Improving the forecast skill of a climate model using machine learning trained on a global 3 km simulation

Abstract:

Vulcan Climate Modeling (a small philanthropically-supported group in Seattle) and NOAA/GFDL are collaborating on a pilot project to use machine learning to develop a skillful parameterization of moist physics for versions of the US operational global weather forecast model, FV3-GFS, with 25-200 km grids, by coarse-graining outputs from a 3 km version of the same global model. Major challenges have included organizing a suitable workflow across multiple institutions and computing infrastructures, designing a suitable coarse-graining approach over topography and diverse land surfaces, and working with a full-complexity operational model written in FORTRAN. We have achieved our initial goal of improving 5-10 day weather forecasts with lower-resolution models by augmenting them with machine learning; results and ongoing challenges will be presented.