A modified deep learning weather prediction using cubed sphere for global precipitation

Deep learning (DL), a potent technology to develop Digital Twin (DT), for weather prediction using cubed spheres (DLWP-CS) was recently proposed to facilitate data-driven simulations of global weather fields. DLWP-CS is a temporal mapping algorithm wherein time-stepping is performed through U-NET. Although DLWP-CS has shown impressive results for fields, such as temperature and geopotential height, this technique is complicated and computationally challenging for a complex, non-linear field, such as precipitation, which depends on other prognostic environmental co-variables. To address this challenge, we modify the DLWP-CS and call our technique "modified DLWP-CS" (MDLWP-CS). In this study, we transform the architecture from a temporal to a spatio-temporal mapping (multivariate setup), wherein precursor(s) of precipitation can be used as input. As a proof of concept, as a first simple case, a 2-m surface air temperature is used to predict precipitation using MDLWP-CS. The model is trained using hourly ERA-5 reanalysis and the resulting experimental findings are compared to two benchmark models, viz, the linear regression and an operational numerical weather prediction model, which is the Global Forecast System (GFS). The fidelity of MDLWP-CS is much better compared to linear regression and the results are equivalent to GFS output in terms of daily precipitation prediction with 1 day lag. These results provide an encouraging framework for an efficient DT that can facilitate speedy, high fidelity precipitation predictions.