Data-Driven Reduced Order Modelling and Transfer Learning for the Structural Optimization of Cruise Ships

In this work we propose an optimization pipeline for the reduction of the structural mass of passenger cruise ships in the preliminary design phase. In simulation-driven design, the large cost of high-fidelity evaluations can be mitigated by leveraging data-driven reduced order models, exploring the design space via multi- and single-objective optimization of the surrogates and only validating the most promising candidates. To tackle the complexity in the formulation of effective optimization problems, we propose a reparameterization procedure that adapts the decision variables to the emerging structural behavior of the ship evaluated by the surrogates. Starting with a coarse problem and iterating optimization and refinement, we show that the final configuration outperforms parameterizations of similar complexity produced by expert designers. We then use a transfer-learning approach to design a one-shot reparameterization procedure, which leverages the results from past optimizations to refine the problem without the need for surrogates. To enhance the optimization via metaheuristic algorithms, we propose a population transfer procedure that accelerates the exploration of promising areas of the design space, leveraging the data from past optimizations. The resulting pipeline is extensively evaluated on multiple test cases of industrial interest, both on simplified midship sections and full-scale ships. Finally, to facilitate the introduction of life-cycle quantities in the design phase, we develop an open-source library for the implementation and evaluation of digital twins based on probabilistic graphical models, and showcase its workflow on a simplified structural health monitoring application. This work has been conducted within the framework of the SHOPMEPA project, an industrial Ph.D. grant co-financed by Fincantieri S.p.A. and the NextGenerationEU initiative.