AERO-ENGINE EXHAUSTS – DATA ANALYTICS TO IMPROVE THE DESIGN SYSTEM

It is expected that the on-going need for fuel burn reduction will guide the choice of ultra- high bypass ratios (BPR) and lower fan pressure ratios for the next aero-engine generation. This will lead to lower specific thrust and greater propulsive efficiency. For a given net thrust, high BPR values will increase the gross to net thrust ratio and, consequently, increase the exchange rate between fuel consumption and net thrust. This will place greater emphasis on the design and performance of the exhaust systems. However, due to the large number of geometric and aerodynamic variables, the design process can be expensive and complex. Hence, this thesis explores new methods to reduce the computation costs. In this work, the 2D axisymmetric GEMINI tool is used to analyse the aerodynamics of separate jet exhausts and generate the geometry. The CFD simulations is computed with 2D Reynolds Averaged Navier-Stokes (RANS) codes. Linear and non-linear dimensionality reduction techniques such as Principal Components Analysis (PCA), Active Subspaces (AS) and Autoencoders (AE) are utilised for feature selection and extraction. As a result, the dominant design variables are identified by the three methods, and they can reduce the design space in two dimensions. Auto-encoders are limited by the reproducibility, hyperparameters, and model loss and the direction of the metric gradient is a combination of the two latent dimensions. On the other hand, both PCA and AS can summarise the trend of the performance metric with the first latent variables, and PCA is able to obtain that with only the first one. If a threshold of the first PC is set, an 85% reduction of the design space can be achieved. Furthermore, the reduced space can be used to identify the optimum design with a new Design Space Exploration, and preliminary design stage costs can be cut by projecting new designs in the latent dimensions.