This code belongs to the workshop paper: https://arxiv.org/pdf/2405.20719
Please cite the paper if you use this code.
Abstract: Global Climate Models (GCMs) typically produce simulation outputs at coarse spatial scales due to the large computational costs associated with solving differential equations at fine spatio-temporal scales. However, high-resolution predictions are crucial for critical applications such as energy infrastructure planning and extreme weather event analysis. This need for detailed and accurate data has spurred significant interest in downscaling methods to bridge the gap between coarse model outputs and the high-fidelity information required for effective decision-making and risk assessment. In this work, we investigate conditional generative methods for climate downscaling, focusing on a conditional normalizing flow and a GAN whith conditional Gaussians as base densities where the flow model outperforms the GAN. We demonstrate the successful performance of both methods an ERA5 water content dataset across various downsampling factors. Additionally, we show that the likelihood based method allows us to assess the predictive uncertainty in terms of standard deviation from the fitted conditional distribution mean.
