Climate model downscaling in central Asia: a dynamical and a neural network approach

High-resolution climate projections are essential for estimating future climate change impacts. Statistical and dynamical downscaling methods, or a hybrid of both, are commonly employed to generate input datasets for impact modelling. In this study, we employ COSMO-CLM (CCLM) version 6.0, a regional climate model, to explore the benefits of dynamically downscaling a general circulation model (GCM) from the Coupled Model Intercomparison Project Phase 6 (CMIP6), focusing on climate change projections for central Asia (CA). The CCLM, at 0.22 degrees horizontal resolution, is driven by the MPI-ESM1-2-HR GCM (at 1 degrees spatial resolution) for the historical period of 1985-2014 and the projection period of 2019-2100 under three Shared Socioeconomic Pathways (SSPs), namely the SSP1-2.6, SSP3-7.0, and SSP5-8.5 scenarios. Using the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) gridded observation dataset as a reference, we evaluate the performance of CCLM driven by ERA-Interim reanalysis over the historical period. The added value of CCLM, compared to its driving GCM, is evident over mountainous areas in CA, which are at a higher risk of extreme precipitation events. The mean absolute error and bias of climatological precipitation (mm d-1) are reduced by 5 mm d-1 for summer and 3 mm d-1 for annual values. For winter, there was no error reduction achieved. However, the frequency of extreme precipitation values improved in the CCLM simulations. Additionally, we employ CCLM to refine future climate projections. We present high-resolution maps of heavy precipitation changes based on CCLM and compare them with the CMIP6 GCM ensemble. Our analysis indicates an increase in the intensity and frequency of heavy precipitation events over CA areas already at risk of extreme climatic events by the end of the century. The number of days with precipitation exceeding 20 mm increases by more than 90 by the end of the century, compared to the historical reference period, under the SSP3-7.0 and SSP5-8.5 scenarios. The annual 99th percentile of total precipitation increases by more than 9 mm d-1 over mountainous areas of central Asia by the end of the century, relative to the 1985-2014 reference period, under the SSP3-7.0 and SSP5-8.5 scenarios. Finally, we train a convolutional neural network (CNN) to map a GCM simulation to its dynamically downscaled CCLM counterpart. The CNN successfully emulates the GCM-CCLM model chain over large areas of CA but shows reduced skill when applied to a different GCM-CCLM model chain. The scientific community interested in downscaling CMIP6 models could use our downscaling data, and the CNN architecture offers an alternative to traditional dynamical and statistical methods.