WeatherNext 3: Increasing resolution and performance of global weather models (arxiv.org)

WeatherNext 3 has been unveiled as a groundbreaking advancement in global weather modeling, addressing critical limitations previously faced by AI-based forecasting systems. Unlike traditional models that generate forecasts every six hours, WeatherNext 3 produces updates hourly by integrating low-latency geostationary satellite data. This allows the model to achieve spatial and temporal resolutions comparable to the best physics-based models, with a remarkable 0.1-degree resolution for single-level variables such as solar radiation and cloud cover. By directly learning to predict satellite-derived precipitation estimates and adapting to local geographical features, WeatherNext 3 significantly reduces forecast errors. The significance of WeatherNext 3 for the AI/ML community lies in its innovative approach that merges operational forecasting with real-time observational data. This new capacity enhances probabilistic medium-range forecasting skills, setting a new standard by moving beyond traditional analysis variables. The model's ability to simultaneously assimilate data and generate accurate predictions showcases the potential for AI technologies to transform weather forecasting—from a segmented process to an integrated system, ultimately improving decision-making in weather-sensitive sectors.