Inferring wall-bounded coherent structures from two-dimensional turbulent fields via SHAP analysis
Samuel Molina-Casino, A. Cremades, Sergio Hoyas, José I. Cardesa, F. Chedevergne, Ricardo Vinuesa
Abstract
The analysis of coherent structures in turbulent flows is essential for understanding and controlling turbulence. Explainable Deep Learning (XDL) has been successfully used to identify the most important flow structures in turbulent channels (Cremades et al., 2024), but its computational cost grows significantly as Reynolds number increases. This study adapts the XDL framework to assess whether observations from time-resolved volumetric data of direct numerical simulations (DNS) can be inferred using two-dimensional (2D) slices. Limiting the analysis to 2D subsets reduces the cost of three-dimensional (3D) post-processing and facilitates its application to planar particle image velocimetry (PIV) experimental data. We employ SHapley Additive exPlanations (SHAP), a game-theoretic method, to identify dynamically relevant flow regions by assigning importance scores. Using DNS data from a turbulent channel flow at a friction Reynolds number of 125, we compare SHAP-based structures derived from 3D and 2D data with classically studied structures, including Q events, streaks, and vortex clusters. Our findings show that SHAP-based structures from 2D data statistically agree with those from full 3D fields and previously reported in the literature. These results underscore the potential of SHAP analyses for studying coherent structures with 2D data, enabling experimental and DNS research at higher Reynolds number.