Development of "Starry data" web system for data curation of published experimental thermoelectric properties

Abstract

Although numerous papers are published each year, most of the experimental data reported in those papers are only available as two-dimensional plot images. Data-driven materials science using the machine learning technologies will be accelerated by gathering those published experimental data into a database. By taking thermoelectric materials as a test case, we attempted to optimize the processes of collection of papers, extraction of numeric data from plot images, and sample-based data storage into a database. By searching with a keyword "thermoelectric", we obtained a list of 47,936 papers. Among these papers, we selected 18,471 papers as possible papers with thermoelectric properties, and succeeded to download 14,835 full-text PDF files. We developed a web system named "Starry data", to assist the sequential data extraction from the images contained in those PDF files. This system also assists materials scientists to annotate experimental samples efficiently, to develop a descriptive database that can be used for machine-learning of the complex, sample-dependent materials properties.