Atmospheric doping of graphene, and the variation in atmospheric composition (e.g., humidity) makes graphene machine learning (ML) models limited to controlled environments (e.g., dry air), and unsuitable for atmospheric applications (e.g., non-invasive medical diagnosis). Herein, using nano-porous activated carbon for atmospheric passivation of the graphene channel, Extreme Gradient Boosting (XGBoost), K-nearest neighbors (KNN), and Naïve Bayes supervised ML models for atmospheric gas identification were developed, with model feature contribution studied using the Game theoretic approach, SHapley Additive exPlanations (SHAP). Unlike conventional graphene-based ML models which only monitor the gas adsorption induced doping and scattering without tuning voltage (Vt) modulation, in this work, the tuning voltage evolution of the gas adsorption induced van der Waals (vdW) complex doping and scattering characteristics was mapped out over time. The accuracy, precision, recall, and F1-score of the developed models in the environments: atmospheric ammonia, atmospheric acetone, ammonia in dry air, and acetone in nitrogen were 100% for XGBoost, 96%, 97%, 97%, 97% respectively for KNN, and 95%, 95%, 96%, 96% respectively for Naïve Bayes, even at single-digit ppb gas concentrations. Hence, our results advance the application of graphene based electronic nose models from conventional controlled environments to actual atmospheric sensing.