A deep neural network physics-based reduced order model for dynamic stall

Dynamic stall is a challenging fluid dynamics phenomenon occurring during rapid transient motion of airfoils where the angle of attack exceeds the static stall angle. Understanding dynamic stall is essential for designing efficient aerodynamic systems in applications such as helicopters, wind turbines, and specific aircraft maneuvers. The highly non-linear unsteady aerodynamic phenomena involved in dynamic stall are challenging to be included in a reduced order model. In this work, a physics-based machine learning framework is developed to fast predict the aerodynamic forces and momentum for a pitching NACA0012 airfoil incurring in light and deep dynamic stall. Three deep neural network architectures of increasing complexity are investigated, ranging from two multilayer perceptrons to a convolutional neural network. The proposed model is able to robustly predict pressure and skin friction distributions over the airfoil for an entire pitching cycle. Periodic conditions are implemented to grant the physical smoothness of the solution both in space and time. Furthermore, an improved loss function incorporating physical knowledge to compute the airfoil loads is presented. An analysis of the dataset point distributions to cover the flight envelope is performed highlighting the effects of adopting low discrepancy sequences, like Latin hypercube, Sobol', and Halton, compared to random and uniform grid ones. The current model shows unprecedented performances in predicting forces and momentum in a broad range of flight conditions.