How do AI Data-driven Models Compare with Physics-based Models in Predicting ARs and Extreme Precipitation in the Western US at Subseasonal Lead Times?

In recent years, purely artificial intelligence (AI)-enabled data-driven models proposed by technological companies (like Google, NVIDIA, and Huawei) have emerged as a cost-effective alternative to the computationally intensive numerical weather prediction (NWP) models traditionally used for weather forecasting. The model training involves learning a function describing the relationship between two consecutive atmospheric states separated in time, typically using global reanalysis datasets produced with state-of-the-art physics-based models, data assimilation algorithms, and a variety of observations. Once trained, these AI data-driven models can generate global forecasts from an initial atmospheric state at a fraction of the time and energy needed by traditional physics-based models. The advancements in AI data-driven modeling present unique opportunities for improving the prediction of global and regional hydroclimate. While the forecast skill of these AI models has been studied at a global scale, their regional skill particularly for extreme weather events like landfalling atmospheric rivers (ARs) has been relatively understudied.

Here, we leverage recent AI modeling architectures based on graph neural networks and transformers–specifically Google DeepMind’s Graphcast (Lam et al., 2023), NVIDIA’s Spherical FourCastNet (Bonev et al., 2023), and Huawei’s Panguweather (Bi et al., 2023)–to explore two key research questions: (1) How well do AI models predict extreme weather phenomena in the Western US, where the presence or absence of a relatively small number of impactful ARs are integral to the water year, contributing as much as half of the annual rainfall? (2) Can AI models help improve subseasonal-to-seasonal (S2S) forecasting of precipitation, a challenging frontier for current dynamical forecasting systems in this region? We evaluate the performance of these AI forecast models for impactful AR episodes from recent winter seasons, as well as in simulating large-scale circulation regimes and modes of climate variability (e.g., MJO) in the Western US out to 4-week lead time. Additionally, we provide a comparative skill assessment of the AI data-driven models against physics-based modeling benchmarks, including ECMWF’s Integrated Forecasting System (IFS) for weather time scales and dynamical models from the World Weather Research Program/World Climate Research Program’s (WWRP/WCRP) S2S Prediction Project database (Vitart et al., 2017) for longer lead times. Our results highlight the potential of AI data-driven models to complement physics-based models in enhancing early warning systems for ARs as well as improving precipitation forecasting at extended lead times.

References

• Bi, K., Xie, L., Zhang, H., Chen, X., Gu, X., & Tian, Q. (2023). Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619(7970), 533-538.
• Bonev, B., Kurth, T., Hundt, C., Pathak, J., Baust, M., Kashinath, K., & Anandkumar, A. (2023, July). Spherical fourier neural operators: Learning stable dynamics on the sphere. In International conference on machine learning (pp. 2806-2823). PMLR.
• Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., ... & Battaglia, P. (2023). Learning skillful medium-range global weather forecasting. Science, 382(6677), 1416-1421.
• Vitart, F., and coauthors, 2017: The subseasonal to seasonal (S2S) prediction project database. Bulletin of the American Meteorological Society, 98, 163–173.