AI for high-resolution climate data: downscaling climate projections and decadal predictions with a deep learning Latent Diffusion Model

Over the past decade, advancements in high-performance computing have led Machine Learning (ML) to play a key role in enhancing Earth System Models (ESMs), enabling progress beyond the current state-of-the-art. Downscaling techniques to generate high-resolution data starting from the results of large-scale models are one of the most promising Deep Learning (DL) applications for ESMs. This approach offers a computationally efficient alternative to numerical dynamical downscaling, particularly for climate projections.  

In this study, we present the application of a state-of-the-art DL model to emulate the dynamical downscaling of 6-hourly climate data, focusing on precipitation and minimum and maximum temperatures. The model is trained to reconstruct fields at a 4 km resolution, starting from dynamical predictors at ~100 km resolution. Training data consists of coarsened ERA5 reanalysis data (Hersbach et al., 2018) as predictors and high-resolution target data from the COSMO-CLM dynamical reanalysis for Italy (Raffa et al., 2021). We utilize 40 years of 6-hourly data (1981–2020) for training. 

This training setup is designed to prepare the model for inference on low-resolution outputs from a selection of diverse climate projections and decadal predictions. The ultimate goal is to generate an ensemble of high-resolution projections that deliver additional insights, particularly into extreme values, at a fraction of the computational cost of regional climate models. 

The DL architecture employed is a recently developed Latent Diffusion Model applied with a residual approach (Tomasi et al. 2024), which has demonstrated exceptional performance in downscaling continuous variables, such as 2-m temperature and 10-m wind speed components. Results are compared against other ML models (e.g., UNET) and available numerical regional climate models for benchmarking. Preliminary results are presented, highlighting (i) the enhancements introduced by the LDM architecture compared to baseline models, (ii) its ability to reconstruct coherent structures and extreme values, and (iii) the added value of the high-resolution data obtained by the application of the LDM to low-resolution climate projections.