Efficient yet realistic ship routing is critical for reducing fuel consumption and greenhouse-gas emissions. However, conventional weather-routing algorithms often produce mathematically optimal routes that conflict with the paths mariners use. This study presents a hybrid approach that constrains physics-based weather routing within an AISderived maritime traffic network (MTN) built from one year of global Automatic Identification System data. The MTN represents common sea lanes as a graph of approximately 10,956 waypoints (nodes) and 17,561 directed edges. Using this network, an optimal low-emission route is computed via graph search and then compared against both a traditional unconstrained route and an advanced weather-routing model (VISIR-2). In a May transitionseason case (Busan–Singapore voyage), the AIS-constrained route reduced fuel consumption and CO₂ emissions by about 1.9% relative to the fastest feasible route, while closely following real traffic corridors (over 90% overlap with actual 2024 AIS tracks). While this 1.9% saving does not reach the high-end potential of an unconstrained, state-of-the-art model like VISIR-2 (which can demonstrate double-digit savings in certain conditions), it is achieved with an increase in transit time of ~6.5 h (≈3.2%). This represents a crucial trade-off, prioritizing operational realism and adherence to real-world traffic corridors over maximum theoretical efficiency.

Efficient and safe maritime navigation in complex and congested coastal regions requires advanced route optimization methods that surpass the limitations of traditional shortest-path algorithms. This study applies Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) reinforcement learning (RL) algorithms to generate and refine optimal ship routes in East Asian waters, focusing on passages from Shanghai to Busan and Ulsan to Daesan. Operating within a grid-based representation of the marine environment and considering constraints such as restricted areas and Traffic Separation Schemes (TSS), both DQN and PPO learn policies prioritizing safety and operational efficiency. Comparative analyses with actual vessel routes demonstrate that RL-based methods yield shorter and safer paths. Among these methods, PPO outperforms DQN, providing more stable and coherent routes. Post-processing with the Douglas-Peucker (DP) algorithm further simplifies the paths for practical navigational use. The findings underscore the potential of RL in enhancing navigational safety, reducing travel distance, and advancing autonomous ship navigation technologies.