There exist different methods of deriving mathematical models for industrial processes. These methods can be grounded in CFD simulations or differential equations of motion of the given process. Mathematical models play a key role in designing controllers that in turn would be able to make the processes behave in the desired way.

The accuracy of mathematical models is often good enough for being used in control design. However, for cases where the modelling error and environmental disturbances are too large, the derived models may not be suitable. This can especially be the case in processes where a big emphasis is set on safety.

In this thesis, a method called Safe Learning is implemented to improve an existing first principle model of cycloidal propellers. By applying the model to the real process and comparing it with the behaviour of the real plant, the uncertainty underlying the system is modelled. This model can be learned on-line and applied. As experiments showed, the Safe Learning algorithm did improve the model of the vessel and its propulsion system and also improved the performance of an MPC that was built.