Instantaneous Reconstruction of Turbulent Channel Flow Using Explainable Deep Learning

Sammanfattning

Turbulent flows govern a wide range of physical and engineering systems, yet the physical mechanisms connecting the different components that constitute it remain only partially understood. This study investigates the immediate spatial relationships between velocity components in a turbulent channel flow with a Reynolds number of friction of Reτ = 125, using a combination of deep learning and explainable artificial intelligence (XAI).A three dimensional U-Net was trained on 10,000 flow field snapshots obtained from Direct Numerical Simulation (DNS) to perform an instantaneous (∆t = 0) reconstruction of the streamwise velocity component (u) from the full velocity field (u, v, w). The model went through 63 epochs and achieved a maximum error of 0.09% at the center of the channel. Following the training, Gradient SHapley Additive exPlanations (SHAP) was applied to 1,000 snapshots to quantify the contribution of each velocity component, at every time step, to the reconstruction. The analysis reveals that the streamwise component is approximately 25 times more important for the model’s reconstruction than both the wall normal- (v), and span-wise (w), components. The resulting three dimensional SHAP structures are concentrated and elongated in the streamwise direction near the wall, and get more infrequent and inconsistent closer to the channel’s center. These findings confirm that the U-Net and SHAP framework successfully captures physically meaningful structures in the flow and that the most informative input regions correspond to areas of strong velocity fluctuation in the near wall layer.