Active learning using uncertainty-driven dynamics for configurational space search of molecules on metal surfaces

Efficient identification of stable and metastable adsorption geometries of molecules on metal surfaces is essential for many applications, such as catalysis and corrosion inhibition. Here, we study the mechanism of corrosion inhibition by benzotriazole (BTA) molecule on Cu (111) surface using an equivariant machine learning interatomic potential with long-range electrostatics, NequIP-LR[1]. We employ active learning (AL) using an uncertainty-driven dynamics (UDD) framework for autonomous searching of the configurational space of the BTA/Cu interface. We use an ensemble of NequIP-LR models for UDD, allowing uncertainty quantification and charge transfer between the BTA and Cu surface. UDD guides molecular dynamics trajectories towards unexplored and high-uncertainty regions. AL allows iterative retraining of the models to learn diverse adsorption geometries, molecular orientations, and charge-redistribution patterns at the Cu surface. Thus, the UDD-AL framework provides a computationally efficient approach to discover BTA adsorption motifs on the Cu surface.

[1] M. U. Maruf et al., J. Phys. Chem. Lett. 2025, 16, 35, 9078–9087.