Comparison of Machine Learning Models Based on Different Ship Hull Form Representations for Hydrodynamic Performance Prediction

Accurate prediction of ship hydrodynamic performance is essential for hull form optimization during the early design stage. This study investigates two approaches for incorporating ship hull geometry into machine learning-based surrogate models using offset data. The first approach represents hull geometry as an RGB image and applies a two-dimensional convolutional neural network (2D CNN), while the second directly utilizes offset curves with a one-dimensional CNN and a temporal convolutional network (1D CNN + TCN) to capture longitudinal hull form variations. Models for predicting the residual resistance coefficient, effective wake fraction, and thrust deduction factor were trained using approximately 440 model test dataset from the Korea Research Institute of Ships and Ocean Engineering (KRISO). Their performance was compared with a baseline multilayer perceptron (MLP) model using only principal ship dimensions. The results demonstrate that incorporating hull geometry significantly improves prediction accuracy over the MLP model. Although the average prediction performance of the 2D CNN and 1D CNN + TCN models was similar, the 1D CNN + TCN model achieved higher hull form sensitivity in propulsion performance prediction. These results indicate that offset-based hull representation provides an effective input for machine learning surrogate models and that TCN-based architectures are advantageous for capturing longitudinal hull form variations.