Abstract: Metallization of the surface in an ultrahigh vacuum environment is a common method for the processing of photocathode. However, multiple factors such as coverage, surface type, and metal type need to be considered. In this study, we employ first principles to calculate the adsorption behavior of ten metal elements from the main group (Li to Cs and Be to Ba) on AlGaN surface with different surface types ((100), (001) and (110)), Al composition (ranging from 0 to 1) and adsorbate coverages (0 to 1 monolayer (ML)). The adsorption energy and surface work function are derived to create a dataset with 600 surface configurations for machine learning. Comparative analysis of seven machine learning models reveals that the extreme gradient boosting (XGB) model shows superiority for this dataset, with determination (R²) of 0.993 and mean squared error (MSE) of 0.010 eV for adsorption energy and R² of 0.953 and MSE of 0.048 eV for work function. Analysis of feature importance and partial dependence indicates that the top three important features for the variation of adsorption energy and work function are the atomic number, surface coverage, and Al component. The established methods exhibit an averaged accuracy of 94.59 % for adsorption energy and 96.11 % for work function when applied to a new surface configuration with 0.625 ML metal coverage. Finally, we explain the physical mechanism of work function variation through the generation of dipole moment and the band bending. This framework enhances the interpretability and transferability of machine learning in the surface modeling and processing of photocathodes.

Key words: First principles, AlGaN, Metal adsorption, Work function, Adsorption energy, Machine learning