Optimisation techniques are increasingly being adopted in industrial design processes, where more and more components are designed through significant use of computational analysis. This is the logic consequence of the increased capacity of numerical modelling techniques to accurately predict the behaviour and performance of the components. Through optimisation techniques, these numerical models are automatically and systematically explored to modify designs such that their performance improves, quite often beating designs developed by skilled engineers while obtained in a much shorter timeframe.
One of the most promising techniques to improve designs is based on sensitivity analysis, which tells how much impact a certain design change will have on performance. These sensitivities can be computed efficiently using the adjoint approach, for which the computational cost is independent from the number of design variables, and thus allows to explore a very rich design space. Through the adjoint approach, the optimisation of very complex machines such as full gas turbines is within reach of the current computational capacity. This would have a significant environmental impact as the efficiency of many energy conversion systems on which we rely today can be substantially improved through advanced use of optimisation techniques. The current situation is however far from this prospect, as still major advancements are needed in adjoint methods.
This project has tackled two major shortcomings, which are perceived as the major bottleneck of application of adjoint methods in industrial use. The first issue relates to the parametrisation of the shape optimisation problem. The current practice is to optimise a discrete shape model, such that the optimal shape is represented by a cloud of points rather than a collection of smooth analytically described surfaces, which is the current industrial practice. Hence, a post-processing step is required to fit the point cloud by smooth surfaces, impairing optimality. In this project, the shape has been considered from the start as described by a geometric model which is guaranteed to have smooth surfaces.
A second issue with respect to current shape optimisation practices is that only single disciplines are considered at once. In many applications, only the fluid problem is considered, and the shape is optimised to reach better aerodynamic performance. In general, the resulting shape will not meet structural requirements and hence needs to be reshaped, adding significantly to the design effort while failing to find a compromise solution. In this work, different disciplines are considered simultaneously.
