Performance of the Arctic CORDEX simulations over land analysed from comparison with in situ observations

The current phase of Arctic CORDEX, facilitated by the PolarRES project, provides an ensemble of five regional climate models downscaling two CMIP6 global climate simulations selected using a novel storyline approach. The simulations are provided by MAR (University of Liège), ICON (Alfred Wegener Institute), RACMO (University of Utrecht), HCLIM (Metrological Institute of Norway) and WRF (NORCE Norwegian Research Centre). The models provide a hindcast simulation driven by ERA5 data for the period 2001-2020 and scenario simulations for the period 1985-2100 driven by global CNRM-ESM2-1 and NorESM2-MM simulations.

We use the Global Summary of the Day (GSOD) land station data set to analyse the performance of the regional models for five different variables: daily minimum, mean and maximum temperature, daily mean wind speed and daily mean sea level pressure. We use percentiles of seasonal distributions as a measure of comparison, as well as direct comparison of daily time series between models and observations. For precipitation, we use the Global Historical Climate Network land station dataset to analyse model performance based on seasonal mean precipitation. 

In the ERA5-driven hindcast simulations, averaged daily root mean square errors show high agreement with station data for both daily mean wind speed (below 2.5 m/s) and daily mean sea level pressure (below 4 hPa), with similar performance for all seasons and percentiles. Except for MAR, temperatures deviate less than 3 K, with the best performance in summer and autumn. The comparison of the percentiles between the ERA5-driven simulations and the averaged daily root mean square errors are in broad agreement.

The ERA5-driven simulations clearly outperform the GCM-driven simulations for all percentiles, seasons and variables in all models for temperature, wind speed and sea level pressure. However, the GCM-driven simulations also outperform the GCMs themselves, again for all percentiles, seasons and variables. Neither the bias patterns of the ERA5-driven simulations nor the bias patterns of the GCMs themselves translate directly into the bias patterns of the GCM-driven RCM simulations, highlighting the added value of the regional simulations. Comparing the pan-Arctic performance of the RCMs, we find that for temperature and mean sea level pressure, the 10th percentile is most often the worst fit compared to observations, with an improvement in performance towards the 90th percentile, independent of forcing, in autumn, winter and spring. For wind speed, however, performance is most often best at the 10th percentile and declines towards the 90th percentile.

For precipitation, the ERA5-driven simulations also tend to be closer to the observations than the GCM-driven simulations, but there are exceptions. For example, in autumn, the NorESM-driven WRF simulation outperforms the ERA5-driven simulation, and the CNRM-driven MAR simulation also outperforms the ERA5-driven MAR simulation.