Development of a forecast-oriented km-resolution ocean-atmosphere coupled system for Western Europe and evaluation for a severe weather situation

To improve high-resolution numerical environmental prediction, it is essential to represent ocean-atmosphere interactions properly, which is not the case in current operational regional forecasting systems used in Western Europe. The objective of this paper is to present a new forecast-oriented coupled ocean-atmosphere system and its evaluation. This system uses the state-of-the-art numerical models AROME (cy43t2) and NEMO (v3.6) with a horizontal resolution of 2.5 km. The OASIS coupler (OASIS3MCT-4.0), implemented in the SurfEX surface scheme and in NEMO, is used to perform the commu5 nications between models. The evaluation of this system is carried out using 7-day simulations from 12 to 19 October 2018, characterised by extreme weather events (storms and heavy precipitation event) in the area of interest. Comparisons with in-situ and L3 satellite observations show that the fully coupled simulation reproduces quantitatively well the spatial and temporal evolution of the sea surface temperature and 10 m wind speed. Sensitivity analysis to OA coupling show that the use of an interactive and high resolution SST, in contrast to actual NWP where SST is persistent and at low resolution, modifies the at10 mospheric circulation and the location of heavy precipitation. When compared to the operational-like ocean forecast, simulated oceanic fields show a large sensitivity to coupling. Forced ocean simulations highlight that this sensitivity is mainly controlled by the change in the atmospheric model used to drive NEMO (AROME vs. ECMWF IFS operational forecast). The oceanic boundary layer depths can vary by more than 40%. This impact is amplified by the interactive coupling and is attributed to positive feedback between sea surface cooling and evaporation. 15

. Simulation domain illustrated by the bathymetry [m] in NEMO (in blue) and by the orography [m] of the AROME model (in greenbrown colors). The lines indicate the boundaries of NEMO-eNEATL36 configuration (red) and of the AROME-Mercator domain (black).
For AROME-Mercator, the grey marine zones are always uncoupled (constant initial SST and null current are used, see text). The dashed lines indicate the boundaries of the actual operational configurations of AROME (AROME-France, 1.3 km-resolution in black) and NEMO over the Iberia-Biscay-Ireland (IBI) region (NEATL36, 1/36°-resolution in red).
(2019a) and based on the exact same developments as previously done in the MESO-NH model (Bouin and Lebeaupin Brossier, 2020a). Thanks to its 2.5 km horizontal resolution the deep convection is explicitly resolved while the shallow convection is parameterized with the Eddy Diffusion Kain Fritsch EDKF, (EDKF, Kain and Fritsch, 1990) scheme. The ICE3 one-moment 110 microphysical scheme of Pinty and Jabouille (1998) is used to compute the evolution of five hydrometeor species (rain, snow, graupel, cloud ice and cloud liquid water). Radiative transfer is based on Fouquart and Bonnel (1980) scheme for short-wave radiation and the Rapid Radiative Transfer Model (RRTM, Mlawer et al., 1997) for long-wave radiation.
The surface exchanges are computed by the SURFace EXternalisé (SURFEX) surface model (Masson et al., 2013) considering four different surface types: land, towns, sea and inland waters (lakes and rivers). Output fluxes are weight-averaged inside 115 each grid box according to the fraction of each respective tile, before being provided to the atmospheric model at every time step. Exchanges over land are computed using the ISBA (Interactions between Soil, Biosphere and Atmosphere) parametrization (Noilhan and Planton, 1989). The formulation from Charnock (1955) is used for inland waters, whereas the Town Energy
A similar coupling algorithm as Rainaud et al. (2017) and is used in this study and is only summarised here and in Table 1 for clarity. AROME-SurfEx sends to NEMO the net non-solar heat flux, the two components of the wind stress and the net freshwater flux computed for the sea tile only, which are then imposed at the surface boundary 160 condition of NEMO. The solar heat flux is also send to NEMO and is used to calculate the penetrative radiation in the ocean.
Contrary to Rainaud et al. (2017, but also Arnold et al. (2020), the possibility of exchanging atmospheric surface pressure was implemented in this study and is also exchanged interactively during the coupled simulations for the inverse barometer approximation. In return, NEMO sends to AROME-SurfEx, the sea surface temperature and the sea surface current components that then enter in the sea surface turbulent fluxes computation and in the atmospheric turbulence 165 scheme.
The remapping files needed to interpolate fields between NEMO and AROME-SurfEx with a distance weighted nearestneighbour interpolation method using four neighbours are created offline using OASIS tools. Where the ocean is masked because outside the NEMO domain (hashed area in Fig. 1), AROME uses a SST constant in time and equal to the one used at the initial time, and the surface currents taken are always equal to zero (see also Appendix A1). at Capet Curig. Heavy rainfall also occurred over Wales (Fig. 2b), in particular inland due to an orographic enhancement, with up to 219 mm in 36 hours recorded at Libanus (Powys) making Callum one of the most severe rainfall events across Wales in 180 the last 50 years (Kendon et al., 2019). Storm Callum had indeed strong impacts due to flooding, also because the wind peak coincided with high spring tides and led to large waves, with some coastal flooding, largely enhanced by the heavy rainfall.
Hurricane Leslie was a large, long-lived, and very erratic tropical cyclone over Atlantic. Followed by the National Hurricane  195 As described in Caumont et al. (2021) and Mandement and Caumont (2020), in the night of 14 to 15 October 2018 the Languedoc region in the south of France, was indeed affected by heavy rainfall caused by a regenerative multi-cellular convective system organised along a convergence line between the moist southerly low-level flow and a quasi-stationary cold front over south-western France along a mean sea level pressure (MSLP) trough that linked Leslie to a low located over Ireland over south-western France. During the evening and night of 14 to 15 October, a low rapidly deepened around the cold front and Portugal and just after Leslie's remnant was absorbed into Michael's remnant, following a brief Fujiwhara (1921) interaction.

210
This 7-day period was chosen as the weather situation encountered is known to foster large air-sea interactions, but also because both ocean and weather forecasts may exhibit a larger sensitivity to coupling in such conditions. This is analysed through different simulations in the coupled and forced modes that are described in the following Section.

Experiments
To evaluate the ocean-atmosphere coupling impact on the atmospheric and oceanic forecasts, four experiments were performed 215 and are detailed below and in Table 2.
The OA experiment is the ocean-atmosphere coupled forecast over 7 days, starting on 12 October 2018 00 UTC. The initial atmospheric conditions comes from the global IFS analysis of 12 October 2018 00 UTC and the lateral atmospheric forcing comes every 6 hours from the global IFS forecast starting on 12 October 2018 00 UTC. The ocean initial fields come from the combination, as described in 2.2, of the CMEMS IBI and GLO analyses (3D daily fields of the 11 October) and OBC for the 220 7 days come from the CMEMS GLO daily analyses. The ocean-atmosphere coupling period is set to 600 s, i.e. the fields are exchanged every 4 NEMO time-steps and 12 AROME time-steps.
The reference experiment for atmospheric forecast (ARO) is similar to the OA experiment except that, as uncoupled, (i) the SST is kept persistent in time and (ii) sea surface currents are not taken into account. Note that this ARO experiment is equivalent to one operational deterministic execution of AROME at Météo-France (called AROME-IFS), but with two adap-225 tations. First, the lateral atmospheric conditions frequency is changed to 6 hrs in order to be able to run over a 7-day period (against 42 to 48h for AROME operational forecasts). This was mandatory due to less frequent forecast outputs available for the longest-term ranges of IFS. And secondly, for consistency with OA, the initial SST field is the combination of the PSY4 and IBI36 SST fields (instead of the ARPEGE SST analysis for AROME-IFS). Thus, comparing ARO with OA allows to evaluate the ocean-atmosphere coupling impact, i.e. the effect of an interactive evolution of SST and the impact of taking currents into 230 account, on the weather forecast.
Two ocean-only experiments were also run. OCE-ifs is the simulation close to the operational mode of IBI36, i.e. the initial conditions consist in the combination of the CMEMS IBI and GLO analyses (3D daily fields of the 11 October) and OBC for the 7 days come from the CMEMS GLO daily analyses (similarly to the ocean component of OA). The atmospheric forcing uses the bulk variables from IFS (2 m-air temperature, 2 m-humidity, 10 m-wind components, rainfall, mean sea level pressure, 235 short-wave and long-wave solar fluxes) and the IFS bulk parametrization available in the NEMO surface scheme (meaning the SST evolution and sea surface currents are taken into account to compute the air-sea exchanges). OCE-aro is an intermediate simulation using the ARO (AROME) bulk variables as atmospheric forcing (the same bulk variables as for IFS are used except for the wind speed which is taken at 5 m, the height of first vertical level of AROME) and the COARE3.0 sea surface turbulent flux parametrization (Fairall et al., 2003) through SURFEX offline. Comparing OCE-aro with OA on one hand and OCE-aro 240 with OCE-ifs on the other permits to disentangle the ocean-atmosphere coupling effect on the ocean forecast from the impact of the atmospheric forcing change.

Simulation results
This section presents an evaluation of the coupled OA simulation (Section 4.1), the respective impacts of the high-resolution interactive atmosphere on the oceanic forecast (Section 4.2) and of the coupled ocean on the atmospheric forecast (Section

Evaluation of the OA coupled simulation
This section describes the OA coupled simulation and presents its evaluation in comparison with observations available at the sea surface and in the boundary layers. are visible with an associated sea surface cooling of up to 2.5 • C persisting during the 7 simulated days (Fig. 3d). This cooling is mainly due to oceanic vertical mixing processes enhanced by the strong wind produced by these storms. At the end of the 7 simulated days, the average temperature over the domain is 0.6 • C colder than initially with local differences varying up to 35% of the initial SST (cooler or warmer depending of the location). The maximum differences are located in the areas of influence of the storms (Atlantic ocean).

260
In Figure 4, the simulated sea surface temperature is compared to satellite observations coming from the Copernicus Marine Service portal (http://marine.copernicus.eu). This L3 SST is obtained from several satellite sensors which are combined together and interpolated on a regular 0.02 • grid, and is available every day with daily average. In order to be able to compare the simulated and observed SST fields, it is necessary to interpolate the simulated SST on the satellite observation grid taking into account the masked areas related to the presence of clouds and therefore where no satellite data is available ( Fig. 4a and   265 4d). Whether at the beginning or at the end of the simulation, the simulated SST is close to the observed SST with a mean bias of less than 0.1 • C. Differences can be noted in the position of oceanic structures, which leads to local differences in SST that  Table 2). In (a), the colour circles represent the SST measured by drifting buoys at that time ; B1 and B2 labels indicate the location of the two drifting buoys used in Figure 5. Black squares in (d) correspond to four extracted areas used for analyse in the next subsections.
can vary by up to ±4 • C. In addition, it can be noted that the simulated cooling present in the Celtic Sea is stronger than the observed one.
Temporal evolution of simulated sea surface temperature is also compared to in-situ observations (drifting buoys) available  The rapid and intense SST variations are also reproduced, as visible for B1 (Fig. 5a), related to the storm Callum, or for the diurnal cycle seen at B2 (Fig. 5b), on 12 and 18 October for example in OA, however with differences in terms of intensity with respect to observations.

280
In spite of local differences, the coupled simulation reproduces thus accurately the mean gradient, mesoscale structures and evolution of SST during the 7 simulated days.   Figure 3d. In (q,r,s,t) ARO wind speed is the same as the OCE-aro one. Since the ARO simulation does not take into account the SSS and SSH, they are not represented in this Figure. With respect to SSH variations (Fig. 6i,j,k,l), they are strongest in the Celtic Sea where the tidal amplitude is higher. The amplitude of these variations reaches 4 m and decreases over the 7 days, in relation to the decrease of the tidal coefficient from 95 on the 12 th to 30 on the 17 th (values for Brest harbour). In the Atlantic Ocean, the variation of SSH is also important with an amplitude of one meter, while its weaker in the North sea, due to a smaller amplitude of the tidal harmonics in this 305 area, leading also to a more variable signal related to interactions between these harmonics. In the Mediterranean sea, the SSH variations have the smallest amplitude (≈ 0.2 m), which are in fact mainly related to the presence of oceanic eddies.
In the coupled experiment (OA), the sea surface currents (SSC) are also exchanged. The spatial and temporal evolution of these currents are important during the 7 simulated days. Their intensity are maximum in the Channel, reaching more than 2 m.s −1 locally, due to tidal currents (not shown). Temporal evolution of SSC in the four extracted areas are presented in Fig   310   6m,n,o,p. SSC are maximum in Celtic and North Seas, reaching more than 0.5 m s −1 with intensities that vary with respect to the tides. For the Atlantic Ocean and Mediterranean Sea boxes, SSC intensity is less important but can reach up to 0.25 m s −1 .
The evolution of the ocean mixed layer is analysed more finely thanks to temporal evolution of temperature vertical profiles (Fig. 7). Black lines in Fig 7 correspond to ocean mixed layer depth (MLD). To compute this mixed layer depth, the potential density field is used: for each grid point, the value at 10m depth is taken as a reference, and the mixed layer depth is obtained

Wind
The OA simulated wind field is examined in Figure 9a,c,e and compared to in-situ wind measurements available in the Coriolis database. It is important to note that the wind observations are set at a height of 10 meters, thus we use a 10-m diagnostic wind 330 from AROME and not the pronostic 5-m wind values.
During the first simulated day (12 Oct., Fig. 9a), the storm Callum moves towards the British Islands, inducing strong wind (above 20 m.s −1 ) over a wide area affecting Portugal to United Kingdom. Locally, wind speed value reaches the maximum value of 41.5 m s −1 in the Celtic Sea. The comparison with data (circles in Fig. 9a) shows that OA overestimates wind speed at that time. Regarding the M1 moored buoy (58.3 • N-0.1 • E, north-east of the coasts of Scotland), however, OA reproduces quite 335 well the first wind peak in the afternoon of 12 Oct., but simulates a too strong and too early second peak on 13 Oct (Fig. 10a).
On 15 Oct. 00UTC (Fig. 9b), OA simulates a wind structure related to the remnants of Michael and Leslie close to Galicia.
The comparison to buoy observations shows a good correspondence, even if wind measurements are mainly localised close to the coasts and miss the stronger wind area. Moderate wind (13 m s −1 ) are also simulated in south-western Mediterranean.
The wind time-series at M2 (36.4912 • N, 6.9611 • W, in the Gulf of Cadix, west of Gibraltar Strait) in Fig. 10b shows the good 340 agreement of the OA simulation in this area. Figure 9c shows that at the end of the simulation (after 6 days), OA still performs well when compared to in-situ observations, for coastal as offshore locations, even if, again, there are no observations where OA simulates its highest wind values. This can also be seen in the latest days in Figure 10a,b.
The Taylor diagram in Figure 10c   correlation is 0.36 on average. This bias on AROME wind speed was already identified in Rainaud et al. (2016) and Léger et al.
(2016), in particular for strong wind situation and when comparing to coastal observing platforms. Further investigation would be needed to understand the origin of such systematic bias, looking into both the AROME physics and the method to diagnose the wind at 10 meters, but is out of the scope of this paper.

Rainfall
In the OA coupled simulation, the accumulated precipitation during the 7 simulated days is shown in Figure 11a. The rain is heterogeneously distributed over the domain. In the Bay of Biscay, it follows the trajectory of Callum with rainfall reaching 200 mm in the two first simulated days (Fig. 11c). In the Aude department (Fig. 11e), where the heavy precipitating event  described in section 3.1 occurred, the simulated accumulated precipitation reaches 300 mm in 1 day as observed, but are 355 located about 100 km to the east of the observed one. This location corresponds to the Massif Central relief (also known as the Cévennes), suggesting that the rapid and moist marine low-level flow is well reproduced, but with a slightly different orientation than observed and thus with a dominant triggering factor related to orographic uplift [whereas it was in fact related to convergence between the south-easterly flow with a cold front (Caumont et al., 2021)]. However, it is important to note that the Mediterranean HPE correspond to forecast ranges between +66h and +90h for AROME, i.e. quite far from the standard 360 AROME forecast operational ranges. Despite the fact that observed and simulated intense precipitation amounts are not located exactly at the same place, the heavy precipitation signature with large values of rainfall amounts in only few hours in the OA forecast, appears very valuable in the context of very early warning of such severe events.

Impact of OA coupling on the oceanic forecast
In this section, we compare NEMO forced simulations (OCE-ifs and OCE-aro) and AROME/NEMO coupled (OA) simulations 365 (Table 2) in order to quantify the effect of OA coupling on the oceanic forecast.

Sea surface temperature, salinity, height and currents
The effect of coupling on the temporal evolution of the oceanic surface field forecast is presented in Fig. 6. First, we can note that AROME simulates stronger winds than IFS (Fig. 6q,r,s,t), which leads to more intense oceanic vertical mixing in the OA and OCE-aro simulations than in the OCE-ifs one (see next section). This effect is clearly visible along the Callum trajectory, 370 in the Celtic or North Sea. Changing the atmospheric forcing of NEMO between IFS and AROME drastically modifies the oceanic response, with a more intense sea surface cooling for simulations using AROME (see OA in blue and OCE-aro in green in Fig. 6c,d). Thus, the effect of changing the atmospheric model to force NEMO is larger than the effect of interactive coupling on the simulated surface fields, in particular for SST and SSS forecast. However, the effect of the ocean-atmosphere coupling on the SST and SSS induces also a feedback, leading to a more important cooling of the surface waters in coupled 375 (OA) than in forced (OCE-aro) simulations. This sea surface cooling enhancement with coupling is in fact related to a lower non-solar net heat flux in OA (not shown), meaning a larger heat loss at night (and a lower diurnal heating) for ocean in OA than in OCE-aro. In fact, the surface cooling rapidly change the atmospheric low-level environment and stability [without significant difference in the wind speed (and wind stress)]. In particular, the coupled simulation represents an amplification loop, as the 2m-specific humidity is progressively lower in OA (than in OCE-aro/ARO). This enhances evaporation, and thus 380 amplify slightly the surface cooling. We can note that this effect of ocean-atmosphere coupling is visible for all boxes after 3 simulated days and differences increase until the end of the simulation (see Fig. 6a,b,c,d). Figure 6m,n,o,p display the impact of atmospheric forcing on the sea surface currents, which are on average less intense in the OCE-ifs simulation than in the OA and OCE-aro simulations, which is explained by weaker winds in IFS than in AROME ( Fig. 6q,s,t), except for the Atlantic box where IFS wind is larger than in AROME (Fig. 6r), but no significant change in SSC 385 A1 Coupling masks between NEMO and AROME Figure A1 presents the masked parts of each domain. The black areas in Figure A1a correspond to where NEMO does not resolve the ocean. In AROME (Fig. A1b), the masked area corresponds to the same unsolved areas of NEMO plus the northern, western and southern extensions. There, in the OA coupled experiment, AROME sees null current and keeps the initial SST field constant in time.

A2 Simulation environment and High Performance Computing characteristics
All the developments are performed using Vortex/Olive python-based framework, used to run AROME operational simulations at Météo-France. This coupling system is running on the new Météo-France supercomputer belenos (https://www.top500.org/system/179853/). In total, this supercomputer has 294 912 cores on 2 307 nodes and a peak performance of approximately 10.5 PFlop/s. Each nodes have a Random Access Memory (RAM) of 256 GB minimum.
530 Table A1 summarises the computational cost of the different simulations presented in this article (Tab. 2).
The coupled simulation runs on 15 nodes and 424 cores corresponding to 12 nodes and 384 cores for AROME, 2 nodes and 32 cores for NEMO and 1 node and 8 cores for XIOS. Simulated time is roughly 12 h for AROME (ARO) and AROME/NEMO (OA) simulations indicating that the effect of OASIS coupler is negligible for this coupled system. The OA simulation CPU cost does not exactly correspond to the sum of the executions of AROME and NEMO/XIOS, as NEMO cores pass some time 535 to wait AROME fields in this configuration. It is indeed superior to the 18 432 CPU hours for one AROME forced (ARO) simulation plus the CPU cost of the oceanic model and the XIOS server for coupled AROME/NEMO (OA) simulation and finally corresponds to a 20 % total additional CPU cost (23 040 CPU hours). Note that simulated time of NEMO simulations alone (OCE-aro and OCE-ifs simulations) are roughly equal to 8.5 h (with 2 nodes and 32 cores for NEMO and 1 node and 8 cores for XIOS) corresponding to CPU cost of approximately 3 280 CPU hours (14.2 % of the CPU cost of the OA coupled 540 system). For the purpose of this comparison, we used the same number of nodes for NEMO simulations alone (OCE-aro and OCE-ifs simulations) as the one used in AROME/NEMO simulations but it can be optimised, for example, by increasing the number of used cores by node.