We first consider the connection between ASO and zonally averaged temperature and geopotential height (Fig. 2a and b) in MERRA and SWOOSH data. Higher values of ASO are associated with elevated geopotential height and warmer temperature over the polar lower and midstratosphere, as well as lower geopotential heights and colder temperatures in the tropical lower and midstratosphere in reanalysis data (Fig. 2a and b). This effect is consistent with previous work that has shown that transport controls lower stratospheric ozone concentrations (Douglass et al., 1985; Hartmann, 1981; Rood and Douglass, 1985; Hartmann and Garcia, 1979; Silverman et al., 2018).

Figure 2Correlation coefficients between ASO and geopotential height anomalies (a, c, e, g, i) and between ASO and temperature anomalies (b, d, f, h, j) in March for each of the examined data sources: (a–b) MERRA/SWOOSH, (c–f) WACCM-R3, and (g–j) NIWA-R3. The black dots represent locations where correlations are statistically significant at the 95 % level (see Sect. 2.2). Correlations of ±0.5 are thin black. The contour interval is 0.065.

The CCMI ensemble-mean (i.e., the mean of all 42 submodels) correlation between March ASO and March geopotential height and temperature is shown in Fig. 3. The CCMI models capture the connection between the two phenomena, and the magnitude is similar to that observed though somewhat weaker in the lower stratosphere. In order to more meaningfully compare the relatively short observational record to the model output, we divide the model output into 41-year chunks (see Sect. 2) and focus on coupling of ASO with polar cap 100 hPa geopotential height and temperature in Fig. 4a. The x axis of Fig. 4a shows the correlation of ASO with 100 hPa polar cap temperature. The reanalysis is indicated with a yellow asterisk, the CCMI multimodel mean with a large black x, and each 41-year chunk of each model is indicated with a single x. In March, the multimodel mean correlation is weaker than that observed, but individual models simulate a tighter coupling than that observed. Strikingly, a different 41-year subsample of the same model can alternately simulate essentially no coupling of ASO with lower stratospheric temperatures (the x's near the bottom left of the distribution) or coupling that is similar in magnitude to that observed. The y axis of Fig. 4a shows the correlation of ASO with 100 hPa geopotential height in March. Models with a tighter connection between ASO and polar cap temperatures also feature a tighter connection between ASO and polar cap geopotential height, and the connection in reanalysis data falls well within that simulated by CCMI models. Results are similar for April as well (Fig. 4b). This validates the fidelity of the coupling between these phenomena in the CCMI models. It is beyond the scope of this paper (which focuses on monthly mean data) to determine the dominant direction of causality, though process-driven studies of observational data have indicated that lower stratospheric ASO anomalies are driven by transport (e.g., Douglass et al., 1985). Anomalous transport can occur both on large spatial scales where the same wind field that advects the ozone across the vortex barrier into the pole also leads to a warmer pole, and also on smaller scales where the causality of the connection between zonal mean temperature and ASO is less obvious. The strength of this connection is quantitatively similar for the period 1970–2010 when ODS concentrations were high and for the period 2052–2092 when imposed ODS concentrations are lower (Fig. S43a and b in the Supplement), suggesting that heterogeneous chemical ozone depletion plays a relatively minor role as compared to transport for the connection between ASO and polar vortex variability.

Figure 3As in Fig. 2 but for the multimodel mean.

Download

Does ASO affect the troposphere? Figure 6a shows the correlation of surface air pressure anomalies with ASO in March. Higher concentrations of ASO are associated with elevated surface air pressure anomalies over the polar cap and over Greenland and with reduced surface air pressure further south in the Atlantic sector; overall the pattern resembles the negative phase of the North Atlantic Oscillation (consistent with Ivy et al., 2017,  the feature over the North Pacific will be discussed in Sect. 4). This relationship can be summarized by computing the correlation of surface air pressure anomalies area-weighted from 80^∘ N and poleward with ASO, and we show the result from MERRA and from all of the CCMI models on the x axis of Fig. 4c. The correlation in MERRA data is 0.41 (statistically significant at the 95 % level), which is larger than in most, but not all, of the CCMI models. CCMI models show a stronger relationship in April instead (Fig. 4d); the mean correlation across all models increases from 0.1 in March to 0.14 in April. While these correlations are statistically significant at the 95 % level, the variance explained is low. This point is illustrated visually in Fig. 5d, e, and f, which compares the ASO anomaly and polar cap surface pressure anomaly for each year of all CCMI integrations. Hence, while the CCMI multimodel mean simulates a weaker connection between ASO and polar cap surface pressure (PS) in March than is observed, the observed association is enveloped by the range of CCMI models, and the relationship between ASO and polar cap PS in the CCMI models is still statistically significant at the 95 % level (in agreement with  Calvo et al., 2015).

Figure 4The connection between correlations of ASO anomalies and anomalies in various metrics of the stratosphere and troposphere. In each subplot, the x and y axis represent correlation of two measures. Each x represents a different 41-year chunk of model output, and reanalysis/observational data is represented by asterisks. The first row compares the correlation of ASO with polar cap stratospheric temperature (T) with the correlation of ASO with polar cap geopotential height (Z) in (a) March and in (b) April. The second row compares the correlations of ASO and polar cap height (Z) with ASO and polar cap surface pressure (PS). The third row compares the correlations of polar cap Z and PS with polar cap T with PS. The fourth and final row is similar to the second row except that polar cap Z is regressed out.

Download

Figure 5A scatter plot of (y axis) polar cap surface pressure with March (x axis; a, b, c) polar cap geopotential height and (x axis; d, e, f) Arctic stratospheric ozone. March polar cap surface pressure is used for (a, b, c), and April polar cap surface pressure is used for (d, e, f), corresponding to the month with strongest coupling. Each year of each CCMI integration is marked with a diamond. All correlations are statistically significant at the 95 % level as given by a two-tailed Student t test, as are the differences in correlation between (a, b, c) and (d, e, f).

What may explain the diversity in the strength of the coupling between ASO and polar cap PS among the models? To provide a possible answer to this question, we repeat on the y axis of Fig. 4c and d the correlation of ASO with polar cap height at 100 hPa (the y axis of Fig. 4a and b). There is a statistically significant relationship between coupling of ASO with 100 hPa polar cap height and with the impact of ASO on the surface. Namely, 41-year model subsamples in which the connection between ASO and stratospheric polar cap height is stronger also simulate a stronger connection between ASO and polar cap PS. Specifically, the correlation between these two is 0.43 in March and 0.58 in April (Fig. 4c and d), though clear outliers do exist. Hence the strength of the coupling between ASO and polar cap PS is dependent on coupling between ASO and polar cap height at 100 hPa. This dependence suggests that there is ambiguity as to whether ASO is indeed the proximate cause for the surface climate anomalies evident on the x axis of Fig. 4c and d. Specifically, ozone anomalies are usually accompanied by anomalies in the Arctic vortex (Fig. 4a and b), and previous work has shown that spring Arctic vortex anomalies independent of ozone can influence surface conditions (Black and McDaniel, 2007; Ayarzagüena and Serrano, 2009; Hardiman et al., 2011). We now demonstrate that vortex anomalies are associated with polar cap PS anomalies in the CCMI models as well and then try to isolate statistically the relative importance of ASO.

We first show that stratospheric polar cap geopotential height and temperature are even more strongly related to tropospheric variability than is ASO. The y axis of Fig. 4e and f considers the correlation of 100 hPa polar cap height with polar cap PS across the CCMI models as compared to that observed. The correlation of 100 hPa polar cap geopotential height with polar cap PS is more than double the corresponding correlation of polar cap PS with ASO for both the multimodel mean of the CCMI models (0.57 as compared to 0.1 in March, and 0.39 as compared to 0.14 in April) and for reanalysis data. Results are similar for the correlation of 100 hPa polar cap geopotential temperature with polar cap PS (on the x axis): the correlation of ASO with polar cap PS is also weaker than that of 100 hPa polar cap temperature with polar cap PS. This effect is highlighted visually in Fig. 5, which compares 100 hPa polar cap height with polar cap PS for each model integration in Fig. 5a–c and ASO with polar cap PS in Fig. 5d, e, and f. It is clear that the connection between 100 hPa geopotential height and surface conditions is far stronger than that between ASO and surface conditions.

It could well be that the apparent connection between ASO and polar cap PS is just a byproduct of the tight connection between polar cap stratospheric geopotential height with both parameters rather than a direct connection. We address this issue by using linear regression to statistically remove the portion of polar cap PS variability and ASO variability that is linearly related to stratospheric polar cap geopotential height for each model and then consider whether there remains any lingering connection between ASO and polar cap PS. The results are displayed on the x axis of the bottom row of Fig. 4. The correlation of ASO with polar cap PS is now negative both for most CCMI models and also for observations – high ozone is associated with lower heights over the pole – opposite of the relationship when Z was not regressed out. This statistical argument indicates that there is no distinct pathway whereby ASO affects the polar troposphere, but rather any effect on the polar troposphere is through its coupling to the dynamics of the lower stratospheric polar vortex.

Figure 6Correlation coefficient between ASO anomalies and surface pressure (PS) or surface temperature (TS) anomalies for MERRA/SWOOSH, WACCM-R3, and NIWA-R3, where R3 indicates the third realization. The left and center columns represent the correlation between ASO and surface pressure. The region of interest discussed in Sect. 4 is marked by the red frame and is defined by 20–50^∘ N and 150–200^∘ W. The leftmost column shows the correlation between March ASO and March surface pressure, and the center column shows the correlation between March ASO and April surface pressure. The rightmost column shows the correlation between the March ASO anomalies and April near-surface temperature. The region of interest discussed in Sect. 4 is marked by a red frame and is defined by 15–40^∘ N and 130–180^∘ W. In all the plots, the black dots mark the areas in which the correlation is statistically significant at the 95 % level. The contour interval is 0.065.

The bottom row of Fig. 4 uses linear techniques to deduce that any influence of ASO on surface climate is mediated through the dynamics of the lower stratospheric polar vortex; however, there could be feedbacks between ASO, the lower stratospheric polar vortex, and surface climate. For example, the x axis of Fig. 7 shows the correlation between polar cap temperature and polar cap surface pressure for the period 1970–2010 in blue, for the period 2011–2051 in green, and for the period 2052–2092 in red. The correlation between polar cap temperature and surface conditions is stronger in the future when ODS concentrations are minimal and heterogeneous chemical ozone depletion less common (among other changes in the climate). Similar results are found for the correlation of polar cap geopotential height and polar cap surface pressure (y axis), and also for the upper troposphere (not shown). Indeed, the correlation of ASO with polar cap surface pressure is also marginally higher in a future with low ODS concentrations than in the historical period (Fig. 5), further suggesting that heterogeneous chemical ozone depletion is not a crucial factor for a strong downward impact. We leave for future work an explanation for this relationship.

Figure 7Correlation between stratospheric variability and polar cap surface pressure (PS), with simulations conducted during the period 1970–2010, 2011–2051, and 2052–2092 in separate colors in (a) February, (b) March, and (c) April. For the x axis, the correlation is conducted with polar cap temperature at 100 hPa, and for the y axis, the correlation is conducted with polar cap geopotential height at 100 hPa.

Download

Xie et al. (2016) recently suggested that Arctic stratospheric ozone anomalies influence North Pacific sea level pressure anomalies which in turn affect subtropical sea surface temperatures, and the subtropical sea surface temperature anomalies modulate ENSO through the seasonal footprinting mechanism approximately 20 months later. Their conclusions were based on the limited observational record and model experiments with one model, and we now consider whether CCMI models capture this association in order to assess the robustness of this effect. The CCMI models included in our study contain both ozone variability and an internally generated ENSO (Fig. 1), hence they are the first multimodel ensemble which simulates the relevant underlying processes.

We begin with the lagged correlation of ASO and ENSO for lags ranging between −40 and 40 months using MERRA/SWOOSH data (Fig. 8a) following Xie et al. (2016) in order to establish context for the CCMI models. ENSO is positively correlated with ASO 7 months later (at lag equal to −7), such that El Niño leads to more ozone 7 months later. The more robust relationship is a negative correlation between ASO and ENSO 20 months later (at lag equal to 20) such that more ASO leads to a La Niña 20 months later, though this relationship is not statistically significant at the 95 % level using a two-tailed Student t test. Xie et al. (2016) noted that the strongest observed connection between ASO and ENSO is March ASO with ENSO 20 months later, and hence we show the lagged correlation between March ASO and ENSO in Fig. 8b. The correlation of ENSO in January with ASO 2 months later in March is 0.2, and the correlation of ASO in March with ENSO ∼18 months later is nearly −0.5. All of these relationships are in agreement with Xie et al. (2016), though we find that this relationship is slightly less statistically significant than Xie et al. (2016).

Do the CCMI models capture these relationships between ASO and ENSO? We first focus on the multimodel mean lagged correlation between ENSO and ASO in Fig. 9. The lagged-correlation function between ASO and ENSO reveals the same general lead/lag behavior as seen in the MERRA/SWOOSH model but generally with much weaker correlations. The positive correlation at lag $=-\mathrm{4}$ (Fig. 9a) exceeds 0.1, corresponding to El Niño leading to enhanced ASO after 4 months, and this effect is highly statistically significant and is consistent with that found in CCMVal models (Cagnazzo et al., 2009). Hence the models are able to capture the forcing of ASO anomalies by ENSO, with El Niño leading to enhanced ASO and La Niña leading to reduced ozone (Cagnazzo et al., 2009).

Figure 8Lagged correlation between ASO and ENSO for (a–b) MERRA/SWOOSH, (c–f) WACCM-R3, and (g–j) NIWA-R3. The left column shows the lagged correlation for all calendar months, and the right column shows the association between March ASO of every year and ENSO of every month across all years after regressing out ODS and greenhouse gas changes as described in Sect. 2. The green lines mark the Student t test 95 % confidence level – correlations that exceed the upper green line or are more negative than the lower green line are statistically significant. The red line in subplot (b) represents the correlation between March polar geopotential height and ENSO.

Download

The lagged correlation also contains a negative peak at a lag of +10 months, and statistical significance is maintained from a lag of 6 months up to a lag of 22 months. This result is in agreement with the feature seen in MERRA/SWOOSH and Xie et al. (2016). However the correlation at positive lags in the multimodel mean is a factor of 5 weaker than that observed; furthermore, the correlation is strongest not 20 months after the ASO anomalies but rather 10 months later. We have also computed the correlation coefficients between March ASO and ENSO in order to focus on the season where the observed relationship peaks (Fig. 9b). A statistically significant negative correlation is found when ASO leads ENSO by 8–14 months. An additional, smaller yet also statistically significant, negative local peak is obtained at a lag of 20 month (at 1 November in Fig. 9b), qualitatively in agreement with the observed relationship but much weaker in magnitude.

Figure 9As in Fig. 8 but for the multimodel mean.

Download

The multimodel mean lagged correlations differ from that observed in two aspects: they are weaker and also peak at earlier lags. This does not necessarily imply that the models are inconsistent with the observed effect, as some models do show a relationship that resembles that observed. To highlight this effect, we compare adjacent 41-year subsamples for a given model. The results are shown in the bottom four rows of Fig. 8. Figure 8c, e, g, and i are constructed analogously to Fig. 9, but they focus on two different 41-year subsamples from the same integration for two models. These two subsamples indicate an opposite lead–lag relationship between ENSO and ASO between adjacent subsamples for the same integration. For example, WACCM run no. 3 over the years 2011–2051 indicates a similar lead–lag relationship to the one observed in MERRA/SWOOSH (Fig. 8c). Yet, over the years 2052–2092 (Fig. 8e), the exact same model integration suggests that any apparent modulation of ENSO by ASO is weak and only appears at much shorter lags. Another example for the intramodel variability is evident in the results of the NIWA-R3 model (Fig. 8g and i). NIWA-R3 simulates a lead–lag relationship that does not resemble the observed one over the years 2011–2051, whereas for the years 2052–2092 the lead–lag relationship shows some similarities. Thus, these two subsamples of the same model are different from each other and therefore give different results when comparing to MERRA/SWOOSH. Similar results are evident if we focus on the lagged correlation of March ASO with ENSO (right column of Fig. 8).

The model spread in the connection of ASO with ENSO 10–20 months later does not appear to be associated with stratospheric dynamical or temperature variability, as the correlation between ASO and stratospheric temperature and geopotential height is qualitatively similar in these experiments. Specifically, these specific subsamples from WACCM (compare Figs. 2c and d to 2e and f) and NIWA (compare Fig.s 2g and h to 2i and j) both show a positive correlation between ASO and polar cap geopotential and temperature. Hence internal tropospheric or oceanic variability appears to be the source of the difference in the connection between ASO and ENSO. That internal tropospheric or oceanic variability can mask the connection between ASO and ENSO over a 41-year period suggests that the observational polar ozone record is too short to reach firm conclusions as to the connection between ENSO and ASO. More specifically, it is possible that internal climate variability could have contributed to the apparent observed effect. The possible role of internal variability can be reduced as much as possible by computing the multimodel mean of the correlations (by averaging over all lagged-correlation coefficients), and as discussed above the multimodel mean correlation is much weaker and peaks sooner than that observed.

Xie et al. (2016) propose a specific mechanism between ASO and ENSO, and we now evaluate whether this mechanism is operating in the CCMI models. Specifically, they argue that higher ASO leads to higher PS over the North Pacific in March and April (red box in Fig. 6a and b), which directly leads to warmer sea surface temperatures in the subtropical North Pacific (red box in Fig. 6c) as the cold continental winds off Eurasia are weakened. This warming of the subtropical North Pacific then leads to a La Niña event due to the seasonal footprinting mechanism (Vimont et al., 2003). It is difficult to deduce any evidence for these effects in the multimodel mean surface pressure and sea surface temperature response (Fig. 10), though the multimodel mean correlation between ASO and ENSO was also weak. Even if we focus on the model subsamples that did succeed in capturing a relationship between ASO and ENSO (e.g., WACCM run no. 3 for the years 2011–2051 and NIWA run no. 3 over the years 2052–2092), there is no better agreement with the observed PS and surface temperature anomalies than in the model subsamples that failed to capture the observed relationship between ASO and ENSO (Fig. 6). The correlation between April surface temperature in the red boxed region of Fig. 6c with March ASO averaged across all CCMI models is 0.02, and while individual model subsamples simulate correlations as large as the observed correlation of 0.3, there is no significant relationship between the models that simulate greater warming in this region in response to enhanced ASO and the relationship between ASO and ENSO 20 months later. Hence there is no evidence that the mechanism of Xie et al. (2016) is present in the CCMI models.

Figure 10As in Fig. 6 but for the multimodel mean.

One might hypothesize that the apparent relationship between ASO and ENSO ∼20 months later is driven by the autocorrelation of ENSO: El Niño (which typically drives higher ASO) is often followed by a La Niña event a year or two later, and one might therefore suppose that the La Niña event 20 months after the high values of ASO is due to internal oceanic ENSO dynamics and does not involve the stratosphere (e.g., Garfinkel, 2017).

We test this hypothesis using Pearl causality as described in Sect. 2.2; briefly, we test whether ENSO has a parent ASO. The results of the causality test are shown in Fig. 11, a heat map of parents of ENSO where only statistically significant parents are indicated. Figure 11a shows results for the PCMCI algorithm, and Fig. 11b shows results for the linear mediation algorithm (see Sect. 2.2).

For the observational data (i.e., SWOOSH 1984–2014) we can see that ENSO has two parents: ASO 20 and 22 months prior to ENSO. This connection is obtained from both MCI and linear mediation, and it is a negative connection. This result is in agreement with both Xie et al. (2016) and our correlation results (Fig. 8a), albeit with lower regression coefficients and correlations: for the MCI step we get a score of −0.17 for ASO(−20) and −0.19 for ASO(−22), and in the linear mediation −0.21 for ASO(−20) and −0.22 for ASO(−22). These regression coefficients are lower than the simple lag correlation of ENSO with ASO (−0.32 in our study and −0.35 in Xie et al., 2016, both for ASO (−20)).

Figure 11Pearl causality analysis of the parents of ENSO using the (a) PCMCI algorithm and (b) using the PC algorithm with linear mediation. See Sect. 2.2 for details of the algorithm. Observations are in the bottom row, and each of the CCMI models is shown separately. Parents of ENSO are shown in color, with darker shades for more robust parents.

Download

In the CCMI ensemble, ASO is not a parent of ENSO over the historical period for any model for lags near −20 (and in fact is a parent with the opposite sign in one model). In contrast, ENSO(−10) is a parent for half of the models, and for these models both the MCI step and the linear mediation give similar results. For later time periods, ASO is a parent of ENSO for some models for lags near −20, such that the observed relationship between ASO(−20) and ENSO is simulated on occasion, though in other models ENSO(−10) is a more important parent. Hence over the period 2011 through 2092 the tendency noted by Xie et al. (2016) is evident in some of the CCMI models, though the magnitude of the connection is much weaker than observed.

The association between ASO and ENSO in the CCMI models does not appear to be related to dynamical changes in the Arctic vortex. The red lines in Figs. 8b and 9b indicate the lagged correlation between March 100 hPa polar cap geopotential height and ENSO. El Niño is associated with higher March geopotential height at zero lag in the CCMI models (though not in observations over this time period;  Hu et al., 2017; Domeisen et al., 2019; Garfinkel et al., 2019), but there is no indication that higher geopotential height (typically associated with high values of ASO) leads to La Niña a year later. Furthermore, Zpole is not a parent of ENSO for any lead between 10 months and 27 months in observations and is also not a parent for most models either. (Recall that polar cap PS is more strongly affected by polar cap geopotential height than ASO.)

Overall, the CCMI models do show a tendency of ASO variability to lead ENSO variability, but the effect is much weaker than that observed. This does not imply that the models are inconsistent with observations, as individual models do capture associations as strong as that observed. Rather, internal climate variability could be contributing to the apparent observed connection between ASO and ENSO.

The CCMI output data were downloaded from the Centre for Environmental Data Analysis (CEDA, 2019; http://data.ceda.ac.uk/badc/wcrp-ccmi/data/CCMI-1/output/). CESM1-WACCM data were downloaded from http://www.earthsystemgrid.org (last access: 17 July 2019). For instructions for access to both archives see http://blogs.reading.ac.uk/ccmi/badc-data-access (last access: 17 July 2019).

The supplement related to this article is available online at: https://doi.org/10.5194/acp-19-9253-2019-supplement.

OH performed all data analysis and helped prepare the manuscript. CIG helped prepare the manuscript and provided overarching guidance and ideas. SZZ performed the Pearl causality calculation. The other coauthors carried out the numerical experiments and provided feedback on the paper.

The authors declare that they have no conflict of interest.

This article is part of the special issue “Chemistry–Climate Modelling Initiative (CCMI) (ACP/AMT/ESSD/GMD inter-journal SI)”. It is not associated with a conference.

This research has been supported by the European Research Council (FORECASToneMONTH grant no. 677756), the Israel Science Foundation (grant no. 1558/14), the DECC/Defra Met Office Hadley Centre Climate Programme (grant no. GA01101), the Bundesministerium für Bildung und Forschung (BMBF grant), and the New Zealand Royal Society Marsden Fund (grant 12-NIW-006).

This paper was edited by Paul Young and reviewed by two anonymous referees.