Precipitation Downscaling Using Dynamical and Neural Network Approaches.

High-resolution climate projections are crucial for assessing the future impacts of climate change. Statistical, dynamic, or hybrid climate data downscaling is often employed to create the datasets required for impact modelling. In this study, we utilize the COSMO-CLM (CCLM) version 6.0, a regional climate model, to investigate the advantages of dynamically downscaling a general circulation model (GCM) from CMIP6, with a focus on Central Asia (CA). The CCLM, running at a 0.22° horizontal resolution, is driven by the MPI-ESM1-2-HR GCM (at 1° spatial resolution) for the historical period 1985–2014 and projections for 2019–2100 under three shared socioeconomic pathways (SSPs): SSP1-2.6, SSP3-7.0, and SSP5-8.5 (Fallah et al., 2025). Using the CHIRPS gridded observation dataset for evaluation, we assess the performance of the CCLM driven by ERA-Interim reanalysis over the historical period.

The added value of CCLM, particularly over mountainous areas in CA, is evident, with a reduction in mean absolute error and bias of climatological precipitation by 5 mm/day for summer and 3 mm/day for annual values (Fallah et al., 2024). While no error reduction is achieved for winter, the frequency of extreme precipitation events improves in the CCLM simulations. Future projections indicate an increase in the intensity and frequency of extreme precipitation events in CA by the century’s end, particularly under the SSP3-7.0 and SSP5-8.5 scenarios. The number of days with more than 20 mm of precipitation increases by more than 90, and the annual 99th percentile of total precipitation increases by over 9 mm/day in mountainous areas.

A convolutional neural network (CNN) is also trained to map GCM simulations to their dynamically downscaled CCLM counterparts. The CNN successfully emulates the GCM-CCLM chain across large areas of CA but demonstrates reduced skill when applied to other GCM-CCLM chains. This downscaling approach and CNN architecture provide an alternative to traditional methods and could be a valuable tool for the scientific community involved in downscaling CMIP6 models (Harder et al., 2023).

In future work, we aim to extend this approach by training a neural network model to map the available GCM-RCM model chains for CORDEX-EU and applying the trained model to decadal prediction ICON simulations. This will enable the production of CORDEX-EU-like regional ICON simulations, bridging the gap between global and regional climate information on decadal timescales. By integrating decadal predictions into the framework, we aim to enhance the usability of regionalized climate data for short-term climate planning and decision-making.

References:

• Fallah, B., Russo, E., Menz, C., Hoffmann, P., Didovets, I., and Hattermann, F. F.: Anthropogenic influence on extreme temperature and precipitation in Central Asia, Sci. Rep., 13, 6854, https://doi.org/10.1038/s41598-023-33921-6, 2023.
• Fallah, B., Menz, C., Russo, E., Harder, P., Hoffmann, P., Didovets, I., and Hattermann, F. F.: Climate Model Downscaling in Central Asia: A Dynamical and a Neural Network Approach, Geosci. Model Dev. Discuss. [preprint], https://doi.org/10.5194/gmd-2023-227, accepted, 2025.
• Harder, P., Hernandez-Garcia, A., Ramesh, V., Yang, Q., Sattegeri, P., Szwarcman, D., Watson, C., and Rolnick, D.: Hard-Constrained Deep Learning for Climate Downscaling, J. Mach. Learn. Res., 24, 1–40, 2023.