Prediction of boundary-layer instability in supersonic flow with roughness and shock using machine learning

Numerical simulations of a Ma=4.8 flat-plate boundary layer with discrete roughness element shows an increase of convective disturbance amplification caused by the roughness. Such increase has important implications for the transition location, but cannot be accommodated by current transition prediction methods. We present a new tool to predict the effect of two-dimensional roughness on boundary-layer perturbations, which can account for roughness geometry and location. This tool is based on machine learning techniques such as Gaussian process regression and neural networks. Training data is derived from datasets obtained using O(100) simulations. Prior to training, time-dependent simulation data is postprocessed for data reduction such that for instance only the perturbation's Fourier amplitude is used for training. The high computational cost of generating the simulation dataset is counterbalanced by the speed of prediction, making the tool suitable for applications related to the operation of (e.g. re-entry) vehicles rather than their design. A main benefit of the tool is that it provides transferability, using the existing data to predict cases that have not been simulated (different roughness geometry, location).