The study focuses on addressing the challenge of designing on-board ammonia decomposition units (ADUs) for carbon-free heavy-duty vehicles. The research introduces a surrogate-enhanced optimization framework that aims to optimize ADU design and its integration with ammonia-hydrogen hybrid powertrains. This framework is a dual-phase hybrid optimization approach that combines a genetic exploration method with a Bayesian refinement technique. By utilizing a non-dominated sorting genetic algorithm for exploration and Bayesian optimization for refinement, the framework achieves significant advancements in system efficiency and ADU efficiency at reduced costs. The study includes a sequential domain decomposition strategy that is driven by genetic algorithm-based Pareto sampling, coupled with surrogate training data accumulation. Subsequently, a Gaussian process-guided refinement process is employed to merge adjacent optimal regions through covariance-based surrogate merging. The experimental validation of the framework demonstrates its effectiveness in delivering balanced system performance across key metrics. The optimized powertrain equipped with the ADU achieves remarkable results, including a system efficiency of 31.24% and an ADU efficiency of 76%, all obtained at minimal system costs. Moreover, dynamic validation confirms a significant energy reduction of over 3.5%, showcasing the practical value of the proposed framework for carbon-free heavy-duty vehicles. This research contributes to the advancement of vehicle technology by providing a solution that enhances efficiency and reduces energy consumption in heavy-duty vehicle powertrains.