Interpretable machine learning reveals the drivers of turbulence anisotropy over complex terrain.

Turbulence anisotropy has recently gained attention for its role in the study of surface layer turbulence. In particular the degree of anisotropy, quantified through the anisotropy invariant y_B, has been successfully introduced as an additional non-dimensional parameter into the Monin-Obukhov Similarity Theory (MOST), and tested on the flux-variance and flux-gradient surface scaling relations. The novel extended MOST relations were shown to explain the observed scatter in the MOST scaling relations both over flat and highly complex terrain, thus allowing MOST to be extended outside of its restrictive original assumptions. The challenge, however, still remains in how to predict y_B for a range of realistic conditions, which would allow to implement the novel scaling relations in numerical models’ surface parametrizations.

In this study we use data from both flat (AHATS) and complex terrain (Perdigao measurement campaign), to understand the drivers of turbulence anisotropy. We use interpretable machine learning techniques considering a wide range of macro and micro meteorological variables, surface heterogeneity, and topographic measures to build a predictive model using a tree-based regression algorithm (i.e. random forest and boosting algorithms). Interpretability techniques, such as variable importance measures, partial dependence plots, and Shapley analysis then allow us to select the variables that influence turbulence anisotropy the most and assess their relation with each other and with our target variable. This approach will ultimately lead to an understanding of the processes behind the emergence of different states of turbulence and its anisotropy, paving the way for robust surface parametrizations in numerical weather models.