The hydrocarbon reserves and permeability of carbonate reservoirs are controlled by the microscopic heterogeneity of pore geometries and constituent components, collectively referred to as pore structure. Nevertheless, the geophysical characterization of pore structures is generally constrained by the intrinsic difficulty of direct observation and the theoretical complexity of the associated rock-physics models. Accordingly, as an extension and integration of conventional measurement-driven and model-driven workflows, this study proposes to construct a pore structure classifier based on an adaptive machine learning framework. Firstly, to improve the data quality of the original core sample set, we developed a tailored data-model hybrid training set, which is labeled based on microscopic dominant pore geometries and characterized by macroscopic elastic property features. In this hybrid set, we extended the range and density of rock-physics measurements by incorporating abundant virtual carbonate samples generated through the differential effective medium theory guided by a random procedure. Secondly, to mitigate potential biases in the classifier’s training procedure, an adaptive dual trade-off weighting system is devised to address the issues of imbalance (cored versus modeled samples) and skewness (majority versus minority samples) in the raw sample distribution. Thirdly, the proposed weighted data-model-driven framework outperforms pure data-driven methods, pure model-driven methods, and unweighted methods in both classification accuracy and model robustness for the pore structure characterization task. Our approach proves effective in characterizing the vertical heterogeneity of pore structure, enabling accurate reserve estimation and providing critical insights for inferring geological history from logging data.