Multilevel Deep Learning for Non-Local Prediction of ABL Flow Fields Across Varying Stability Regimes

Understanding and predicting wind flow structures within the atmospheric boundary layer (ABL) across varying stability conditions remains a key challenge in atmospheric science and environmental modeling. Although large-eddy simulation (LES) provides high-fidelity insights, it is computationally prohibitive for near-real-time or large-scale applications. To address this, we propose a deep learning framework that predicts the two-dimensional wind flow fields in the u, v, and w components at a target height z, using corresponding flow fields at three levels above z. By integrating vertical flow correlations and continuity principles, our approach captures essential turbulent features while reducing input dimensionality and eliminating the need for full 3D simulations.

A modified convolutional neural network (CNN) forms the core of this framework, capturing complex spatial and temporal patterns from high-resolution LES datasets. Mass conservation is embedded in the training process to ensure physically consistent results. Preliminary results indicate that the model preserves large-scale turbulence features and captures the influence of higher-elevation dynamics, although smaller high-frequency turbulent features require further refinement. To address this, our ongoing work includes adopting a scale-specific approach to explicitly handle the diverse turbulent length scales observed in the ABL. We are also incorporating multitemporal dynamics and attention mechanisms into the architecture of our model to better account for long-range dependencies over time, thereby enabling the model to adapt to different stability regimes and transitions among them. To enhance interpretability, we will employ explainable AI (XAI) tools such as SHAP and GradCAM, revealing how specific regions in the input influence the emergence of particular turbulent footprints in the predicted flow. These insights guide improvements in both model design and understanding of atmospheric processes governing ABL flow development.

This research underscores the transformative potential of deep learning in boundary layer meteorology. By significantly reducing computational demands while retaining essential flow dynamics, our model enables real-time, high-resolution predictions of ABL flows. This scalable and efficient framework opens new possibilities for diverse applications, including weather forecasting, wind energy optimization, and environmental analysis.