Data-driven parameter and model order reduction for industrial optimisation problems with applications in naval engineering

In this work we study data-driven reduced order models with a specific focus on reduction in parameter space to fight the curse of dimensionality, especially for functions with low-intrinsic structure, in the context of digital twins. To this end we proposed two different methods to improve the accuracy of responce surfaces built using the Active Subspaces (AS): a kernel-based approach which maps the inputs onto an higher dimensional space before applying AS, and a local approach in which a clustering induced by the presence of a global active subspace is exploited to construct localized regressors. We also used AS within a multi-fidelity nonlinear autoregressive scheme to reduced the approximation error of high-dimensional scalar function using only high-fidelity data. This multi-fidelity approach has also been integrated within a non-intrusive Proper Oorthogonal Decomposition (POD) based framework in which every modal coefficient is reconstructed with a greater precision. Moving to optimization algorithms we devised an extension of the classical genetic algorithm exploiting AS to accelerate the convergence, especially for highdimensional optimization problems. We applied different combinations of such methods in a diverse range of engineering problems such as structural optimization of cruise ships, shape optimization of a combatant hull and a NACA airfoil profile, and the prediction of hydroacoustic noises. A specific attention has been devoted to the naval engineering applications and many of the methodological advances in this work have been inspired by them. This work has been conducted within the framework of the IRONTH project, an industrial Ph.D. grant financed by Fincantieri S.p.A.