Generative learning for km-scale atmospheric simulation of global to regional

Recent advances in global machine learning weather models trained on ERA5 reanalysis data are set to revolutionize weather predictions (see ECMWF blog post). These models demonstrate skill akin to leading numerical models (IFS and GFS), but with order 10,000 times speedup. However, both ML and numerical global models do not adequately resolve phenomena such as tropical cyclones and large convective systems. This necessitates a robust downscaling approach to bridge the divide between global-scale forecasts and regional-scale weather events.

In numerical weather forecasting, downscaling is split into dynamical downscaling, where a high-resolution regional model is nested in a low-resolution global model, and statistical downscaling, which maps the relationship between high and low-resolution atmospheric states at a given time. My presentation will focus on Nvidia Research's latest endeavors in developing both dynamical and statistical downscaling ML models. These models are designed to seamlessly integrate with global ML models, creating a comprehensive global-to-regional ML forecasting system.

In ML statistical downscaling, we learn a mapping between the 25km ERA5 and a 2km regional WRF analysis. Drawing inspiration from Reynolds decomposition, we employ a two-step approach. In essence, a deterministic regression (based on UNet) is trained to predict the mean state, followed by a generative “diffusion” model to forecast the distribution of potential perturbations. This method reduces the – otherwise large – distribution shift between ERA5 and WRF, simplifying the training of a generative model.

Similarly, in ML dynamical downscaling, a 2km resolution autoregressive regional ML model is coupled with a 25km global ML model. Employing a similar “Reynolds” decomposition, here a Vision Transformer type architecture (swin2) is trained to timestep the mean high-resolution state, while a generative “diffusion” model predicts possible fine-scale perturbations. Both ML downscaling models are trained on data from operational numerical simulations done with the data-assimilating WRF system in Taiwan and the CONUS region (HRRR).

Our results demonstrate these models' exceptional capability to emulate WRF simulations at a fraction of the cost. They accurately intensify and compress tropical cyclones to match WRF-simulated structures, sharpen weather fronts, and adjust diurnal cycles to account for small-scale effects. These findings pave the way for deploying extensive ensembles of cost-effective, high-resolution simulations, particularly for extreme weather events.

For statistical downscaling see - https://arxiv.org/abs/2309.15214