Magnetic and superconducting phase diagrams and transition temperatures predicted using text mining and machine learning

Callum J Court,Jacqueline M Cole

doi:10.1038/s41524-020-0287-8

Abstract

Predicting the properties of materials prior to their synthesis is of great importance in materials science. Magnetic and superconducting materials exhibit a number of unique properties that make them useful in a wide variety of applications, including solid oxide fuel cells, solid-state refrigerants, photon detectors and metrology devices. In all these applications, phase transitions play an important role in determining the feasibility of the materials in question. Here, we present a pipeline for fully integrating data extracted from the scientific literature into machine-learning tools for property prediction and materials discovery. Using advanced natural language processing (NLP) and machine-learning techniques, we successfully reconstruct the phase diagrams of well-known magnetic and superconducting compounds, and demonstrate that it is possible to predict the phase-transition temperatures of compounds not present in the database. We provide the tool as an online open-source platform, forming the basis for further research into magnetic and superconducting materials discovery for potential device applications.

Highlights

Driven materials discovery is costly, inefficient and largely reliant on scientific intuition[1,2]
Materials informatics has demonstrated the effectiveness of machine learning for property prediction and materials discovery[2,3,4,5,6]
We freely provide the complete database in the RESULTS Case study of perovskite manganites: reconstructing phase diagrams We begin with a case study of the perovskite-type oxides

Summary

Introduction

Driven materials discovery is costly, inefficient and largely reliant on scientific intuition[1,2]. Materials informatics is an emerging field of research that aims to enhance this materials discovery process through computational methods. The majority of such projects are high-throughput computational methods; examples include the Harvard Clean Energy Project[8] and the Materials Project[9], focussed on the discovery of photovoltaic and battery materials, respectively. Computationally expensive, these approaches present significant savings in time and cost compared with experimentally driven research, thereby decreasing the timeline of materials discovery from decades to months. High-throughput projects that integrate computational and experimental data are rare, but afford actual materials discovery where they do exist[10]

Methods

Results

Conclusion