[Purposes] To improve vehicle merging efficiency under mixed traffic conditions, this study proposes a merging control method based on physics-informed reinforcement learning. [Methods] A mixed traffic freeway merging environment is constructed on the Highway-env simulation platform, and the merging area is modeled as a multi-agent system. A multi-agent proximal policy optimization (MAPPO) algorithm with embedded physical-information constraints is designed. A static–dynamic hybrid action-masking mechanism is developed to filter out invalid actions that violate traffic safety rules. In addition, physical rationality penalty terms are incorporated into the reward function to guide agents to comply with vehicle motion laws, reduce abnormal behaviors, and improve training efficiency and interpretability. [Findings] Compared with the uncontrolled scheme, the proposed method achieves significant improvements in merging success rate, average vehicle speed, and operational stability. Specifically, the average vehicle speed increases by up to 12.89%, and the merging success rate improves by up to 9.44%. Compared with reinforcement learning methods without physical-information constraints, the proposed approach accelerates policy convergence and effectively reduces the risk of falling into local optima. [Conclusions] The proposed control strategy significantly enhances traffic efficiency in merging areas across various traffic density scenarios and provides a novel and effective solution for mixed-traffic freeway merging control.