Machine Learning Accelerated Calculations of Alkali Metal Adsorption on AlGaN: Mixed Adatoms, Multiple Coverage, and Variable Components

Surface activation is pivotal for achieving aluminum gallium nitride photocathode with low surface affinity, yet optimizing composition, adsorbates, and coverage is highly intricate. Using the (001) surface as a model, we constructed 625 configurations spanning Al component (0–1), coverage (0–1 ML), and coadsorption of main-group atoms (Li–Cs). Density-functional theory calculations of adsorption energy and work function served as the learning targets. Atomic and structural descriptors were standardized via statistical functions and one-hot encoding. Among seven algorithms, XGBoost achieved the highest accuracy (adsorption energy: R 2 = 0.982, MSE = 0.0006; work function: R 2 = 0.902, MSE = 0.0185). Feature analysis identified atomic number, atomic work function, and Al component as the dominant descriptors. After hyper-parameter optimization with the DBO algorithm, XGBoost predicted two out-of-sample surfaces (0.625 ML coverage and 0.125 Al component) with average accuracy of 94.9 and 98.3% for adsorption energy and 97.7 and 95.4% for work function, respectively. This study demonstrates the interpretability and transferability of machine learning in photocathode surface engineering and offers a new pathway for optimizing surface processes in photocathodes.