UT-MeteoGAN: A Next-Generation AI Model for High-Resolution Numerical Weather Prediction

We present UT-MeteoGAN, an advanced AI-based model designed for rapid, high-resolution numerical weather prediction. Leveraging the power of the Swin Transformer architecture and adversarial loss mechanisms, UT-MeteoGAN enhances the accuracy and spatial resolution of weather forecasts beyond what traditional Super Resolution Convolutional Neural Networks (SRCNN) can achieve. A key innovation in our approach is the integration of data assimilation directly into the model's training pipeline, which includes real-time station observations to increase forecast reliability. The model utilizes a global atmospheric model as input and learns to generate high-resolution gridded forecasts, effectively compensating for the lack of detailed observational data in some regions across the globe. In a comprehensive case study focused on the continental United States, UT-MeteoGAN generated 36-hour lead forecasts at a 1-km spatial and 1-hour temporal resolution. The model's input came from the global model Graphcast, while its high-resolution training target was NOAA's Analysis of Record for Calibration (AORC), supplemented with hourly data from NOAA’s Global Historical Climatological Network (GHCNh) for station observations. Results demonstrate that UT-MeteoGAN consistently produces forecasts that match or outperform those generated by existing models, delivering high accuracy with rapid computational turnaround. The development and operational deployment of UT-MeteoGAN are powered by the cutting-edge NVIDIA H100 GPUs. This model represents a significant advancement in numerical weather prediction, offering a scalable, high-precision solution that can be adapted to various geographic regions. The successful implementation over the continental United States illustrates UT-MeteoGAN's potential to rapidly, augment meteorological forecasting on a global scale as well as applying it for high impact weather ever as was done for recent (2024) as well as legacy hurricanes in the US Gulf of Mexico domain.