Researchers have developed WeatherBridge, a novel adapter designed to improve the accuracy of pretrained weather forecasting models. This adapter bridges the gap between operational numerical weather prediction (NWP) initial states and the ERA5 dataset, which is commonly used for training. WeatherBridge adapts heterogeneous NWP inputs to ERA5 reference states without altering the original forecasting model, utilizing a Swin-UNet to predict residual corrections and incorporating source embeddings for spatial scale awareness. The system's training objective includes ERA5 reconstruction and a one-step forecast error alignment, optimizing for both initial state fidelity and downstream forecast compatibility. Evaluations showed WeatherBridge significantly reduces root-mean-square error (RMSE) in forecasts from various NWP systems, demonstrating its effectiveness in improving predictions from diverse operational inputs and its transferability across different pretrained models. AI
IMPACT Enhances accuracy of AI-driven weather forecasting by adapting diverse operational inputs to training data distributions.
RANK_REASON Academic paper detailing a new method for adapting AI models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
