Abstract: Adaptive mesh refinement (AMR) is a key technique that dynamically adjusts spatial grid resolution during computation, thereby improving the efficiency of computational fluid dynamics (CFD) simulations while maintaining numerical accuracy. However, the key parameters of traditional AMR methods, such as refinement intensity and refinement frequency, are mainly determined based on empirical knowledge, making it difficult to achieve an optimal balance between accuracy and efficiency under complex flow conditions. Therefore, this study proposes a reinforcement-learning-based intelligent dynamic AMR method to enable efficient parallel simulation in large-scale CFD solvers. The proposed method designs the reward function with the objectives of improving accuracy and reducing computational cost, and updates the policy network using the proximal policy optimization algorithm, thereby transforming traditional heuristic parameter tuning into a data-driven intelligent decision-making process. Specifically, the reinforcement learning agent dynamically adjusts the sensor weights that determine the refinement and coarsening thresholds according to the evolving flow field, replacing fixed empirical parameters with state-dependent decisions. To validate the effectiveness of the proposed method, the framework is embedded into the high-performance solver m-AIA and tested on standard benchmark cases, including flow past a circular cylinder and steady and unsteady flows around a NACA0012 airfoil. The results show that, compared with traditional AMR, the proposed method reduces the average computational time by 37.21% and the average number of grid cells by 29.82% while maintaining comparable accuracy. In regions with strong flow gradients, the method maintains necessary local refinement and enables smooth transitions of grid levels as the flow evolves, which helps reduce unnecessary grid reconstruction overhead and improve the stability of the computational process. Generalization tests further demonstrate that, within a certain range of Reynolds numbers and angles of attack, the proposed method can maintain good adaptive performance without retraining, showing favorable robustness and engineering application potential. Therefore, the integration of reinforcement learning and AMR can further improve the computational efficiency of adaptive mesh methods and provide a reference for the development of intelligent CFD adaptive simulation approaches.