The purpose of this secondment was the study of the lift and drag coefficients for the flow around an obstacle. The main idea is to perform an Artificial Neural Network that let us to provide a prediction of the Lift and Drags coefficients, depending on the geometry.
We started considering as the obstacle, a simplified 2D geometry of a Volkswagen’s car. Optimad team provided us the geometry details and the geometrical parameter variation, for which we have computed some snapshots. In Figure 1, we show the mesh of three different geometries within the parametrical set.
Figure 1: Mesh of the 2D geometry for different geometrical parameter values.
Due to the complexity of obtain the Lift and Drags coefficients for each value of the parameter, the idea is to compute the Lift and Drags coefficients for a certain number of parameter values, and then perform an Artificial Neural Network to predict the value of the Lift and Drag coefficient for any parameter value in the parametrical set.
We started to compute some snapshots of the flow around car, considering a VMS-Smagorinsky model. The flow is clearly turbulent, so at the beginning, we considered that the Lift and Drags desired was a mean value of the Lift and Drags coefficients from 1 to 10 second. We have seen that the Lift and Drags coefficients may not be accurate as expected since the high-fidelity solution presents some unexpected recirculation at the bottom of the domain, before the obstacle. We think that, since the problem is highly turbulent, we need to improve the high-fidelity model, adding a wall law, that let us to model properly the flux near the floor, and near the walls of the obstacle.
Moreover, with Maria Strazzullo from Politecnico di Torino, we have been studying the setup of a reduced order modelling for an optimal control problem in a cavity, where the target is to reduce the fluid vorticity in the cavity by considering controlled jets in some parts of the boundary. Due to the high computational cost of solving a high-fidelity controlled problem, for the reduced order setting, we intend to consider a multi-fidelity approach, that could let us avoid solving the controlled problem for every parameter value in the offline phase.