Study of interscale interactions for turbulence over the obstacle arrays from a machine learning perspective

Investigating interactions between large- and small-scale motions is essential for understanding turbulence over rough boundaries. The present work applies XGBoost models to predict the spatial distribution of ejections and sweeps and quantify their statistical dependence on scale-decomposed velocity fields. Based on large eddy simulation, the models are trained and validated at 20 horizontal planes in turbulence over two types of obstacle arrays. At each height, a default XGBoost model X-0 and four comparison models ( X-uL,X-wL, X-uL,X-wS, X-uS,X-wL, and X-uS,X-wS) are trained. The model X-0 is trained by the set with four scale-decomposed velocity fields {u(L), u(S), w(L), w(S)}, where u and w are the streamwise and vertical velocity fluctuations and subscripts L and S refer to above-canyon and sub-canyon scales, while the comparison models are trained by subsets of the scale-decomposed velocity fields. The results indicate that the model X-0 predicts the spatial distributions of both ejection and sweep events well, with the structure underestimation being less than 8% within the canopy layer and 3% above it. Along the vertical direction, the relative importance of scale-decomposed velocity fields on the prediction of ejections and sweeps is quantified by the feature importance and prediction errors. The feature importance profiles reveal that both sweeps and ejections are most strongly related to w S within the canopy, but ejections have a stronger dependence on u(L) well above the canopy. For the comparison models, those trained with w S (namely, X-uL,X-wS and X-uS,X-wS) give better predictions within the canopy layer, whereas those trained with u(L) (namely, X-uL,X-wS and X-uL,X-wL) perform better above the canopy. This study shows that a machine-learning-based approach can be designed to quantify the relative importance of different scale-decomposed velocity fields on predicting ejections and sweeps and to detect vertical changes of such relative importance.