GenEPS: A Generative Foundation Model for Probabilistic Weather Forecasting

Accurate weather forecasting is essential for a broad range of socioeconomic activities. While emerging data-driven models match numerical weather prediction accuracy with reduced computational cost, their deterministic nature overlooks uncertainties in initial state estimates, model systematic biases, and stochasticity arising from unresolved subgrid physical processes. This obliviousness results in over-confident deterministic predictions that render uncertainty quantification inaccessible, thereby limiting their utility for risk-based decision-making.

To address these challenges, we present the Generative Ensemble Prediction System (GenEPS), a framework that systematically explores uncertainties in initial states, model formulations, and model stochasticity. GenEPS functions as a foundation model that has explicitly learned the probability distribution of high-dimensional atmospheric states. It provides a plug-and-play solution for ensemble forecasting with arbitrary deterministic models. Specifically, GenEPS utilizes deterministic forecasts as conditions to perform generative sampling, producing an ensemble of states projected back into the realistic atmospheric phase space defined by ERA5. This stochastic sampling process quantifies uncertainties in initial conditions and forecast dynamics while ensuring physical consistency. Crucially, by treating each step as a re-initialization within the valid state space, the framework decouples state evolution from specific model formulations, enabling seamless cross-model integration to mitigate systematic biases.

By explicitly representing all three sources of uncertainty, GenEPS outperforms state-of-the-art numerical ensemble predictions and data-driven predictions when evaluated against ERA5 reanalysis data using both deterministic and probabilistic metrics. GenEPS also enhances extreme event predictions, offering physically consistent forecast fields. These advances establish a new paradigm in ensemble forecasting through multi-model generative integration, combining a surging number of data-driven weather forecasting models and potentially numerical models, to achieve more reliable predictions.