Deep-learning downscaling of CMIP6 heatwave projections for Greece and comparison with a regional climate model

Abstract. The Mediterranean region is a climate change hotspot with an increasing frequency and intensity of heatwaves. In Greece, complex topography creates microclimates, so heatwave properties vary at scales that global climate models (GCMs) do not resolve. In this study, we downscale five CMIP6 GCMs to about 5 km with a deep learning (DL) model trained on regional reanalysis, bias correct the output, and project heatwaves for 2051–2080 under high (SSP5-8.5) and moderate (SSP2-4.5) emissions. Under SSP5-8.5, the number of heatwaves per season rises from about one in 1985–2014 to about seven, and they last about a week longer. Even under SSP2-4.5, a season with as many heatwaves as the record-warm 2024 would be below average. The increases are largest over northern Greece and smallest over Crete. This pattern reflects the GCM warming and persistence and the local variability of the reanalysis used for bias correction. The bias correction changes the projected patterns more than the choice of downscaling method. Bias-corrected interpolation yields nearly the same 30-year statistics as the DL downscaling, although the latter gives more accurate daily fields. The regional climate model ICON-CLM projects fewer and shorter heatwaves, mainly because it warms 13% less than its driving GCM.