A deep multi-agent reinforcement learning framework for climate-aware aircraft trajectory planning considering air traffic complexity

The aviation sector is facing increasing pressure to reduce its environmental footprint, necessitating effective and immediate solutions. Flight planning has emerged as a short-term measure to mitigate aviation-induced climate effects. However, existing literature has primarily focused on individual trajectory optimization, overlooking interactions between flights and their implications for the air traffic management (ATM) system. This study addresses this gap by introducing a two-stage framework to reconcile climate-optimal flight trajectories with the operational requirements of the ATM system. First, individual aircraft trajectories are optimized to mitigate their climate impact. Subsequently, a distributed multi-agent reinforcement learning (MARL) method is proposed to compensate for the adverse effects these trajectories may impose on the ATM system manageability by mitigating air traffic complexity. The proposed approach leverages the multi-agent proximal policy optimization algorithm and adapts it to address the scalability issue in real traffic scenarios. The effectiveness of the proposed framework is validated through a case study using real traffic data over European airspace. Experimental results show that the proposed approach balances environmental objectives with air traffic manageability, achieving up to 24 % mitigation in climate impact while maintaining air traffic complexity at levels comparable to standard business-as-usual trajectories, with a 1.8 % increase in operational costs.