A machine learning approach to predict the magnetic property of metal-doped graphene

Metal-doped graphene materials have a wide range of applications. In particular, this study seeks to predict the magnetic moment of metal-doped graphene. Although density functional theory (DFT) provides accurate predictions, it is a computationally expensive approach, and due to being a quantum mechanical modelling method, does not produce human-readable models of molecule-level properties in terms of atomic-level properties and interactions. To aid the search for such predictive models, this research implements a machine-learning approach that couples a multi-objective genetic algorithm (MOGA) to the sure independence screening and sparsifying operator (SISSO). The MOGA will begin with an initial data set of candidate variables and magnetic moments calculated using DFT, and then optimize based on model goodness of fit and parsimony. Initial results on small monovacancy graphene supercells are promising, and indicate a relationship between the magnetic moment and other key features, such as d-orbital electron configurations, dopant bond length, and dopant height out of plane, as well as higher-order nonlinearities due to band gap, and binding energy. This presentation will showcase the extension of this approach to larger supercells, as well as other related configurations.