Using state-of-the-art generative neural networks for high-resolution NWP ensemble emulation

Convection-permitting numerical weather prediction (NWP) models are a useful tool to forecast high-impact phenomena such as storms and heat waves. Ensemble prediction systems based on such models can then be used to quantify the associated forecast uncertainties. However, high-resolution ensemble forecasts come with a high computing cost; this limits the size of operational ensembles, and potentially reduces their relevance in critical situations.
In this work we propose a new and computationally efficient way to synthesize additional ensemble members of the kilometer-scale AROME model,based on deep generative models. For that purpose, state-of-the-art style-based generative adversarial networks (StyleGAN) and denoising diffusion probabilistic models (DDPM) are examined and compared on the joint generation of 10-meter wind and 2-meter temperature forecasts. We first show that these neural networks, once properly trained, create physically consistent, realistic, multivariate ensemble members. We then propose several methods to condition the generated members on the NWP model outputs 'of the day', in order to produce large hybrid physical/statistical ensembles at a small numerical cost. In particular, we show how to leverage the latent representations learnt by the networks to control the resulting ensemble statistics. A thorough evaluation of the resulting hybrid ensembles is proposed to assess their relevance and the new information they add with respect to small, pure NWP ensembles. Finally, we compare large, deep-learning-generated ensembles to a 875-member AROME forecast on a situation corresponding to a weather alert on the Mediterranean region, and evaluate the ability of generative approaches to provide an acceptable approximation of this large NWP ensemble.