Improving turbulence control through explainable deep learning (2504.02354v2)

Abstract: Turbulent-flow control aims to develop strategies that effectively manipulate fluid systems, such as the reduction of drag in transportation and enhancing energy efficiency, both critical steps towards reducing global CO$_2$ emissions. Deep reinforcement learning (DRL) offers novel tools to discover flow-control strategies, which we combine with our knowledge of the physics of turbulence. We integrate explainable deep learning (XDL) to objectively identify the coherent structures containing the most informative regions in the flow, with a DRL model trained to reduce them. The trained model targets the most relevant regions in the flow to sustain turbulence and produces a drag reduction which is higher than that of a model specifically trained to reduce the drag, while using only half its power consumption. Moreover, the XDL model results in a better drag reduction than other models focusing on specific classically identified coherent structures. This demonstrates that combining DRL with XDL can produce causal control strategies that precisely target the most influential features of turbulence. By directly addressing the core mechanisms that sustain turbulence, our approach offers a powerful pathway towards its efficient control, which is a long-standing challenge in physics with profound implications for energy systems, climate modeling and aerodynamics.

• The paper integrates Deep Reinforcement Learning (DRL) with Explainable Deep Learning (XDL) using SHAP values to identify and target coherent flow structures for more effective turbulence control.
• The integrated DRL-XDL approach achieved 33.3% drag reduction in simulations, outperforming other methods while using significantly less power.
• The study demonstrates that SHAP values provide deeper insights into critical flow features sustaining turbulence, enabling the development of more precise and energy-efficient control strategies.

Improving Turbulence Control through Explainable Deep Learning

The paper presented in "Improving turbulence control through explainable deep learning" investigates a novel approach to controlling turbulent flow, a fundamental challenge in classical physics and a significant concern in engineering applications such as transportation and energy systems. This research integrates Deep Reinforcement Learning (DRL) with Explainable Deep Learning (XDL) to devise more effective strategies for managing turbulence, particularly with respect to drag reduction—an advance with potential implications for reducing carbon dioxide emissions on a global scale.

Key Contributions and Findings

1. DRL and XDL Integration: The research leverages DRL to identify flow-control strategies, informed by XDL. By utilizing SHAP values—a method derived from cooperative game theory that attributes importance to different input features—the paper highlights the coherent structures crucial to flow dynamics. This integration allows the DRL model to target these structures, enhancing the model’s efficiency and effectiveness in reducing drag.
2. Enhanced Drag Reduction: Through training in a simulated environment, the DRL model achieved a drag reduction performance superior to models purely focused on minimizing drag. Notably, the proposed SHAP-based DRL approach reduced drag by 33.3%, which surpasses other traditional structure-based and directly drag-targeted control strategies. This performance is coupled with a reduced power consumption, using roughly half the power compared to the direct drag reduction model.
3. Understanding Coherent Structures: The paper explores the relevance of different coherent structures, including Q events and streaks traditionally identified in turbulent flows, demonstrating that SHAP values provide a more comprehensive insight into which features significantly sustain turbulence. Consequently, targeting these features results in more precise control strategies.
4. Power Efficiency: A significant finding of this paper is the efficiency of the SHAP-based control in terms of energy input. By modulating the turbulence more strategically, this approach achieves substantial drag reduction with reduced energy input—showcasing not only improved performance but a noteworthy gain in power savings, which is particularly crucial for large-scale applications.

Implications for Future Research

The paper suggests several pathways for future exploration. The combination of DRL and XDL in turbulence control shows potential for broader applications beyond wall-bounded turbulence, such as in biological systems or active matter. Moreover, improvements in the computational framework, like employing volumetric data or upgrading the DRL policy architecture, could further enhance the performance of these control strategies.

Furthermore, the research opens avenues for experimental validation, where the SHAP-based DRL approach could be directly implemented on experimental data. This capability could substantially dilute the computational demands often associated with high-resolution simulations and foster more practical applications in industrial and real-world scenarios.

In conclusion, the paper provides a compelling case for the integration of DRL with XDL in turbulent flow management, underlining a pathway to not only understand but effectively influence the mechanisms underpinning turbulence. As such, it offers substantive contributions to the fields of fluid dynamics and machine learning, with implications for sustainable technological development.