Flow-field reconstruction for dynamic stall of a NACA0012 airfoil using sequence-conditioned PINNs

Abstract

This study investigates spatiotemporal flow-field reconstruction for the unsteady pitching dynamic stall of a NACA0012 airfoil using a sequence-based physics-informed neural network (PINN). During dynamic stall, the formation and shedding of the leading-edge vortex, wake transport, and aerodynamic hysteresis all exhibit strong history-dependent behavior. Therefore, using only instantaneous spatial coordinates and the current angle of attack is insufficient to accurately represent the unsteady flow evolution. To address this limitation, a temporal modeling branch is introduced into the PINN framework, and two models, namely PINN-LSTM and PINN-Transformer, are constructed. The historical flow-field sequence, spatial coordinates, and airfoil kinematic features are used as inputs to predict the velocity, pressure, and vorticity fields at the next time step. The results demonstrate the feasibility of sequence-based PINNs for spatiotemporal reconstruction of dynamic stall flows and show that LSTM and Transformer models have different levels of applicability under different stall intensities and physical evaluation metrics.