Diagnosing Extreme Precipitation from Atmospheric Fields with a UNet Architecture on a HEALPix Mesh

This study examines the minimal information required for a data driven model to accurately diagnose precipitation, both extreme and typical values. The model diagnosis precipitation from large-scale ERA5 reanalysis data at resolutions as fine as 0.25 x 0.25 degrees latitude and longitude. The model uses a U-Net architecture with data represented on a HEALPix mesh. ERA5 precipitation, though a model-generated product, is used for verification. As shown in the figure, we obtain excellent results for 6-h accumulated precipitation at coarse 1 x 1 degree resolution using 7 spherical shells of data, including temperatures at 2 m and 850 hPa, geopotential heights at 3 levels, and total column water vapor. Notably, the model employs a log-transformed precipitation variable field in training, following Rasp & Thuerey, 2021 as well as a custom loss function to accommodate the inherent skewness of precipitation data.

Accurate representation of the feedbacks of the tropical convection shown in the figure on large atmospheric scales in global GCM’s is often thought to require very high-resolution convection models. Yet evidently the machine learning approach is capable of efficiently diagnosing the ERA5 representation of such precipitation with minimal coarse-resolution data.

The utilization of AI techniques allows for the representation of nuanced patterns and relationships with a fraction of the computational costs of conventional methods. Our diagnostic model’s performance is assessed as a function of the number of input variables and their spatial and temporal resolution. This work provides insights into the relative importance of different atmospheric features for predicting extreme precipitation events with machine learning models. The model’s performance on various rainfall intensity levels is also assessed, with an emphasis on the most extreme events.