The overall aim of RefMap is to promote a new approach for a greener aviation using AI for the trajectory optimisation. Although RefMap focuses on both commercial aircrafts and UAVs, only the latter cases are presented. The neural network with Deep Reinforcement Learning (DRL) is employed to track and choose the drone trajectory in a simplified urban environment simulated using the high-fidelity LES. 

Here we follow the entire development from the setup of the simulations used to identify and physically understand the motivations behind the setup of the DRL. The analysis of the turbulent structures around the buildings, here modelled as wall-mounted, square cylinders shows that the most hazardous region for a drone corresponds to the detached-flow region that envelops the rear obstacle. 

This region also displays the highest stress gradient for which the NN has been trained to avoid. The navigation of the drone in the 2D slice is a significant example of the potentiality of this tool as, in real time, as the path proposed avoids the hazardous region until to reach the target.