Documento completo qui
Ludovico Petrucci
Formula One is the pinnacle of Motorsport: every year each Team that participates to the competition builds and develops its own car, trying to push to the limit its performance on track. A lot of effort is dedicated to the aerodynamic development: Computational Fluid Dynamics (CFD) is one of the most used tool to develop the car. Each development is tested in wind tunnel and track to check for correlation of the expected performances. Very often Particle Image Velocimetry (PIV) measurements are captured in wind tunnel sessions to get two dimensional snapshots of the velocity field around the model. Despite it is very useful to be able to visualise the velocity filed in the key areas of the car with a certain accuracy, no conclusions can be done on the pressure field. In literature they are available many numerical methods that allow to get an approximation of the pressure given the velocity components, mainly solving the momentum equations of the Navier-Stokes equations. Unfortunately, most of the times the PIV snapshots capture highly three dimensional phenomena and it is not possible to reconstruct the pressure due to the lack of information about the velocity component normal to the PIV plane, for which it is not possible to compute the spatial derivative.
The present thesis will explore the capability of Convolutional Neural Networks (CNNs), a class of Supervised Neural Networks, to learn predicting the pressure flow field from the velocity one in a two dimensional slice. Due to the huge amount of CFD simulations that every week is performed, the training and the development of the network will use two dimensional slices of the velocity and pressure field around the car from these data, trying to replicate the PIV measurement snapshots obtained in wind tunnel. For what it concerns the neural networks, they have been tested a class of architectures called Autoencoders, developing each part of the network to improve its learning capability. The approach is a standard Supervised Learning where two dimensional velocity slices are input to the network and prediction for the pressure field at the same slices is obtained.