We select the proxy data with the following criteria. Firstly, the proxy data should have a strong imprint of climate. In case of the MXD, tree-ring width (TRW) and BI data from tree-ring chronologies, this means high correlation to local temperature, while we look for a clear imprint of atmospheric circulation in the δ18O-based records. Secondly, the proxy data needs to have 0–1-year dating uncertainty and a well-defined seasonality. To achieve consistent performance of the reconstruction, we only select proxy records that span the whole time frame of the reconstruction (CE 1241–1970) (Fig.  and Table ).

The dating uncertainty of anchored tree-ring chronologies covering the past millennium is generally assumed to be zero. Furthermore, the sample replication of both the conventional and δ18O_cell tree-ring records featured in this study is stable throughout the reconstructed interval. The ice core data are dated using layer counting and synchronised using volcanic historically identified reference horizons, and for the past millennium the dating uncertainty is estimated to be maximum 1 year for the GICC05 chronology originally used for the high resolution ice core data . The volcanic dating horizons make it unlikely that the age scale of an entire ice core record shifted in time. However, one cannot exclude that some records are off by up to one year in some intervals between dating horizons, although the validity of the GICC05 time scale for the interval covering our reconstruction was confirmed by comparison to the updated GICC21 chronology .

We use the isotope enabled version of the climate model ECHAM5/MPIOM run for the past 1200 years forced by natural and anthropogenic forcings. This is the same model simulation used in SEA18 and SEA20.

The atmosphere component is run in a T31L19 configuration corresponding to 3.75° × 3.75° horizontal resolution and 19 vertical layers, while the ocean component, MPIOM GR30L40, is set to 3° horizontal resolution with 40 layers. The model features isotope fractionation during all phase changes, including kinetic fractionation during evaporation and in mixed-phase clouds.

The variability of δ18O_cell is governed by several different processes, some of which have a high degree of co-variability. From the following analysis, we assume that the main control of δ18O_cell is the isotopic composition of precipitation forming the groundwater, which comprises the source water for formation of xylem sap used to form cellulose during the growing season. Depending on the hydrological setting at the site of the tree-ring chronology, there can be a lagged isotope signal from precipitation formed prior to the growing season, or from climate driven intra-annual variability. The lag can be due to groundwater recharge during spring snow melt or sap formed during late winter and spring . Our tests indicate that both δ18O_cell records used in this study are sensitive to weather variability (SLP, T2m, precip. amount) during the growing season (Fig. S1 in the Supplement), as well as the weather variability of the winter preceding the growing season (Fig. S2). The records ENG and CZEC are mainly correlated to variability connected to the winter NAO, which can be explained by the JJA Palmer drought severity index (pdsi) at these sites being sensitive to the DJF NAO. This is due to moisture availability of the growing season being affected by the preceding winter precipitation, as well as land-atmosphere feedback mechanisms where weather patterns promoting summer drought are enhanced by low soil moisture of the previous winter season . Based on these findings we use the δ18O_cell records as representing a mixture of summer and winter variability. The impact of large-scale circulation on weather variability and δ18O in precipitation weakens in spring, while the maximum correlation between δ18O_cell (ENG and CZEC, Table ) and climate variables are seen in the growing season . We use JJA for representing the growing season and the extended winter season (November–April) to represent the winter preceding the growing season. Using the extended winter accounts both for impact of the winter circulation patterns, precipitation and composition groundwater for early spring sap formation. In addition, this choice of seasonal dependency the δ18O_cell is linked to the Greenland ice core data through the extended winter δ18O, and to the conventional tree-ring data through the co-variability of JJA temperature and δ18O. While a mechanistic approach to connect the δ18O_cell to δ18O in precipitation is possible through forward modelling e.g., this would also introduce new parameters with uncertainties specific to each site. Since the untreated δ18O_cell in the first place has a clear imprint of atmospheric circulation for summer and winter, we choose to assume that the variability of δ18O_cell primarily is governed by a mix of δ18O in precipitation for both seasons.

We use a variety of data ranging from instrumental data to climate field reconstructions to evaluate our climate reconstruction. For assessing spatial coverage of skill, we use the 20th Century Reanalysis version 3 (20CRv3) , which is only constrained by observations of SLP and sea surface temperature (SST). We furthermore compare to the HadCRUT5 gridded temperature and the GPCC gridded precipitation . As a supplement to evaluating the reconstruction with the 20CRv3 and HadCRUT5 temperature we use long-term observed temperature from Greenland , Iceland , England , Denmark and Sweden , to achieve a longer overlapping time interval for the evaluation. To evaluate reconstructed SST we use three different datasets ERSSTv5 , COBE2 , HadISST that have different properties due to different treatment of observational data and missing values. In addition to purely observation-based datasets, we compare our to the climate field reconstructions the Last Millennium Reanalysis (LMR) in annual (LMR v2.1) and seasonal resolution (LMR Seasonal) , as well as the ModE-RA reanalysis of monthly resolution covering 1421–2008. In addition to proxy data, ModE-RA is assimilated based on instrumental data and historical documents.

We use the analogue method to assimilate proxy data using an isotope enabled climate model. This setup can be described as assimilation with a static ensemble or static model prior. Building on experience of using this method effectively for previous reconstructions (SEA18/20), we expand the assimilated proxy network with new records and update the methodology for calculating the ensemble mean and introduce a variance correction. Our reconstruction is not calibrated against observations, and only relies on matching the model output with proxy data. For each site we evaluate the model output against the proxy data, linking simulated JJA temperature to conventional tree-data (MXD/BI/TRW), and simulated seasonal δ18O with ice core δ18O and δ18O_cell (Table ). We use different evaluation criteria of the model data for the summer and winter season. For the summer season we use a χ2-measure for the goodness of fit: 1χsum2(t)=∑i=1nm(n,t′)-p(n,t)2σmp, where, m is the modelled anomaly and p is the proxy anomaly at a given site, n is the number of proxies, t is the year of the reconstruction, t′ is the year of the model run, and σmp is the combined standard deviation of the model and proxy data. Dividing by σmp normalizes the data, which is effectively the same as normalizing all time series (proxy data and model output for proxy sites) prior to the matching procedure.

The imprint of large-scale circulation on the isotopic composition is very strong for winter and the amplitude of the δ18O anomalies bears a signal in it-self. Part of this signal is the gradient between proxy sites . Since we are interested in preserving as much of the signal in the proxy data as possible, we do not normalize the data when fitting the model data to the proxy data. The measure for goodness of fit for winter is then: 2χwin2(t)=∑i=1nm(n,t′)-p(n,t)2, where, m is the modelled anomaly and p is the proxy anomaly at a given site, n is the number of proxies, t is the year of the reconstruction and t′ is the year of the model run.

The workflow for the annual and seasonal reconstructions is illustrated in Fig. . For the seasonal reconstructions we reconstruct the summer and winter separately using Eqs. () and (), respectively, while for the annual mean and monthly reconstruction, the proxies for summer and winter are matched simultaneously using Eq. (). The monthly data from the best matching model years are then extracted, taking into account that the ice core data for winter are representing months November–April when the annual mean for the calendar mean is calculated (i.e., November–December are assigned to previous calendar year compared to the matched winter proxy data centred in January). As we have no direct constraints on monthly data, we mainly include the monthly reconstruction to be able to investigate the seasonality of SAT25 and compare with other datasets.

In this study we calculated the ensemble mean using a logarithmic weighting function. Due to the bias-variance trade off, a high number of ensemble member reduces noise and increases skill in term of correlation but causes decrease in variability, as well as skewing the variability towards the main mode and causing mixed secondary modes in the SLP field. Before arriving at our current approach, we tested different ways to calculate the ensemble mean: constant number of ensemble members with equal weighting (as SEA18/20), variable number of ensemble members based on match with proxy data, also in combination with a minimum number of ensemble members. Here we calculate the weighted mean based on the χ2 fit (Eqs.  and ), which yields a smoother behaviour of reconstruction with less year-to-year difference in the statistical properties of the reconstruction compared to a variable number of ensemble members, while still allowing the best matching models years to have the most weight. The weighting is calculated as the logarithm of the normalized χ2-distance, where the least important ensemble member receives zero weighting (Fig. S3). The χ2-distance as a function of number of ensemble members is similar to an exponential decay function and we apply the logarithm to achieve a more equal weighting of the ensemble members. Otherwise, the result would be skewed too much towards a few of the best fitting ensemble members, which results in more noise in the reconstruction. With the applied weighting 50 % of the weight goes to the best matching 30 % of the ensemble members.

In our results we show time series as anomalies from the mean of the full-length data series, and monthly data are anomalies with respect to the mean annual cycle. We use Pearson correlation (r) and significance levels are determined using Student's t test. For correlation maps we list the maximum correlation, mean significant correlation and number of grid cells with significant correlation below the figure title. The number of grid cells with significant is a rough indication of the spatial coverage of the skill, which will depend on the spatial auto-correlation, which in turn depends on season, latitude (area of grid cell) and type of climate variable. For the point-wise correlation maps all data from reanalysis, observations and climate field reconstructions are re-gridded to the 3.75° × 3.75° grid used for SAT25.

Before comparing to observed climate we evaluate the coherence between the reconstruction and the assimilated proxy data, and assess how many model analogues per year to use in the ensemble reconstruction. When determining number of ensemble members to include for each year there are several factors to consider. A high number of ensemble members reduces noise by smoothing the variability spatially and temporally, but as mentioned above, the smoothing also makes it more difficult to separate the main modes of climate variability. For example, with a high number of ensemble members, the East Atlantic and Scandinavian blocking patterns cannot be separated and more of the variability is assigned to the NAO. Conversely, a low number of ensemble members fail to span both the imperfect fit of the model to the proxy data, as well as the confounding noise of the proxy data.

We first consider the match between the reconstructed signal at the sites of the proxy data. There is high correlation between the reconstruction and all proxy records, with a match to ice core δ18O and δ18O_cell in the range r ∼ 0.65–0.9, and on average slightly better correlation between reconstructed temperature and conventional tree-ring records, though in similar range (r ∼ 0.6–0.9). We tested the match with proxy data with increasing number of ensemble members. The maximum correlation to proxy data is achieved with 10–100 ensemble members per year, however the match to the proxy data only degrades slowly with up to 300 ensemble members per year due to the weighting with respect to χ2-distance to the proxy data (Fig. S4). Using 300 ensemble members introduces spatial smoothing which filters out noise in particularly for areas far from the proxy sites (see Sect. ). With this high number of ensemble members, we on one hand lose information on modes of atmospheric circulation, as mentioned in the introduction, but on the other hand gain skill for basin-wide variability for the North Atlantic region. We therefore provide two versions of the reconstruction, one using 150 ensemble members and one using 300 ensemble members. For simplicity we focus on the version with 300 ensemble members as the differences are subtle and will be explored in future work.

Previous seasonal reconstructions (SEA18/20) only used ice core data, or introduced tree-ring data after a pre-selection of model analogues had been made using ice core data. Here we fully incorporate the tree-ring data. The match with ice core records can be achieved regardless if the tree-ring data is included or not, indicating that the analogue sampling pool is sufficiently large to provide relevant analogues, and that further constraining the climate variability with records in Europe does not hinder a good match over Greenland.

To further test the role of the tree-ring and ice core data, we made two new reconstructions based on our annual reconstruction. One based only on ice core data (RECON_IC) and once based only on tree-ring data (RECON_TR). We then extracted the data for tree-ring sites from RECON_IC and the ice core sites from RECON_TR. The results show that the prediction of tree-ring variability from RECON_IC has weak skill, although some correlations are significant (see Figs. S5–S8). The best correspondence is seen for Scandinavian tree-ring records (TAA, EFmean) and winter δ18O correlated to CZEC δ18O_cell. Conversely, the prediction of ice core variability from RECON_TR also has weak skill (Figs. S9 and S10), with the best correlation to GRIP δ18O (r = 0.33) (Fig. S9). This underlines the different properties of the proxy data due to differences in seasonality, region and what is recorded by the proxies. In broad terms, the main strength of the ice core data is to record large-scale atmospheric circulation for winter, while the tree-ring data capture temperature changes on a more local scale, but the wider availability of tree-ring data ensures a reasonable spatial coverage to capture a large-scale summer signal. Our main conclusion from this test is that the properties of the different sets of proxy data are complementary and constrain different seasons and regions of the reconstruction.

The uneven distribution of proxy data concentrates skill and variance near the sites of the proxy data, and calculating the ensemble mean dampens the overall variance. We bias correct the variance of the reconstruction by rescaling the standard deviation of each grid point with the corresponding standard deviation of the ECHAM5-wiso/MPIOM run. Our motivation is that the climate model acting as the model prior is the best estimate of the variance in locations not constrained by proxy data. This very simple correction improves the match between reconstructed variance compared to the variance for SLP, T2m and precipitation from 20CRv3 (Figs. S11 and S12).

To test the long-term stability of the reconstruction we investigated temporal variations of the match with proxy data. The average fit of model analogues is fairly constant over time, indicating that the quality of the climate reconstruction is consistent throughout 1241–1970. One reason for this consistency is the strict selection of proxy records that span the whole reconstruction, which means that there are no methodological differences throughout the reconstruction. We found no clear preference with respect to selection of model analogues throughout the reconstruction. This is consistent with results for earlier versions of the reconstruction (SEA18/20).

SEA25 is not calibrated to observations (see Sect. ). The comparison to observed climate in this section is therefore an independent measure of skill. In addition, since we use a constant number of proxy data and no change in methodology throughout the reconstruction, we should, in theory, reach the same level of skill for any time period of the reconstruction.

Compared to the JJA and DJF mean for 20CRv3, the reconstruction shows coherent patterns of skill for SLP and T2m, while the skill for precipitation is patchier (Fig. ). Standout features are widespread significant correlation for the winter SLP and strong correlation over the eastern North Atlantic and Northern Europe for summer T2m (max. r = 0.71). In winter the level of correlation is tightly connected with the skill for reconstructed circulation patterns. This is especially clear for the winter SLP and T2m which resemble the spatial pattern of NAO-type variability, while summer SLP and T2m which resemble the spatial pattern of Scandinavian blocking-type variability. The weaker skill for summer circulation is due to a combination of less well constrained large-scale circulation during summer and biases in the JJA SLP patterns of the ECHAM5/MPIOM model . An alternative to Fig.  replacing the 20CRv3 data with HadCRUT5 for temperature and GPCC for precipitation can be found in the Supplement (Fig. S13). These results are comparable to the comparison with 20CRv3, although the correlation to HadCRUT5 is somewhat better.

The correlation patterns for the annual mean are, not surprisingly, a combination of features seen for summer and winter (Fig. a–c). While it is less pronounced than for DJF, the signature of NAO-type variability is still evident in the spatial patterns of the correlation. The maximum correlation (r = 0.68) is achieved for T2m with lower correlations for SLP and precipitation than for winter, indicating that auto-correlation in temperature aids higher skill in some areas, while the seasonal information is lost in the annual mean SLP and precipitation due to shifting circulation patterns and different impact of circulation on T2m and precipitation on a seasonal scale.

On monthly time scales the skill of the reconstruction is lower with at maximum correlation of 0.34 for T2m (average for all months) (Fig. d–f). Despite lower skill, the correlation is still significant over large areas, due to the larger sampling pool (12 times as more sample points compared to annual data). The spatial patterns of the correlation are similar to that of the annual mean reconstruction. Reviewing the skill for individual months, correlations peak in February and August for T2m with maximum correlation with 20CRv3 of 0.55 and 0.70, respectively (Fig. S14). Again, the comparison the HadCRUT5 monthly temperature shows somewhat higher correlation than to 20CRv3 (Fig. S15). For SLP the imprint of large-scale circulation is again evident during winter months (max. corr. = 0.38 for February) with lower and more localized skill during summer (not shown). A summary of the SAT25 correlation to monthly 20CRv3 data can be found in Fig. S16. The skill for the monthly reconstruction varies regionally due to the seasonality of the proxy data and seasonal shifts in climate patterns, leading to summer months and winter months not having high skill at the same location. For example, this is part of the reason for the annual average monthly temperature correlations to remain low, although they peak above 0.6 for temperature during July and August (Fig. S16).

In addition to the comparison with 20CRv3 we compare to observed temperature from Greenland, Iceland, Central England, Denmark, and Sweden which are among the longest observed high-quality temperature records and span our study area well. For summer the high skill of the reconstruction seen in Fig. b is confirmed for sites in Sweden and England (Fig. ). It appears that there are periods of over and underestimated reconstructed temperature which could be a sign of underestimated multi-decadal variability, or it could be due to a different long-term trend in the reconstruction. However, for Uppsala, which is close by Stockholm and has a longer observational record, the reconstruction shows the highest correlation to observed summer temperature (r = 0.56). It is also reaffirming that the reconstruction shows good correlation to the longest existing instrumental record, the Central England temperature (r = 0.5). The lower skill for Greenland sites during summer is similar to earlier studies, because Greenland ice core summer δ18O is more sensitive to climate variability east of Greenland, also illustrated by good correlation to Iceland summer temperature (r = 0.54) . Due to a bias of the model prior, the variance in Southern Greenland is underestimated for summer, which can be seen in the reconstructed time series from Qaqortoq (Fig. ) and in the map of the standard deviation of reconstructed temperature (Fig. S11).

The highest skill during winter is seen for Greenland winter temperatures (r ∼ 0.6), although we also see good performance for Stockholm (r = 0.46), despite having no proxy site nearby (Fig. ). Compared to summer there are only the two δ18O_cell proxy sites available for the winter season besides the Greenland ice cores. However, this is partly counteracted by the stronger impact of large-scale circulation on regional climate variability enabling good skill also in the Eastern part of the study area. We observe that the variance for DJF is approximately double of the JJA variance. This is captured well for most sites by the reconstruction, also showing that the variance correction is effective also compared to instrumental data.

For annual mean data the correlation is more consistently in the range of 0.5–0.6 (Table ), again, combining the characteristics of the summer and winter season, as well as capturing amplitude of the temperature variance (Fig. S17).

The monthly reconstruction has correlations of ∼ 0.3 at all sites but for Central England which has lower correlation (r = 0.21). The lower skill for Central England is likely due to the winter variability being less coupled to the NAO. As the reconstruction is not constrained at monthly resolution, we cannot expect high skill on a month-to-month basis. What we do see is the reconstruction capturing part of the modulation of the annual cycle seen in the observations. For example, for the cold winters during the 16th and 17th century in Scandinavia (Fig. S18).

The annual reconstruction, which targets the calendar year, enables direct comparison to more common annual reconstructions and is useful for analysing indices of sea surface temperature which are often defined on annual mean data. Our reconstruction uses a high number of ensemble members (300, ∼ 25 % of the available model analogues) introducing spatial smoothing to filter out noise in particular for areas far from the proxy sites. Using this high number of ensemble members, we on one hand lose information on modes of atmospheric circulation, as mentioned in the introduction, but on the other hand gain skill for basin-wide variability for the North Atlantic region (Fig. ).

In this section we compare to climate field reconstructions of annual, seasonal and monthly resolution. The Last Millenium Reanalysis (LMR) SLP, temperature and precipitation amount is available in annual resolution , as well as seasonal resolution for temperature . Both the annual and seasonal resolution LMR span the entire time period of our reconstruction. The ModE-RA reanalysis is one of the few monthly resolution paleo-reanalysis datasets and covers 1421–2008 .

The annual mean SLP, temperature and precipitation of our reconstruction correlates well with the annual mean LMR data if we compare the full time period of the reconstruction (CE 1241–1970) (Fig. a, b, c), showing similar or better correlation than to the 20CRv3. For seasonal temperature there is a strong coherence between SAT25 and LMR for JJA temperature with maximum correlation of 0.9 over Northern Europe (Fig. d), while there is no correlation over Greenland. The correlation during winter reaches 0.59 (Fig. e), which is just below the DJF correlation of SAT25 to 20CRv3.

For the comparison to seasonal means of the ModE-RA reanalysis of the full overlapping time period (1421–1970) we see similar correlations as to the 20CRv3 for JJA SLP and temperature, while the coherence to JJA precipitation is weaker (Fig. a, b, c). The correlations to the ModE-RA DJF mean are lower than for the comparison to the correlations to DJF 20CRv3 (Fig. d, e, f). Comparing our reconstruction and ModE-RA temperature in monthly resolution for the full overlapping time period 1421–1970, we see comparable results as for correlation to 20CRv3 for the summer months, showing large areas of coherent correlation over the North Atlantic (Fig. ). In contrast, the winter months show lower correlation than to 20CRv3 between the two reconstructions.

Next, we investigate the long-term variations of co-variability between SAT25 and the LMR and ModE-RA climate field reconstructions by looking at the correlation in 121-year time windows, providing similar properties as the comparison with the overlapping time period with the 20CRv3 (1850–1970). The results are summarized in Fig.  by plotting the maximum correlation and number grid points with significant correlation for annual and seasonal means for six time windows covering 1250–1970 for the LMR data, and four time windows covering 1490–1970 for the ModE-RA reanalysis (see Figs. S19 to S24 for all correlation maps). The maximum correlation shows larger excursions than the mean significant correlation, and is used in Fig.  as an effective indicator of strength of the co-variability between the different datasets. For reference, we also plot the JJA and DJF maximum correlation and number grid points with significant correlation from the 20CRv3 comparison. The correlations to LMR remain rather stable over time, although generally highest in the most recent time window (1850–1970) and lowest in the earliest part of the reconstruction (1250–1370) (Fig. a). While more variable, this pattern is also reflected in the number of grid points with significant correlation, although the relation between SAT25 and LMR annual SLP is relatively stable (Fig. b). Apart from the moderate long-term trend of decreasing coherence between SAT25 and LMR going back in time, the intervals covering 1490–1730 display somewhat variable correlation between the reconstructions. For example, the correlation to LMR DJF temperature is lowest for this time frame, coinciding with a shift in the correlation to LMR annual SLP, while the correlation to LMR JJA and annual temperature remains high.

The results for the stability of the correlation to the ModE-RA reanalysis contrast the results of the comparison to the LMR, as they show a clear decrease in coherence going back in time (Fig. c and d). This is most prominently seen as an almost complete loss of coherence between SAT25 and the ModE-RA DJF variability, going from correlations similar to the 20CRv3 in the most recent interval (1850–1970), to very little correspondence in variability for the earliest interval (1490–1610).