Porous materials such as metal–organic frameworks (MOFs) are widely studied for applications in catalysis and gas adsorption, including carbon dioxide capture and hydrogen storage for energy and environmental challenges. Accurate simulation of gas adsorption in these materials remains challenging because most approaches rely on force fields (FFs), which lack first-principles accuracy. Here, we present a hybrid sampling approach, complemented by an active-learning-style refinement step, to generating machine-learning potentials (MLPs) tailored for grand canonical Monte Carlo (GCMC) simulations of gas adsorption. The training dataset combines snapshots from FF-based GCMC sampling, evaluated by single-point density functional theory (DFT) calculations, with trajectories from DFT-based molecular dynamics (MD). We demonstrate this approach for hydrogen adsorption in MOF-303, achieving accurate energy and force evaluation, as well as robust GCMC simulation including on-the-fly quantum corrections for nuclei. The MLP-based calculation also revealed four primary and four secondary adsorption sites formed near N, O, C, and –OH motifs, highlighting the material’s structural heterogeneity and adsorption behavior. The resulting MLPs enable calculation of adsorption isotherms, isosteric heat values, and diffusion coefficients across temperatures. This hybrid-sampling framework bridges DFT-level accuracy and larger-scale simulation capability than $abinitio$ MD (AIMD), and can in principle be applied to other porous materials and gas types to support the design of materials for gas storage and separation.