Shapley-Additive-Explanation-Based Stall Flutter Control: from Neural Interpretation to Explainable Policy Design

In recent decades, the widespread success of deep learning in aerodynamics has intensified the need for interpretable models to ensure trustworthy engineering deployment. This study develops a rule-based, transparent control policy derived from Shapley additive explanations (SHAP) to replace a trained deep reinforcement learning (DRL) controller for stall flutter suppression via trailing-edge morphing. SHAP analysis of the DRL controller reveals two critical control intervals within the pitching cycle and identifies compensatory–cooperative interactions between pitch-related and morph-related inputs. Additionally, the contributions of the morphing input exhibit piecewise linear dependence, while pitch-related inputs follow sinusoidal patterns. Based on the interpreted patterns, an explainable control policy is constructed and implemented in fluid–structure–control simulations. The policy is evaluated under extrapolative conditions of airspeed ([Formula: see text], corresponding to [Formula: see text]), structural damping ratio (4–6%), stiffness ([Formula: see text]), airfoil thickness (0–14%), and airfoil camber (0–2%). The simulation results demonstrate that the proposed explainable control policy achieves complete flutter suppression or a higher suppression rate than the DRL controller across all tested conditions. These findings provide a practical pathway from model interpretation to interpretable control law design in complex aeroelastic problems.