¿Can regional climate models reduce global model large-scale biases? A climate ensemble approach for downscaling cut-off lows using the HCLIM model

Dynamical downscaling with regional climate models (RCMs) may inherit biases from the boundary conditions of a reanalysis or a general circulation model (GCM), leading to flawed results in the higher-resolution output. Typical examples of large-scale biases in CMIP5/6 models are the “too equatorial and too zonal” North Atlantic storm track (Schemm, 2023) or the bias in location and frequency of cut-off lows (COLs) (Pinheiro et al., 2022). These large-scale deviations may be even more noticeable when the geographical domain of interest is small, such as islands (Adinolfi et al., 2025). Even when using reanalysis as boundary conditions, some events may be lost after downscaling (Lavin-Gullon et al., 2021).

Although the primary objective of an RCM is to develop the smaller scales, if the RCM integration domain is large enough it can resemble the behaviour of a global model and the finer resolution potentially may correct some of the large-scale bias (Diaconescu & Laprise, 2013)

This study focuses on the optimal configuration for simulating COL events in convection-permitting scales using ensembles. The geographical area of interest is Southern Europe -including the Mediterranean and the eastern Atlantic coast-, one of the three areas of higher COLs occurrence worldwide (Nieto et al., 2005). Frequently the presence of a COL is related to high precipitation, accounting for about 80% of the extreme rainfall in the Valencia region of Spain (Nieto, 2021).

A convection-permitting resolution experiment is proposed, using the habitual double nested approach in RCMs, but defining an intermediate domain size appropiate to capture the possible weather regimes in an ensemble. The ensemble consists of several runs with different initial conditions, each lagged by one day. Results are shown for some COL specific events over Canary Islands and Iberian Peninsula using the regional model HCLIM (Belušić et al., 2020), demonstrating that in certain cases this approach can better reproduce an event that is not properly captured by the large-scale model.

Although a climate ensemble approach requires more computational resources, it can be appropriate for shorter runs, as in certain pseudo global warming approaches or event-based downscaling. It can also be used for longer simulations by taking advantage of hybrid techniques: physical models for the intermediate runs (that require relatively low HPC resources) and ML/AI emulators for the convection-permitting resolution.