Semiempirical and interpretable machine learning of the oxygen interaction barriers in thousands of the two-dimensional materials

We present a combined semiempirical and machine learning approach to predict oxygen interaction barriers in 4036 two-dimensional (2D) materials from the C2DB database. Using the Extended Hückel Method (EHM), calibrated to reproduce the known oxygen barrier on graphene, we computed barrier energies along multiple adsorption paths. These values served as targets for supervised learning models based on descriptors from C2DB and Matminer. Among the tested models, XGBoost delivered the best performance, with SHAP analysis revealing that electronic features, such as electronegativity and valence electron count, are key predictors of barrier height, highlighting the underlying nonlinear relationships between material features and adsorption behavior. This framework enables efficient and interpretable screening of oxygen reactivity in 2D systems, supporting the design of oxidation-resistant and functional surfaces. These findings underscore the role of nonlinear science in materials discovery and highlight how combining semiempirical modeling with interpretable machine learning can efficiently capture complex surface interactions in 2D materials.