Towards a Dynamic Data Driven AI Regional Weather Forecast Model

Abstract

The advent of long-term reanalysis datasets such as ECMWF ERA 4/5 has enabled the development of AI-driven machine learning models for weather forecasting. The major benefit of AI as an approach is its ability to reduce computational forecast time from tens of hours to tens of seconds, thereby enabling a variety of new applications ranging from extreme regional weather event forecasting to first-responder aid for wildfires, severe storms, floods, oil spills, tornadoes and other extreme events in real time. Today, several operational weather forecast centers are evaluating these models as compliments or alternatives to their existing models. However, similar efforts in applying AI/ML approaches to mesoscale weather forecasting have lagged behind due to a lack of a reanalysis for current operational regional weather forecast models. Recently, the ECMWF made publicly available the Copernicus European Regional ReAnalysis (CERRA) at spatial resolutions of 11 km (0.10) and 5.5 km (0.050) from 1984 to the present. We present the first demonstration of a successful AI regional forecast at 5.5 km spatial resolution employing the Nvidia FourCastNet (FCN) model with its Adaptive Fourier Neural Operator (AFNO) and transformer self-attention modeling approach. We describe the training of a regional FourCastNet model on the NASA Center for Climate Studies (NCCS) Adapt cluster at the Goddard Space Flight Center using five years of CERRA reanalysis data at 3-hour intervals for five variables across four pressure levels. We show the RMSE forecast errors of a 5.5 km implementation trained on 5 years of data improved for all variables but one over a forecast trained on three.