GENERALIZATION PROPERTIES OF MACHINE LEARNING BASED WEATHER MODEL DOWNSCALING

Mayur Mudigonda, Alok Singh, Brian L. White, Adrian Albert

Abstract:

Modern numerical weather models utilize methods from computational fluid dynamics to simulate meterological variables, but are resolution-constrained due to
the high computational cost of solving atmospheric PDEs over fine grids. However, many topics of interest in atmospheric modeling, such as turbulent wind
flow, are difficult to observe outside of very fine spatial scales. Several statistical
methods have been developed for downscaling gridded wind maps, but most use
crude schemes such as bilinear interpolation.
In this work, we analyze machine learning based techniques for this problem. The
techniques considered here are similar to image super-resolution (SR) models,
which have been successfully applied to natural images. In particular, we consider
the Enhanced Super Resolution GAN model (ESRGAN) Wang et al. (2018) and
analyze its transferability and generalization properties.
We find that training on random regional grids beats all other approaches, even
when compared against models trained specifically on a region. Adding topographical data as input speeds and stabilizes training dramatically, but does not
significantly boost accuracy