The first long-range simulation of the general circulation of the atmosphere dates back to 1956 . Today it is still the primary tool for global climate projections. However, due to limited computer resources, the current horizontal resolution of 100–200 km is still too coarse to resolve sufficient detail for regional climate projections. Finer spatial resolution that resolves the land surface heterogeneity can be obtained by nesting a regional climate model (RCM), over a smaller domain, within a coarse-resolution global climate model (GCM). This is also referred to as dynamical downscaling. A GCM or global reanalysis data product provides the large-scale meteorological and surface fields to the RCM as initial and lateral boundary conditions. The global features are thus translated into regional and local conditions over the region of interest . Hence, RCMs enable climate simulations over a smaller domain with finer-scale horizontal resolution and with less expensive computational cost than running a GCM at the same resolution.

Since the late 1960s, the numerical weather prediction (NWP) community has used high-resolution limited area models. The numerical approach was first used for a regional climate simulation by . Their climate simulation used the NWP model in forecasting mode with 3–5 daily reinitialisations of the initial conditions. To be able to run them without these short-term reinitialisations, the regional climate community applied monthly to multidecadal simulations, with only a single initialisation of the large-scale fields and frequent updates of the lateral boundary conditions . These long-term continuous simulations required improvements in the representation of physical processes in the RCMs. The continuous simulation is still the most common in the RCM community . However, by applying the continuous approach, the simulated large-scale fields deviate from the driving lateral boundary conditions .

The accuracy of the dynamical downscaling has improved by using short-term reinitialisations to reduce systematic errors . However, only few in the RCM community adopted this method, mainly because of its higher computational costs. Most studies dealing with the evaluation of reinitialised versus continuous climate simulations, examined only short time periods of 1 month to 1 year. Using a daily reinitialisation, showed improvements in the prediction of precipitation for a case study of a large flooding event in the Elbe river catchment in August 2002. Changing the period of reinitialisation, from monthly to 10 daily, a reduction in systematic errors has been shown for precipitation when using the 10-day reinitialisation . Even in a 20-year RCM simulation forced by reanalysis data, the sequence of events was better preserved by using daily reinitialisations .

A model approach with short-term reinitialisations demands additional simulation time at each reinitialisation start. This time is required to reach dynamical equilibrium between the lateral boundary conditions and the internal model physics and dynamics . Beyond 24 h small perturbations in the initial conditions of the atmosphere have only limited impact on the simulation potential . In contrast to the atmosphere, the surface takes a longer time to reach dynamical equilibrium with the overlaying atmosphere, from a few weeks to several seasons, depending on the depth of the soil layer.

The surface interacts with the climate through the soil moisture and soil temperature, by influencing the surface energy budget . The soil moisture controls the partitioning of the incoming energy into latent and sensible heat flux. The soil moisture limitation on the evapotranspiration is largest during summer . The availability of soil moisture for evapotranspiration is determined by the 2 m temperature . As the land surface–atmosphere interactions play a crucial role in the representation of the current and future climate , it is important to validate the model with observations. In-site measurements can provide valuable estimates of the surface energy fluxes. More specifically, FLUXNET establishes a global network of eddy-covariance towers measuring these fluxes .

The objective of this study was to evaluate the simulation potential of three regional climate downscaling approaches with different update frequencies of the initial conditions: (1) a continuous simulation of both the atmosphere and the land surface, (2) a simulation with daily reinitialisations for both the atmosphere and the land surface and (3) a simulation with daily reinitialisations of the atmosphere while the land surface is kept continuous. We used the ALARO model to dynamically downscale the European Centre for Medium-Range Weather Forecasts (ECMWF) Interim Re-Analysis (ERA-Interim; ). Within this study, ALARO was coupled to the Météo-France SURFace EXternalisée land surface model (SURFEX; ). We evaluated the mean 2 m temperature and mean daily total precipitation by comparing with the 0.22∘ ECA&D E-OBS dataset , and the surface energy fluxes by comparing with the FLUXNET database . The analysis covered a 10-year period from 1991 to 2000, for a domain encompassing western Europe.

The models, experimental design and observational datasets are described in Sect. 2. The results for the mean surface parameters are covered in Sect. 3. Section 4 demonstrates the results with respect to the surface energy budget. Finally, conclusions are given in Sect. 5.

The spatial distributions of the 5-year daily maximum BR, at the time of maximum H and LE, of CON, DRI and FS were compared to FLUXNET observations, for the summer period only (Fig. a–c). The corresponding spatial distributions of the 10-year daily maximum BR of CON, DRI and FS were evaluated with respect to the results for the 5-year period (Fig. d–f). The mean diurnal cycles of the surface energy fluxes are illustrated over the 5-year summer period from 1996 to 2000 for the FLUXNET stations of Vielsalm and Collelongo and their corresponding model grid points (Fig. , Table ).

The daily maximum BR showed a strong gradient of increasing values towards the south of the domain (Fig. a–c). Recall that when the value is lower (higher) than 1, the latent heat flux (LE) is higher (lower) than the sensible heat flux (H). Southern Europe is characterised by dry summers, with a strong control of the soil moisture on the evapotranspiration , and thus lower LE than H. Large differences existed between the three downscaling approaches, particularly for the continental part of the domain. This is in agreement with the larger differences between the simulation modes in 2 m temperature for the continental subdomains compared to the southern subdomains. Therefore, the 2 m temperature is correlated to evapotranspiration by the sensitivity of LE on the soil moisture .

CON had relatively low BR values of 0 to 1 for the continental area, DRI showed BR values of 0.5 to 1 and FS had the highest values of 2 to 3. The larger values presented by FS were in agreement with the warm and dry summer bias. Comparison with the FLUXNET observations showed a general underestimation of BR for CON and DRI, whereas FS estimated well the BR for most of the sites. The FLUXNET sites show typically a closure imbalance of 20 % , whereas it was assumed that the BR is well estimated by the eddy-covariance system . Though this validation was based on five summer periods only from 1996 to 2000, it was still robust as indicated by the corresponding BR for the 10-year summer period from 1991 to 2000 (Fig. d–f, Table ). Considering the two FLUXNET sites Vielsalm and Collelongo, DRI indicates the best agreement with the observed BR (Fig. b), with values of 0.78 and 1.35 compared to 1.12 and 1.32 respectively (Table ). Despite higher BR values in FS, estimates corresponding to Vielsalm and Collelongo were located in regions where the BR was comparatively low, resulting in average values of 0.61 and 0.83 respectively (Table ).

The model performed well for the simulation of the daily cycle of the net radiation (RN; not shown), even though the model underestimated the values of RN by ca. 5–10 % (Table ). The model generally underestimated H, and overestimated LE (Fig.  and Table ). The ground heat flux (G) showed improbably high values compared to the observed ones (Table ). G is dependent on the soil temperature, which was largely overestimated by the land surface model (not shown). The standard version of ISBA, the nature tile of SURFEX, aggregates soil and vegetation properties for each grid cell . The net radiation is directly transferred to the ground, causing an inaccurate partitioning of the incoming energy into turbulent and ground heat fluxes . An additional parameterisation for the leaf litter on the surface soil impacts this distribution . We suggest to include an explicit formulation of the canopy layer and potentially a parameterisation for the forest litter layer . The implementation of these explicit formulations in ISBA outperformed the representation of the soil temperature of the original ISBA model . They showed that the original ISBA model overestimated the G flux amplitude with several tens of W m-2 during both daytime and nighttime. However, using the distinct surface energy budgets resolved part of the overestimated G by intercepting most of the downward solar radiation, leaving more energy available for turbulent fluxes. Consequently, less net radiation reaches the forest surface, reducing the energy available for the soil conductance .

For Vielsalm, H was simulated well by DRI and FS during nighttime and daytime, whereas CON underestimated H during daytime (Fig. a). The daily maximum H by CON was only 118 W m-2, as compared to 151 and 139 W m-2 for DRI and FS respectively (Table ). This validation was only done for five summer periods from 1996 to 2000, but the corresponding daily maximum values for the 10-year summer period from 1991 to 2000 indicate that the 5-year period was representative for the validation of the fluxes (Table ). The LE was overestimated by all simulations, but the difference in the observations was smallest for DRI, due to the frequent land surface reinitialisations, compared to the highest values for CON. The daily maximum BR was lower than 1 for all downscaling approaches (Table ). This means that they all simulated a higher latent than sensible heat flux. Still, DRI and FS showed higher values for BR than CON. Therefore, the partitioning of the surface energy fluxes was better represented by DRI and FS for the station of Vielsalm.

For Collelongo, the model underestimated H during daytime and overestimated H during nighttime, except for DRI, which demonstrated a good agreement with the observations. Consequently, the model overestimated LE during daytime, except for DRI. The daily maximum H for DRI of 247 W m-2 was close to the observed value of 253 W m-2, whereas CON and FS simulated much lower values of 159 and 197 W m-2 respectively (Table ). CON showed the largest LE bias. Regarding BR, the simulation by DRI with a value of 1.35 was in very good agreement with the observations. The DRI simulation resulted in the least biased partitioning of the surface energy fluxes at Collelongo. However, FS showed that frequent atmospheric reinitialisations can add value.

In summary, the model presented a good spatial variability of BR, and the agreement with observations was highest for FS. However, when focusing on the sites that represent at least 50 % of the land cover at the corresponding model grid cell, it was DRI that performed well for H at Vielsalm and for LE at Collelongo. For Collelongo, this resulted in the least biased simulation of the partitioning of the surface energy fluxes, translating into a better BR estimate. The use of a daily reinitialised atmosphere and land surface improved the correct partitioning of the surface energy fluxes, whereas the continuous land surface initialisation as in FS could not improve the representation of the surface energy fluxes. The high values of G were caused by the large overestimation of the soil temperature and could be solved by implementing additional parameterisations for the canopy layer and soil surface layer. This implementation could alter the energy budget available for the turbulent fluxes, but this lies outside the scope of this study.

An assessment of three downscaling approaches has been performed using the regional climate model ALARO-0 coupled to the land surface model SURFEXv5, with lateral and initial boundary conditions from ERA-Interim. The simulations were applied for a 10-year period from 1991 to 2000, for a western European domain. The performance of ALARO-0 with SURFEX has already been validated for NWP applications , and here we present an evaluation for long-term climate simulations.

We compared the commonly used approach of a continuous climate simulation with two alternative methods of frequently reinitialising the RCM boundary conditions, combined with either a daily reinitialised or continuous land surface. The use of a daily reinitialised atmosphere outperformed the continuous (CON) approach for winter and summer 2 m temperature, and deteriorated the summer precipitation. However, the use of a continuous land surface (FS) with a daily reinitialised atmosphere improved the summer precipitation relative to the full continuous approach. Furthermore, it improved the winter 2 m temperature, whereas it resulted in a neutral impact on the summer 2 m temperature and the winter precipitation, despite a slight deterioration over the Mediterranean. The SSTs were reinitialised daily together with the atmosphere, as compared to the monthly updated SSTs in the continuous approach.

The seasonal cycle of the 2 m temperature and precipitation was different for three selected subdomains that covered large climate variability. Both the temperate climate of mid-Europe and the continental climate of eastern Europe indicated more seasonal variability than the Mediterranean climate of the Iberian Peninsula. The simulation of the 2 m temperature had improved when applying daily reinitialised atmosphere with continuous land surface, despite an overestimation of the summer 2 m temperature. Precipitation biases were larger and are perhaps associated with the tendency for ERA-Interim to be wetter than E-OBS, and the low spatial coverage by the observations in some regions. It was clear that the agreement for the precipitation between the model and the observations was highest during summer, while other seasons showed stronger deviations.

During summer, the interaction between the land surface and the overlaying atmosphere is largest. The 2 m temperature interacts with the soil moisture and influences the partitioning of the surface energy fluxes. The daily reinitialisation of the atmosphere improved the correct partitioning of the latent and sensible heat flux, although the biases were still quite large to be conclusive. Still, this approach outperformed the use of a continuous simulation. For a more comprehensive analysis, future research will consider including more FLUXNET stations. A more in-depth analysis on the interaction between 2 m temperature, precipitation and surface energy fluxes can reveal soil–moisture–temperature coupling , but this lies outside the scope of this study.

In conclusion, this study demonstrated that the approach of a daily reinitialised atmosphere was superior over the full continuous approach. The use of a continuous surface next to a daily reinitialised atmosphere improved the winter temperature and summer precipitation. We recommend using FS in a set-up with GCM forcing for climate simulations with ALARO-0.

The used ALADIN codes, along with all related intellectual property rights, are owned by the members of the ALADIN consortium. Access to the ALADIN System, or elements thereof, can be granted upon request and for research purposes only. The used SURFEX codes are freely available, together with the ECOCLIMAP database, at http://www.cnrm-game-meteo.fr/surfex///spip.php?rubrique8.

This study is based on large datasets written in .FA and .lfi format. The relevant output is exported to R datasets. Due to licensing restrictions, this model output is not made publicly available. However, for the purpose of the review, the data can be made available for the editor and reviewer upon request by contacting Julie Berckmans.