Machine-Learning-Assisted Materials Discovery from\nElectronic Band Structure

Traditional methods of materials\ndiscovery, often relying on intuition\nand trial-and-error experimentation, are time-consuming and limited\nin their ability to explore the vast design space effectively. The\nemergence of machine learning (ML) as a powerful tool for pattern\nrecognition has opened exciting opportunities to revolutionize materials\ndiscovery. This work explores the application of ML techniques to\nassist in the discovery of materials using band structure data. The\nelectronic band structure, which describes the energy levels of electrons\nin a material, holds vital information regarding its electronic and\noptical properties. The band structure data of 63,588 materials, including\nmetals and insulators, have been retrieved from the Materials Project\ndatabase. The data were grouped into 85 batches based on the band\npath in the first Brillouin zone. Three ML clustering algorithms were\ntrained on the band structure data after performing feature selection\nand engineering, followed by noise reduction. The models were validated\nby comparing the materials’ properties in a cluster.