Designing floating wind turbine (FWT) substructures involves navigating high-dimensional, nonlinear, and tightly constrained design spaces, where conventional manual or rule-based methods often struggle with scalability and efficiency. This study proposes a novel deep reinforcement learning (DRL)-based optimization framework that autonomously explores and optimizes preliminary substructure designs. Leveraging the Soft Actor-Critic (SAC) algorithm, the agent continuously interacts with a coupled simulation environment to identify high-performing configurations through trial-and-error learning, without explicit gradient information or exhaustive search. The environment integrates geometry parameterization, hydrodynamic coefficient computation via potential flow theory, and fully coupled dynamic response simulations using established tools such as Capytaine and OpenFAST. By interpreting platform performance indicators, including motions, structural responses, and stability metrics, as reward signals, the agent learns to balance trade-offs across conflicting objectives under physical constraints. Results demonstrate that the proposed approach achieves efficient convergence in complex continuous action spaces and enables fully automated design iterations with minimal human input. This work offers a promising direction for intelligent early-stage design of FWT foundations, especially in scenarios requiring adaptability, generalization, and multi-objective performance.
Chen et al. (Fri,) studied this question.