Do AI Models Produce Better Atmospheric River Forecasts than Physics Based Models? A Quantitative Evaluation

The recent surge in using state-of-the-art machine learning (ML)-based techniques for weather prediction systems like Google's Graphcast, Nvidia's Fourcastnet, and Huawei's Panguweather have demonstrated remarkable accuracy and efficiency compared to traditional numerical weather prediction (NWP) methods. As these methods integrate into our forecasting toolbox, it’s crucial to assess the capability of current machine learning models in simulating high-impact weather phenomena. They perform as well or better than the European Centre for Medium-range Weather Forecasts (ECMWF) system on most metrics. Yet, no comprehensive study has looked at their ability to predict atmospheric rivers (ARs). This study will examine the AR prediction skill of these new ML-based models on the U.S. West Coast at lead times up to 10-days. Skill and accuracy will be evaluated during a few key case studies as well as over long-term averages. These results will then be compared with the ECMWF's state-of-the-art High-Resolution Forecast (HRES) model.