Artificial Neural Networks for surrogating high-fidelity simulations of turbulent flows

In many problems represented by a mathematical model, the knowledge of the solution at different values of specific parameters is necessary and one way to proceed is to compute a high-fidelity simulation; however, since a high-fidelity simulation is very computational demanding, this is often too time consuming. In these situations, surrogate models can be very useful. The aim of this work is to show viability of Machine Learning as a surrogate model in a specific setting. The objective is to obtain a reconstruction of a high-fidelity solution of a fluid dynamics problem, depending on both a specific Reynolds number and a specific time, without executing the relative simulation, with the aim of accelerating the case where thousands of simulations are required. In fact, since a Machine Learning based approach will require a conspicuous number of simulations for the training, a trade-off between the number of high-fidelity solutions actually needed and those needed for the training has to be taken into account.