles modeling and simulation have remained limited, especially in regards to characterizing how macroscopic properties depend on the heterogeneous microstructure. An improved ability to model and understand these multiscale and anisotropic effects will be critical in designing future materials, especially given rapid improvements in the enabling technologies of additive manufacturing and active metamaterials. In this review, we discuss recent progress in the data-driven modeling of dynamical systems using machine learning and sparse optimization to generate parsimonious macroscopic models that are generalizable and interpretable. Such improvements in model discovery will facilitate the design and characterization of advanced materials by improving efforts in (1) molecular dynamics, (2) obtaining macroscopic constitutive equations, and (3) optimization and control of metamaterials.

The high theoretical capacity of Li-O₂ batteries attracts a lot of attention and this field has expanded significantly in the last two decades. In a more general way, the large number of articles being published daily makes it difficult for researchers to keep track of the progress in science. Here we develop a text mining program in an attempt to facilitate the process of reviewing the literature published in a scientific field and apply it to Li-O₂ batteries. We analyze over 1800 articles and use the text mining program to extract reported discharge capacities, for the first time, which allows us to show the clear progress made in recent years. In this paper, we focus on three main challenges of Li-O₂ batteries, namely the stability-cyclability, the low practical capacity and the rate capability. Indeed, according to our text mining program, articles dealing with these issues represent 86% of the literature published in the field. For each topic, we provide a bibliometric analysis of the literature before focusing on a few key articles which allow us to get insights into the physics and chemistry of such systems. We believe that text mining can help readers find breakthrough papers in a field (e.g. by identifying papers reporting much higher performances) and follow the developments made at the state of the art (e.g. by showing trends in the numbers of papers published—a decline in a given topic probably being the sign of limitations). With the progress of text mining algorithms in the future, the process of reviewing a scientific field is likely to become more and more automated, making it easier for researchers to get the 'big picture' in an unfamiliar scientific field.