Explainable deep reinforcement learning reveals energy-efficient control strategies for turbulent drag reduction

Researchers developed a method combining multi-agent deep reinforcement learning with explainable AI techniques to optimize drag reduction in turbulent flows, achieving 34.44% drag reduction with only 0.43% energy input—significantly outperforming traditional opposition control methods.

This research represents a significant advancement in applying machine learning to fluid dynamics optimization, demonstrating how explainable AI can guide reinforcement learning toward physically intuitive and energy-efficient solutions. The study trained multiple agents using SHAP (SHapley Additive exPlanations) attributions from neural networks predicting fluid dynamics parameters, rather than directly optimizing for wall-shear stress. The most successful approach combined skin-friction coefficient and wall-pressure fluctuation predictions, revealing that the learned policy activates control primarily at near-zero wall pressure—aligning with known turbulent structure dynamics.

The practical significance lies in the dramatic efficiency gains: compared to opposition control baselines, the method achieves 49.41% better drag reduction while using 48.52% less energy. Against direct wall-shear-stress optimization, it reduces actuation costs from 5.90% to 0.43% normalized input power. This demonstrates that explainable AI isn't merely interpretable—it produces superior performance by constraining learning toward physically meaningful strategies.

Beyond fluid dynamics, this work establishes a template for leveraging XAI in reinforcement learning across engineering domains. By using neural network attributions as reward signals, researchers create a feedback loop that encourages agents to discover control strategies that correspond to dominant physical mechanisms. The finding that optimal policies operate on timescales matching near-wall turbulent structure lifetimes validates this approach.

Future applications could extend this methodology to aerodynamic design optimization, pipeline efficiency, and marine vessel drag reduction—industries where even marginal efficiency improvements generate substantial economic returns. The technique also suggests broader potential in robotics and autonomous systems where human-interpretable decision-making is critical.

• →SHAP-guided reinforcement learning achieves 34.44% drag reduction using only 0.43% normalized input power, outperforming traditional opposition control by 49.41%
• →The optimal control policy activates predominantly at near-zero wall pressure, aligning with known turbulent flow physics rather than learning arbitrary patterns
• →Explainable AI techniques guide reinforcement learning toward physically intuitive solutions while simultaneously improving performance metrics
• →The method reduces actuation costs by 92.7% compared to direct wall-shear-stress baseline approaches while improving drag reduction
• →Temporal analysis reveals the learned policy operates on timescales comparable to near-wall turbulent structure lifetimes, suggesting discovery of fundamental flow control principles