Abstract: [Objective] Deep reef-shoal carbonate reservoirs are characterized by complex depositional origins and strong heterogeneity. Meanwhile, constrained by the limited availability of deep subsurface data and insufficient data resolution, conventional approaches that primarily rely on data and statistical characteristics remain considerable uncertainties in predicting reef-shoal facies distribution and the fine-scale characterization of reservoir architecture. [Methods] In this study, deep reef-shoal carbonate reservoirs in the Upper Permian Changxing Formation of the Puguang area (northeastern Sichuan Basin) are selected as a case study. Using the iDRS-CarbSIM software, stratigraphic forward modeling (SFM) was conducted to reconstruct the carbonate depositional system. The optimal stratigraphic forward model served as training images for integrated modeling with multiple-point geostatistics, while a large-scale training dataset generated via parameter perturbation was used to develop intelligent geological models using deep learning methods. [Results] Results demonstrate that the SFM model precisely characterizes spatial-temporal distribution of reef-shoal facies during 4 stages of the Late Permian Changhsingian Stage in the Puguang area and reconstruct the entire process of depositional evolution. By integrating quantitative analysis based on SFM model and previous qualitative analysis, controls on carbonate platform formation and reef-shoal development are revealed. The integrated modeling approach combining SFM with multiple-point geostatistics generates a high-resolution geological model, in which the geological consistency and spatial continuity of reef-shoal distribution were substantially enhanced, compared to conventional geological model approaches. Furthermore, the integration of SFM and deep learning approach enables geological knowledge-driven intelligent modeling, which significantly improves modeling efficiency and enhances the consistency between the generated models and observational data. [Conclusion] The study highlights that SFM can provide a quantitative geological basis for favorable sedimentary facies prediction of deep reef-shoal carbonate reservoirs. It can also serve as reliable geological constraints for three-dimensional geological model optimization. Collectively, this approach has important implications for the fine-scale characterization of similar deep facies-controlled reservoirs.