Google DeepMind's WeatherNext AI Model Improves Cyclone Forecasts, Open-Sourced

WeatherNext, an AI weather forecasting model released jointly by multiple global meteorological agencies, has achieved technical progress in tropical cyclone prediction. Developed by a research team in collaboration with meteorological departments from several countries, the model's findings have been published in an academic journal. The model combines global weather dynamics data with a database of historical weather systems to predict key indicators such as tropical cyclone path, intensity, and wind structure. Its forecast lead time is improved over previous models, with three-day prediction accuracy equivalent to that of traditional two-day forecasts.

Tropical cyclones are among the most destructive natural disasters globally. Over the past five decades, weather-related disasters have caused more than 700,000 deaths worldwide and economic losses of US$1.4 trillion. Traditional forecasting models rely on macro-scale global data to track the movement of weather systems, while predicting storm intensity and core structure requires high-resolution local data. Forecasters often have to trade off between track and intensity predictions. Through architectural innovation, WeatherNext combines global weather dynamics with data from nearly 5,000 historical weather systems for training. It makes predictions using data resolution roughly a hundred times coarser than traditional models, and can rapidly run multiple ensemble forecasts on a single processor to help assess the probability distribution of extreme weather.

The model participated in actual disaster prevention work during the 2025 storm season, successfully assisting relevant meteorological agencies in predicting the intensification process and landfall path of Hurricane Melissa. Currently, the model's code and weights have been officially open-sourced, including a lightweight version that can run on a free platform. The forecast results have been integrated into public weather platforms for global researchers and the public to access.