Using machine-learning tools, the scientists identified important features to characterize materials that exhibit a metal-insulator transition. Credit: Northwestern University and the Massachusetts Institute of Technology
The work could allow scientists to accelerate the discovery of materials showing a metal-insulator transition.
An interdisciplinary team of scientists from Northwestern Engineering and the Massachusetts Institute of Technology has used artificial intelligence (AI) techniques to build new, free, and easy-to-use tools that allow scientists to accelerate the rate of discovery and study of materials that exhibit a metal-insulator transition (MIT), as well as identify new features that can describe this class of materials.
One of the keys to making microelectronic devices faster and more energy efficient, as well as designing new computer architectures, is the discovery of new materials with tunable electronic properties. The electrical resistivity of MITs may exhibit metallic or insulating electronic behavior, depending on the properties of the environment.
Although some materials that exhibit MITs have already been implemented in electronic devices, only fewer than 70 with this property are known, and even fewer exhibit the performance necessary for integration into new electronic devices. Further, these materials switch electrically due to a variety of mechanisms, which makes obtaining a general understanding of this class of materials difficult.
“By providing a database, online classifier, and new set of features, our work opens new pathways to the understanding and discovery in this class of materials,” said James Rondinelli, Morris E. Fine Professor in Materials and Manufacturing at the McCormick School of Engineering and the study’s corresponding primary investigator. “Further, this work can be used by other scientists and applied to other material classes to accelerate the discovery and understanding of other classes of quantum materials.”
“One of the key elements of our tools and models is that they are accessible to a wide audience; scientists and engineers don’t need to understand machine learning to use them, just as one doesn’t need a deep understanding of search algorithms to navigate the internet,” said Alexandru Georgescu, postdoctoral researcher in the Rondinelli lab who is the study’s first co-author.
The team presented its research in the paper “Database, Features, and Machine Learning Model to Identify Thermally Driven Metal–Insulator Transition Compounds,” published on July 6, 2021, in the academic journal Chemistry of Materials.
Daniel Apley, professor of industrial engineering and management sciences at Northwestern Engineering, was a co-primary investigator. Elsa A. Olivetti, Esther and Harold E. Edgerton Associate Professor in Materials Science and Engineering at the Massachusetts Institute of Technology, was also a co-primary investigator.
Using their existing knowledge of MIT materials, combined with Natural Language Processing (NLP), the researchers scoured existing literature to identify the 60 known MIT compounds, as well as 300 materials that are similar in chemical composition but do not show an MIT. The team has provided the resulting materials – as well as features it’s identified as relevant – to scientists as a freely available database for public use.
Then using machine-learning tools, the scientists identified important features to characterize these materials. They confirmed the importance of certain features, such as the distances between transition metal ions or the electrostatic repulsion between some of them known, as well as the accuracy of the model. They also identified new, previously underappreciated features, such as how different the atoms are in size from each other, or measures of how ionic or covalent the inter-atomic bonds are. These features were found to be critical in developing a reliable machine learning model for MIT materials, which has been packaged into an openly accessible format.
“This free tool allows anyone to quickly obtain probabilistic estimates on whether the material they are studying is a metal, insulator, or a metal-insulator transition compound,” Apley said.
Reference: “Database, Features, and Machine Learning Model to Identify Thermally Driven Metal–Insulator Transition Compounds” by Alexandru B. Georgescu, Peiwen Ren, Aubrey R. Toland, Shengtong Zhang, Kyle D. Miller, Daniel W. Apley, Elsa A. Olivetti, Nicholas Wagner and James M. Rondinelli, 6 July 2021, Chemistry of Materials.
Work on this study was born from projects within the Predictive Science and Engineering Design (PS&ED) interdisciplinary cluster program sponsored by The Graduate School at Northwestern. The study was also supported by funding from the Designing Materials to Revolutionize and Engineer our Future (DMREF) program of the National Science Foundation and the Advanced Research Projects Agency – Energy’s (ARPA-E) DIFFERENTIATE program, which seeks to use emerging AI technologies to tackle major energy and environmental challenges.