How Realistic are Idealized Copper Surfaces? A Machine Learning Study of Rough Copper-Water Interfaces

Abstract

Copper is a highly promising catalyst for the electrochemical CO$_2$ reduction reaction (CO2RR) since it is the only pure metal that can form highly added-value products such as ethylene and ethanol. Since the CO2RR takes place in aqueous solution, the detailed atomic structure of the water-copper interface is essential for unraveling the key reaction mechanisms. In this study, we investigate copper-water interfaces exhibiting nanometer-scale roughnesses. We introduce two molecular dynamics protocols to create rough copper surfaces, which are subsequently brought into contact with water. From these interfaces, we sample additional training configurations from machine-learning-interatomic-potential-driven molecular dynamics simulations containing hundreds of thousands of atoms. An active learning workflow is developed to identify regions with high spatially resolved uncertainty and convert them into DFT-feasible cells through a modified amorphous matrix embedding approach. Finally, we analyze the local environments at the interface using unsupervised machine-learning techniques. Unique environments emerge on the rough copper surfaces absent from model systems, including stacking-fault-induced configurations and undercoordinated corner atoms. Notably, corner atoms consistently feature chemisorbed water molecules in our simulations, indicating their potential importance in catalytic processes.