In this study, we used an ensemble of meteorological models as a forcing condition for the ocean simulations. To reproduce the meteorological characteristics of Medicane Ianos, we performed a dynamic downscaling of the operational deterministic analysis of the European Centre for Medium-Range Weather Forecasts global Integrated Forecasting System (ECMWF IFS). Two different horizontal grid resolutions, 10 km (about 0.1∘) and 2 km (about 0.02∘), over a domain covering the central Mediterranean region (the black box indicated in Fig. a), were used. The deterministic ECMWF IFS atmospheric model (cycle 47r1) has approximately 9 km horizontal resolution and 137 vertical levels, of which 20 are below 1000 m (see the full documentation available at https://www.ecmwf.int/en/forecasts/about-our-forecasts/evolution-ifs/cycles/summary-cycle-47r1, last access: 14 February 2023). The operational analyses are based on the incremental form of the 4-dimensional variational data assimilation scheme (4D-Var; ) and are available at 6 h intervals (i.e. at 00:00, 06:00, 12:00 and 18:00 UTC).

According to several authors, operational forecasts benefit from the combination of different models by considering different physical parameterizations, numerical schemes, model resolutions and forcings . Moreover, a variety of independent models is suitable to explore the uncertainty associated with model errors . In this study, dynamical downscaling is supplied by an ensemble of regional models implemented by different research centres and made available within the research initiative fostered by the MedCyclones COST Action (Table ). They include some of the most widespread atmospheric mesoscale models used in academic and operational applications. The models have been applied at the horizontal grid spacing of 10 and 2 km, which are approximately the most common resolutions adopted in operational centres for numerical prediction over large domains and over regional areas, employing, respectively, parameterized or explicitly resolved convection. The reader is referred to the references listed in Table  for a more detailed description of the atmospheric models.

The regional models have different characteristics, parameterizations and implementations, but they share a common setup considering the following constraints:

5 d run covering the period from 15 (00:00 UTC) to 20 (00:00 UTC) September 2020;

The objective of this study is not to mimic an operational forecasting system but to analyse the level of uncertainty due to the meteorological mesoscale models, associated with this extreme event that is strongly sustained, especially in the initial phase, by deep convection. The common setup of the atmospheric models minimizes the error in large-scale dynamics due to the boundary conditions (using the same analysis fields). Due to the common integration domain, the different results are mainly ascribable to the characteristic of the models. Moreover, to allow a fair comparison of the results at 10 and 2 km grid spacing, the domain and initial/boundary conditions are also the same for all simulations.

The sea-level and wave patterns during Medicane Ianos were simulated using the finite element SHYFEM hydrodynamic model (https://github.com/SHYFEM-model/shyfem, last access: 14 February 2023; ) coupled with the unstructured spectral wave model WWMIII . The considered interactions between waves and surge were (1) the contribution of waves to the total sea levels through the wave setup and wave set-down; (2) the influence of tides and storm surge on wave propagation, affecting the refraction, shoaling, and breaking processes; and (3) the effect of sea-level variation and currents on the propagation, generation, and decay of the wind waves. The coupled model has already been applied to simulate hydrodynamics and waves in the Mediterranean Sea and in the Adriatic Sea during extreme meteo-marine events .

Both the hydrodynamic and wave models run on the same computational domain that covers the whole Mediterranean Sea through an unstructured grid consisting of approximately 200 000 triangular elements (Fig. a). The use of elements of variable sizes is fully exploited to create a seamless transition between the open sea and the coastal waters following the medicane path. The ocean models solve the combined large-scale oceanic and small-scale coastal dynamics in the same discrete domain by subdivision of the basin into triangles varying in form and size. Mesh resolution varies from 12 km in the open sea to a couple of kilometres along the cyclone path (Fig. b) and up to a few hundred metres along the western coast of Greece, where the cyclone landed (Fig. c). Model bathymetry over the Mediterranean Sea is obtained by a bilinear interpolation on the model grid of the European Marine Observation and Data Network (EMODnet) 2020 dataset .

The numerical experiments were forced by 10 m winds and mean sea-level pressure (MSLP) provided by the ensemble of meteorological models described in Sect. . ECMWF IFS 6 h analyses were used to cover the wave–current model grid outside the atmosphere domain. Also, meteorological forcings from the ECMWF IFS operational analysis over the whole current–wave model domain are considered for benchmarking. All numerical experiments cover the period from 10 to 20 September 2020 with a spin-up period of 5 d (from 10 to 15 September 2020) for which we used ECMWF IFS 6 h analysis as forcing. Since we are interested in examining only the meteorological contribution of the cyclone, we imposed no sea-level Atlantic Ocean boundary conditions at the Gibraltar Strait. The 2D barotropic hydrodynamic model simulates the sea level induced by the storm surge (driven by MSLP and wind) and waves, without considering tidal and baroclinic dynamics.

Medicane Ianos formed in the central Mediterranean Sea over the warm waters of the Gulf of Sidra on 15 September 2020, developing from the interaction between an upper tropospheric precursor and a cluster of convective storms. Then the cyclone moved northward, progressively deepening over the Ionian Sea. On 17 September 2020, Ianos evolved into a powerful medicane, and on 18 September, it made landfall on the western coast of Greece (near the Lefkada, Kefalonia and Zakynthos islands), accompanied by high coastal waves and storm surges, determining coastal damages and flooding . The cyclone's path, as defined by the 6 h ECMWF IFS analysis, is indicated with a red point line in Fig. .

Since this study aims at evaluating the sea conditions driven by Medicane Ianos, we focused the meteorological analysis only on the mean sea-level pressure and wind over the sea, leaving out the physical processes behind the genesis and evolution of the medicane, as well as its effects overland (e.g. rainfall and damaging winds). Surface winds derived from the Advanced SCATterometer (ASCAT) on board Metop satellites provided by the Royal Netherlands Meteorological Institute present an overview of the cyclone structure over the Ionian Sea (Fig. ). On 16 September, the cyclone was located over the south-western Ionian Sea with winds above 20 m s-1 surrounding the centre. On 17 September, the cyclone intensified, and in the evening it approached the western Greek coast, thus travelling over a distance of about 300 km in 23 h at an average speed of 3.6 m s-1. At this stage, southerly winds up to 25 m s-1 hit the Zakynthos and Kefalonia islands. Ianos then remained over the Ionian islands during the night between 17 and 18 September, maintaining high wind intensity.

The results of the different atmosphere models are presented in Fig.  in terms of the cyclone's path and along-path minimum pressure. Generally, the 2 km models produced tracks closer to analysis and a deeper cyclone with respect to the 10 km ones. While the location of landfall over Greece is shifted southeastward by all models compared to the analysis, both error and spread in tracks are about twice larger in 10 km than in 2 km simulations. Furthermore, a large spread is found in the simulated cyclone intensity and is also larger in 10 km than in 2 km simulations. The high-resolution runs systematically overestimate the maximum intensity compared to the IFS analysis, although the latter underestimates the actual cyclone intensity (MSLP values of 984.3 and 989.1 hPa were recorded in the landfall region; ). These results stress the high sensitivity of Medicane Ianos to the resolved physical processes as also pointed out by . Such uncertainty in the resolved track and intensity of the cyclone leads naturally to diverse estimations of ocean impacts, which are assessed in the following sections.

Clearly, the strong surface winds swirling around the cyclone centre produced a severe sea state with high waves. Indeed, the wave model forced by ECMWF IFS analysis simulated significant wave heights (SWHs) with values up to 5 m in the Ionian Sea (Fig. ) and patterns replicating the surface wind fields detected by the scatterometers (see Fig. ). It is worth noting that waves generated by Medicane Ianos propagated over the whole Ionian Sea, determining maximum SWH values higher than 1.5 m along the southern Italian, the northern African and the western Greek coasts.

Satellite-based SWH observations confirmed the simulated values in the Ionian Sea. We considered SWH data acquired by altimeters (Jason-3, Sentinel-3A, Sentinel-3B, CryoSat-2, SARAL/AltiKa, CFOSAT and Hai Yang 2B) passing over the central Mediterranean Sea from 16 to 18 September 2020 (Fig. a) and retrieved from the E.U. Copernicus Marine Service . Along-track (at a spatial resolution of 7 km) satellite SWH observations were used to assess the capacity of the wave model in reproducing the sea state when the medicane was moving from the central Ionian Sea to the Greek coast. In this study, we consider normalized standard deviation, centred RMSE (root mean square error of the simulated values with respect to the observations after subtracting the respective mean fields), bias (difference between the mean of simulation results and observations) and R (correlation coefficient between model results and observations) as the metrics for measuring the model performance.

The statistical values shown in the Taylor diagram (, Fig. b) and summarized in Table  indicate that the wave simulations forced with the 10 km meteorological fields have a slightly lower RMSE with respect to the ones that use the 2 km forcings (on average 1.0 m for 10 km and 1.1 m for 2 km). On the contrary, R is higher for the 2 km simulations (on average 0.84 for 2 km and 0.74 for 10 km). Generally, both the 10 and 2 km sets of simulations have a standard deviation larger than observed, and the 2 km numerical experiments tend to overestimate the measured SWH (the average bias for the 2 km runs amounts to 0.6 m). It is worth noting that none of the 10 and 2 km forcing simulations performs better than the one driven with the 6 h ECMWF IFS analysis, which has an RMSE of 0.4 m and R equal to 0.92. This emphasizes the challenge of accurately simulating the medicane evolution several days ahead, even on a limited area forced by the analysis at its lateral boundaries.

The simulated sea levels forced by ECMWF IFS analysis show a circular pattern on 16 and 17 September following the medicane path with peak values in the centre of the cyclone of about 20 cm (Fig. ). Such a distribution is mostly induced by the contribution associated with the sea-level isostatic adjustment to the air pressure variations (the so-called inverse barometric effect, which amounts to 1 cm of sea level per 1 hPa of air pressure). Wind and waves raised the sea levels along the northern African coast with values up to 10 cm. Between 17 and 18 September, the cyclone hit the western Greek coast, raising by more than 15 cm the sea levels near the Lefkada, Zakynthos and Kefalonia islands.

The 1 Hz (7 km) interval along-track sea surface heights recorded by altimeter Sentinel-3B (passage of 17 September 2020 20:00 UTC; magenta line in Fig. b) and Sentinel-3A (passage of 18 September 2020 09:00 UTC; green line in Fig. c), retrieved from the E.U. Copernicus Marine Service , were used for visual model comparison. We considered the filtered sea-level anomaly (SLA) data uncorrected from the dynamic atmospheric component (i.e. SLAunc=SLA+DAC). It is important to stress that altimeters register the total sea surface height that is the result of different phenomena, including also the baroclinic dynamics, which are not accounted for in the storm surge model . However, as shown by , the storm surge induced by the inverse barometric effect and wind stress associated with a medicane represents the largest contribution to the temporary rise of the sea level during such events. Indeed, barotropic models have been widely used for simulating the sea-level impact of different sea storms (, and references therein). Figure  shows that the along-track observations have higher spatial variability with respect to the model results. However, the rise in sea level towards the Ionian islands (of the order of 20 cm on 17 September and 10 cm on 18 September) is captured by both the storm surge model and the altimeters. As for the waves, the model results display a large scatter of the simulated sea levels.

While approaching the coast, the strong winds associated with Medicane Ianos raised the sea levels near the Ionian islands with the consequent flooding of several coastal locations . The sea-level data acquired at the tide gauges of Zakynthos, Katakolo and Kyparissia (indicated with red dots in Fig. ) and provided by the UNESCO/IOC sea-level station monitoring service (http://www.ioc-sealevelmonitoring.org/, last access: 14 February 2023) were processed with the Python UTIDE tidal harmonic analysis tool based on the least-squares fitting to separate the tidal and the atmospheric (storm surge) contributions to the total sea level. Observations clearly indicate a significant storm surge contribution associated with the medicane on 17 and 18 September with peak values of 0.19, 0.27 and 0.16 m at Zakynthos, Katakolo and Kyparissia, respectively (Fig. a, b and c). The only available wave monitoring station in the area of interest is the buoy of Pylos (provided by the POSEIDON monitoring, forecasting and information system for the Greek seas; https://poseidon.hcmr.gr/, last access: 14 February 2023), which recorded a maximum significant wave height of 4.7 m on 18 September at 03:00 UTC but probably missed the storm's peak (Fig. d).

The analysis of the numerical results highlights that even if the simulation forced by ECMWF IFS operational analysis correctly reproduced the wave conditions in the open sea, it strongly underestimated the peak values of both the sea level and the wave height at the coast. The storm surge from both the 2 and 10 km forcing experiments shows a large spread at the coast with some simulations strongly overestimating the peak values. Similarly, the simulated significant wave height reaches maximum values ranging from 2 to more than 12 m at Pylos. Such a large spread in model results is mostly related to the complexity of the coastline. Indeed, the response of the sea to an intense cyclone with a highly variable wind pattern in space and time is strongly dependent on the specific meteorological and morphological conditions of the considered coastal location. Therefore, even a small deviation in the simulated cyclone path or intensity could greatly change the impact on the sea conditions. However, given the long duration (5 d) of the simulations, it is not surprising to see such a large spread. Indeed, the results presented in Fig.  show high variability in the medicane track and intensity in the different meteorological models that consequently simulated different locations (spanning over about 250 km of coastline) and timing of cyclones' landing.

In order to assess the uncertainty associated with the two sets of simulations (10 and 2 km grid resolution), we computed their ensemble mean and standard deviation (considered here as a measure of the ensemble spread), and we report in Table  the computed peak values and timing. We have to point out that the maximum value of the ensemble spread does not always coincide with the maximum value of the ensemble mean. These results highlight that the 2 km experiments generally produce a larger ensemble mean and spread with respect to the 10 km ones, but there is no clear indication that one of the simulation sets provides a more accurate reproduction of the maximum observed sea-level and wave values. The simulated maximum values of the ensemble mean are found to occur shifted by 2 to 10 h with respect to the observed peaks with a spread in the timing ranging from 3.4 to 6.7 h, depending on the location. It results in the fact that the 2 km experiments have a lower spread of the peak time (4.2 h as the average of the selected stations) with respect to the 10 km ones (6.2 h), thus reducing the uncertainty in the time occurrence of the storm peak.