Towards Cost-effective Convection-Permitting Simulations for Ireland using Deep Learning

High-resolution convection-permitting climate simulations are essential for assessing climate risk at the national scale, but their high computational cost limits ensemble size and uncertainty sampling. In Ireland, national climate projections rely on multi-GCM, multi-RCM ensembles dynamically downscaled at considerable expense. Using the HARMONIE-Climate (HCLIM) model, we show that precipitation and temperature characteristics are comparable between 3 km and 5 km resolutions for ERA5-driven downscaled simulations produced using a nested GCM → 12 km HCLIM → CPM HCLIM approach over an Ireland–UK domain. This indicates that intermediate-resolution simulations can serve both as a cost-effective approach and alternatively as a basis for refinement using deep learning. Building on this result, we present initial findings from a deep learning model developed to emulate ≤3 km fields from 5 km ERA5-driven simulations, with a view to assessing whether this approach can provide high-resolution convection-permitting simulations more cost effectively for use in national climate projections.