 Thank you. I am Irene Maviglia from CMCC in Bologna, Italy, and I'm going to talk about the Atlantic Multidigradal Variability, which is examined here in a set of multi-century pre-industrial simulations performed with different couple general circulation models. The focus is on the assessment of AMV stationarity. Here I briefly show you two illustrative examples of no stationarity found in NIO time series on the left, based on observations, and in ENSO time series on the right, based on a model. Regarding NIO, RLM Valun noticed the no stationary behavior in the time evolution. If you look at the bottom panel, you can see that the large part of NIO variance comes from the, at biennial periods, comes from the early part of the record, if you see this gray area at the top right corner, while the variability between six and ten years is more prominent during the latter half of the record. This indicated that there is a shift in the associated to the period of the largest NIO variability. Regarding ENSO, Wittenberg, in this very long simulation, identified different regimes of variability that are highlighted here in colored boxes. So we can see periods with very intense events, periods with hard-to-yanny variability, and periods with, that are very irregular or regular, both in amplitude and periods. So there are periods that resemble the temporal sequence of the observational record that is reported on the bottom, but there are also other periods that may not contain the same oscillation. So, however, we don't know about the Atlantic multi-decadal variability stationarity. So the focus of this research is on the AMV. We've already seen these pictures. So these are the observed AMV, and the average of the North Atlantic surface temperature in the black box gives us the time series on the top. And we can see that the multi-decadal component of variability spectrum is only marginally resolved because observations last 100, 150 years at most. So the question is, are these observed oscillatory changes truly representative of the long-term AMV behavior that can emerge looking at longer datasets? So we know from proxy analysis from Sanger et al. 2009 that Atlantic assistee had very different spectral characteristics in the past because only the last periods show power at multi-decadal time scale similar to the observed AMV here, while there is no such significant multi-decadal power for other past periods epochs. So we can wonder if in the future, also in the future, AMV will have different spectral characteristics. We can start so with our research, looking at the phenomenology, so the first evidence of no-stationarity in AMV time series. We looked at the low frequency and internal variability, so we selected only the longest and pre-industrial simulation that our analyzed models are listed in the table. And on the left side, you can see the time series, the corresponding time series. And already from a visual inspection, we can guess an stationary behavior. In fact, in analogy with Wittenberg's approach, I pointed out at different epochs with different characteristics that we can identify in the AMV time series. So for example, we can see periods with mostly warm skewed events in red, periods with moderate, nearly sinusoidal events in green, and periods with intense and longer period events in purple, and small amplitude events in yellow. So time-evolving view of AMV spectral features highlights a strong dependence to the selected time interval, revealing the no-stationarity as a prominent feature of AMV. Here, AMV autocorrelation is diagnosed for chunks of moving 200-year-long windows. And for most of the models, we can see that the dominant autocorrelation time scale changes with the period. And here, I show you two limit cases. So for most of the model, we found that the AMV is the no-stationarity. For example, like in the MPI-ESMLR model, where a 50-year time scale characterizes the initial part of the record, later shortening down to a 25-year-long time scale, as indicated by these sloping contours. On the other hand, the NOR-ESM-1M model shows an almost stationary AMV throughout the whole length of the series. So we need a statistical test in order to objectively assess this detected no-stationarity. And the question now is, does this AMV modulation arise by chance? So the null hypothesis is a statistically stationary AMV, with no underlying cause more than a round of outcomes of flipping a coin. So in order to check this, we create a large set of analytical time series with the same spectrum than the true one, but with random phases in each of the Fourier modes. And we use this distribution to compute a confidence band for auto covariance and see if there are true values that exceed this confidence band. These are the results. So all the points, the circles above the red line are the models that pass the test. So for most of the models, 10 out of 11 models, we can say that the AMV time series is not stationary. That means that these models, display epochs, characterized by auto covariance, which significantly deviates from the auto covariance of the whole time series. Now we are looking at the mechanisms. So in particular, what drives this detected AMV no-stationarity? And we expected changes in AMV-AMOCK relationship and changes in teleconnection patterns to understand the role of internal ocean-only processes and coupled atmosphere-ocean interaction processes. A widely accepted candidate we have already seen to explain the Atlantic multidegadal variability is the AMOCK. In fact, lagged correlation, AMV-AMOCK correlation for the entire time series shows that for all the models, the maximum of lagged correlation occurs when AMOCK leads AMV by a few years. You can see the year of maximum correlation in red. So it ranges between one year to six years. And this is coherent with the mechanism that an increase in the overturning drives the warming of the Atlantic. But if we look at this relationship in a time-evolving view, so again for chunks of 200-year time window, we can see that AMV-AMOCK correlation undergoes significant fluctuations with time. And alternating period with higher correlation and periods with very low correlation. In particular, I highlighted here in these green boxes a detected decrease in AMV-AMOCK correlation that may suggest that even if AMOCK does generally play an important role in Atlantic multidegadal variability, but there are also other factors that may contribute. So in order to understand this, we look at the role of the atmosphere. Atmosphere-ocean interactions processes are studied here with the maximum covariance analysis, MCA, applied to North Atlantic assist from the oceanic side and global sea level pressure as a representative of atmospheric surface circulation. So large uncertainties among the models do not allow to associate specific MCA patterns to a specific AMV regime of variability. Nevertheless, for most of the model, AMV autocorrelation presents a shorter time scale around 20 years that seems to alternate with a longer time scale around 60 years. And two modes can be identified looking at the MCA homogeneous and the corresponding heterogeneous maps. So the shorter mode is associated with a triple in the North Atlantic SST that corresponds to a strong signal in North Atlantic NIO-like forcing. Whereas when AMV presents lower frequency time scale, this mode features a monopoly AMV-like in North Atlantic SST and a weaker signal in North Atlantic with respect to mode one. So this suggests that in this case, the dynamics could be especially inside the ocean. This is more evident in the MPISMP model, so we choose this model as case of study. And this model, its North Atlantic SST spectrum presents two peaks at these multi-decadal time scales. So we applied the same statistical test in order to understand which time scale contributes most to the non-stationarity of the entire time series. So we pass band filter the time series in these two gray band regions, so from 20 to 40 and 40 to 80 years. And these are the results, so the long time scales contributes most to the non-stationarity of the entire time series with more than 15% of values that exceed the confidence band. So we turn from the statistical to the physical point of view. These two time scales are associated with the two modes of variability. Here you can see the lagged correlation between mixed layer depth, sea surface temperature, sea surface salinity and subsurface density averaged over the Labrador sea. So these are Labrador sea anomalies correlated with the North Atlantic SST index for the short time scale here on the left and the long time scale on the right. And here we can see the differences. So starting from the short time scale we can see that the cooling of SST, the blue curve at lag minus 13 leads an increase in mixed layer depth and density, so the red and the pink curve. And this in turn leads to AMOC becomes stronger and so warmer water, more warm waters are transported to the Labrador sea regional convection sites and so since the temperature feedback is negative this allows the oscillation we see for the short time scale. The situation is different for the long time scale because here temperature is in face both with mixed layer depth and density curve but since we know that warm waters prevent the deep water formation so in this case the salinity that is the green curve is more effective as a feedback for driving the density anomalies. And this could explain also why here the time scale is longer because salinity feedback is positive so it tends to maintain the original state and also because the temperature anomalies are rapidly damped by heat flaxies with respect to salinity anomalies. Concluding in most of the analysed models AMV exhibits an ostational behaviour, the relationship, AMV-AMOC relationship is intermittent, alternating period with higher and lower correlation. In most of the models short term time scales alternates with longer time scale and this is found also in observations. If you look at the plot here, the central England temperature record we can see these two time scales significantly above the red noise spectrum. And for MPISM model in particular these two time scales correspond to different modes where temperature feedback prevails at short time scale and salinity feedback prevails at long time scale. So the message is the the stationarity detected here suggests that the character of the observed AMV may undergo significant changes in the future so we should be careful in relying only on the observations because they cannot be truly representative of the long term behaviour of AMV. And again the importance of reducing a model inconsistency is even more important in the light of the stationarity detected here. Thank you.