Asymmetric leading-edge design mechanisms and multi-agent reinforcement learning optimization for wide-incidence turbine blades

Maintaining low aerodynamic loss over a wide incidence range remains a major challenge for variable-speed power-turbine rotor blades. This study investigates the section-level aerodynamics of a representative transonic rotor-blade cascade and proposes a physics-informed asymmetric leading-edge (LE) design strategy that combines mechanism analysis, surrogate modeling, and optimization. Wind-tunnel experiments and Reynolds-averaged Navier–Stokes simulations are first used to characterize the incidence-dependent loss mechanisms and to establish a validated aerodynamic database. The results show that positive incidence is primarily associated with suction-side LE separation and its downstream mixing loss, whereas severe negative incidence is governed by pressure-side LE flow misalignment and separation-related penalties. Guided by these mechanisms, an asymmetric LE parameterization is developed to decouple the suction-side and pressure-side geometries, enabling separation control at positive incidence while limiting frictional penalties and adverse pressure-side effects at negative incidence. To efficiently explore the resulting design space, a Graph Attention Network (GAT) surrogate is coupled with multi-agent reinforcement learning (MARL), in which three cooperative agents are assigned to −50°, 0°, and +30° incidence conditions. The optimized design is further validated experimentally. Compared with the baseline, the optimized asymmetric blade reduces the total pressure loss coefficient by 29.29% at −50°, 13.73% at 0°, and 46.64% at +30°, demonstrating substantial aerodynamic performance gains across a wide range of incidence. These results establish an experimentally supported and physics-grounded framework for rapid off-design LE design of turbine rotor blade profiles under wide-incidence conditions.