The operational medium-range deterministic weather forecasting can be extended beyond a 10-day lead time

Abstract

Given the complexity of the atmospheric system, current numerical weather prediction models struggle with accurate forecasts. Here we present FengWu, an Artificial-Intelligence-driven global medium-range forecasting system employing multi-modal and multi-task learning to simulate atmospheric dynamics at 0.25° spatial resolution across 13 pressure levels. To decrease the error accumulation problem, a replay buffer mechanism has been implemented with high computational efficiency. These enhancements allow FengWu to outperform deterministic forecasts produced by European Centre for Medium-Range Weather Forecasts High-Resolution Model, Pangu-Weather, and GraphCast. Additionally, to address predictive uncertainty, we develop FengWu-Ensemble, using conditional diffusion model that generates reliable multi-member forecasts based on deterministic predictions. Comparative evaluations against the Integrated Forecasting System Ensemble show that FengWu-Ensemble achieves superior performance across multiple meteorological variables and evaluation metrics. These results indicate that FengWu holds strong potential for improving both deterministic and probabilistic weather forecasting. A new deep learning-based global medium-range weather forecasting model outperforms existing machine learning models and the European Centre for Medium-Range Weather Forecasts model, producing accurate global weather forecasts beyond ten days.