Machine Learning-Guided Population Initialization for High-Dimensional Structural Optimization of Very Large Crude Oil Carriers

This study presents a two-step optimization framework that combines a machine learning (ML)-based prediction model with metaheuristic optimization for the structural design of very large crude oil carriers. The framework aims to minimize the midship cross-sectional area while satisfying the Common Structural Rules (CSR) and associated strength constraints. In the proposed approach, an ML-based prediction model generates a well-informed initial population, which is subsequently refined through metaheuristic optimization. The ML model is trained using 20,000 datasets derived from accumulated engineering design records. The effectiveness of the framework is demonstrated through a high-dimensional optimization problem involving more than 200 design variables. Numerical results show that conventional random or uniform initialization often leads to poor convergence, whereas ML-based initialization significantly improves convergence behavior and computational efficiency. The proposed method also exhibits robust performance for extrapolated design cases beyond the training data, indicating its potential for practical large-scale ship structural design.