Learning-augmented optimization and control of long-haul mobility propulsion systems

Decarbonizing the long-haul transportation sector is a critical global challenge. This paper introduces a hierarchical, learning-augmented framework for co-designing an ammonia-hydrogen hybrid electric powertrain as a viable carbon-free propulsion solution. At the core of the proposed framework is the tight coupling between online control and offline design. In the online supervisory control layer, a deep reinforcement learning (DRL) agent is trained to make real-time decisions on power split and on-board hydrogen production. In the offline propulsion system optimization layer, a DRL-augmented adaptive non-dominated sorting genetic algorithm (RL-ANSGA) is employed to solve the multi-objective component sizing problem, with each candidate design evaluated in high-fidelity co-simulation under a fixed, pre-trained DRL energy management policy. Results demonstrate that the optimized ammonia-hydrogen vehicle outperforms conventional diesel and diesel-hybrid counterparts in energy efficiency and well-to-wheel carbon emissions.