The interactions between the atmospheric gases and the halide perovskite materials are receiving attention in these years before the extensive industrial deployment of halide perovskite materials. In this manuscript, we combine first-principles calculation and machine learning techniques to evaluate the interactions between the atmospheric gas molecules and a two-dimensional Ruddlesden–Popper halide perovskite Cs2PbBr4 surface based on the adsorption energies and automatically design advanced molecular descriptors for the target output. The impacts of density functionals are considered while an accurate machine learning model (r = 0.954 and R 2 = 0.951) is obtained based on the XGBRF ensemble algorithm. Importantly, the symbolic regression automatically finds an effective hybrid descriptor that exhibits high correlation with the target output that is comparable with the machine learning model; the symbolic regression-derived descriptor is mathematically simple and chemistry-aware, which complements the debatable ‘black-box’ machine learning model. Both feature importance ranking and symbolic regression indicate the importance of the functional-dependent energy levels of the perovskite systems and the amide/hydroxyl functional groups of the molecules. The present study highlights the viability of combining density functional theory and machine learning techniques to model the low-dimensional perovskite structures under the atmospheric conditions.