Proposal of a Reusable Deep Reinforcement Learning AI–Based Structural Optimization Method for the Ship’s Midship Section

This study proposes an efficient methodology for lightweight design of a ship hull’s midsection, taking into account variations in principal dimensions, by leveraging the agent‐reuse capability of a Double Deep Q Network (DDQN) based deep reinforcement learning framework. First, a baseline ship model is optimized using DDQN, with plate thickness and stiffener geometry as design variables and maximum von Mises stress as a constraint, thereby training an agent that encapsulates the optimal search strategy. The trained agent is then directly applied to multiple design cases featuring different principal dimensions. Numerical experiments demonstrate that reusing the trained agent reduces the number of required finite element analysis (FEA) evaluations by an average of 76.4% while maintaining or improving structural performance. These results indicate that the proposed approach enables rapid, practical lightweight structural optimization of ship hulls within limited design timeframes.