Data-knowledge-driven dynamic stall modeling guided by stall patterns and semi-empirical model

Dynamic stall often causes unsteady loads and negatively affects the aerodynamic performance of the aircraft. Thus, accurate modeling of dynamic stalls is crucial for aircraft design. With the development of machine learning, the existing data-driven dynamic stall models always rely on extensive, costly training data but lack physical knowledge, which limits their generalizability and interpretability. Therefore, this study proposes a data-knowledge-driven dynamic stall modeling procedure. First, by exploring the aerodynamic damping and the evolution of the moment coefficient, three distinct stall patterns are identified. A transitional stall state, which significantly differs from both deep stall and light stall, is proposed to assist neural network modeling. Subsequently, a deep neural network with P-based stall degree classification and force component is developed, which integrates the proposed stall patterns and the Leishman-Beddoes dynamic stall model. This model provides a unified approach to predict dynamic stall aerodynamics across different degrees of dynamic stall. Compared to a purely data-driven neural network, incorporating expert knowledge improved the generalization accuracy by 50%. Moreover, physical insights significantly reduce the reliance on high-precision training data of the neural network.