Scale-dependent biases in Alpine sub-daily areal precipitation extremes: added value of convection permitting models

Intense sub-daily precipitation is a key driver of flash floods and debris flows. Convection-permitting models (CPMs) have shown improved representation of short-duration and localized precipitation extremes compared to coarser regional climate models (RCMs), yet their evaluation is typically performed at the native grid scale, neglecting hydrologically relevant spatial aggregations. Here, we assess how well CPM simulations represent areal precipitation extremes over Switzerland across durations from 1 to 24 h and spatial scales from ∼ 10 to 5000 km 2 . We use 20 years (2005–2024) of hourly precipitation from Switzerland's high-resolution radar–gauge product as reference and analyse simulations from the CORDEX Flagship Pilot Study on Convection Phenomena, including nine CPMs (2–3 km resolution) and seven driving RCMs (12–25 km resolution). CPMs reproduce the observed spatial organization of short- and long-duration precipitation extremes more realistically than RCMs over complex terrain. For 1–3 h extremes, CPM bias in 20-year return levels strongly depends on the spatial scale, shifting from a ∼ 15 % underestimation at native resolution to near-zero bias at ∼ 400 km 2 and to ∼ 20 % overestimation at ∼ 4000 km 2 . RCMs consistently underestimate 20-year return levels across all spatial scales, with biases ranging from ∼ 40 % underestimation at native resolution (∼ 144 km 2 ) to ∼ 10 % underestimation at the largest aggregation scales (∼ 5000 km 2 ). For longer durations (6–24 h), both ensembles show an approximately scale-invariant positive bias of ∼ 15 %. Biases also increase with elevation, particularly for daily extremes, highlighting challenges in modelling complex terrain and in radar–gauge observations. Overall, CPMs offer important added value for representing short-duration extremes relevant to flash floods and debris flows in small to medium-size Alpine catchments (10–10 3 km 2 ), but scale-dependent biases must be accounted for when translating CPM outputs into flood-risk assessments at these scales.