Introduction

In the early half of September 2023, Cyclone Daniel struck the central Mediterranean Sea, with Greece and Libya being particularly affected. Daniel originated from a low-pressure system around 4th September and evolved into a Mediterranean tropical-like cyclone, impacting both the northern and southern Mediterranean shores before dissipating around 12th September. Regrettably, Cyclone Daniel has been identified as the deadliest storm in the recorded history of the region1. The exact death toll remains undetermined but is estimated to range between 4000 and 10,000 casualties2,3, with estimated damages surpassing $20 billion4,5.

The genesis of this cyclone can be attributed to an omega block centered in southern Europe. This weather pattern not only led to Cyclone Daniel but also triggered an intense heatwave in central Europe and heavy rainfall in Spain6. Daniel set records with its precipitation, with Greece experiencing an accumulation of over 700 mm in just 24 h (Zagora, 759 mm). Libya, on the other hand, saw a record highest daily rainfall exceeding 400 mm in Al-Bayda (414 mm)7. Unfortunately, Libya suffered the most due to significant infrastructure failures and the collapse of at least two dams3,7.

These remarkable rainfall rates require a substantial amount of atmospheric water vapor. In the complex interplay of factors contributing to the development and intensity of weather systems like Cyclone Daniel, Sea Surface Temperature (SST) stands as a primary driver8. Although these cyclones do not need high SSTs as Tropical Cyclones9,10, warmer SSTs provide energy and moisture through surface enthalpy fluxes, which help intensify these storms. The latent heat flux from the Mediterranean surface serves as the primary source of water vapor and energy for severe weather systems in the region8,11,12, including cyclones such as Daniel13. According to the Clausius-Clapeyron relationship, evaporation depends on the temperature of the liquid phase’s surface. Thus, SST modulates the water vapor available for precipitation and has the potential to influence the intensity of tropical-like cyclones, such as Daniel. While horizontal advection can supply moisture during extreme precipitation events in the Mediterranean14, its role is less critical than locally generated moisture during the warmer months and heavy rainfall episodes15,16. Advection mainly affects the development of purely baroclinic disturbances and influences systems such as mid-latitude extra-tropical cyclones, atmospheric rivers17, and even the early stages of tropical-like cyclones18, especially during the cold season19. However, in late summer Mediterranean cyclones, it takes a back seat to locally driven processes such as surface enthalpy fluxes, and hence SST is more influential during these atmospheric phenomena8,9. Therefore, the interactions between the ocean mixed layer and the atmosphere are crucial in supplying moisture and instability during intense rainfall events20, including tropical-like cyclones21, underscoring the potential influence of Mediterranean Sea Surface Temperature.

Over the past three years (2021–2024), the sea surface temperature in the Mediterranean has consistently exceeded normal levels, particularly during the warm months of 2022 and 202322,23. In the summer of 2023, an exceptional SST anomaly of 3 °C with respect to the period 1985–2005 was recorded across much of the Mediterranean, with some locations, such as the coasts of Greece and Northern Africa, experiencing values exceeding 5.5 °C23. Therefore, it is expected that the anomalously high SST of the Mediterranean played a crucial role in the intensification of Storm Daniel and contributed to the unprecedented precipitation amounts recorded during the event.

It is worth noting that this idea was previously posited to investigate the role of global warming in the formation of the Derecho observed in the western Mediterranean in August 2022, which impacted Corsica and northern Italy12. Global warming and the consequent high SSTs have also been proposed as factors modifying summer precipitation extremes in central Europe24, the characteristics of tropical cyclones in other regions25,26,27,28,29, and the intensity of other high-impact Mediterranean events11,30,31,32. Therefore, given the recent and projected increase in SSTs, particularly in the Mediterranean33, it is expected that tropical-like cyclones in the region will become more intense. The intensification of Mediterranean tropical-like cyclones with climate change has been proposed in various studies with different methodologies34,35,36. It is thus urgent to quantify the contribution of recent SST increases to intensifying these storms in the context of current climate and observed trends to better understand the potential risk the region may undergo in the next decades.

In this study, we perform a series of numerical experiments (see Methods section) over Central Mediterranean (Fig. 1) using a mesoscale atmospheric model at convection-permitting scales (~3 km). We conducted two sets of 16-member experiments initialized at different times to determine the contribution that recent anomalously high Sea Surface Temperatures in the Mediterranean had on the severity of storm Daniel, the deadliest storm in the Mediterranean to date. One of the sets is driven from the closest-to-reality information available (ERA5 Reanalysis) and is considered the control ensemble, while for the second set, we modify the SST fields. We use a novel method to remove the signal of long-term global near-surface temperature changes from SSTs, considering its effects on both the trend and the variability (see “Methods”). While we focus primarily on the influence on precipitation amounts, we also analyze the effect on mean sea level pressure, 10 min wind speed, surface enthalpy fluxes, and precipitable water to quantify the role of SST on the storm intensification.

Fig. 1: SST differences between experiments.
figure 1

Mean SST in (a) CTL and (b) CFA experiments during the storm period (1–12 September 2023) for the 0100 ensemble members (all members of each ensemble have the same SST field). c Mean SST difference between CFA and CTL during the storm. Black polygons delimit the two focus regions centered in Greece (37°–41°N, 20°–25°E) and Libya (30°–34°N, 20°–25°E).

Full size image

Results

Model evaluation

Evaluating the model performance during high precipitation events such as Storm Daniel is challenging because of the sharp gradients in rainfall spatial patterns, the sparse distribution of rain gauges, and the limitations of gridded datasets (i.e., their spatial and temporal resolution). This is further complicated by stations often going offline under severe weather conditions and the reluctance of National Weather Services (NWSs) to share available observations, especially during high-impact events. Fortunately, Dimitrou et al. (2024)37, from now referred as DI24, published information from 36 rain gauges in Greece, which constitutes an invaluable reference to investigate the event and evaluate the model outputs. We have also gathered additional near-surface observational information such as products based on remotely sensed measurements and reports from official organisms (see “Methods”). This combination of data sources provided a detailed picture of the storm at the surface, which illustrates the extraordinary rainfall rates produced by the storm and served us to confirm that the model rainfall amounts are consistent with the actual intensity of the event.

First, our model outputs are directly compared with observations during the Greek phase (Fig S1). At each location, we extracted the nearest 9 grid points, covering approximately 81 km2, for all members of the CTL ensemble spanning the period 4th–7th September (CTL2200 to CTL0312), resulting in 99 (9 grid points × 11 members) model estimates of accumulated rainfall at each station. Although models at convection-permitting resolutions are often prone to overestimating rainfall38, our CTL ensemble generally produces rainfall amounts that either match or underestimate observed values in most cases. However, there are notable discrepancies. For instance, the maximum total rainfall recorded in DI24 database for Makrinitsa (1235.4 mm) is markedly underestimated, with model estimates reaching at most 1031.9 mm. Similar discrepancies are observed in other locations, where the model consistently underestimates the most intense rates in the observations. For example, the model estimates fall below the observed values in Anilio (482 mm vs 692.6 mm), Portaria (704 mm versus 884.5 mm), Volos (491 mm versus 617.4 mm), and Zagora (1065 mm versus 1095.6 mm). Conversely, the model overestimates rainfall in two stations, with the lowest end of the range across ensembles and nearest grid points systematically exceeding the observed values. In Smokovo, the minimum model estimate is 249 mm compared to 240.6 mm in the observations, while in Skopia, the model produces a minimum accumulated rainfall of 256 mm, whereas the observations recorded 241.3 mm. For the remaining 30 stations from the DI24 dataset, observed rainfall amounts generally fall within the range of the model estimates. Differences between the model and the observations may be due to multiple factors including, but not limited to, topography misrepresentation, spatial discrepancies, model spatial resolution, and the extraordinary nature of the event itself.

In Libya, we approximated the location of stations and extended the grid-point extraction to include a larger sample (16) and accommodate the inherent uncertainty in these approximations. Using this approach, our model systematically underestimates observed rainfall near Al-Bayda, Al-Marj, and Marawah (Fig. S2), while observations are within the range of the model accumulated rainfalls in the remaining locations (near Derna, Benghazi, and Al-Qayqab). Differences are pronounced in Al-Bayda, where even the highest model estimates (215 mm) fall significantly below the observed record (414 mm). Nonetheless, these comparisons should be interpreted with caution given the uncertainties in the measurement locations.

In addition to in-situ measurements, we used gridded observational products that combine satellites and sometimes rain gauges, to provide another layer of model evaluation. Each satellite precipitation dataset has inherent uncertainties stemming from sources such as the spatial and temporal resolution, the sensor types, the retrieval algorithms, and the input data. For example, PERSIANN, GPCP, and CHIRPS provide gridded estimates, but their lower temporal resolution may miss short-duration high-intensity events. Moreover, CHIRPS only provides information over land. However, their strength lies in their broad spatial coverage and their long-term, stable records, making them useful for certain climatological studies. On the other hand, higher frequency datasets like GSMaP, IMERG, and CMORPH capture finer temporal scales (e.g., hourly), which makes them better suited for studying specific events. However, high-frequency products often suffer from increased uncertainty in retrieval algorithms and are prone to errors in complex terrain and coastal areas, and in regions with sparse observations. Using multiple datasets helps mitigate these uncertainties in the model evaluation. Overall, their spatiotemporal coverage represents a major advantage because it lets us determine the realism of the model precipitation in areas where there are no direct observations.

The total accumulated amounts generated by the model (CTL0100) during the period 2nd to 12th of September yield a realistic depiction of the storm’s spatial structure when compared to gridded datasets (Fig. S3). Foremost, it must be noted the wide disparity across satellite products. Daily gridded observations (PERSIANN, GPCP, CHIRPS) consistently show lower precipitation records, diverging from both in-situ observations and satellite-derived products at higher frequencies. This is potentially attributable to their spatiotemporal resolution, or the data incorporated in their generation, and it stresses the need for using multiple datasets to incorporate that uncertainty in the model evaluation. Considering the amounts registered by rain gauges, gridded datasets at higher frequencies and showing more intense rainfall rates, such as GSMaP, IMERG, and CMORPH, seem more plausible.

Comparing the model with these gridded products (Fig. S4 and Table S1), which offer information at sub-daily frequency, shows similar values in Thessaly (Greece), exceeding 400 mm over the course of the storm. This highlights the model’s notable ability to generate very high rainfall rates, as it occurred during Storm Daniel. Similarly, in northern Cyrenaica (Libya), the model captures rainfall amounts of nearly 300 mm along the coast, which is coherent with most gridded products, although the intensity and extent of rainfall varies across observational datasets.

From a time perspective, the CTL ensemble produces region-averaged timeseries that are generally aligned with gridded observations. For example, rainfall averaged over the central Mediterranean, where the storm developed, is comparable between the CTL ensemble and the range of observations (Fig. 2, see also Table S2 for comparison metrics). It is also remarkable that the impact of initializing the model at different dates and times is limited. In fact, ensemble members from CTL2200 to CTL0418 display a very similar evolution of precipitation until September 8th. Even beyond this point, most differences arise from ensemble members initialized after the Greek phase, and thus they are directly influenced by the storm’s presence at a mature stage in the initial conditions.

Fig. 2: Effect of SST on the temporal evolution of Storm Daniel rainfall.
figure 2

Timeseries of precipitation from all ensemble members (dots), the ensemble mean (black line), and gridded observations (dashed grey) over the region limited by a red square (30°–42°N, 10°–30°E) in the inset map (top right) for (a) CTL and (b) CFA ensembles. Bottom panels show the rainfall as horizontal bars for each ensemble member separately, each of them identified by the day and time of initialization. c Comparison between the two ensembles and bottom panel shows the difference in rainfall between CFA and CTL ensemble members.

Full size image

Focusing on the regions that were most affected (Greece and Libya) we can identify the two distinct phases of the storm. Even on these smaller regions, the model timeseries are, for the most part, within the range of precipitation estimates of observational datasets (Figs. S5, S6). The main difference occurs during the Libyan phase, when the model generates less rainfall than the gridded observations and the peak is produced approximately 24 h before. The only exception being the member initialized during the Libyan phase (CTL0918) when much of the model’s rainfall is influenced by initial conditions. The spread across ensemble members during this second phase of the storm is also much larger especially when it comes to the timing of the precipitation peak. This is expected given that most experiments are initialized at least 10 days before this peak. In the Libyan region, there are contributions from the two phases (Fig. S6), but it is the latter that produces the most precipitation. Both phases are captured by the CTL ensemble.

In summary, we have shown that the model tends to underestimate the extraordinary amounts of rainfall detected in observations, with overestimations only at very specific locations. Overall, it accurately captures the spatial structure of the precipitation fields and their temporal evolution, generating precipitation values close to the unusual rates produced by Storm Daniel.

Effect of local SST on Storm Daniel’s rainfall

According to our experiments, the anomalously high SST in the central Mediterranean has a direct impact on precipitation. Removing the signal correlated with long-term global near-surface air temperature changes from local SST leads to substantially lower precipitation throughout the duration of the storm, but especially during its second phase, as shown by results from CFA (Fig. 2). We recall that CFA SST follows the observed chronology of SST and thus contains the same temporal evolution, including marine heatwaves, but without the contribution from increased global near-surface air temperature. In fact, the CTL ensemble mean (depicted in blue) consistently remains above the CFA ensemble mean, underscoring the influence of local SST on precipitation throughout the storm’s duration. Notably, the differences between CTL and CFA are predominantly positive for individual members (horizontal bars, Fig. 2c), further supporting this idea. While this result is also discernible in Greece (Fig. S5), it is particularly clear in Libya (Fig. S6), where the warm local SST played a crucial role in driving the observed intense rainfall. Specifically, the CFA simulations yield much lower precipitation amounts over Libya compared to CTL. The CFA ensemble mean produces precipitation rates below 0.2 mm h−1 in Libya, while the CTL ensemble reaches region-averaged rates of 0.6 mm h−1. This difference is evident not only during the second phase, which impacted the region the most but also during the initial phase, which also contributed to rainfall accumulation in Libya (Fig. S6c).

At the domain-average scale, removing the signal of global air temperature changes from the local SST reduces total accumulated precipitation during the storm (1st–12th September) by 42%, from an ensemble mean of 9.14 mm gridcell−1 in CTL to 5.33 mm gridcell−1 in CFA. This difference is observed over most of the storm extension (Fig. 3), but especially in Thessaly and Epirus and northeastern Peloponnese (Greece), over the central Mediterranean, and along the northern coast of Libya. Interestingly, in Central and Western Greece, there are areas where the effect is opposite according to the ensemble means, but the magnitude of the differences pale in comparison to the reduction detected in other regions. Focusing in the two target regions and based on the total accumulated rainfall by the ensemble means, CFA yielded 17% less precipitation (1099 mm gridcell−1 compared to 910 mm gridcell−1) over Greece and a substantial 81% reduction over Libya (379 mm gridcell−1 compared to 72 mm gridcell−1).

Fig. 3: Spatial impact of SST on Storm Daniel precipitation.
figure 3

Ensemble means accumulated precipitation during the episode (1–12 September 2023) for the (a) CTL and (b) CFA experiments. c Difference in accumulated precipitation during the storm between CFA and CTL ensembles. Stippling indicates non-significant differences according to a two-sample Kolgomorov-Smirnov test at the 99% confidence level.

Full size image

Despite the undeniable importance of the impact that total accumulated rainfall had throughout the storm’s duration, it is equally important to estimate the influence of local SST on the most intense rainfall rates occurred during the event. The anomalously high local SST had a direct effect on the extreme rainfall hourly rates, according to our numerical experiments (Fig. 4). Virtually all ensemble members from CFA generated smaller upper percentiles than CTL members for both regions. There are only a few exceptions in Greece, but they are ensemble members initialized during or after the Greek phase (0600, 0706, 0812, 0918), and thus the peak of the storm was not captured by the simulated period. For instance, the impact of the local SST on the two highest percentiles we calculated (99.9th and 99.99th) is mostly in the range −20% to −5% in Greece, and −90% to −40% in Libya. We also calculated the ten most intense hourly rates in each ensemble member, which better reflect the extreme nature of this event. In Greece, the median of these rates across all ensemble members is reduced from 101 mm to 95 mm. On the other hand, in Libya, the effect is much stronger, and it is reduced from 149 mm to 82 mm. At daily timescales (not shown), the median of the ten most extreme events across members (10 events × 16 members) in Greece varies from 550 mm to 462 mm. Notably, the highest daily value in all members is slightly larger in CFA (716 mm) than in CTL (670 mm), but once again, the influence of initializing just before the peak of precipitation should be factored in. In Libya, the median of most extreme values goes from 272 mm in CTL to 128 mm in CFA. Furthermore, the highest daily value also decreases drastically from 623 mm in CTL to 319 mm in CFA.

Fig. 4: Effects of SST on rainfall extremes during Storm Daniel in Greece and Libya.
figure 4

For Greece (see red square in bottom right inset map): a Extreme percentiles of hourly precipitation for all members of CTL and CFA ensembles, b top 10 most intense values of hourly precipitation for each ensemble member, and the boxplot of all top 10 events across members for CTL and CFA experiments, and c the relative difference between CTL and CFA extreme percentiles. d–f As (a–c) but for Libya (see red square in bottom right inset map).

Full size image

These findings suggest that anomalously warm SSTs linked to increasing global near-surface air temperatures in the central Mediterranean impacted both regions, albeit with a more pronounced effect observed in Libya compared to Greece, where the influence was more limited. The distinct effect of local SST on total accumulated precipitation over the two regions can be explained by understanding the sources of moisture that fed each phase of the event. We used the WAM2Layers, a Lagrangian back-tracking algorithm39,40, and ERA5 data to track rainfall sources in each of the two regions under study. This allows us to determine the origin of moisture that produces precipitation in a region and pinpoint areas that contribute the most to each phase of the episode through evaporation. Although ERA5 may not adequately represent extreme precipitation due to its relatively coarse spatial resolution, it yields notable accumulated amounts both in Greece (625 mm) and Libya (170 mm off the coast, and 150 mm on land) during the storm. Moreover, ERA5 offers the advantage of tracing moisture beyond the regional model domain and the duration of individual experiments. It also keeps consistency across datasets because we used ERA5 to drive the regional model. As such, it enables us to estimate the relative contributions to precipitation over the two regions from ocean and land areas, both local and remote (Fig. 5).

Fig. 5: Moisture sources of Storm Daniel’s precipitation.
figure 5

Moisture sources of rainfall accumulated in (a) Greece (3-9 September 2023) and (b) Libya (7–12 September 2023), according to ERA5 and the WAM2Layers back-tracking algorithm. Solid black lines are the domains where the accumulated rainfall is considered (target regions), dashed black lines delimit our regional model domain, and dark and light red contours are the 1.0 mm and 0.1 mm evaporation isolines, respectively.

Full size image

For both regions, we extended the tracking back to 22nd August 2023, when the first simulation started (Table 1). For the Greek region, we could track 84% of the rainfall that ERA5 produces between 3rd and 9th September. Only 36% of all the moisture that produced the event originated over the sea within the model domain and during the period covered by any of the ensemble members. Thus, only that amount could potentially be affected by local SST changes, although in absolute terms the Aegean and Black Seas contributed with substantial amounts of moisture (Fig. 5a). If we include land areas, then a total of 53% of the moisture was originated within the domain, and the rest is coming from areas beyond the model domain, hence provided to the model by the boundary conditions. All ocean areas, regardless of where they are located, contribute to 56% of the moisture that became rainfall in the Greek region.

Table 1 Percentage of precipitation tracked back to sources for Greece and Libya (Fig. 5)
Full size table

By contrast, we could track up to 91% of the moisture that led to precipitation in the Libyan region between 7th and 12th September, of which 59% comes from local SST and 74% from all ocean areas. This means that large part of the precipitation generated during the Libyan phase had its origin in evaporation occurred over the central Mediterranean and the Black Sea (Fig. 5b), while the precipitation in the Greek phase was mostly generated elsewhere (Fig. 5a). Interestingly, the contribution from all land areas to the Greek phase was 27%, as opposed to only 17% in the Libyan phase, which suggest that a significant portion of the rain occurred in Greece had its origin over the continent and thus not directly affected by local or remote SSTs.

In summary, the local SST, which is a major driver of evaporation over marine regions, had a strong potential to impact the second phase of the storm. On the other hand, the extraordinary rates registered in Greece were largely governed by remote water vapor sources (boundary conditions) or were already in place (initial conditions) once the simulations had started. This also links to the high rates obtained over Libya from 0918 ensemble members (CTL0918 and CFA0918), because in both cases an active system with tropical characteristics is already in place in the initial conditions, thus the little impact of modifying SSTs.

Removing the global near-surface air temperature change signal from SST in the central Mediterranean reflects in the surface enthalpy flux (Fig. 6a–c), which can be a key factor in the invigoration of cyclones with tropical characteristics8,18. Miglietta and Rottuno (2019)8 identified a variety of mechanisms in the development of Mediterranean tropical-like cyclones and established three categories based on which factors dominate. Storm Daniel seems to be a mix: during the early stages, Bora and Etesian winds are the primary drivers for Wind Induced Surface Heat Exchange (WISHE) type mechanism, while during the maturing phase, there is a direct link between surface heat fluxes and the vortex circulation (Fig. S7).

Fig. 6: Influence of SST on surface enthalpy fluxes and precipitable water.
figure 6

Ensemble mean average surface enthalpy fluxes for (a) CTL and (b) CFA during the storm (1–12 September 2023), and (c) Difference between the CFA and CTL. d–f As (a–c) but for total precipitable water. Stippling indicates non-significant differences according to a two-sample Kolgomorov-Smirnov test at the 99% confidence level.

Full size image

The surface enthalpy flux, measured as the sum of sensible and latent heat fluxes at the surface, is affected particularly in the Ionian and Libyan Seas (Fig. 6c), where the SST differences between CTL and CFA were larger. Surface sensible heat flux can be both positive and negative depending on the direction of the heat transfer and the relative thermal difference between the ocean surface and the atmosphere above. Thus, the diurnal cycle and whether the sea takes up heat from the air depends on this thermal difference throughout the day. In fact, this could be regarded as a limitation in studies that alter SSTs, because there is an imbalance between the SST and the atmosphere above, especially during initialization. To minimize possible initialization imbalance effects on our results, we conducted an ensemble (see “Methods”). Furthermore, the consistency we found across the ensemble for precipitation (Fig. 2) suggests that the model may become relatively insensitive to potential imbalances in the initial conditions. It is beyond the scope of this study to analyse the diurnal evolution of surface sensible heat fluxes, but it is worth noting that the net contribution from a warmer sea surface temperature is, as expected, positive. This constitutes, on average, a net source of heat that could further intensify the storm, although the response of the storm to changes in sensible heat is complex and yet to be fully understood. On the other hand, the role of latent heat flux is more direct and easier to interpret. Moreover, most of the change in surface enthalpy fluxes is due to differences in latent heat flux, since the contribution to enthalpy fluxes from sensible heat flux over the sea is relatively small (<20%), according to our experiments (Fig. S8) and in agreement with previous findings8,18,21.

A warmer sea surface produces an increase in latent heat flux (Fig. S8), which is equivalent to higher evaporation rates. This injects moisture in the atmosphere which becomes available for precipitation processes. The column-integrated water that is available for precipitation, also known as precipitable water, increases near these areas of enhanced surface enthalpy fluxes (Fig. 6d–f), which are in turn dominated by enhanced evaporation. In our ensembles, these differences are clearly visible off the coast of Libya, which explains the higher precipitation amounts obtained with the CTL experiments. The differences in precipitable water between the two ensembles in the Aegean Sea are small compared to the Libyan Sea, and thus the impact of warmer SSTs due to global near-surface air temperature changes is not as marked.

The storm intensity is often measured through the minimum Mean Sea Level Pressure (MSLP) and the maximum windspeed at 10 m above the ground. We have calculated the minimum MSLP near the cyclone center for each experiment. To ensure we analyse values near the cyclone and that MSLP corresponds to relatively large-scale features, we tracked the cyclone center by smoothing MSLP fields using a Gaussian filter with a σ of 6 timesteps (6 h) and 20 grid points in each direction (~60 km). The motion of the cyclone is tracked by looking for the lowest local minima in a 100-grid points radius in the next time step. If no local minimum is located, the distance is increased in 50-grid intervals up to 200 grid points. If the deepest local minimum is above 1012 hPa or no local minimum is detected, the cyclone is considered dissipated. With these criteria, 13 out of 16 members of the CTL ensemble make landfall in Libya, while only 2 members of the CFA ensemble reach the Libyan coast, one of which dissipates at the shoreline. The tracks and the evolution of minimum MSLP (Fig. 7) reveal that the CTL ensemble tends to produce deeper and longer-lasting cyclones. MSLP from the CFA ensemble is systematically above the CTL ensemble after 06UTC 5 September, and they were very similar before. This supports the idea that SST played a role intensifying the storm during its second phase (7–12 September), but the effect on cyclone Daniel’s during the Greek phase (3–7 September) is evident only towards the end of the phase. Both ensembles reach their minimum on the 10 September with the experiment initialized at on the 9th at 18UTC (997.2 hPa in CTL0918 and 998.7 hPa in CFA0918), thus they are strongly influenced by atmospheric initial conditions, which are the same in both ensembles. The next minimum in the CTL ensemble corresponds to CTL0418 on 11 September 00UTC (997.7 hPa), while the cyclone dissipates in all other CFA experiments and produces no other minima of the second phase. The next minimum in the CFA ensemble is reached during the first phase (5 September 13UTC, CFA2806, 1000.8 hPA) and is comparable to most of the CTL ensemble members.

Fig. 7: Cyclone tracks and minimum Mean Sea Level Pressure.
figure 7

a Cyclone Daniel tracks in each CTL (blue) and CFA (red) ensemble member. Numbers indicate where the storm is located at 00UTC on that day calculated as the mean location of the ensemble tracks. b Minimum MSLP timeseries for each of the simulations. All minima are located within ~120 km from the center of the cyclone identified using the smoothed MSLP.

Full size image

The differences in minimum MSLP and maximum 10 m wind speed between the two ensemble means (Fig. 8) further support that the high SSTs registered in the central Mediterranean before and during the Storm Daniel contributed to enhancing its intensity. Along the tracks shown in Fig. 7a, especially after impacting Greece, the CTL ensemble produces lower minimum MSLP values than the CFA ensemble. The spatial pattern of MSLP is similar between the two ensembles, but the difference between CFA and CTL ensemble means exceeds 2 hPa in multiple locations. Apart from the differences near the eastern border of the domain, the effect of SST is more evident over the region where Storm Daniel becomes a tropical-like cyclone, between Libya and Italy, where the ensemble means differ as much as 2.26 hPa. As for the maximum winds, the CFA ensemble produces lower values than the CTL ensemble, although the differences are not as large considering the typical range of wind speeds. On average across members, the CFA produces a maximum windspeed in the central Mediterranean of 26.0 m s−1 versus 27.3 m s−1 in the CTL ensemble. In the central Mediterranean, the differences between the two ensembles are generally negative (Fig. 8f), indicating that the CTL ensemble tends to produce higher wind speeds near the surface. The largest difference between the two ensembles occurs in this region and reaches −4.9 m s−1. It must be noted that windspeed, like all other near-surface variables, is saved every hour, and those values are instantaneous. Therefore, if we were to examine wind speed differences at higher frequencies (i.e., wind gusts), the differences would likely be larger, but we cannot affirm in which direction. However, our model experiments do support the hypothesis that removing the signal of global near-surface air temperature changes from local SSTs influences the storm characteristics, suggesting that the anomalously high SSTs contributed to intensifying the storm, especially during its second phase.

Fig. 8: Influence of SST on minimum mean sea level pressure and maximum 10 m wind speed.
figure 8

Ensemble mean of minimum MSLP calculated at hourly frequency for (a) CTL and (b) CFA ensembles during the storm (1–12 September 2023), and (c) Difference between CFA and CTL ensemble means. d–f. As (a–c) but for maximum 10-m wind speed. Stippling indicates non-significant differences according to a two-sample Kolgomorov-Smirnov test at the 99% confidence level.

Full size image

Discussion

Our study hypothesized that the anomalously warm Mediterranean Sea surface temperatures (SSTs) during the summer of 2023 contributed to intensifying Storm Daniel, which occurred in early September and significantly impacted Greece and Libya. We conducted two sets of simulations using the regional climate model WRFv4.5.1 at a spatial resolution of approximately 3 km (0.0275°). Both sets used boundary conditions from ERA5 to replicate real atmospheric conditions. The first set used SST from ERA5, while the second set included modified SST for which the signal correlated with long-term global near-surface air temperature changes since 1940 was removed. This methodology allowed us to estimate the influence of global temperature trends on the SST, and in turn, how exceptionally high local SST affected the storm.

Our findings reveal key insights into links between elevated SSTs and Storm Daniel’s development, intensity, and precipitation characteristics. The anomalously high SSTs in the Mediterranean during summer 2023, with temperatures up to 5.5°C above normal in some areas, provided additional water vapor and latent heat, enhancing the storm’s intensity and rainfall. Our numerical experiments support the hypothesis that these high SSTs significantly increased precipitation rates, particularly during the storm’s second phase affecting Libya, when it acquired tropical-like characteristics. Without the unusually warm SSTs, accumulated precipitation amounts in Libya would have been considerably more moderate. Similarly, sub-daily extreme rainfall rates would not have reached the high values observed in either Libya or Greece.

In Greece, the effect of high SSTs on precipitation was less pronounced for two key reasons. First, the contribution from local sea areas to moisture is weaker in the Greek phase compared to the second phase, when the storm developed tropical-like characteristics. During the first phase, the storm was driven by baroclinic disturbances and thus the surface fluxes may not be as important. As a result, the potential for increased rainfall due to high SSTs was lower in Greece. Additionally, our counterfactual SST product showed only a limited impact over the Aegean Sea, unlike the differences seen in Libya.

Comparisons with rain gauge data and gridded observations indicated that our model effectively captured the observed rainfall patterns and intensities, despite a tendency to underestimate the most extreme precipitation rates from stations, which could partially be due to the model’s spatial resolution. The overall alignment with in-situ and satellite-derived observations demonstrates the model’s capability to replicate the fundamental dynamics of the storm.

Our study underscores the role of recent global warming in Mediterranean tropical-like cyclones. The counterfactual analysis, which removed the global warming signal from SSTs, suggests that current climate trends may have contributed to amplifying the intensity of such storms. However, it should be noted that our study focuses on a single, probably unique event, with two phases that are likely governed by different mechanisms.

Further research should delve deeper into the interactions between SST anomalies, particularly marine heatwaves, and cyclone dynamics in the Mediterranean. It is also necessary to investigate the effect on other similar storms so that our findings can be generalized. Efforts should also include boundary conditions adjustments to fully isolate the impact of SST changes. Improving model validation and understanding of extreme precipitation events could benefit from integrating higher-resolution observational datasets and enhancing data sharing among meteorological agencies. We identified limitations in this study, particularly in the context of cyclone intensity, that suggest opportunities for future research. For instance, the impact of atmospheric boundary conditions on storm characteristics could be better addressed using a global model with a stretched grid over the region of interest, such as MPAS, which requires only initial conditions and SST. Another promising approach is employing coupled models to capture fine-scale air-sea interactions and their effects on storm development, as these systems typically produce SST cooling during the storm’s passage not represented interactively in the model. Additionally, examining the contribution of soil moisture, especially given its role in the Greek phase of the storm, is necessary to understand how changes in soil conditions might affect storm characteristics.

In conclusion, the anomalously high SSTs observed in the Mediterranean during the summer of 2023 were a critical factor in the development Storm Daniel and its unprecedented rainfall intensities. As Mediterranean SSTs continue to rise in response to global warming, understanding and preparing for the implications of such extreme weather events is increasingly crucial.

Data and methods

Observations

Observations were primarily obtained from Dimitriou et al. (2024)37. They compiled rain gauges from the National Observatory of Athens and the Greek Ministry of Environment and Energy, among others. Thirty-eight stations form the database, although one does not have valid records for the period 4-7th September and thus the total accumulated rainfall during the storm is not registered. Precipitation measurements are provided every day. We also used reports obtained from the World Meteorological Organization and other scientific publications41, which contained scattered data on measured precipitation rates. For example, in Libya, according to ECMWF42, there are no available precipitation observations on the World Meteorological (WMO) Global Telecommunication System (GTS), and only media reports could be used to estimate rainfall intensities. Other than the extraordinary value registered in Al-Bayda (414 mm), which was declared the national daily record by Libyan authorities, ECMWF42 compiled information from various sources and reported rainfall amounts in five other places. However, the exact location of these in-situ measurements and the period covered are unclear, and thus the comparison with the model should be interpreted qualitatively. Despite being incomplete or partial, this information is very valuable for evaluating the most extreme precipitation rates generated by the model. Overall, they highlight how exceptional the storm was. In Greece, at least 3 stations (Zagora, Portaria, and Makrinitsa) registered amounts above 700 mm on 5th September alone, and 10 stations reported more than 500 mm over the course of the Greek phase of the storm (4th–7th September). In Libya, other than the daily national record of 414 mm in Al-Bayda, reported rainfall amounts include: 42 mm near Benghazi, 73 mm near Derna, 170 mm near Al-Qayqab, 185 mm near Al-Marj and 240 mm near Marawah during the second phase (8th–12th September).

In addition to in-situ measurements, we also resorted to data from satellite-derived products and gridded observations to offer a spatiotemporally comprehensive description of the event. We used widely used products such as GSMaP43, GPM IMERG v7B Final Run (IMERG)44, CMORPH v145, PERSIANN46, GPCP v1.347 and CHIRPS v2.048.

Model experiments

We used version 4.5.1 of the Weather Research and Forecasting (WRF) Model49 to conduct numerical simulations of Storm Daniel within the central Mediterranean region. We built two ensembles of simulations composed of sixteen members each, varying their initial conditions. The ensemble members are initialized thirty hours apart, from 22nd August to 9th September 2023, and all ended on 12th September 2023. Members are initialized thirty hours apart so that they start at different times of the day too. Most members encompassed the entire lifecycle of the cyclone, except those starting after the 4th of September, which capture it only partially. We chose to include them to account for the role of initial conditions during the second phase of the storm affecting Libya, once the driving data provides conditions for an already mature storm.

Both ensembles were identical in their model configuration except for the SST (Fig. 1). The control (CTL) ensemble was driven by ERA550 both at the lateral and the lower boundary conditions, including SST. The second ensemble, which we named counterfactual (CFA), was also fed by ERA5 for all fields except SST, which was obtained from a data-driven method aimed at removing the signal linked to global surface air temperature increase since 194051 (see Sea Surface Temperature counterfactual estimates). The boundary conditions were updated every 3 h in both sets of simulations, although SST is effectively updated daily because ERA5 SST remains constant throughout each day, and it is only updated every 24 h. Each simulation is named with the acronym of the ensemble followed by the day and hour of initialization. For example, the member of the counterfactual ensemble initialized on the 30th August 2023 at 18UTC is identified as CFA3018. The proposed experimental design let us determine the contribution from recent anomalously warm SST on the storm characteristics. However, it is worth noting that it comes a limitation that needs mentioning: the atmospheric initial and lateral boundary conditions are not modified in the CFA ensemble, thus there is still an underlying signal from the anomalously high SSTs in the atmospheric information provided to the model. On one hand, this may lead to an underestimation of the effects that high SST may have. We have included ensemble members initialized well before the start of the storm to mitigate this. On the other hand, this may also introduce some imbalance at the borders between the atmospheric boundary conditions and the solution provided by the model, which is affected by the modified SST. The interplay between SST and the atmosphere is complex, and the available water for precipitation directly depends on evaporation, which in turn depends on the SST, the partial pressure of vapor, the salinity, and wind, among other factors. In addition, moisture may be advected by the atmosphere. While in this study we only analyze the role of one of these factors, we also explored adjusting the atmospheric conditions using an approach like Pseudo-Global Warming52 but removing the observed signal, and we found the influence of adjusting the atmospheric boundary conditions was only minor compared to modifying SSTs.

The spatial configuration of the model was primarily based on the CORDEX-FPS initiative for Convective Phenomena at high-resolution over Europe and the Mediterranean53,54. We adapted the domain location and size, and the boundary conditions (ERA5 instead of ERA-Interim) to suit the specific requirements of our study. Many of the decisions in terms of physics schemes and dynamical core options were influenced by the ongoing CORDEX-HRSST initiative, aimed at examining the impact of high-resolution SST on precipitation estimates. Furthermore, they draw upon existing literature and the experience gained from preliminary tests conducted with the model in the Mediterranean Sea region. For example, tests within the context of the CORDEX-HRSST initiative guided us to use a single domain (Fig. 1) directly driven by the boundary conditions (ERA5) because we found no overall improvement from using an intermediate domain (not shown). This is further supported by previous studies that found models able to generate small-scale features and no signs of deterioration even when directly driven by boundary conditions55,56. Thus, we used a single domain operating at an approximate spatial resolution of 3 km (0.0275°), which enables the explicit representation of convection to a large extent. The domain extends 891 by 761 grid points (an area of circa 2700 by 2100 km), which is large enough to keep the borders away from the area of interest. In addition, the buffer zone to specify the boundary conditions was increased from a default of 5 grid points to 10 grid points to make the transition smoother. The geographical data that defines topography, land use, land-sea mask, and other static fields was set to “modis_15s + default”. The vertical is resolved with 45 hybrid-coordinate levels with the atmosphere top located at 50 hPa. A sample configuration file with all model options is provided in the supplementary material for reproducibility.

Sea surface temperature counterfactual estimates

The counterfactual Sea Surface Temperature corresponds to a hypothetical SST climate for which the signal correlated with long-term global surface air temperature changes has been removed while preserving the observed short-term variability. Since the increase in globally averaged surface temperature since the mid 20th century has been attributed to anthropogenic forcing57,58, we obtain an SST estimate without the influence of the human-induced global warming and, hence, we can quantify its effect on different aspects of the climate system, such as the intensity of storms as we do here. In fact, Marcos et al. (2024)51 analyzes the specific case of the marine heatwave (anomalously high SST sustained in time) in the Mediterranean during the summer of 2023, just before the occurrence of storm Daniel.

Counterfactual SST fields are built using daily SST data over the Mediterranean Sea and global monthly 2 m air temperature from ERA5 at 0.25° spatial resolution from 1940 to the present. At every grid point, daily SST time series are separated into calendar-day records (i.e., one for each day of the year, containing 83 values each corresponding to the period 1940–2022). Calendar-day time series are then modeled as Gaussian distributions whose means and variances are allowed to vary with long-term global surface air temperature changes. A Kolmogorov-Smirnov test has been applied to all calendar-day time series to ensure they follow a normal distribution. Long-term global surface air temperature has been calculated using a low-pass filter with a cut-off period of 10 years applied to globally averaged monthly surface air temperature. The temporally varying parameters of the distributions for each grid point and calendar day were inferred using a Bayesian framework. Once these are calculated and assuming that global temperature remains unchanged since 1940, the counterfactual SST records are built locally through quantile mapping between the observed (ERA5) and the counterfactual probability distributions at every grid point. Note that only those changes correlated with global temperature variations over ~10 years are removed in the counterfactual SST, while shorter-term climate variability (for example, linked to large-scale modes, such as ENSO) remains as well as regional climate variability. Importantly, this approach does not alter the chronology of the observed SST variability and only accounts for the fraction of this variability correlated with low-frequency changes in global temperatures. Further details on the method may be found in Marcos et al. (2024)51 and Mengel et al. (2021)59.