From Single-Parameter Reinforcement Learning to Integrated Multi-Parameter Optimization: A Data-Driven Design Framework for Airship Aerodynamics

This study presents a reinforcement learning (RL)-based framework for the aerodynamic optimization of the Lotte airship, combining mid-fidelity dynamic simulations with adaptive learning strategies. To address the complex nonlinear coupling between the hull shape and tail configuration, a staged, data-driven optimization strategy is developed. In the first stage, single-parameter RL experiments are conducted to independently analyze the aerodynamic sensitivity of key design variables. This conceptual stage isolates pure aerodynamic potential, focusing on the unconstrained optimization of the hull’s Bézier parameterized profile, alongside the individual sensitivities of empennage area, longitudinal shift, lift slope factor, and efficiency. These experiments yield a comprehensive sensitivity map, clarifying each parameter’s independent influence on the average lift-to-drag ratio (L/D¯) of the airship. In the second stage, the obtained sensitivities are utilized to structure an integrated multi-parameter optimization scenario. Crucially, this unified environment integrates the hull and tail while enforcing rigorous longitudinal trim constraints via a dynamic bisection search. This forces the RL agent to balance system-level aerodynamic recovery against inevitable trim drag penalties. The proposed framework is implemented in MATLAB R2023b using the SILCROAD airship dynamics environment and trained by the Deep Deterministic Policy Gradient (DDPG) algorithm. Results demonstrate that the initial single-parameter sensitivity extraction not only accelerates algorithmic convergence but also significantly improves the interpretability and physical validity of the final trimmed full airship configuration. This hierarchical approach establishes a systematic path from isolated parameter understanding to holistic, physics-informed aerodynamic design, offering a transferable methodology for future autonomous airship optimization.