Global performance benchmarking of artificial intelligence models in atmospheric river forecasting

The rapid advancement of artificial intelligence models is expanding the frontiers of weather forecasting, underscoring the need for comprehensive evaluations in specific applications to ensure their effective use and guide future development. Here we benchmark five state-of-the-art artificial intelligence models for forecasting atmospheric rivers, assessing both meteorological fields and atmospheric river-related metrics across global and regional scales. Results show that FuXi achieves the best performance at a 10-day lead time for meteorological fields and atmospheric river forecasts globally. However, regional assessments reveal that the model incorporating numerical components—NeuralGCM—performs better in predicting atmospheric river intensity. Two case studies along the North and South American coasts further highlight NeuralGCM’s superior ability to predict atmospheric river shapes and intensities at 10-day lead times. Nonetheless, accurately predicting atmospheric river landfall locations beyond one week remains a challenge, emphasizing the need for refinement of these models for region-specific forecasts in future applications. AI models show promise in forecasting atmospheric rivers, with FuXi excelling globally at 10-day lead times for specific humidity and horizontal winds, while NeuralGCM performs better regionally in predicting intensity, according to a comparison of five state-of-the-art models.