Machine learning and symbolic regression for adsorption of atmospheric molecules on low-dimensional TiO2

Abstract

The adsorption of atmospheric gas molecules and the low-dimensional TiO2 species are relevant for the atmospheric photochemistry and energy-related photoelectrochemistry. In this manuscript, an accurate and interpretable machine learning model for the adsorption energies of the atmospheric and green-house molecules on a representative low-dimensional TiO2 surface is constructed, while the symbolic regression based on the genetic algorithm is employed to automatically design relevant molecular descriptors for the target adsorption outputs. Different algorithms are compared and the random forest algorithm achieves highest accuracies (r = 0.98, MAE = 0.29 and R2 = 0.96) to describe the adsorption energies of the atmospheric gas/TiO2 interface subjected to various external electric fields. The scientific analysis is performed to visualize the relationships between the input features and output target, including the impacts of the elemental chemical compositions and external electric field of the atmospheric gases on the adsorption energies; the beneficial effects of the NH/OH groups in the atmospheric molecules for the surface adsorption and the negatively effects of the atomic van der Waals volumes as well as the number of the heavy atoms are revealed. Importantly, a highly relevant and chemical-aware hybrid molecular descriptor is automatically designed via the symbolic regression process, which complements the machine learning model. This study highlights the machine learning and symbolic regression for the fundamental understanding of the structural interactions between the gas molecules and the low-dimensional metal oxide surfaces.