Revealing the drivers of turbulence anisotropy over flat and complex terrain: an interpretable machine learning approach

Abstract: Turbulence anisotropy was recently integrated into Monin-Obukhov Similarity Theory (MOST), extending its applicability to complex terrain and diverse surface conditions. Implementing this generalized MOST in numerical models, however, requires understanding the key drivers of turbulence anisotropy across various terrain conditions. This study therefore employs random forest models trained on measurement data from both flat and complex terrain and including upstream terrain features, to predict turbulence anisotropy. Two approaches were compared: using dimensional variables directly or employing non-dimensional groups as model input. To address cross-correlation among features, we developed a new selection method, Recursive Effect Elimination. Finally, interpretability methods were used to identify the most influential variables. Contrary to expectations, variables related to terrain influence were not found to significantly impact turbulence anisotropy. Instead, non-dimensional groups of common turbulence length, time and velocity scales proved more robust than dimensional variables in isolating anisotropy drivers, enhancing model performance over complex terrain and reducing location dependence. A ratio of integral and turbulence memory length scales was found to correlate well with turbulence anisotropy in both daytime and nighttime conditions, both over flat and complex terrain. During the day, a refined stability parameter incorporating both the surface and mixed layer scaling emerged as the dominant driver of anisotropy, while at night, parameters related to rapid distortion were strong predictors.