Predictability of precipitation extremes in the Western United States and the U.S. Gulf Coast using AI weather models

Extreme precipitation events, such as those caused by hurricanes and atmospheric rivers, pose significant risks to infrastructure and human populations. Accurate forecasting is essential for effective disaster preparedness and long-term climate adaptation, particularly in regions prone to intense rainfall. While numerical weather prediction (NWP) models are the current standard, they are computationally expensive. Recent progress in artificial intelligence (AI) offers a promising alternative through fast, data-driven AI-based weather prediction (AIWP) models. This study evaluates the performance of four advanced AIWP models: Pangu-Weather, GraphCast, FourCastNet version 1, and FourCastNet version 2. The models are assessed in their ability to forecast extreme precipitation across the Contiguous United States (CONUS), with particular focus on the Western United States and the U.S. Gulf Coast region. In a first step, performance is evaluated based on native model output, with historical data from the PRISM Climate Group and forecasts from the state-of-the-art NWP model, the Medium-Range Weather Forecasts’s Integrated Forecasting System (IFS), serving as references. In a second step, all models are subjected to a neural network-based post-processing procedure. This step aims to improve the raw precipitation forecasts and, for models that do not natively predict precipitation, to generate it from other output variables. Results show that the model performance depends on region, lead time, and precipitation intensity. In several cases, the AIWP models matched or exceeded the performance of the IFS. GraphCast showed particularly strong and consistent performance, reducing RMSE by approximately 15% over CONUS across lead times of up to seven days compared to IFS. Post-processing improved some forecasts and enabled precipitation prediction for models lacking native output, though postprocessed forecasts generally exhibit reduced spatial detail relative to native precipitation outputs. The study highlights AI’s potential to complement traditional NWP systems, offering a path toward faster and more efficient weather prediction.