Rapid prediction of molecule arrangements on metal surfaces via Bayesian optimization

The spatial arrangement of molecule adsorbates on a metal surface is very difficult to predict via first-principles calculations and standard optimizing algorithms. In this Letter, we show that a machine learning technique called Bayesian optimization can optimize the arrangement of two medium-sized aromatic adsorbates on a copper (111) surface within tens of density functional theory energy evaluations. The methodology reported here is therefore a step toward first-principles structure predictions for chemically modified surfaces, without the need to first specify the arrangement of molecule adsorbates from experimental data.