Characterizing turbulence structures in convective and neutral atmospheric boundary layers via Koopman mode decomposition and unsupervised clustering.

This article focuses on characterizing coherent turbulence structures in atmospheric boundary layer (ABL) flows using large eddy simulations (LES), Koopman mode decomposition (KMD), and unsupervised machine learning techniques. Eight LES cases of convective, neutral, and unsteady ABLs were analyzed, revealing that as the ratio of buoyancy to shear production increases, turbulence structures transition from roll vortices to convective cells, with corresponding changes in sweep, ejection, and inward/outward interaction events. KMD effectively decomposed the complex nonlinear ABL dynamics into a small number of modes associated with physical forces such as pressure gradient, Coriolis, and buoyancy, enabling accurate reconstruction of flow fields with only about 5–10% of the modes. To manage the large number of Koopman modes, a novel data-driven approach combining K-means clustering with convolutional neural networks (CNNs) was developed, which classified modes into distinct, rotation- and displacement-invariant clusters representing different turbulence scales and structures. The study demonstrates that integrating LES, KMD, and machine learning provides a systematic framework for identifying and reconstructing key coherent turbulence patterns in ABLs, with potential applications in meteorology, wind energy, and climate modeling.