A Paradigm Shift in Precipitation Modeling: Moving Beyond Numerical Models

Abstract

Accurately representing surface precipitation in weather and climate modeling is crucial to practical operational use of these models. Presently, global numerical weather prediction (NWP) models struggle to recreate the precipitation variable due to unresolved physical processes at the subgrid level as well as use of poorly constrained microphysical parameterizations. The advent of machine learning in the field of weather and climate prediction has proven to be beneficial to advance modeling efforts and has the potential to also target weak points in NWP such as estimating the precipitation field. Training a deep learning model using satellite data, we can bypass the parameterizations traditionally used by NWP to produce precipitation, achieving a field that more closely matches observations than the widely used ERA5 reanalysis dataset. The resulting model can compute precipitation from only ten ERA5 input fields and is able to better capture extremes while also improving the issue of overproduction of light precipitation in the ERA5 product when evaluated against the IMERG satellite dataset. The machine learning model is also used to produce precipitation from NWP forecast fields, improving on the forecasted precipitation of the NWP model. This work supports future development of deep learning models that meet the needs of current weather and climate modeling.