Identifying regions of importance in wall-bounded turbulence through explainable deep learning.

Area of Science:

• Fluid Dynamics
• Turbulence Research
• Computational Physics

Background:

• Wall-bounded turbulence remains a significant unsolved problem in classical physics.
• Understanding coherent structures is crucial for tackling turbulence.
• Existing methods require new perspectives for deeper insights.

Purpose of the Study:

• To explore interactions among energy-containing coherent structures in wall-bounded turbulence.
• To apply an explainable deep-learning method to analyze turbulent flow dynamics.
• To identify the importance of different flow structures in predicting turbulent behavior.

Main Methods:

• Utilized a U-net architecture for predicting velocity fields in turbulent channel flow simulations.
• Employed SHapley Additive exPlanations (SHAP), a game-theoretic algorithm, to assess structure importance.
• Applied the explainable AI framework to both simulated and experimental turbulent flow data.

Main Results:

• The study's findings align with existing literature on turbulent flows.
• Revealed that the most influential structures in turbulence are not necessarily those contributing most to Reynolds shear stress.
• Successfully identified and ranked flow structures based on their importance scores in an experimental database.

Conclusions:

• The developed explainable deep-learning framework provides novel insights into wall-bounded turbulence.
• This approach can identify key flow structures beyond their contribution to Reynolds shear stress.
• The methodology holds potential for advancing fundamental understanding and developing new flow control strategies.