Global Surrogate Modelling of Gas Turbine Aerodynamic Performance

The compressor and turbine aerodynamic performance of a gas turbine engine depends on complex three-dimensional unsteady interactions between static and rotating blades. Each component is made up of a series of rotor and stator blades, whose overall performance is dependent on blade geometry, manufacturing and operating conditions. This in turn leads to a large number of independent parameters which control compressor and turbine aerodynamic performance. Conducting high fidelity parametric analysis within such a large domain and constraint space is not feasible, even when using state-of-the-art computational resources. Surrogate modelling techniques have been widely used within the industry to overcome such complex, multi-dimensional analysis problems. That is, the compressor and turbine have been broken down to single blade stages, each blade stage and operating condition represented parametrically and surrogate models developed to predict performance of each stage. A basic surrogate model is then used to predict the performance of the compressor or turbine system. The surrogate models that are currently being used are based on limited experimental data and are open to interpretation. Also, they seem to reflect outdated technology and the construction and enhancement of these models based on experiment are very time consuming and costly. This research combines the latest Design and Analysis of Computer Experiment (DACE), Computational Fluid Dynamics (CFD) and surrogate modelling techniques to predict the performance of compressors and turbines. A number of different surrogate modelling techniques are evaluated and optimised for accuracy and efficiency. The surrogate models are further enhanced using parametric space exploration and subset regression analysis. The surrogate models are then validated with CFD generated compressor and turbine maps. Data mining or knowledge extraction studies are presented to mine the large number of simulations not only for prediction but also to gain an understanding of the underlying processes that drive compressor and turbine performance. Advances to the current state-of-the-art in a number of areas are also presented. The first is the development of a blade geometry mapping techniques that allows the use of DACE space filling designs to irregularly shaped design envelopes. The second is the development and refinement of adaptive sampling techniques using Gaussian process regression also known as Kriging. This research shows that it is possible to create accurate surrogate models of turbine blade aerodynamic performance for up to fifteen dimensions. Increasing the number of dimensions results in issues associated with the `curse-of-dimensionality'. Feature extraction and data mining techniques used in assessing the importance of blade performance parameters point to a similar weighting of the blade feature importance, however, the study also showed that it is not possible to eliminate lower ranked features without significantly degrading the surrogate model accuracy. Tackling the `curse-of-dimensionality' using advanced adaptive sampling and multi-fidelity approaches has shown promising results. In a limited study, it was shown that an accurate surrogate model of a seven-dimensional parametric model could be constructed with less then ten high-fidelity equivalent simulations