To construct self-defined aerosol models and enhance the accuracy of aerosol optical depth (AOD) retrieval, a novel satellite-based method for estimating the volume percentages of aerosol components based on a stacking model was proposed. First, a dataset was constructed utilizing MODIS(MODerate resolution imaging spectroradiometer ) data and ERA5 meteorological data. To address the challenge of limited training samples, Tab-DDPM (tabular denoising diffusion probabilistic model) was employed to generate synthetic samples. Subsequently, a stacking model integrating random forest, XGBoost, and CatBoost was established to accurately estimate the volume percentages of aerosol components. Based on these estimated proportions, self-defined aerosol models were constructed to facilitate improved AOD retrieval. Experimental results indicate that compared with single machine learning models, the estimation accuracy of aerosol components is significantly improved by the stacking model. Validation against ground-based AERONET(AErosol RObotic NETwork) observations in the Beijing and Xianghe regions demonstrates that AOD retrieved based on self-defined aerosol models exhibits high consistency, with correlation coefficients r of 0.80 and 0.83, root mean square errors (RMSE) of 0.23 and 0.20, and mean bias errors (MBE) of 0.09 and 0.03, respectively. Compared with the traditional continental aerosol model (which yields RMSEs of 0.29 and 0.28), systematic errors are significantly reduced by the proposed method. Furthermore, in comparison with the official MODIS AOD product, although the correlation coefficients are slightly lower (official product r≈0.90), superior error control capability is demonstrated by the proposed method in terms of RMSE (official product values of 0.29 and 0.22) and MBE. The results demonstrate that the model mismatch problem is effectively addressed, and the accuracy of AOD retrieval is significantly enhanced by the proposed method.