Variable Fidelity and Surrogate Modelling in Simulation-Based Design

摘要

With the aim of increasing the confidence of ship designers in the Simulation Based Design, allowing the integrated (and corrected) use of some of the best standard methodology into more complex design optimization processes, this paper analyze the integration of existing, relatively simple but reliable software (e.g. potential flow solvers) with more advanced, but more expensive, high-fidelity software (e.g. RANS solvers) into a non-heuristic procedure. The goal is to demonstrate the possibility of correcting the low-fidelity to a useful approximation of the high-fidelity results. To pursue this aim we explore the Variable Fidelity / Variable Physics approach. The overall speed-up of the optimization procedure is increased by the combined use of surrogate approximations, i.e. algebraic, inexpensive models based on a training subset computed with the high-fidelity solver and able to mimic their response. These approximations are used to reduce the overall optimization time, while their accuracy is improved during the optimization process. Numerical examples of growing complexity are presented, including a sailing yacht design problem.