Towards ANN Based Digital Twins of Ship Propulsion Systems

In shipping, the choice of the right routing and speed offers the opportunity to act more sustainably from both an economic and an ecological point of view. Reinforcement Learning (RL) agents could be suitable for this task. However, as a learning environment the agents require the most detailed, accurate, and fast representation of reality possible. This paper describes approaches to build such an environment using neural networks (NN) trained with both simulation and real-world data. It is shown that simple feed-forward networks can reproduce data created by 1D flow simulation sufficiently accurate. By examining the differences between simulation and measured data, the simulation could be improved. Since NNs trained with vessel data only are limited in their generality, approximating nets trained with simulation data to vessel data using Transfer Learning (TL) was investigated. Initial results for this approach show good quantitative results, but only in the data region where vessel and simulation data overlap. The paper provides an overview of the necessary steps towards Digital Twins for ship propulsion systems.