Item

Abstract

Machine learning models have shown great success in predicting weather up to two weeks ahead, outperforming process-based benchmarks. However, most existing approaches do not fully integrate data assimilation into the prediction workflow and often suffer from error accumulation in autoregressive computation, limiting their applicability to predictions beyond the medium range and climate projections. We introduce Generative Assimilation and Prediction (GAP), a unified deep generative framework that connects data assimilation, weather forecasting, and climate simulation through a shared probabilistic representation of atmospheric states. GAP learns the climatological distribution from reanalysis data, and adapts it to different tasks by incorporating observational, predictive, and forcing constraints. Evaluated across a range of weather-climate tasks, GAP produces probabilistic weather forecasts competitive with leading forecasting systems, using its own assimilated initial states derived from 5% synthetic perfect observations, with particular skill in capturing extreme events such as atmospheric rivers, heatwaves and the Madden-Julian Oscillation. At seasonal timescales, GAP shows promising predictive skill when constrained by oceanic boundary forcing. Moreover, GAP supports stable millennial-scale climate simulations that reproduce atmospheric variability from daily weather fluctuations to decadal oscillations. These results suggest that generative modeling of the climatological distribution offers a viable foundation for seamless weather and climate prediction, thereby opening a path toward integrated, observation-to-climate workflows.