We use daily 10 m winds, sea-level pressure (SLP), surface temperature and omega velocity from the Last Millennium Ensemble Project that carried out simulations over the 850–2005 period with the Community Earth System Model (CESM) coupled atmosphere–ocean model (Otto-Bliesner et al., 2016). The model horizontal resolution in the atmosphere and ocean is ∼2 and ∼1∘, respectively. We present results from (a) an ensemble of 12 simulations (archived runs 2–13; run 1 omitted because it did not archive daily wind fields before 1700) that consider all known historical forcings (greenhouse gases, solar variability, volcanic, land use and orbital) and (b) a twin ensemble of 5 members but considering volcanic forcing only. Volcanic forcing in CESM follows the ice-core–based reconstruction of Gao et al. (2008), in which zonally uniform fixed single-size stratospheric aerosols are prescribed in the three layers in the lower stratosphere above the tropopause (Otto-Bliesner et al., 2016). Given our focus on volcanic signatures, those two sets of simulations are merged into a grant ensemble (CESM-LME hereafter) of 17 members spanning over the 850–2005 period, and all results are based on this merged dataset. From 850 to 1850, we analyse all strong tropical and high-latitude eruptions according to the classification of Stevenson et al. (2016), while over the 1851–2005 period we additionally consider Krakatau (1883), Santa Maria (1903), Novarupta (1912), Mount Agung (1963), El Chichón (1982) and Pinatubo (1991) (all selected eruptions are listed in Table 1). However, we find that CESM-LME simulates significant responses for the pre-20th-century eruptions only (bold eruption years in Table 1), presumably because the pre-20th-century volcanic forcing was much larger than the more recent events. To demonstrate the consistency of our findings with different volcanoes, we discuss results for the strongest tropical eruptions of Samalas (1258) and Tambora (1815) and the high-latitude eruption of Laki (1762), while over the 20th century we consider Pinatubo only, a selection based on the magnitude of the volcanic forcing in the model (note that in the CESM-LME, Samalas and Laki erupt in 1258 and 1762 and not in the true year 1257 and 1783, respectively). However, similar signatures are found for most of the pre-20th-century eruptions as shown in the figures in the Supplement. The Laki eruption serves as an example to demonstrate the amplified sensitivity of Etesian wind response to NH eruptions. This is further elaborated in Sect. 3.1 by comparing differences for all SH, tropical and NH pre-20th-century eruptions.

CESM-LME signatures for Pinatubo are evaluated against version 3 of the NOAA–CIRES–DOE 20th Century Reanalysis project (20CR hereafter), which reconstructs past climate over the 1836–2015 period by assimilating historical air pressure observations into a global atmospheric model (0.5∘ horizontal resolution) and specifying sea ice and sea-surface temperatures at the model surface boundary (Slivinski et al., 2019). Dafka et al. (2016) reported an overall consistency in the representation of Etesians in different reanalyses compared with station wind observations, and similarly our results do not critically depend on the choice of the reanalysis dataset (not shown). This is additionally confirmed by considering the updated L-days dataset of Carapiperis (1951), which is an independent observation-based index for the number of Etesian days from 1892 to 2006 that describes Etesian outbreaks when the northerly wind in Athens exceeds the local wind breeze.

We calculate daily wind speed (WSP) and wind direction (WDIR) during the late summer (July and August, JA hereafter) at a fixed grid point (37.5∘ N, 25.0∘ E) in the central Aegean Sea (red star in Fig. 1) so as to select the months typically demonstrating the strongest wind speeds under the influence of monsoonal convection over northern India (Tyrlis and Lelieveld, 2013). Following the methodology of Logothetis et al. (2020), a day with Etesians occurs when the following criteria are satisfied simultaneously: (a) WDIR is northerly between NW (315∘) and NE (45∘), (b) daily WSP exceeds its long-term median, and (c) criteria (a) and (b) are fulfilled for at least 2 consecutive days. The latter criterion filters out intermittent disturbances unrelated to the semi-persistent synoptic system of Etesians winds (Hochman et al., 2019). Finally, we calculate the number of Etesian days (NED) per year as the sum of all Etesian days in the JA season. Our definition of the NED index is similar to the “northerly wind index” of Gomez-Delgado et al. (2019) based on ship-log observations with the addition that our method identifies days with moderate to strong wind speeds only, which better correspond with the historical L-days index (Poupkou et al., 2011). This is validated by the significant correlation between the NED from the 20CR dataset and the L-days index over the common 1892–2006 period (r=0.59, p<0.01 based on a two-tailed t test; see also Fig. 2c). Other methodologies for calculating days with Etesians (e.g. surface pressure gradients) give consistent results (e.g. Dafka et al., 2016).

Despite differences in the period considered and the horizontal resolution that could impact the representation of the wind speed climatology (Kotroni et al., 2001), the CESM model simulates realistic mean wind and SLP fields in the summer EMed (Fig. 1a). A comparison of the probability density functions of the northerly WSP in the central Aegean Sea (red star in Fig. 1a) finds comparable mean and higher statistical moments, with median values applied to the classification methodology of Etesians (second criterion) of 6.8 and 7.1 m s-1, respectively (Fig. 1b). The NED over the last millennium in one arbitrarily selected CESM-LME run varies from 3 to 50 d, with a median of 21 d, and similar statistics are found in the other runs. NED in the 20CR dataset ranges from 9 to 48 d, and the L-days index demonstrates a minimum of 3 and maximum of 42 d (Fig. 1c). However, the PDF of NED in the observations is skewed to higher NED values compared to CESM-LME, likely related to the coarser horizontal resolution of the CESM model compared to 20CR.

We first analyse the ensemble-mean time series of the northerly WSP (criterion a; see Sect. 2.2) and NED, respectively, over the last millennium, as simulated in the CESM-LME (black lines in Fig. 2). We find that the most notable deviations in the ensemble-mean WSP and NED are simulated in volcanically active years. Major volcanic eruptions, as noted by increased outgoing NH (0–90∘ N) clear-sky short-wave radiation at the top of the atmosphere (grey lines in Fig. 2) that is used as a proxy for stratospheric aerosol loading, frequently reduce the northerly WSP and consequently the NED up to 2 years after the eruption. For Samalas, the largest eruption in the last millennium, CESM-LME simulates negative WSP anomalies exceeding -1.3 m s-1 in the 2 post-eruption years, while the ensemble-mean NED hardly exceeds 10 d in the summer of 1259. Interestingly, the second-strongest WSP and NED reduction is found for the NH Laki eruption, even surpassing changes associated with Kuwae, the second-strongest eruption in the last millennium, and all other tropical eruptions (e.g. Tambora).

Figure 2c zooms over the common 1836–2006 period to compare CESM-LME with the 20CR and L-days. We caution about the first years in 20CR given the scarcity and quality of observations, but this early period lacks any large eruption; 20CR does not simulate any significant NED reduction after Krakatau, broadly consistent with CESM-LME. On the other hand, the Pinatubo eruption reduces NED in the summer of 1992 (and wind speed; not shown) as evidenced in both the 20CR and L-days datasets. It is interesting to note that the absolute minimum NED anomaly in the 20th century is found in the summer of 1913, with only 8 and 3 d, respectively, in 20CR and L-days, which could be associated with the Novarupta (Katmai) eruption in Alaska (Hildret and Fierstein, 2012). However, this postulation is not supported by the CESM-LME runs, possibly because of the unrealistically weak forcing imposed in the model. Such an underestimation of the NH volcanic forcing is commonly found in many forcing datasets used in model intercomparison activities (Toohey et al., 2019). A comparison of the NH clear-sky, top-of-atmosphere, outgoing SW radiation finds about 5-times-stronger anomalies in the 20CR compared with CESM-LME, further supporting this possibility (not shown).

The previous discussion highlighted a tendency for reduced NED in volcanically active periods. This is further substantiated with a superposed epoch analysis of NED in the 5 post-eruption years (years 1 to 5) compared to the pre-eruption 5-year average (years -5 to -1; see Fig. 3). To facilitate the comparison, anomalies are given in percentages. For completeness, Fig. S1 in the Supplement presents modelled responses for all selected eruptions from 850 to 2005. After Samalas, the NED declines in all individual runs (17 thin grey lines) with maximum anomalies up to -90 % peaking at year +1. The large number of realisations in CESM-LME facilitates the detection of volcanic signatures versus natural variability (Stevenson et al., 2016), and we detect a significant response exceeding 2 standard deviations of the previous 5 years. The absolute minimum NED in the summer of 1259 is 3 d, found in two runs, essentially describing a summer without Etesians. Similar summers are also simulated after Tambora, but with higher intra-ensemble spread regarding the timing of the peak reduction, given that NED in individual runs reduces in either year -0 or year +1. As for strong tropical explosive eruptions, the NH Laki effusive eruption causes a significant NED decline, with the strongest ensemble-mean reduction of -60 % found in year -0. In individual model runs, NED anomalies are as large as those of Samalas, with magnitudes up to -80 % (Fig. 3b). Because of the high latitude and the relative low altitude of the eruption, the lifetime of the Laki influence is relatively short; NED returns to the pre-eruption conditions the following year. Yet, Pinatubo does not significantly impact the ensemble-mean NED (Fig. 3d), despite the fact that some runs show strong reductions with amplitudes exceeding -60%. The peak response of -20% at year -0 is not significant and in addition is suspiciously early as it peaks just 2 months after the eruption date (June 1991). In observations, NED drops by 20 %–40 % in the summer of 1992 in both 20CR and L-days, but the signal is neither significant based on a t test nor exceptional as NED also drops by about the same magnitude at year -4. Hence, we conclude that the observed NED reduction in the summer of 1992 is not significant and might not be related to the volcanic forcing, which is consistent with the model results. This is further discussed in the following section.

As the historical volcanic forcing is larger than in recent events (e.g. see grey lines in Fig. 2), it is not surprising that the CESM simulates a robust decline in the ensemble-mean NED for volcanoes prior to 1900 only, while signatures are insignificant over the 20th century. In addition, not all strong eruptions impact Etesians in the same way because of interhemispheric differences in the forcing. This is demonstrated in Fig. 4, which shows NED anomalies separately for all SH, tropical and NH eruptions from 850 to 1900 (bold eruption years in Table 1). For the SH eruptions, only Kuwae causes a significant NED anomaly at year +1 but nevertheless of considerably weaker magnitude compared to strong tropical and NH volcanoes (Fig. 4a). Tropical eruptions typically reduce NED in year +1, as previously discussed for Samalas and Pinatubo (Figs. 3 and S1). NED also reduces after all NH eruptions, with a multi-eruption average anomaly of -30 %, larger than the mean response of all tropical eruptions. Given the generally weaker amplitude of NH eruptions compared to strong tropical eruptions, CESM-LME suggests an amplified sensitivity of Etesians to NH volcanoes. This is consistent with studies that show disproportionally stronger climate forcing between NH high-latitude and tropical eruptions of an equal magnitude (Liu et al., 2016; Toohey et al., 2019).

CESM-LME simulates a decline in NED of considerable magnitude in the post-eruption summers, which additionally is found to be sensitive to the hemisphere of the volcanic forcing. To understand these key findings, we need to investigate the large-scale circulation changes in relation to the ISM. As a first step, we analyse large-scale surface temperature, SLP and wind anomalies using monthly mean data.

Previous analyses of the CESM-LME simulations have identified the coldest annual NH temperatures after major volcanic eruptions, with magnitudes generally stronger than in the reconstructions, possibly related to uncertainties in the specified volcanic forcing (e.g. Otto-Bliesner et al., 2016). Likewise, the strongest summer (JA) cooling in southern Europe and northern Africa is simulated after Samalas, with anomalies exceeding -3.5 K in the EMed, Balkans and Levant (Figs. 5a and S2 for all eruptions). A similar albeit weaker cooling is also simulated in the post-eruption summers after the Laki and Pinatubo eruptions, which can be explained by the reduction in the incoming radiation by the volcanic aerosols in the stratosphere. The considerably weaker sea surface temperature anomalies in the EMed are related to the heat capacity of water that dumps a response to an intermittent forcing. The direct radiative cooling is superimposed on dynamical signatures associated with changes in the large-scale circulation and regional land–sea contrasts. These dynamical signatures can be isolated by subtracting the zonal mean temperature response (e.g. about -2 K in the 30–40∘ N band). This approach isolates a spatial pattern that is characterised by an amplified cooling over the land masses in the EMed, Balkans and Levant (Fig. S3). In the extreme case of Samalas, the dynamical cooling pattern contributes to about -1.5 K in the total cooling simulated in the EMed.

The surface cooling in the post-eruption summers is associated with an increased SLP in the EMed and Anatolian low, exceeding 300 Pa in the case of Samalas and Laki (contours in Figs. 5 and S2). This indicates a weakened SLP gradient over the Aegean Sea, which is associated with reduced wind speeds as evidenced by the southerly anomalies of about 1 m s-1 (arrows in Figs. 5 and S2). Hence, fewer days with Etesians are expected under such surface conditions as found in Fig. 3. The positive SLP anomalies extend throughout the Middle East and the Arabian Peninsula. The north-easterly wind anomalies in the Arabian Sea oppose the prevailing south-westerlies in the summer season, which is associated with a weakened ISM. This can also explain the anomalous surface warming simulated over India given that a reduction in the cloud amount and increased downward short-wave radiation in years of reduced ISM can cause positive surface temperature anomalies (Dogar and Sato, 2019). This pattern is robust and is simulated for all strong NH and tropical eruptions prior to 1900 (Fig. S2). We note that the strongest ensemble-mean surface warming over India is found after Laki, with temperature anomalies exceeding +1.5 K (Fig. 5b). Likewise, there is also some evidence in CESM-LME for north-west wind anomalies in the Arabian Sea after Pinatubo in the summer of 1992, but the simulated surface cooling in the EMed is considerably weaker compared to 20CR (Fig. 5e). In 20CR, SLP anomalies in the Anatolian low are trivial after Pinatubo, which is consistent with the insignificant observed NED response shown in Fig. 2d.

This pattern of forced response in the temperature and surface winds shown in Fig. 5, which has also been simulated with other models (e.g. Dogar and Sato, 2019), suggests a possible connection to the ISM. This possibility is investigated by analysing the omega velocity (expressed in -Pa s-1 to denote an upward direction) fields at 200 hPa (Figs. 6 and S4 for all eruptions). The climatology of omega velocity in JA is characterised by ascending motions over the Bay of Bengal, India and Nepal (positive contours in Fig. 6), linked to the monsoonal activity, while subsidence prevails over the region of the EMed (Rizou et al., 2018; Logothetis et al., 2020). These two opposite vertical motions are connected in the summer months, linking the Indian and South Asian summer monsoons to the circulation in the EMed (Rodwell and Hoskins, 2001; Tyrlis et al., 2013). CESM-LME simulates negative anomalies in the region of the Indian monsoon after the volcanic eruptions of Samalas, Pinatubo and Laki, indicating a significant reduction in the upward motion and a waning monsoon activity, which can also be inferred by the reduced precipitation (not shown). The anomalous descending in the ISM region is paired with positive anomalies over the EMed, indicating a reduced subsidence in the post-eruption years. A comparable pattern, although of weaker magnitude, is obtained at 500 hPa (not shown). Anomalies are significant at p<0.05 based on a two-tailed t test, and the strongest changes exceeding -0.04 -Pa s-1 (0.02 -Pa s-1) in the ascending (descending) branch are found for the Samalas eruption.

Previous studies detected a substantial decrease in precipitation over land after the Pinatubo eruption, associated with a reduction in the ISM and positive surface temperature anomalies over India (Trenberth and Dai, 2007). This is consistent with the surface temperature and wind anomalies found in 20CR (Fig. 5e) and broadly supports the pattern of warming in India and cooling in the EMed found in CESM-LME. Yet, the surface temperature response after Pinatubo is considerably weaker in CESM-LME compared to 20CR. The reduced ISM in the summer of 1992 causes negative anomalies in the ascending region, but the observed signature in the descending branch in the EMed is negligible and insignificant (Fig. 6e). This is consistent with the weak SLP anomalies shown in Fig. 5e, associated with insignificant anomalies in the observed NED.

Despite the inconsistent signatures after Pinatubo, CESM-LME simulates similar patterns of omega velocity anomalies for all tropical and NH eruptions prior to 1900 (Fig. S4). Moreover, we identify an almost linear relationship between changes in ISM strength and NED anomalies (Fig. 7). Following Logothetis et al. (2022), the ISM strength is approximated by the omega velocity anomalies at 200 hPa (-Pa s-1) averaged over the region of the strongest mean ascending motion (black boxes in Fig. 6). As previously discussed, all selected eruptions cause negative omega velocity anomalies over the ascending region, and this significantly correlates with the corresponding NED anomalies (r=0.8, p<0.01 according to a two-tailed t test) in the first post-eruption year. A linear regression calculates a positive slope of about 2.3 d per 0.01 -Pa s-1 increase, significant with p<0.01 based on a two-tailed t test. If we additionally consider the three SH eruptions before 1900, the linear regression shows a steeper slope of about 3.5 d per 0.01 -Pa s-1, mainly attributed to the outlier of the 1341 eruption, which shows positive NED and omega anomalies (Figs. S1, S4). This linear relationship, therefore, suggests that the reduction in NED can mainly be explained by the response of the ISM to volcanic forcing through the ISM–EMed teleconnection.