Real-Time Detection and Characterisation of Trapped Lee Waves: From Deep Learning Model to Operational Deployment in the Iceland-Greenland Region

Trapped lee waves and the associated turbulent rotors pose significant hazards for aviation and land-based transport. While high-resolution numerical weather prediction (NWP) models can capture these phenomena, there is a lack of automated, reliable tools for detecting and characterizing lee waves in model output. Traditional spectral methods, although useful, are often computationally expensive and require domain-specific tuning. This study presents an operational pipeline based on deep learning (DL) for the real-time detection and characterisation of trapped lee waves. 

Building on previous research demonstrating the effectiveness of a DL model (LeeWaveNet) for segmenting and extracting wave characteristics from vertical velocity fields in the UK, we have deployed this model in a production environment for the Iceland-Greenland region using an operational NWP model for Iceland and Greenland, named Harmonie-IG. The model was containerized and hosted on a server at Danish Meteorological Institute, receiving Harmonie-IG output as input and producing near-real-time predictions of lee wave activity. The results, including key characteristics such as wavelength, orientation, and amplitude, are stored in an AWS S3 bucket and visualized through an automated post-processing pipeline. 

To assess the generalizability and reliability of the model beyond the training domain, we performed an extended validation over Norway using the Danish ReAnalysis dataset (DANRA). In particular, we compared the predicted orientation of the lee waves with the observed wind direction, demonstrating promising correlation and physical consistency. This validation step is a key milestone towards confirming the robustness of the model when applied to new geographies and datasets. 

This work highlights not only the potential of DL in advancing the operational use of AI in meteorology, but also showcases a complete MLOps workflow—from inference at scale to data management and visualization. Our pipeline demonstrates that with the right infrastructure, deep learning models can be effectively integrated into real-time forecasting systems, providing timely and accurate identification of hazardous atmospheric features.