Machine-Learning Screening of Inorganic Compounds for Defect Spin Properties

An ensemble regression algorithm predicting refractive index, band gap and magnetic susceptibility utilizing elemental composition based physical descriptors is developed to screen for inorganic compounds with color defect spin properties at low computational cost. Mean absolute error (MAE) values of 0.46 for refractive index, 0.65 eV for bandgap utilizing gradient boost regression and 2.95e-3 cm3/mol for magnetic susceptibility with random forest regression were obtained via 70:30 train-test splits. When screened for threshold values of bandgap >2 eV, refractive index > 2 and magnetic susceptibility < 10e-5 cm3/mol, notable binary compounds with defect spin properties include GaN and SiC among 348 compounds demonstrating model utility.