For many shipbuilders, the majority of a vessel's lifecycle greenhouse gas emissions occurs during its operational phase, commonly referred to as downstream emissions. For Feadship, this challenge is particularly pronounced, with downstream emissions accounting for 94% of a superyacht’s carbon footprint. Addressing this majority requires accurate and efficient methods to predict propulsion energy use during the design stage - a task hindered by the limitations of existing prediction methods, which are either time-intensive or prone to significant errors. This is affecting the design process in two ways: an increased risk of incorrectly proportioned energy and power systems, and limited exploration of design space. Data-driven methods, based on machine learning algorithms, have been proposed in the literature. However, these methods expose two key gaps in the literature: their performance under extrapolation conditions and their limitations when applied to small datasets.

This thesis addresses these challenges by developing hybrid modeling approaches that combine physical insights through a physical model with data-driven techniques, enabling improved predictive accuracy under extrapolation scenarios and with small datasets. Three modeling approaches are tested - physical models (PMs), data-driven models (DDMs), and a combination of the two that forms hybrid models (HMs) - with a shared prediction target, namely calm-water ship resistance. Four datasets were assessed: Dataset 1 (CFD resistance), Dataset 2 (CFD power), Dataset 3 (towing tank tests), and Dataset 4 (speed-power trials), resulting in the selection of computational fluid dynamics (CFD) resistance data as the basis for training the models.

Instead of directly learning from the CFD resistance data, it appears more effective when the data-driven model learns to apply corrections to the output of a PM. Where traditionally these corrections were based on the naval architect's experience, they are now driven by data, offering fast and accurate alternative to existing methods. This philosophy is embodied in this study through a newly developed parallel HM, which achieves superior performance by learning how to apply these corrections to the PM's output automatically. During interpolation, the new HM demonstrates a mean average percentage error (MAPE) of 3.8%, outperforming the best available PM (6.7%) and the best DDM (8.9%). For extrapolation, standalone DDMs, including the best interpolator, failed dramatically, with MAPE values exceeding 180% in some cases. The new HM maintains average errors within 12% across scenarios. And with less data, the new HM consistently outperformed the best DDM, with its competitive edge most pronounced at low data availability (10% of CFD observations). By advancing these methodologies, the study not only enhances early-stage design confidence but also contributes to future steps towards automated design optimization.