Weather forecasting, which predicts the state of the atmosphere only up to about 2 weeks in advance, often does not involve the development of the ocean state. Hence, the ocean surface condition is often kept constant during the forecasts in weather prediction models or just slightly adjusted with a climatological trend for the forecast period. Thus, in ICON-LAM, there is only option to update the sea surface temperature (SST) and sea-ice cover monthly. For ICON-CLM, we want to have a flexible option to update of SST and sea ice from external data at a user-defined interval. For this purpose, an option for flexible update frequencies of these boundary conditions was implemented in ICON-CLM. Time-dependent SST and sea-ice data can now be read from external data files and are fed to ICON-CLM with shorter intervals than 1 month (e.g. 1-hourly or 6-hourly). The user can select this option of frequent update of SST and sea ice via namelist settings. The external SST and sea-ice data must be prepared and remapped to the ICON grid.

Similarly, the greenhouse gas (GHG) values are usually kept constant in weather forecast models, because the changes during the forecast period are negligible. In climate projections, however, it is necessary to use the time-dependent GHGs as provided by corresponding GHG scenarios. Such an option was already available in ICON but only in combination with the ECHAM physics package. The corresponding read routine was therefore extended so that it works for the NWP physics as well. Some additions to the NWP radiation scheme were made with respect to the GHG vertical profile with a new option to retrieve the profile from external gas data. A file that contains yearly values of CO2, CH4, N2O and chlorofluorocarbons (CFCs) for all years of the experiment needs to be prepared in advance. These features of the time-dependent SST and GHG were largely based on the corresponding implementations in ICON-A .

In the NWP configuration of ICON-LAM, the number of soil layers is always constant with eight layers. The depths of half soil layers are also fixed at values between 5 mm and 14.5 m. However, for climate simulations in domains other than Europe (e.g. Africa, Asia) or to achieve better simulation of the soil variables for the European domain, it is usually reasonable to adjust these soil parameters. Therefore, an option for a flexible number and depth of the soil layers has been implemented in the ICON-CLM code.

The input/output of ICON-CLM has also been adjusted to have more flexibility. In NWP mode, the precipitation data are accumulated from the start of the forecast till the end without any reset. This is suitable for short weather forecasts, but for long climate simulations, this procedure is inconvenient and could, in the worst case, cause problems due to data imprecision. Furthermore, the maximum and minimum 2 m temperature values are calculated for 6-hourly intervals in NWP applications, while for climate simulations the standard for these output variables is usually 24 h. To control this flexibility extension, new namelist parameters were introduced in ICON-CLM.

At the lateral boundaries, ICON-LAM requires, by default, information on cloud liquid water content and cloud ice water content from the global forcing data. These input fields are usually available if the ICON-CLM lateral boundary conditions are taken from reanalysis data like ERA-Interim. But if global climate projections are used as lateral boundary conditions, these fields are usually not provided. Thus, the model code has been adjusted so that if cloud liquid water content and cloud ice content are not available in the lateral boundary data, these variables are initialized with zero.

Table provides an overview of some differences between COSMO-CLM and ICON-CLM.

In order to facilitate long-term climate simulations, we developed a runtime infrastructure called starter package and a separate evaluation tool. Both are provided along with the ICON-CLM model source code. The ICON-CLM_SP starter package contains a run routine, a climatological test suite, all necessary utilities and configure scripts for different supercomputing environments. At the moment, two system settings for Nec-Aurora (DWD) and Atos/Bull (DKRZ) are supported and tested (note that our ICON-CLM simulations in this paper were done on the DWD Cray XC40; this computer was replaced by the Nec-Aurora afterwards). Settings for other machines could be easily added if necessary.

The run routine in ICON-CLM_SP, called “subchain”, was adapted from the routine of the existing COSMO-CLM package. The “subchain” contains five subroutines for input preparation (prep), converting input data (conv2icon), ICON-CLM job management (icon), archiving (arch) and postprocessing (post) of the model output. Subroutine “prep” copies and checks all the global forcing data as input for “conv2icon”. Then “conv2icon” preprocesses and interpolates the initial data and the lateral, lower and upper boundary data onto the ICON-CLM model grid for the current model simulation. Subroutine “icon” does the job management for ICON-CLM model. After that, all model output data are compressed by “arch” and some post-processing steps like the provision of time series of selected output variables are done in “post”. Usually the simulations in ICON-CLM are done per calendar month with restarts at the end of each month; however, the simulation duration can be easily customized by changing in the subroutine “icon”.

A climatological test suite (CTS) was also created based on the CTS from COSMO-CLM. In the CTS, 5-year test simulations can be done automatically with “subchain”. The users can choose one simulation as a reference. The test simulations then will be compared with the reference simulation with respect to observational data (E-OBS and Climate Research Unit (CRU); see Sect.  for more details) by an extra subroutine called “eval”. At the end, the results are visualized with standardized plots. This CTS was built for the purpose of testing different versions of model source codes or different setups of the same model version. Hence, it is a very helpful tool for model development and tuning.

Besides the “eval” subroutine in CTS, a separate evaluation tool called “ETOOLS” was also adapted from the COSMO-CLM evaluation tool. This tool provides comparisons of the simulation results with observation datasets and creates standardized plots to visualize the results. In order to facilitate the transition from COSMO-CLM to ICON-CLM for the users, both ICON-CLM_SP and ETOOLS were created such that the “look and feel” as well as the usage of the software packages are as similar as possible to the corresponding packages that exist for the COSMO-CLM model. The output structure of ICON-CLM or post-processed time series from “subchain/post” are also similar to those of COSMO-CLM for the same reason. On this account, users should be able to use all existing scripts and programmes that were developed for COSMO-CLM output also for ICON-CLM data.

All of the ICON-CLM tests described afterwards and the evaluation run were performed at R2B8 resolution (approximately 10 km) on a domain (Fig. a) completely covering the EURO-CORDEX domain . The model atmosphere is divided into 70 vertical layers and the model top is at a height of 30 km or 10 hPa. The soil in ICON-CLM contains eight layers down to a depth of 14.5 m. ICON-CLM was driven at the lateral and lower boundaries by ERA-Interim at 6-hourly intervals. The model atmosphere was initialized with ERA-Interim data. The soil temperature and soil moisture were taken from a previous simulation which is long enough to ensure that the spin-up of the soil has been completed. The monthly Tegen aerosol climatology and the ozone climatology from Global Earth-system Monitoring using Satellite and in situ data (GEMS) were used in our simulations. For the upper boundary, as described in Sect. , there are two options: (1) using the driving data and nudging gradually in the relaxation zone; (2) damping the vertical wind beneath the upper boundary. To assess the impact of these two options, two 10-year simulations (1979–1988) were done with the same setup, with and without global data nudging. Analysis from these 10-year runs resulted in very minor differences on surface variables and none of the options showed any advantage over the other. For the evaluation run, we chose the option with nudging data at the upper boundary with ERA-Interim data, as later we wanted to compare the results with those from a COSMO-CLM run using a similar nudging option. The nudging zone started from the height of 12 km to the model top of atmosphere (30 km). Variables which are nudged within this layer are the horizontal wind and the thermodynamic variables (air pressure and temperature) with nudging coefficients of 0.04 and 0.075, respectively.

In order to find a suitable model configuration for ICON-CLM at R2B8 resolution, an optimized namelist setup was used, namely the setup from ICON-NWP for R3B7 with nested domain on R3B8 grid (approximately 13 and 6.5 km, respectively). The tuning parameters were taken over from the global settings. In this setting, the Tiedtke–Bechtold convection parameterization scheme and the Rapid Radiation Transfer Model (RRTM) radiation scheme were used. These setups were checked to make sure that they were appropriate for climate applications and were used in all simulations.

Former simulations (not published) with COSMO-CLM showed that in some cases the model results depended on the chosen model time steps. There was one particular time step that led to larger biases in precipitation and surface pressure, especially over the Alps and the southwestern area of the EURO-CORDEX domain. This issue has been analysed by the CLM-Community but is still not fully understood yet. To ensure that such a dependency of the results on the model time step is not present in ICON-CLM, different fast physics/advection time step (hereafter time step) choices were tested. At R2B8 (approximately 10 km) resolution, the time step should not exceed 120 s for stability reasons. With the common model and experiment setups described above, we carried out multiple 1-year simulations for the year 1979 with time steps of 60, 80, 90, 100 and 120 s. Figure  shows the biases compared to reference data of 2 m temperature, total precipitation, MSLP and total cloud cover. The biases were averaged for each month and for the Alpine region (subregion denoted AL in Fig. c). Colours show the different time step experiments. The biases for all variables from any particular time step experiment were small and did not stand out from the rest. Similar results were found for all other subregions and therefore are not shown here. The annual and seasonal biases of these multiple 1-year simulations were also very similar to all time steps (not shown here).

Because there is no big difference in the model results depending on the choice of time step, for experiments at the R2B8 spatial resolution, we chose the time step of 90 s due to the computational efficiency and stability, the time step of COSMO-CLM at similar horizontal resolution is 100 s. An evaluation run was carried out for the EURO-CORDEX domain at the R2B8 resolution. The simulation period is 1979 to 2000. The model and experiment setups were the common setups as described above, this evaluation run is later referred to as ICLM-REF.

The results of ICLM-REF were compared to the reference experiment of the recent recommended version of COSMO-CLM (v5.0_clm9). This COSMO-CLM simulation is later referred to as CCLM-REF. This COSMO-CLM setup showed the best performance for the EURO-CORDEX domain in an intercomparison with a large number of setups which was performed within the COPAT project (COordinated Parameter Tuning, a project providing a thorough evaluation of a large number of COSMO-CLM configurations to come up with a recommended version) in the CLM-Community. The simulation period is 1979 to 2000; the model resolution is 0.165∘, also about 10 km like in ICLM-REF. The initial, lower and lateral boundary data are taken from ERA-Interim. A sponge layer with Rayleigh damping in the upper levels of COSMO-CLM domain was used. The damping was done against the external boundary values, similarly to the nudging at the model top in ICLM-REF simulation, and thus the results from both experiments are comparable.

For model assessment and evaluation, output fields from six variables were analysed, namely 2 m temperature, daily maximum and minimum values of 2 m temperature, MSLP, total precipitation and total cloud cover. Monthly average values of these variables were calculated and used for further analysis. For total precipitation, the monthly accumulated amounts were calculated. Parts of the evaluation were carried out based on seasonal averages. The following definitions and abbreviations of the seasons are used in this paper: winter – December, January, February (DJF); spring – March, April, May (MAM); summer – June, July, August (JJA); and autumn – September, October, November (SON). Results were averaged for eight subregions as already used in the Prediction of Regional scenarios and Uncertainties for Defining EuropeaN Climate change risks and Effects (PRUDENCE) projects (described by ). These subregions are the British Isles (BI), Iberian Peninsula (IP), France (FR), Mid-Europe (ME), Scandinavia (SC), Alps (AL), Mediterranean (MD) and Eastern Europe (EA) (shown in Fig. c).

For the evaluation of mean values of 2 m temperature, daily maximum and minimum 2 m temperature, MSLP and total precipitation, the E-OBS dataset was used as reference data. E-OBS is a 0.25∘ gridded daily dataset covering all of Europe. The data are available over land from quite recently back to 1950. In order to compare 2 m temperature data from different datasets (ICLM-REF, CCLM-REF and E-OBS), a height correction was performed. The model temperature values at E-OBS 2 m height were calculated based on the differences between model and E-OBS surface elevation and the moist adiabatic lapse rate (0.0065 K m-1). The reference cloud data, which were used for assessment of the model cloud cover, are CRU time series (TS) data . This is a monthly gridded dataset at 0.5∘ resolution, available globally over land area. The dataset covers the period from 1901 to 2013.

For all ICON-CLM simulations in this paper, the outputs were written in NetCDF format and on the rotated lat–long grid as in CCLM-REF. Because this rotated lat–long grid is finer than E-OBS and CRU grids, these data were remapped to the regular lat–long grids with the same spatial resolution as the observational data for the purpose of comparison. The E-OBS and CRU datasets contain data only over land; therefore, the evaluation in this paper (e.g. areal averaged fields) were done using only land points.

The evaluation within COPAT for 2 m temperature, MSLP and cloud cover was done using E-OBS version 10.0 and CRU version 3.22 as reference data. Therefore, in order to compare our evaluation to the one from COPAT, we also used the same versions of the datasets. The comparison period is 20 years from 1981 to 2000, the same as the evaluation period in COPAT. The reference total precipitation data were taken from E-OBS version 12.0 because this dataset (among versions from 10.0 to 17.0) shows the fewest missing data for precipitation over the area of Poland.

Some important climate indices (listed in Table ) were calculated from ICLM-REF, CCLM-REF and E-OBS 2 m temperature and total precipitation data and averaged over the period 1981–2000. The number of days that fulfils the definition (in Table ) was counted for each horizontal grid cell, then averaged over a subregion.

Root-mean-square error (RMSE) was calculated from the model and observed monthly values (monthly aggregated values for precipitation and monthly mean values for other variables): 3RMSE=1N∑i=1n(Si-Oi)2, where Si and Oi are the model (ICLM-REF or CCLM-REF) and reference data (E-OBS or CRU TS) monthly values, respectively, averaged over the subregions considered at the ith month; N is the total number of months in the evaluation period (1981–2000).

To compare the spatial variability of the data, spatial standard deviation (SD) was also calculated from time-averaged fields of model and observed data. The SD ratio (SDmodel/SDobservation) was calculated for the ease of comparing the spatial variation of the two model data with respect to the observation. The model with SD ratio closer to 1 better represents the spatial variation of the observation data.

The results of air temperature from ICLM-REF were also compared to those of other regional climate models (RCMs) within the EURO-CORDEX experiments. A thorough evaluation of several EURO-CORDEX ensembles is presented in . One ensemble, called EUR-11, has quite a similar setup to our ICLM-REF setup. The horizontal resolution of the RCMs in this EUR-11 ensemble is about 12 km; these simulations were also driven by ERA-Interim. The simulation period, however, differs from ours: 1989–2009 instead of 1981–2000. Nevertheless, this comparison for sure gives us some knowledge about ICON-CLM performance relatively to other state-of-the-art RCMs. In Figs. 5 and B1 in , some similar analysis to ours were done for nine different RCMs, seasonal means of temperature bias are shown for eight PRUDENCE subregions. These biases were also calculated against E-OBS data like in our evaluation, though an older version of E-OBS data was used in their paper. These figures were used to compare to our figures from ICLM-REF.

After the technical adaptation in the ICON model source code to enable long-term climate simulations, a number of technical tests was performed. First, the influence of different domain decompositions for parallelization on the simulation results was tested on the Cray XC40 supercomputer at DWD. The domain of ICON models is split by a built-in geometric subdivision according to the number of processors used for calculation. Tests with ICON-CLM were done using 120, 240, 480 and 960 computer processors; each test simulated the year 1979. As these tests are 1-year long each, they contain multiple monthly restarts of the model. The model outputs were checked for all climate variables evaluated in this paper. Results were binary identical, independent of the numbers of processors used for the simulating. ICON-CLM showed the ability to produce the same results and to restart properly with a variety of domain decompositions.

Repeating tests starting from the same restart state were also carried out. ICON-CLM was restarted at the time point on 1 January 1980 at 00:00 UTC multiple times from the same restart file on the Cray XC40. Each simulation was carried out for 1 month from 1 to 31 January 1980. Model outputs were checked for all model calculation steps and for all climate variables evaluated in this paper. In addition, the restart files created at the end of each test for the simulation time on 1 February 1980 at 00:00 UTC were also compared. The results from these repeating tests showed binary identical values. Two additional restart tests were also done. In the first test, ICON-CLM was run for 2 months from 1 January to 29 February 1980 without restarting in between. And in the second test, ICON-CLM also simulated these 2 months but was restarted on 1 February 1980 at 00:00 UTC. The model outputs were compared, with a focus on the period after the restart of the second test. Results showed that restarting did not introduce any difference in model diagnostics.

Short tests up to few months were also done on a different computing system, the Atos/Bull at the German Climate Computing Centre (DKRZ). ICON-CLM showed the ability to run stably on at least two machines. Performance tests were done for ICON-CLM and COSMO-CLM on the Atos/Bull. The two models were run in 1-month simulations on the European domain. The horizontal resolution is roughly 50 km; the same number of computer processors was used. ICON-CLM was about 15 % faster than COSMO-CLM in these tests.

Initially, a longer simulation with ICON-CLM (30 years) was planned and carried out. But since the data from CCLM-REF are only available to us in the period 1981–2000, only data in this period were taken for evaluation.

As the limited-area models are only forced by the global data at their boundaries, inside the limited domain, the regional climate models can more or less freely develop their own circulation. One of the major concerns in regional climate modelling is to what extent the regional climate models modify the large-scale information from the global data .

We want to test if ICON-CLM can be able to reproduce the large-scale atmospheric condition from the ERA-Interim forcing data. The geopotential at 500 hPa data from ICLM-REF and ERA-Interim were averaged at different temporal scales: 6-hourly, daily, monthly, seasonal and yearly. Due to the different grids and ERA-Interim lower resolution, ICLM-REF data were remapped to ERA-Interim grid. The spatial correlation coefficient was calculated for each time slot within the evaluation period (1981–2000).

Time series of the correlation coefficient for the different temporal scales are shown in Fig. . For all of the timescales we considered, results showed a higher correlation for longer timescales; the correlation is lowest for 6-hourly data and strongest for yearly data. For the 6-hourly and daily means, the mean correlation coefficient over the whole time period (1981–2000) was high: 0.925 and 0.928, respectively. But the correlation for some time slots dropped below 0.8, causing the minimum values of correlation coefficient only over 0.5. ICON-CLM showed correlation coefficients above 0.85 for longer timescales from monthly to yearly scales. With this result, we can conclude that ICON-CLM reproduced well the large scale of the driving data for timescales from monthly to yearly and partially for timescales from 6-hourly to daily.