The SMHI Large Ensemble (SMHI-LENS) with EC-Earth3

The Swedish Meteorological and Hydrological Institute used the global climate model EC-Earth3 to perform a large ensemble of simulations (SMHI-LENS). It consists of 50 members, covers the period 1970 to 2100 and comprises the SSP1-1.9, SSP3-3.4, SSP5-3.4-OS and SSP5-8.5 scenarios. Thus, it is currently the only large ensemble that allows for analyzing the effect of delayed mitigation actions versus no mitigation efforts and versus earlier efforts leading to similar 10 radiative forcing at year 2100. We describe the set-up of the SMHI-LENS in detail and provide first examples for its application. The ensemble mean future changes of key variables in atmosphere and ocean are analyzed and compared against the variability across the ensemble members. In agreement with other large ensemble simulations, we find that the future changes in the near surface temperature are more robust than those for precipitation or sea level pressure. As an example for a possible application of the SMHI-LENS, we analyse the probability of exceeding specific global surface warming levels in 15 the different scenarios. None of the scenarios is able to keep global warming in the 21st century below 1.5  ̊C. In SSP1-1.9 there is a probability of approximately 70 % to stay below 2  ̊C warming while all other SSPs exceed this target in every single member of SMHI-LENS during the course of the century. We also investigate the point in time when the SSP5-8.5 and SSP5-3.4 ensembles separate, i.e. when their differences become significant, and likewise when the SSP5-3.4-OS and SSP4-3.4 ensembles become similar. Last, we show that the time of emergence of a separation between different scenarios 20 can vary by several decades when reducing the ensemble size to 10 members.

extended simulations with constant forcing we then select 9 initial states separated by at least 2 years, and from each of the remaining 4 simulations we select 8 initial states, yielding a total of 50 new initial states.
To check the ensemble spread after initialization and whether it captures the full intra-model variability, we compare the ensemble spread in SMHI-LENS in year 1970 (annual mean) to EC-Earth3 historical simulations for CMIP6 (23 members) which have been integrated independently for already 120 model years at this point (Fig. 1). The differences between the two 100 ensembles but these are not significant (at the 5% level) neither for the means nor for the variances of the two ensembles and therefore the two ensembles can be considered independent samples of the same distribution.

Simulations
The 50 members of the historical ensemble were started in 1970 from the 50 initial conditions, using the forcing provided for the historical experiment for CMIP6. The historical simulations run until the end of 2014 followed by several scenarios that 105 cover the 2015-2100 period with forcings according to ScenarioMIP. The following scenarios are included in the large ensemble: • SSP5-8.5 is a high-end scenario that yields a strong warming signal, marking the upper end of a plausible evolution of the climate • SSP5-3.4-OS is an overshoot scenario with a strong warming until 2040 (using the same forcing as SSP5-8.5 until 110 2040), followed by a curbing and net-negative emissions after 2060 resulting in a radiative forcing of 3.4 W m-2 in 2100. The difference between SSP5-8.5 and SSP5-3.4-OS will tell about the efficacy of mitigation measures that set in around the mid-century. Following the CMIP6 protocol, the SSP5-3.4-OS experiment branches off from the SSP5-8.5 experiment in 2040 which means results for SSP5-3.4-OS are only available for the 2040-2100 period.
• SSP4-3.4 also has a radiative forcing of 3.4 W m-2 in 2100, but without the peak and decline of  Differences between these two scenarios can tell about the impact of a previous overshoot and possible nonreversible effects when the forcing at the end of the century is similar.
• SSP1-1.9 is the low-end scenario addressing the needs of the Paris Agreement to reach the 1.5 degree warming level, marking the lower end of a plausible evolution of the climate.
All these scenarios except SSP5-8.5 are from Tier-2 of ScenarioMIP. The wider EC-Earth community is planning to provide 120 between 20 and 30 members of the Tier-1 scenarios, and therefore it was considered more valuable to extend the EC-Earth contribution to CMIP6 with Tier-2 scenarios. Furthermore, the selection of scenarios for the large ensemble was guided by questions about the impact of mitigation and overshoot. In addition, the low-and high-end scenarios span the full range of possible futures.

Data output 125
Limitations on storage capacity do not allow us to save the full model output as it has been done for CMIP6. Instead we select a subset of variables from ocean and atmosphere, and save only daily and monthly means. Tables 1 and 1 list the variables for atmosphere and ocean, respectively. All data from the large ensemble are CMIP6 compliant and are available from any ESGF data portal as part of the CMIP6 data holding. Realisation_id's r101 to r150 from the EC-Earth3 model have been reserved for the large ensemble. 130 The limited output does not allow for any in-depth analysis of extreme events such as strong storms or an extreme precipitation event on sub-daily timescales. We therefore plan to re-run selected periods with full output and, for this purpose, have saved the full model state on Jan 1 of each year, for each member and for each scenario.

Results
The aim of this work is to provide an overview of SMHI-LENS and we therefore focus only on main characteristics of major 135 variables. To benefit from the large number of ensemble members, we not only look at ensemble means but also at the ensemble spread as a measure of the internal variability, both in global mean timeseries as well as in the analysis of regional climate change patterns. More detailed studies with the data from SMHI-LENS are in preparation.

Timeseries
Timeseries of global annual mean temperature and precipitation are displayed in Fig. 2, together with timeseries of AMOC 140 and the Arctic minimum sea ice extent. The ensemble spread is illustrated by the shaded area that shows the full spread, minimum to maximum of the ensemble. The scenarios continue the historical experiment after 2014 with little differences among the different scenarios. They start diverging first around year 2040 for three out of four variables considered here.
The exception is the Arctic sea ice minimum where the reduction in SSP5-8.5 is stronger than in the other scenarios already after year 2030 (Fig. 2d). The temperature timeseries (Fig 2a) shows the anticipated warming of the different scenarios with 145 a strong warming signal in SSP5-8.5 that keeps increasing throughout the 21st century while SSP1-1.9 first overshoots slightly and then stabilises around the mid-century at a level only slightly higher than the present-day climate. Increasing temperatures lead to a more vigorous hydrological cycle with increased global precipitation (Fig 2b) and a decrease in the Arctic sea ice minimum (Fig 2d). We also find a distinct impact on the AMOC that first weakens compared to present-day conditions but partly recovers in all scenarios except for the high end SSP5-8.5 scenario (Fig 2c). The ensemble spread in 150 AMOC is high for the historical period and until the mid-21st century, afterwards there is a tendency towards a reduction in the ensemble spread in all but the SSP5-8.5 scenario. The mitigation measures as represented in SSP5-3.4-OS lead to more or less immediate impacts on global mean temperature and precipitation while an imprint onto the AMOC becomes visible with a delay of approx. 20 years.

Regional patterns 155
The ensemble mean annual mean 2m air temperature (T2m) averaged over 1995-2014 shows the well-known north south gradients with minimum values below -20°C in the polar regions and up to 30°C in the tropics (Figure 3 a). The typical discrepancies from the zonality, for example the tongue of warm air in the northeastern North Atlantic and North Pacific and colder T2m over the parts of the northern hemispheric continents are well reproduced. Details on biases in the mean climate in EC-Earth3 are provided by Döscher et al. (submitted). The standard deviation of T2m, averaged over 1995-2014, across 160 model members shows substantial internal variability with largest variability near the ice edges of the North Atlantic Arctic sector where one standard deviation reaches values of 3 K and more. Also, mid and high latitude regions of northern hemispheric continents and the ice regions around Antarctica experience high internal T2m variability. In subtropical and tropical areas, one standard deviation of T2m variability is generally below 0.5 K.
The ensemble mean temperature change until the middle of the 21st century shows a clear Arctic amplification with the 165 largest warming rates in regions where even winter sea ice disappears, especially in the Barents and Kara Seas. Here, warming exceeds 5 K in all scenarios until 2040-2059, and reaches even more than 10 K in the SSP5-3.4-OS and SSP5-8.5 scenarios. Over the continents, the warming is generally larger than over the oceans, and is smallest over the mid-latitude oceans of the southern hemisphere with warming rates below 1 K.
The general warming patterns are similar in the different scenarios. The warming until 2040-2059 is somewhat more 170 pronounced in SSP5-3.4-OS and SSP5-8.5 compared to SSP4-3.4 and SSP1-1.9. The difference between the scenarios is increasing until the end of the century. While especially SSP5-8.5 shows an accelerated T2m increase until 2080-2099, T2m in SSP1-1.9 does not increase any more compared to 2040-2059. The T2m increase in SSP5-3.4-OS is small after 2040-2059 compared to SSP5-8.5, and is similar to the one in SSP4-3.4 by the end of the century. This shows the impact of the strongly decreasing greenhouse gas emissions in SSP5-3.4-OS after 2040. 175 North Atlantic, SLP biases of up to 2 hPa exist and over parts of the Antarctic, SLP is up to 2 hPa too high compared to ERA5-reanalysis data. The SLP variability across members is generally largest in mid and high latitudes of both hemispheres, and one standard deviation of SLP variability reaches here up to around 1 hPa (Figure 5 b). In the tropics, the SLP variability is small and one standard deviation is below 0.2 hPa. 225 The change of SLP until 2040-2059 is small and not significant at the 95% significance level in many areas. The change in SLP is asymmetric, more pronounced in southern hemispheric mid-latitudes and some subtropical and tropical regions where changes are positive and can reach up to 1 hPa in SSP1-1.9 and SSP3-3.4-OS and up to 1.5 hPa in SSP5-3.4-OS and SSP5- and CMIP5 projections (Vecchi et al. 2006(Vecchi et al. , 2007Bayr et al. 2014).
In polar regions, SLP generally decreases in all scenarios in both hemispheres. The spatial SLP change pattern remains similar in 2080-2099 compared to 2040-2059. However, as for T2m and P, the amplitude of SLP-change in SSP5-8.5 is strongly enhanced compared to the period 2040-2059. In polar regions, SLP is reduced by more than 3 hPa and it is increased by up to 3 hPa in southern hemisphere mid-latitudes. In contrast to the other SSPs, SSP5-8.5 shows significantly 240 increased SLP in most northern hemispheric ocean regions as well. Despite these larger changes until 2080-2099 in SSP5-8.5, the variability strongly dominates over the mean change in all polar regions and in most of Eurasia, North Africa and North America as well as over the North Atlantic. In SSP1-1.9, the mean change is only robust across model members in larger parts of the area between 10° N and 40° S.

Probability of exceeding specific surface warming levels 245
An important question of climate adaptation is the likelihood for passing a specific surface warming level (SWL). The large ensemble allows for a quantitative estimate of the probability of surpassing a given temperature. It is common practice to express the warming relative to pre-industrial levels, in other words the difference between the global mean temperature in the future scenarios and the global mean pre-industrial temperature. The pre-industrial temperature is computed as the ensemble mean of 23 realisations of the historical EC-Earth3 experiments for the 1850-1870 period that have been published 250 on the ESGF. For each year we then compute the fraction of the SMHI-LENS members that exceed a given warming threshold.
The probability for exceeding three different SWLs in the four scenarios is shown in Fig 6. All members of SSP5-8.5 exceed SWL3 after 2060 (Fig 6a). SSP5-3.4-OS that branches off from SSP5-8.5 after 2040 reaches only about 20 % probability for https://doi.org/10.5194/gmd-2020-428 Preprint. Discussion started: 17 February 2021 c Author(s) 2021. CC BY 4.0 License. exceeding SWL3 for the period 2060-2080 and lower probability thereafter, demonstrating clearly the impact of the 255 mitigation that is underlying this specific scenario. SWL2 and SWL1.5 are tightly linked to the Paris Agreement that aims at avoiding warming above 2 or 1.5 degrees. Our results with the 4 scenarios used here reveal that only SSP1-1.9 is likely to keep the warming below 2 degrees (Fig 6b). There still is an almost 40% probability for exceeding SWL2 even in SSP1-1.9 around the middle of the century after which the probability becomes lower again. In the other scenarios the likelihood to pass SWL2 reaches 100% around year 2040 in SSP5-8.5 and about 20 years later in SSP4-3.4. The more ambitious 1.5 260 degrees warming target cannot be reached by any of the scenarios used here, the likelihood to exceed SWL1.5 reaches 100% before 2040 with little difference between the scenarios (Fig 6c). The future analysis of SMHI-LENS will include a more thorough investigation of the impact from an overshoot in the climate trajectory. (1) where denotes the ensemble mean, the std deviation and the number of members in each ensemble. The difference between the two ensembles with 50 members each is significant at the 95% level when t exceeds * (0.95,49) = 2.009 for the two-sided 95% confidence level and 49 degrees of freedom. 275

Separation of scenarios
We apply Eq (1) to the annual temperature means of the SSP5-8.5 and SSP5-3.4-OS experiments to compute the t-score in each gridpoint and for each year. The t-scores are then smoothed with a 5-yr running mean. Figure 7a displays the year after which the smoothed t-scores become larger than * (0.95,49), indicating the year after which the differences between SSP5-8.5 and SSP5-3.4-OS have diverged enough for their difference being statistically significant. Similarly, Fig. 7b shows the year after which the difference between SSP5-3.4-OS and SSP4-3.4 is not significant any longer, telling when the two 280 scenarios have converged.
The differences in annual mean temperature between SSP5-8.5 and SSP5-3.4-OS emerge in most regions between 2050 and 2060, with the exception of Antarctica and the Southern Ocean, Africa south of the Sahara, India and central Australia where the differences become significant after 2060 (Fig 7a). The temperature differences between SSP5-3.4-OS and SSP4-3.4 show larger spatial variability (Fig 7b). There is a hint of a North-South gradient in the year when the difference between these two scenarios ceases to be significantly different. In the Northern Hemisphere the last year with a significant difference occurs during the 2060-2080 period in most gridpoints, with notable exceptions in Northern Canada and Greenland. In the Southern Hemisphere the temperature differences are significant until 2080-2100 over large areas of the Oceans, Africa and Antarctica. Over South America and Australia the temperature difference between SSP5-3.4-OS and SSP4-3.4 ceases to be significant in the 2070-2080 period. 290 How does this result depend on the ensemble size? The t-score that is used to assess if the temperature differences between 2 scenarios are significant is proportional to the square root of the ensemble size. Furthermore, the * value for testing significance depends on the degrees of freedom that in turn depend on the number of ensemble members. Let us now assume that for each of the scenarios used here we have a hypothetical ensemble with the same mean and variance as the large ensemble, but only 10 members. The t-scores for the difference between two scenarios are first scaled by √5 and then 295 compared to * (0.95,9) = 2.262 to assess significance at the 95% level. This reflects the larger uncertainty of all estimated parameters, given the smaller sample size. The results for the time of detection of significant differences between SSP5-8.5 notably the Northern Hemisphere continental areas do not show any significant temperature differences between these two scenarios during the 21st century if only 10 ensemble members were available. And even in regions where differences between SSP5-3.4-OS and SSP4-3.4 would still be significant, the differences would stop being significant several decades ahead of the time when it happens with 50 members, thus reducing the period where the two scenarios can be considered to 305 be distinct from each other. This would be a clear drawback for any studies of the impact from the overshoot in SSP5-3.4-OS as the number of available years for such an analysis would be limited. Fig. 7 is a clear example for the need of sufficiently large ensembles when assessing differences between certain scenarios to assess the impacts of mitigation measures.
The analysis of the emergence/cessation of significant differences between different experiments could be expanded to all scenarios, this would however be beyond the scope of the present paper to provide an overview over SMHI-LENS and will 310 be saved for future studies.

Discussion and conclusions
Here we have presented an overview of the SMHI Large Ensemble that consists of 50 members done with the EC-Earth3 model. We described the process of creating a large set of initial conditions for 1970 starting from 6 members of the ensemble of the historical experiment that in turn had branched off at various points in time from the piControl experiment. 315 The future projections, following the ScenarioMIP-protocol, have shown the anticipated results: a strong warming with SSP5-8.5, an overshoot in the warming with SSP5-3.4-OS in the middle of the century followed by a negative warming trend https://doi.org/10.5194/gmd-2020-428 Preprint. Discussion started: 17 February 2021 c Author(s) 2021. CC BY 4.0 License. towards the end of the century, a continuously increasing warming with SSP4-3.4 reaching the same level of warming as SSP5-3.4-OS towards 2100, and a limited warming with SSP1-1.9. Not surprisingly, the projections in the large ensemble are in line with other CMIP6 results, the advantage of the large ensemble being that it allows us to better quantify the impact 320 of internal variability on the changes and thus derive results subject to reduced uncertainty.
When comparing the mean future change against the variability of the change across the ensemble we have found that the future changes in the near surface temperature are significant almost everywhere but not for precipitation or sea level pressure. This result agrees qualitatively with earlier studies involving large ensembles yet there are regional differences between SMHI-LENS and large ensembles from other models. Deser et al. (2012) show in agreement to our results that the 325 mean temperature change signal is much more robust than P and SLP change signals. For P and SLP, they found similar regions with large and small ratios between mean change and internal variability as this study. Both regions and amplitudes of standard deviation of T2m, SLP and P trends agree relatively well with our results. Compared to results from the MPI-  2018)), but only few studies so far have analysed the probability itself for passing a specific warming level. We show that none of the scenarios used here is able to keep global warming in the 21st century below 1.5 degrees. In SSP1-1.9 there is an approximately 70% probability for the warming to stay below 2 degrees warming while all other SSPs exceed this target during the course of the century. SSP5-8.5 is the only one of the used scenarios to definitely pass even a 3-degree warming. SSP5-3.4-OS has a 20-40% chance to exceed SWL3 345 temporarily during the 2050-2090 period, but at the end of the century the risk of warming beyond this threshold is very small. For comparison, based on the CMIP5 model ensemble, Jiang et al. (2016) show that the probability to exceed the 2 °C global warming level before the year 2100 is 26, 86, and 100% for the Representative Concentration Pathways 2.6 (RCP2.6), 4.5 (RCP4.5), and 8.5 (RCP8.5) scenarios, respectively, with the median years of 2054 for RCP4.5 and 2042 for RCP8.5.
To demonstrate the importance of a sufficiently large ensemble we look at the point in time when the differences between 350 the SSP5-8.5 and SSP5-3.4-OS ensembles become significant, and when the SSP5-3.4-OS and SSP4-3.4 ensembles become These results are just examples for what kind of analyses and risk assessment are possible with a large ensemble. In the future it is planned to extend this kind of work to regional warming signals, the frequency of occurrence of extreme events (e.g. heat waves), and other variables (e.g. precipitation, sea-ice).