Accomplishment and challenge of materials database toward big data^*

The collection of materials data started from the 1880s. At that time, the data sets were published as handbooks. Beilstein collected data of properties, spectra, and preparation of organic compounds from literatures, and published the first edition of his handbook in 1881.^[1] Up to 1989, the 4th edition of Beilstein handbook was published in 503 volumes (over 440000 pages). Since 2009, the content has been maintained and distributed by Elsevier information systems in Frankfurt under the product name "Reaxys".^[2]

Similar work has been done by Landolt and Börnstein^[3] since 1883. The current edition of Landolt–Börnstein handbook includes 195000 pages. More than 72000 element systems, 150000 chemical substances, and 530000 substance–property pairs are included. SpringerMaterials^[4] is a web platform, which provides data service based on Landolt–Börnstein handbook.

The Journal of Physical and Chemical Reference Data^[5] is published by American Institute of Physics (AIP) Publishing since 1972. One source of the journal contributions is the national standard reference data system (NSRDS), which was established in 1963 to coordinate the production and dissemination of critically evaluated reference data in physical sciences. National Institute of Standards and Technology (NIST) coordinates several data evaluation centers, located in universities, industrial, and government laboratories, as well as within NIST.

In Japan, National Research Institute for Metals (NRIM) published the first NRIM creep data sheet in 1966 and fatigue data sheet in 1976. This work has been continued after the reorganization of NRIM into National Institute for Materials Science (NIMS) in 2001. And NIMS published the first NIMS corrosion data sheet in 2002. Up to now, 57 volumes of creep data sheet, 123 volumes of fatigue data sheet, and 4 volumes of corrosion data sheet have been published.^[6–8]

Besides these large-scaled comprehensive chemical and physical data collections, many handbooks on specific material properties have also been published. For example, ASM handbook^[9] on mechanical properties, and Thermophysical Properties Research Center (TPRC) handbook^[10] on thermos-physical properties.

The data handbooks or journals are featured by high-quality data evaluated and selected by experts with high reputation in these fields. The editors did not only put those data together, but also classified and indexed them. For example, Beilstein developed the Beilstein system to classify compounds according to their constitutional features.^[1]

The computerized database started in the 1960s. In 1970, Codd^[11] published an important paper to propose the use of a relational database model, and his ideas changed the way people thought about databases. In his model, the schema or logical organization of the database is disconnected from physical information storage, and this became the standard principle for database systems. Since then, various relational database system products have been developed, such as MS SQL Server, DB2, Allbase, Oracle, etc.

The database enables us to store, update, and search a large amount of data, quickly, accurately, and securely. It also makes it very easy to transform and reorganize the data, and to generate a new data set for new purposes.

Since the 1990 s, some materials databases have been developed, e.g., Pauling File project,^[12] which was launched in 1995 as a collaboration between Japan Science and Technology Agency (JST) and Material Phases Data System (MPDS), a Swiss company. Since January 2016, the data has been copyrighted by NIMS and MPDS. The first goal of Pauling File project is to create and maintain a comprehensive materials database for non-organic (no C–H bonds) solid state materials, covering phase diagrams, crystallographic data, diffraction patterns, and physical properties. Up to now, the Pauling File data have been contained in 12 products. AtomWork^[13] and AtomWork Adv.^[14] are two database systems developed by NIMS. AtomWork is a free database opened to the Internet as a part of NIMS materials database system MatNavi,^[15] which contains the Pauling File data before 2001. AtomWork Adv. is a fee-based database released in 2018, which contains the full Pauling File data and will be updated annually. The present version of AtomWork Adv. contains 42406 phase diagrams, 303885 crystal structures, 550507 x-ray diffraction patterns, and 365517 materials properties, extracted from 141490 publications.

In these years, first principle computation has become an efficient method to generate materials data such as electronic structures and properties. Many databases have been established in this way. For example, Novel Materials Discovery (NOMAD) repository^[16] is developed as a joint project of Fritz-Haber-Institut (FHI), Humboldt-Universität zu Berlin (HUB), and Max Planck Computer and Data Facility (MPCDF). Materials Project^[17] is an integrated framework of data and analysis tools developed by University of California at Berkeley and Lawrence Berkeley National Laboratory.

In 2015, a Japanese national project "Materials Research by Information Integration Initiative (MI²I)" has been launched. The purpose of this project is to accelerate material exploration and development by combining materials data and data science method. As the fundamental of this project, a data platform MI²I-DPF^[18] has been constructed. MI²I-DPF is an integration of materials data, data application tools, and computational environment. MI²I-DPF is contemporary opened to the project members and the members of the industrial consortium. The users can access and download nearly a million data entries from 6 main databases of MatNavi through application programming interface (API), and use the tools to analysis or do simulation with the data on the platform. An example of the research results of this project using this system is as follows.

Various material interfaces exist in devices and materials such as composites, alloys, and ceramics. Especially for nanostructured materials, the properties of interfaces play important roles to determine the properties of the materials. However, interface is one of the most difficult issues in material science, because it is affected by numerous equilibrium and nonequilibrium factors.

Interfacial thermal resistance (ITR) is caused by scattering or reflection of phonons and electrons at the interface. Two physical models,^[19] acoustic mismatch model (AMM) and diffusion mismatch model (DMM), have been set up to calculate ITR of an interface based on the physical properties, concretely, density, acoustic velocity, and unit cell volume of the materials at two sides. However, comparing the calculation results to the experimental data, we can see that the accuracy of the calculation is not satisfied (Fig. 2).^[20] The reasons of the deviation between theory and experiment can be considered as: (i) the property data used in the calculation has large uncertainty, because properties such as acoustic velocity obtained by different methods or reported by different papers usually have large deviation; (ii) the descriptors used in AMM and DMM models are not sufficient. Except the above physical properties, ITR may be affected by many other chemical and materials factors. Some experimental studies have confirmed that the chemical bonding and interfacial structure are factors with significant influences on ITR.^[21,22]

Zoom In Zoom Out Reset image size

We try to establish a data model^[20] to predict ITR with higher accuracy from two approaches: (I) avoiding to use physical properties with high uncertainty; (II) introducing descriptors other than physical properties. To find an appropriate set of property descriptors, we check the correlations among 11 heat transfer related properties, and select 4 properties which are relatively independent and have less uncertainty. The optimized property descriptor set includes specific heat, melting point, density, and unit cell volume. We also introduce a new descriptor film thickness, since ITR measurements are usually done on samples with a film deposited on a substrate, and film thickness is one of the most basic parameters of film deposition. Although there is no clear theoretical explanation of the influence of film thickness on ITR, the data show an obvious correlation between them. With the new descriptor set and machine learning methods, we obtain much better prediction of ITR than AMM and DMM (Fig. 3).^[20]

Zoom In Zoom Out Reset image size

This example shows the potential of data science in material study. Because it neglects the complicated physical and chemical process between material composition, material structure and material property, it is especially suitable for studies on subjects which cannot be modeled by physical and chemical models.

Data shortage is the biggest difficulty when we apply data science methods to material science. Although the importance of materials database has been well noticed today, data accumulation takes long terms. Nevertheless, for materials, the correlations among composition, structure, and properties allow one descriptor to be substituted by another one or more descriptors, which provides us a possibility to select descriptors with relatively sufficient and high quality data.

The first problem when we try to combine data from different data sources is to identify the similarity between two materials, because in most cases, the data are obtained on different samples. Scientifically speaking, material can be defined by its chemical composition and structure. However, the characterization of a materialʼs structure is not an easy task, because it covers a large range of scales from atomic to macroscopic scale. Practically, many people use the process conditions to identify a material, however, the conditions are equipment dependent. MatML^[23] is a specification from NIST designed for the interchange of materials information. It uses chemical composition and processing conditions to describe a material. Most databases use only chemical composition or chemical formula to identify a material; in these case, materials with different structures are not distinguishable. Based on our experiences with the data of single crystal, ceramics, alloys, and polymers, we have developed a material identification system according to the fundamentals of materials science. In this system, materials are identified at four different levels: chemical system, compound, substance, and material. Figure 4 shows the four levels and the identifiers of each level.

Zoom In Zoom Out Reset image size

Chemical system is the fundamental of all materials. It indicates the element or elements of which the materials is composed. Compound is the second level, which identifies a material at molecular level. For most inorganic materials, the compound can be defined by its chemical formula. However, for organic or polymer materials, molecular structure must be specified. The third level is substance. At this level, the state (gas, liquid, or solid) of a compound should be defined. For solid state, the crystalline state and crystal structure should also be given. In most cases, a substance corresponds to a phase in a phase diagram. The forth level is material. To define a material, many types of information are needed, for example, the material form, size, micro-structure, process condition, etc. It is difficult to define a general descriptor set applicable to all types of material. Since samples prepared by different labs or under different processing conditions are usually different in micro-structure and properties, we treat different samples as different materials, except that the data providers indicate that they are identical at the material level. The four-level system allows us to distinguish each individual sample, meanwhile, to cluster materials according to their common physical and chemical features. Even two samples are different at the material level, if they belong to the same substance, compound, or chemical system, we can still find the relationship between them.

The above material identification system provides a possibility to compare and link materials data at the level of material, substance, compound, and element. However, many practically used materials are composed by multiple substances or materials. In order to make such complex materials comparable, we have designed a hierarchical structure for material description, which can be used to describe materials from atoms to composites. As an example, the hierarchical structure of the SiC/Ti alloy composite is shown in Fig. 5.

Zoom In Zoom Out Reset image size

Since material identification is a common issue when integrating data from different data sources, a system which manages the identification record of materials in different resources (Fig. 6) is efficient for this purpose. To construct such a system, a standardized format of material identification record is necessary. We have designed an XML format based on the four-level material identification method. In this format, we prepare 9 types of materials: atom, molecular, cluster, single phase material, multi-phase material, composite, interface, organic, and polymer. A similar hierarchical structure of Fig. 5 is used to describe a complex material. For example, a multi-phase material is composed of multiple single phase materials, and a composite is composed by multiple materials. As an example, the XML schema for single phase material is shown in Fig. 7.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size