Machine Learning Applications To Turbulence

Sammanfattning: By using a fully convolutional neural network, accurate predictions are made of turbulent flows in an open channel. The flow fields were predicted at two different heights y+ = 15,50 in the wall-normal direction. The effect of different input parameters of the neural network have been studied by generating two neural network models. The first model takes the shear stress in the span- and streamwise directions aswell as the pressure, while the second model only uses the shear stresses. The statistics of the models were calculated and the pressure was determined to be an important parameter for predictions at the distance y+=50 from the wall. Closer to the wall, at y+ = 15, the pressure had a smaller effect on the error, and in the spanwise direction the prediction made by the two-input model had a lower error.