Subseasonal predictability of Southwest US rainfall in AI weather prediction models

Predicting the amount and distribution of precipitation that the Southwest US region will receive during the peak of the rain season (October to March) is a great challenge to climate scientists. The subseasonal forecast time scale (2 to 6 weeks) is of particular interest for water managers and stakeholders, however forecast skill is notoriously low in long-range forecast systems over the region. Since AI weather prediction models now have equal skill to the best dynamical forecast systems for weather forecasting (up to 7-10 days), whether they perform as well for subseasonal prediction is an outstanding research question.

Here we present an overview of a few AI weather models that have been adapted for use in the context of subseasonal prediction. The models are exclusively trained on historical reanalysis (i.e., real-world) data, and subseasonal reforecasts have been run over the 2018-2023 test period. The performance of the models in predicting North Pacific large-scale atmospheric circulation patterns a few weeks in advance is evaluated, and compared to state-of-the-art traditional subseasonal forecast models. The AI models exhibit promising skill that equals the best dynamical models, although it remains relatively low for real-world applicability. An advantage of the AI models is their low computational cost which allows for running very large forecast ensembles, a requirement to capture the probability distribution of possible future outcomes and improve decision-making in the face of extreme weather events.