Abstract
Automatic optimisers can play a vital role in the design and development of engineering systems and processes. However, a lack of available data to guide the search can result in the global optimum solution never being found. Surrogate models can be used to address this lack of data and allow more of the design space to be explored, as well as provide an overall computational saving. In this thesis I have developed two novel long-term prediction methods that investigate the use of ensembles of surrogates to perform predictions of aerodynamic data. The models are built using intermediate computational fluid dynamic convergence data. The first method relies on a gradient based learning algorithm to optimise the base learners and the second utilises a hybrid multi-objective evolutionary algorithm. Different selection schemes are investigated to improve the prediction performance and the accuracy of the ensembles are compared to the converged data, as well as to the delta change between flow conditions. Three challenging real world aerodynamic data sets have been used to test the developed algorithms and insights into aerodynamic performance has been gained through analysis of the computational fluid dynamic convergence histories. The trends of the design space can be maintained, as well as achieving suitable overall prediction accuracy. Selecting a subset improves ensemble performance, but no selection method is superior to any others. The hybrid multi-objective evolutionary algorithm approach is also tested on two standard time series prediction tasks and the results presented are competitive with others reported in the literature. In addition, a novel technique that improves a parameter based surrogates learning through the transfer of additional information is also investigated to address the lack of data. Transfer learning has an initial impact on the learning rate of the surrogate, but negative transfer is observed with increasing numbers of epochs. Using the data available for the low dimensional problems, it is shown that the convergence prediction results are comparable to those from the parameter based surrogate. Therefore, the convergence prediction method could be used as a surrogate and form part of an aerodynamic optimisation task. However, there are a number of open questions that need to be addressed, including what is the best use of the surrogate during the search?