A Reference-based Design Approach

Abstract

Design decisions which are made in the preliminary ship design phase have a significant influence on the performance and total cost of a ship. These design decisions are mostly made with very little knowledge of the ship design problem. In fact, it is the personal experience of the naval architect which plays a significant role in this phase of the design process. In the world of rapidly increasing possibilities with artificial intelligence it is hard to imagine that these decisive design choices are based on a naval architect’s personal experience. Especially when one takes into account the large capital and operational costs of ships. The development of C-Job’s Maritime Intelligence Tool (MIT) in 2019 has shown that reference data can better be exploited in this phase of the ship design process. As a result, theoretical reductions of the resistance and weight of the vessel go up to 19% and 10% respectively, using this tool. When the availability of reference data is limited, the trustworthiness of this tool cannot be guaranteed. This is especially unfavorable at the boundaries of a design space, as it is expected that novel and innovative ship design can be found here. Thus, in order to support naval architects in all regions of a design space, a solution must be found. First, research is done into design approaches in the preliminary ship design phase. In this research, naval architects of C-Job with different backgrounds were interviewed. During these interviews it became clear that time, budget and customer ambitions are important motives in this phase. As a result, a lot of ship design decisions in the preliminary ship design phase are based on the naval architect’s personal experience. The more the design is developed, the more insight is gained into the complexity of that ship design. As a result,more design decisions could be based on this gained insight, instead of the personal experience. During these interviews, challenges were identified and discussed that the naval architects face, before the Maritime Intelligence tool can be used in practice. Based on these challenges a list of tool requirements was determined and potential solutions were sought. Three solutions were found to be promising for this thesis. These were serial hybrid modelling, parallel hybrid modelling and constrained black box identification. The parallel hybrid model is chosen, primarily because of the independent operations of the data-driven sub-model (Black Box model) and the knowledge-driven sub-model (White Box model) in a parallel hybrid model. There are two requirements for parallel hybrid modelling. The first requirement is a method to estimate a ship design parameter. The second requirement is the availability of the true values of the same design parameter of reference vessels. These were both only available for the design parameter lightship weight. In the proposed parallel hybrid model, the white box model is used to estimate the lightship weight. Thereafter the black box model is trained to predict the difference between this estimation and the actual lightship weight, based on reference data. The proposed parallel hybrid model is subjected to multiple experiments to assess the performance. The R2-score and 10-fold cross validation are used to determine the performance. First the performance of the white box, black box and parallel hybrid model is discussed. Thereafter, the relation between the availability of reference data and the predicting capability is researched. The final experiment was to research the performance of the three different models in interpolation and extrapolation gaps. Based on these experiments it was concluded that for a training data set of 50 reference vessels or smaller, the parallel hybrid model was the best model. For larger training data sets, the black box and parallel hybrid model performed similarly. For interpolation and extrapolation problems the white box model should be chosen. Additionally, a method is presented to update the used white box model. As a result, it is expected that high performance scores can also be obtained without the use of artificial intelligence tools.