The objective of this thesis is to explore the possibilities of utilising machine learning to predict the axial velocities in a nominal wake field of a single-propeller vessel, given a database with only basic parameters describing hull geometries, such as length, beam, and block coefficient. The nominal wake field is of primary importance for the propeller design process, as it represents the influence of the hull on the flow field that a propeller encounters. Furthermore, propeller-hull interactions can be derived from it.

Five different machine learning models have been developed, trained and evaluated. Visual inspection of predictions made by all five models revealed that none of the models has successfully captured the underlying physical phenomena that drive the wake field: all predictions show highly generalised wake fields. This finding was supported by a feature
importance study, which showed that the features that supposedly contribute the most to understanding the hydrodynamic phenomena that drive the wake field do not have the highest relative importance.

The limited performance of the developed models is primarily attributed to two factors: the limited dataset size and a lack of feature informativeness, meaning that the used features contain insufficient information about the problem to let a model effectively capture the underlying physical phenomena. To a lesser extent, computational limitations also play a role.

The increased focus on the reduction of emissions in the maritime sector has been the driving force behind many technological advancements related to marine engineering. Among those is the invention of the gate rudder system, which, while still at an early stage of development, promises improvements in propulsion efficiency and ship manoeuvrability. This thesis investigates the concept from the perspective of time domain simulation in order to pave the way towards identification of the working principle of the gate rudder power savings.

First, an extensive literature review is presented. A number of subjects related to the project was studied, starting from the principles of ship manoeuvrability, which comprises of ship hull hydrodynamics, hydrodynamics of control surfaces and regulations on manoeuvrability of ships. Afterwards, the current state of research into the gate rudder concept is investigated. Then, techniques of modelling of ship motion are laid out, with a focus on mathematical modelling of ship motion and rudders. Machine learning is identified as a potentially useful tool in the research, therefore a dedicated section consists of a brief introduction of the concept, some example regression algorithms and practical techniques utilized in conjunction with their use. Finally, the concept of time domain simulation in the context of maritime engineering is investigated and the options of simulator software available for the project are reviewed.

Initial decisions made for the execution phase of the project are described, including the choice of Manwav as the preferred time domain simulation software and S175 container ship as the benchmark vessel to carry out model tests on. An effort to tune the ship's model in Manwav to better match its performance found in literature is also described at the end of the chapter.

A physics-based gate rudder model is adapted into Manwav from a scientific paper. It is based on a mathematical model of a rudder derived from the MMG methodology, and the goal of its implementation was to pave the way in terms of code structure in Manwav and to serve as a reference point to the focus of the project, which is the data-driven model.

The data-driven gate rudder model uses a machine learning regression algorithm to build a set of rudder force coefficient predictors using a CFD dataset containing the values of force coefficients within a range of rudder and drift angles of a vessel equipped with a gate rudder system. A well-behaving model is successfully built using support vector regression and deployed in Manwav.

Both gate rudder models are put to test in the case study, where three ships with different rudder models are seen performing a set of benchmark manoeuvres, comprising of turning circles and zig-zag manoeuvres at different vessel speeds and rudder deflection settings. While the physics based model was found to be generating abnormal vessel behaviour, the data-driven model performed as expected, providing some insight into a potential impact a gate rudder system might have on a real ship's manoeuvrability and power consumption.

The thesis is concluded with a discussion on the data driven model's applications, along with limitations that prevent it from reaching its true potential. What follows is an extensive list of recommendations that go into detail on how the project could be improved in the future. The overall conclusion from the project is that a relatively limited amount of CFD data was successfully used to produce a rudder model that provides simulation possibilities otherwise beyond the reach of high fidelity methods due to scenario complexity and computational cost associated with them. With not a lot of additional effort, the model will make it possible to test whether or not gate rudder power savings reported in literature come from sailing in realistic weather conditions.