On surrogate methods in propeller optimisation

•We discuss several meta-models to replace simulations in optimisation design tasks.•Attempts were made to recognise and constrain certain types of cavitation.•Cavitation pattern could be reduced and transformed towards attached sheet cavities.•Kriging interpolation results are most accurate.•Trends in blade geometry variation can be predicted similarly by all met-models.

In marine propeller design, tools for propeller performance evaluation are often time consuming and automated optimisation of the blade geometry is thus not conducted. This paper discusses several response surface methods to replace the main part of the needed computations: the Kriging predictor, standard and with input improvement; the feed forward neural network; the cascade correlation neural network; and a mixed version. Optimisation assignments are performed by applying each of the surrogates to find the best solution in a multi-objective propeller design task including advanced constraints on cavitation. The final performance regarding geometry trends and degree of improvement are evaluated. Further, an approach is presented on a practical application of minimum computational effort by combining a response surface method to fill the design space and calculations in a local search method.