Testing Machine Learning Methods for Downscaling in the 3D RTMA project

In order to provide background fields at the resolution at which the En3Var-based 3D Real Time Mesoscale Analysis (RTMA) runs, the products of the Rapid Refresh Forecast System (RRFS) forecast have to be downscaled from 3 to 2.5 km. This brings the realism to the produced analyses through better adjustment to terrain and surface conditions. We are developing a dynamical downscaling capability to accomplish that end. One of the major constraints of the 3D RTMA project is the urgency to deliver timely analysis products to stakeholders for nowcasting and other applications. With the idea to deliver much needed efficiency in producing downscaled results, we plan the development of a machine learning (ML) version of the downscaling procedure. In the final stages, the dynamically downscaled fields obtained from running a 2.5-km version of the RRFS will be used for training and testing of developed ML algorithm. However, at this point, we start the development using the 2-m temperature background fields from NCEP’s operational 2D RTMA system. These fields are obtained via an empirical downscaling of the model forecast using NCEP’s “Smartinit” package. Two possible candidates for the ML-based downscaling are being explored: Generative Adversarial Networks (GANs) and Reservoir Computing (RC); these techniques will be compared to determine the most optimal choice for the ML-based downscaling. This paper will present and discuss preliminary results of this study.