JAMIP: an artificial-intelligence aided data-driven infrastructure for computational materials informatics

Abstract

Materials informatics has emerged as a promisingly new paradigm for accelerating materials discovery and design. It exploits the intelligent power of machine learning methods in massive materials data from experiments or simulations to seek new materials, functionality, and principles, etc. Developing specialized facilities to generate, collect, manage, learn, and mine large-scale materials data is crucial to materials informatics. We herein developed an artificial-intelligence-aided data-driven infrastructure named Jilin Artificial-intelligence aided Materials-design Integrated Package (JAMIP), which is an open-source Python framework to meet the research requirements of computational materials informatics. It is integrated by materials production factory, high-throughput first-principles calculations engine, automatic tasks submission and monitoring progress, data extraction, management and storage system, and artificial intelligence machine learning based data mining functions. We have integrated specific features such as an inorganic crystal structure prototype database to facilitate high-throughput calculations and essential modules associated with machine learning studies of functional materials. We demonstrated how our developed code is useful in exploring materials informatics of optoelectronic semiconductors by taking halide perovskites as typical case. By obeying the principles of automation, extensibility, reliability, and intelligence, the JAMIP code is a promisingly powerful tool contributing to the fast-growing field of computational materials informatics.

Introduction

Currently, materials science research is entering a new paradigm featuring materials informatics, which applies machine learning techniques to massive materials data [1], [2], [3], [4] to accelerate materials discovery and design. It was first promoted in part by the Materials Genome Initiative [5] and emerged in practice as the result of the fast development of experimental materials synthesis, characterization approaches, and theoretical materials simulation through available computation resources that yield massive materials data [1]. By utilizing the recognized predictive power of materials informatics, encouraging breakthroughs have been made, including the design of new materials, the identification of new relationships [6], [7], and the prediction of new principles [8], [9]. In the fields of both experimental and computational materials informatics, there is an urgent need to develop powerful facilities/infrastructures to meet research requirements under this new paradigm.

Generating massive materials data of the composition-structure-property relationship is essentially the first step for materials informatics studies. In the field of computational materials science, there are generally three materials data generation techniques: the first is through direct calculations of single or a few materials. This is used frequently in the old paradigm when theorists want to study some material properties observed experimentally or to explore new physics in some previously unstudied materials. Such data are dispersedly distributed in literature and may be gathered manually or by automatic text extraction approaches [10], [11]. The second is through the high-throughput (HT) calculations, where extremely large numbers of materials spanning different chemical compositions are automatically calculated based on prototype structures [12], [13]. Along this direction, a series of computational infrastructures are developed and used in high-throughput calculations [12], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24]. The accumulated computational materials data lead to the formation of databases that are accessible for query and research purposes [25], [26], [27], [28], [29]. The third is through crystal structure searches. These seek new stable and metastable material structures under the typical chemical composition. This method complements high-throughput calculations. It explores the potential energy profile of materials, usually by combining artificial intelligence global optimization methods with energetic calculations [25], [30], [31], [32], [33], [34]. Because in structure search studies, the object of the greatest concern is the lowest-energy ground-state structure, the generated materials data are usually regarded as intermediate products, and do not accumulate in materials informatics studies.

Machine learning techniques act as an engine that triggers/accelerates materials discovery in computational materials informatics [35], [36], [37]. Various machine learning models are already available for use in computational material informatics, such as supervised learning, unsupervised learning, and active learning [38]. Supervised learning requires an artificially constructed training process, which then understands the mapping relationships between material descriptors and material properties. Using this model, both the underlying physical laws between descriptors and the material properties can be mined [39], [40]. Additionally, it can be used to reverse-engineer new materials to uncover new materials [41], such as predicting energy-related materials for catalysis, batteries, solar cells, and gas capture [42], [43], [44]. For instance, Chen et al. [44] reported a powerful machine learning model with the extreme gradient boosting regression (XGBR) algorithm to predict the Gibbs free energy changes of CO adsorption in the dispersed metal-nonmetal codoped graphene electrocatalysts. Unsupervised learning, by contrast, is generally used to uncover differences among materials or chemical systems in terms of unmarked “descriptors”. Clustering algorithms are the most widely used for this purpose. For example, Chen et al. [45] used the K-mean clustering algorithm to analyze the oxygen diffusion pattern, hopping statistics, and site occupation within crystals. In addition, in the field of materials discovery, active learning, and Bayesian optimization, with the self-optimizing characteristics of model, allow us to search the potential materials space for high-quality materials in a limited data range, and contribute to the material experimental and computational design [46].

The software/infrastructure for computational materials informatics needs to meet at least the following requirements, but is not limited to them: (1) adaptability for diverse materials discovery and design: the software should flexibly deal with different material systems from inorganic and organic to organic-inorganic hybrid systems, and different materials variations/combinations, such as alloy or defect, surface or interface, and heterostructure or superlattice. To fit the potential complexity of different materials with different properties, computational workflow modules should have high scalability and ensure flexible combinations among different types of computing tasks. (2) Efficiency and reliability in data generation and management: a high-level automation framework to initialize, run, and analyze large-scale high-throughput calculations is essential. Monitoring computational tasks in real-time and reporting/correcting potential errors are important. The software should have tools to efficiently extract and collect results and store them in a self-contained database. (3) Orientation by materials functionality: the software should be oriented by the functionality of materials (e.g., photovoltaic, thermoelectric, ferroelectric, and catalytic), which is central to materials discovery. Rational design of workflow modules needs to be sufficiently considered to effectively calculate the functionalities of different material systems. (4) Synthetical integration of data generation, storage, processing, and learning: it is desirable to integrate data generation, storage, processing, and learning modules with a unified infrastructure framework, which greatly facilitates the data fluxion and conversion procedure, as well as the efficiency of data mining. In the integrated framework, newly calculated data motivated by the feedback from data-learning procedures may accelerate knowledge accumulation.

In this paper, we report on an artificial-intelligence-aided data-driven infrastructure that we have been developing to fulfill the above requirements for computational materials informatics. It is an open-source Python-integrated framework named the Jilin Artificial-intelligence-aided Materials-design Integrated Package (JAMIP). The code is integrated by intimately connected units of Data generation (e.g., high-throughput materials calculations and screening), Data collection (e.g., automatic data extraction, management, and storage) and Data learning (e.g., integrated feature engineering, and machine learning based data mining functions). Below, we describe the JAMIP code framework and its usage in carrying out materials-informatics-related research on optoelectronic semiconductors by taking halide perovskites as instances. The code and manual describing its detailed usage options are freely available for download at www.jamip-code.com.