Revealing the drivers of turbulence anisotropy over complex terrain with interpretable machine learning.

Turbulence anisotropy has gained attention in recent years for its role in surface layer atmospheric turbulence. The degree of anisotropy y_B, retrieved from anisotropy invariant analysis, has been introduced as an additional non-dimensional parameter into the flux-variance, flux-gradient and spectral surface scaling relations of Monin-Obukhov Similarity Theory (MOST). This new approach allowed to explain the observed scatter in the scaling relations both over flat and highly complex terrain, thus extending MOST outside of its original range of validity. However, how to predict y_B in a range of realistic terrain and stability conditions, remains an open question.

In this study we use observational data from two atmospheric surface layer experimental campaigns over flat (METCRAX II) and complex terrain (Perdigao), to reveal the drivers of the degree of anisotropy y_B. To do so, we adopt interpretable machine learning techniques considering a multitude of variables, both related to macro and micro meteorology, terrain roughness and heterogeneity, as well as topography. We train a random forest regressor on data from multiple meteorological towers, considering the spatial distribution of the measurements, and employ interpretation techniques such as permutation variable importance and SHAP analysis to extract the best predictors of y_B and their interactions. The results highlight that two non-dimensional groups including the information on stratification, drag and turbulence memory explain over 90% of the variance in flat terrain, and 70% over complex mountainous terrain. These results pave the way to not only understand better the processes that lead to different states of turbulence anisotropy but also allow the implementation of the new scaling framework into surface parametrizations for numerical models.