Abstract: Permeability prediction for tight sandstone reservoirs usually uses physical and fitted models. However, physical models face challenges in obtaining accurate petrophysical parameters, while completely data-driven fitted models show poor performance in heterogeneous reservoirs. In response to these challenges, we propose a coupled approach integrating physical and fitted models. An ensemble learning model Stacking is employed to predict flow zone indicator (FZI) that characterizes reservoir flow units. The Kozeny-Carman model is combined with a discrete rock typing (DRT) method for reservoir classification. An improved particle swarm optimization (IPSO) algorithm is then utilized to synchronize the parameters of the physical and machine learning models dynamically. As per a case study in Jianghan oilfield, the adaptive IPSO-Stacking model, established by coupling a physical model with a machine learning fitted model, achieves an accuracy of 98% in FZI prediction and 93% in permeability estimation, demonstrating robust predictive ability. The IPSO method is superior to conventional meta-heuristic optimization algorithms in adjusting physical and fitted model parameters. Through IPSO iteration, the derived DRT empirical coefficients exhibit enhanced adaptability. The dual-driven IPSO-Stacking model achieves high-precision permeability prediction for tight sandstone reservoirs through collaborative optimization by integrating physics-based and data-driven approaches.