Physics Informed Neural Network model for wind field prediction in urban spaces for small Unmanned Aerial Systems.

In recent years there has been significant interest in using Unmanned Aerial Systems for various applications in urban spaces, including disaster management, law enforcement, delivery/catering services, and Advanced Air Mobility. However, Unmanned Aerial Systems are highly sensitive to the abrupt wind patterns generated in urban spaces due to obstacles like buildings and other structures. It is especially difficult for Small Unmanned Aerial Vehicles owing to their lightweight and smaller overall structure. Safe Wind-Aware navigation is thus an essential part of small Unmanned Aerial Systems operation and deployment. Although Computational Fluid Dynamics solvers could provide accurate solutions, they are computationally expensive and cannot be used for real-time or close to real-time wind predictions. Non-intrusive Data-driven, Reduced Order Models could offer a viable alternative for wind-filed predictions by relying on offline training using these high-fidelity solutions. However, they are not easily generalizable for different flow conditions since they depend highly on the training data.

In contrast, Physics Informed Neural Networks (PINN) incorporate known physics into training the reduced order model by using loss functions based on governing equations. Furthermore, PINNs also enables easy data assimilation from sparse Spatio-temporal observations into the model. In this work, we aim to utilize PINNs to generate a generalizable reduced order model for wind-field predictions in a typical urban environment, using limited high-fidelity Large Eddy Simulation (LES) data.