Ocean and atmosphere influence on the 2015 European heatwave

During the summer of 2015, central Europe experienced a major heatwave that was preceded by anomalously cold sea surface temperatures (SSTs) in the northern North Atlantic. Recent observation-based studies found a correlation between North Atlantic SST in spring and European summer temperatures, suggesting potential for predictability. Here we show, by using a high-resolution climate model, that ocean temperature anomalies, in combination with matching atmospheric and sea-ice initial conditions were key to the development of the 2015 European heatwave. In a series of 30-member ensemble simulations we test different combinations of ocean temperature and salinity initial states versus non-initialised climatology, mediated in both ensembles by different atmospheric/sea-ice initial conditions, using a non-standard initialisation method without data-assimilation. With the best combination of the initial ocean, and matching atmosphere/sea-ice initial conditions, the ensemble mean temperature response over central Europe in this set-up equals 60% of the observed anomaly, with 6 out of 30 ensemble-members showing similar, or even larger surface air temperature anomalies than observed.


Introduction
The importance of skilful seasonal forecasts is highlighted by the devastating socio-economic impacts of extreme summer conditions over Europe (Ciais et al 2005, Zampieri et al 2017. The heatwave of 2003 resulted in more than 70 000 deaths (Robine et al 2008) and similar numbers apply to the Russian heatwave of 2010 (Grumm 2011). In the summer of 2015, Europe experienced a heatwave ranking third warmest on record, surpassed only by the summers of 2003 and 2010 (Russo et al 2015). The heatwave began in late June in Western Europe and then spread towards Southern and Eastern Central Europe, with several cities reporting record high temperatures (Sippel et al 2016). The exceptionally warm conditions led to extreme drought over most of central Europe (Orth et al 2016) with several heat-related deaths (Muthers et al 2017, Urban et al 2017, Výberči et al 2018. With future climate projections estimating mean summer temperatures over Europe to increase by 0. 6°-1.5°C in 20166°-1.5°C in -20356°-1.5°C in (Kirtman et al 2013, heatwaves will become more common (Meehl and Tebaldi 2004, Schär et al 2004, Ballester et al 2010, Lhotka et al 2018.
To date winter forecasts yield greater seasonal prediction skill than summer forecasts (Scaife et al 2014, Stockdale et al 2015, O'Reilly et al 2017. However, several recent studies have noted a connection between the winter/spring sea surface temperatures (SSTs) in the Atlantic Ocean and conditions over Europe (Feudale and Shukla 2011, Gastineau and Frankignoul 2015, Ossó et al 2018. In the months leading up to the exceptionally warm summer of 2015 an anomalous 'cold blob' was present in the northern part of the North Atlantic (Josey et al 2018). It was argued that the resulting anomalous gradient in SST phase-locked a meander in the jet stream, leading to the warmer temperatures over Europe (Buchan et al 2014, Duchez et al 2016. Here, we investigate whether the 2015 European heatwave can be re-forecasted using an alternative initialisation method that does not make use of dataassimilation. The rationale behind this choice is that the ocean is an almost purely adiabatic system below the surface mixed-layer. As a consequence, the slowly varying ocean circulation contains both surface and subsurface temperature and salinity anomalies, with the potential for subsurface anomalies to be mixed up to the surface where they interact with the atmosphere and may contribute to skilful predictions (e.g. Grist et al 2019). Hence our choice to test a gentler initialisation method where neither ocean nor atmospheric initial conditions are nudged towards observations by artificial sources and sinks. In this study we focus on the link between ocean and atmosphere initial conditions and how it may affect the summer re-forecast. A set of ensemble experiments using the coupled climate model HadGEM3-GC2  are performed using various ocean and atmosphere initial conditions. In contrast to previous studies the ocean is initialised using an anomaly initialisation technique whereby anomalies from a forced ocean-only simulation are introduced to the coupled model restart. In addition, this initialisation method allows for separating the influence of ocean, sea-ice, land-surface (soil moisture) and atmosphere initialisation. Studies have shown that land surface properties such as soil moisture also influence heatwaves (e.g. Quesada et al 2012, Ardilouze et al 2017), and may therefore provide an important source of predictability for heatwaves. However, since our experiments require a coupled high-resolution prediction model, we decided to focus all our available computational resources on ocean, sea-ice and atmosphere initialisation only.

Methods and data
2.1. Model setup To investigate the potential link between patterns of anomalous North Atlantic SST and European summer heatwaves we conduct a set of experiments using the coupled climate model HadGEM3-GC2 , consisting of atmosphere, ocean, sea-ice and land-surface models. The ocean configuration is Global Ocean 5.0 (Megann et al 2014), which is based on NEMO v3.4 (Madec 2015) and uses the ORCA025 (nominally ¼°) tripolar grid configuration. The atmosphere model is the Met Office unified model with the Global Atmosphere v6.0 (Walters et al 2017) that has a horizontal resolution of N216 (approximately 60 km) and 85 vertical levels. This model setup is used in the seasonal forecasting system, DEPRESYS at the Met Office but with the addition of data assimilation (Dunstone et al 2016).

Initialisation approach
We use a new method to initialise our experiments, taking anomaly fields from a forced ocean-only simulation (Garry et al 2019) with the same ocean model as used in the coupled model. Thereafter, these anomalies are added to the climatological mean of the ocean in the coupled model to generate an initial state. By introducing the anomalies in this way, we minimise initial shocks and model drift and we avoid the need for drift correction, data assimilation or nudging towards a particular field. This approach is in contrast to previous studies where the ocean is either initialised using the full fields from a forced ocean-only experiment (Matei et al 2012) or anomalies based on observations (Magnusson et al 2013, Smith et al 2013. We generate 5 different ocean initial conditions: a control ocean initial condition where climatological three-dimensional (3D) ocean temperature and salinity averaged over April/May from 1981 to 2010 are taken from a historical+RCP4.5 simulation of the coupled model (CLIM) and four experiments where 3D temperature and salinity anomalies on 1st May from a forced ocean-only simulation (driven by an atmospheric reanalysis product; Garry et al 2019) are added to the CLIM ocean initial condition. The temperature anomalies from the ocean initial conditions are computed as the difference between the forced ocean-only simulation restart file and April/May means averaged from 1981 to 2010. The ocean anomaly fields are taken from four years: 2015, 2014, 1993 and 1984, (figures 1(b) and 2). For each of these initial states we run a 30-member initial condition perturbation ensemble, adding spatially varying white noise. The amplitude of the white noise was chosen to be equal to the standard deviation of differences in SST in consecutive 5 days means from the NEMO hindcast, which is a typical timescale of mesoscale variability in the ocean. The noise was linearly tapered to zero over the upper 50 m of the ocean and then added to the ocean initial state, to prevent the anomalies from disappearing quickly. The perturbed SSTs immediately transfer the noise to the atmosphere leading to sufficient spread in the ensemble. Focusing on the ocean initial state we decided it was more appropriate to add a realistic noise pattern to the ocean initial fields. Any convective instabilities arising from the initialisation were removed from the 3D temperature and salinity initial condition fields using the same method as in the NEMO ocean model code. Each ensemble member was then computed for 5 months, starting from 1st May.
The atmospheric initial conditions (which in our setup include the states of atmosphere, sea ice and land) are taken on 1st May in the years of 2015, 1992 and 1986 in the coupled HadGEM3 historical+RCP4.5 simulation. Note that, apart from the external forcing, atmospheric states for each year in HadGEM3 bear no resemblance to the climate of the real world in the corresponding year. Similarly, the ocean states in these HadGEM3 runs are unrelated to the ocean state in the ocean-only forced run of that same year. Therefore, in the remainder of this paper we will refer to the three atmospheric initial conditions as A (2015), B (1992) and C (1986).
We only vary the atmospheric and sea ice states in the initial conditions A, B, and C. Soil moisture also has an influence on heatwaves (Quesada et al 2012, Ardilouze et al 2017. However, here we initialise soil moisture from its climatological mean  in all experiments as this facilitates isolating the effect of ocean and atmosphere/sea-ice on the re-forecasts. This also allows us to keep computational cost within the limits of available HPC resources. Even without varying the initial state for soil moisture exploring the impact different combinations ocean and atmosphere initial conditions described above we have generated a total of 450 ensemble members across 15 different experiments.
To analyse the influence of the ocean anomalies on simulated temperatures over Europe, the ensemblemean atmospheric response in the CLIM experiment is subtracted from the ensemble-mean atmospheric response of the anomalously initialised runs for each ocean initial state and for each atmospheric initial condition separately, showing data averaged over JJA. We emphasise that in each case, the same atmospheric initial condition is used in both the initialised ensemble and the CLIM ensemble and that only the ocean state differs between the two ensembles. It is also important to note that the external forcing used in these experiments is based on 1978 for all ensemble members, therefore the only impact seen from external forcing is the fingerprint they have left on the ocean initial conditions. The main focus is on the impact of ocean initial conditions, mediated by a small set of different atmosphere and sea-ice states, on the exceptionally warm summer of 2015.

Significance tests
The Kolmogorov-Smirnov test is used throughout the manuscript when testing the difference between two means. A significance level of 5% was used throughout this study. To test whether a correlation value is significant a Student's t-test is applied.

Ocean state and the 2015 heatwave
To assess the impact of ocean initial conditions on the June, July and August (JJA) temperature anomalies over Europe we run seasonal re-forecasts for JJA, starting from four different ocean initial conditions , 2014, 1993, 1984see Methods). These ocean initial states preceded the second warmest summer since 1980 (2015), an average year (2014), and two colder than average years (1984 and 1993), with 1984 being the coldest since 1980. (figure 1(b)). The summer of 2015 also had the least precipitation since 1980; 2014 slightly wetter than average; 1984 slightly drier than average; and 1993 approximately average (figure 1(c)). Three of those 4 initial ocean states featured a warm eastern tropical Pacific with El Niñolike conditions (1993,2014,2015) and one initial state (1983) with La Nina-like conditions (figure 2). To assess how the atmospheric initial conditions affect the impact of ocean initialisation we use the three arbitrarily chosen atmospheric initial conditions from the model's historical+RCP4.5 simulation (A, B, C), which also include sea ice and climatological soil moisture values for land. The same atmospheric conditions are applied in both the experiment and the control ensemble, CLIM, so their effect is only indirect by the way they influence the evolution of the ocean anomalies and how these anomalies feed-back to the atmosphere.
Our series of re-forecast ensembles show that both the initial ocean and atmosphere conditions in May influence the JJA temperature anomalies (figure 3).
When considering the entire Atlantic-European-African (A-E-A) domain (figure 1(a)) especially over the North Atlantic and Africa the JJA SAT anomaly patterns remain similar when atmospheric initial conditions are varied. Concentrating on the 2015 ocean conditions (figure 3, top row) the tripole pattern in the North Atlantic with warm anomalies between 15°N and 45°N, flanked by cold anomalies to the north and south, is present for all atmospheric initial conditions. Similarly, the warm anomalies observed over much of Africa and the Middle-East are also simulated with all three atmospheres. This shows that in our experiments the ocean initial state in May imposes a strong constraint on the large-scale JJA SAT anomaly patterns, and on average this ocean state leads to pronounced JJA anomalies (figure 3, fourth column), which are significantly correlated with the JJA temperature anomaly patterns observed in the real world in 2015, 2014, 1993 and 1984 (figure 3, fifth column; figure 4(b)). It is also evident that there is a sensitivity to atmospheric conditions. Over the European region (defined in figure 1(a)) significant differences between the different ensembles arise. This is most evident in the ensemble using the 2015 ocean and atmospheric state C, where cold JJA SAT develop over much of Europe ( figure 3, 3rd column, top), in contrast to the ensembles using atmospheres A and B. However, averaging over JJA anomalies obtained for the oceanic conditions 2015, 2014, 1993, and 1984 (figure 3, bottom row) gives rise to weaker mean anomalies than averaging over atmospheric initial conditions A, B, and C ( figure 3, 4th column), over the larger A-E-A. This is not always the case over the central European box where the average over initial condition A gives a stronger JJA temperature signal than the average over the oceanic conditions 2015. This suggests that the atmospheric initial condition plays a role in obtaining the correct sign of the anomalies over central Europe. However, our results also show that to capture the magnitude of the heatwave getting the correct ocean initial conditions is essential (figures 3 and 4(a)). In general, it is difficult to see how the impact of atmospheric initial conditions could have a systematic effect on the re-forecasts, as the control ensemble always uses the same atmospheric initial states, unless the four ocean initial states have a signal in common. As can be seen in figure 2 and figure S1 (available online at stacks.iop.org/ERL/14/114035/mmedia) this common signal projects on an El Niño signal (see discussion).
Despite the sensitivity to atmospheric initial conditions, the JJA average SAT across all ensemble members and atmosphere initial conditions A and B shows that the predicted temperatures over Europe rank in the same order (2015,2014,1993,1984) as the observations, although with a weaker amplitude, and differences between ocean initial conditions with the majority of the SAT anomalies are not statistically significantly different from 0 ( figure 4(a)). Over the A-E-A region, the spatial patterns in re-forecasts and observations show several similarities (figure 4(b) squares), with all but one ensemble mean having a positive pattern correlation with the observations (figures 3 and 4(b)). When averaged over all ocean initial conditions,  35°E, 45-53°N). The first column is ERA5 data referenced to the period 1981-2010, the next three columns show each atmospheric initial condition separately, the fifth column is the average of all three atmospheric initial conditions, and the last column shows an average over all ocean initial conditions for each atmospheric initial condition. For the model simulations the dot shows the median value, shading indicates the range between the first and third quartiles and the thin line covers range of values. A black circle was added if the ensemble mean is statistically significantly different from 0; (b) Pattern correlation between ERA5 anomaly and model experiments over the Atlantic-European-African region (80°W-60°E, 0°−75°N, squares) and the larger European region (10°W-40°E, 30°N-70°N, triangles). Each column separated by a line represents a different atmospheric initial condition (same as columns 2-5 in (a)) and the colours indicate the year for the ERA5 anomaly and ocean initial condition used in the computation. The last column shows the correlation with the 2015 ERA5 JJA SAT anomaly and mean over all ocean initial conditions. predictions with atmospheric initial state A give anomalously warm temperatures over central Europe, but the amplitude is about half of what we find when atmospheric initial state A and ocean initial state for 2015 are used together (figure 3, 1st column top and bottom panel, figure 4(a)).

Successful re-forecast of summer 2015
For all atmospheric initial states our summer 2015 reforecasts show anomaly patterns over the A-E-A region that are positively correlated with the observations (figures 3 and 4(b)), with ensemble mean pattern correlations between prediction and observation ranging from 0.32 to 0.61 ( figure 4(b), squares). However, only the prediction using atmospheric initial state A was able to simulate the large spatial-scale warm anomaly over Europe (figure 3 top-left), with statistically significant warm JJA temperature anomalies over much of Europe. The ensemble-mean temperature anomaly in the central European box is 1.04°C, which is more than half of what was observed (1.79°C) and statistically significantly different from 0 (figure 4(a)), with 6 of the 30 ensemble members having temperature anomalies equal to or warmer than the observations. The ensemble-mean pattern of the 2015 summer temperature anomalies using atmosphere initial condition A closely matches the observed pattern of anomalous temperatures in JJA over the European region (pattern correlation of 0.68; figure 4(b), red triangle). Europe also experienced very dry conditions with record low precipitation falling over the central Europe box region in 2015 ( figure 1(c)). The combination of the 2015 ocean initial condition and atmosphere initial condition A was also able to capture this exceptionally low precipitation over Europe with a statistically significant signal ( figure S3).
The pattern correlations over the A-E-A region and the European region using the 2015 ocean initial condition and Atmospheric initial condition A (red square and triangle in the atmosphere initial condition A column in figure 4(b)) are similar to the pattern correlation computed from the average of all ocean initial conditions using atmosphere initial condition A and the 2015 ERA5 anomaly (pink square and triangle in the all ocean column in figure 4(b)). However, when considering pattern regressions instead we find that the amplitudes, and therefore regression values, are much weaker in the 'all ocean' column ( figure S4). The consistent large-scale (A-E-A) temperature signal found when ocean conditions are kept the same whilst varying the atmospheric initial conditions suggest that ocean initial conditions are key to the successful seasonal re-forecast. At the same time our results also show that combining with the 'correct' atmospheric initial conditions matter. This implies that, when initialising a coupled seasonal forecasting system with ocean anomalies obtained from a forced ocean run, the anomalously warm SAT over central Europe in 2015 can be successfully modelled, but only when the 'correct' atmospheric initial state is used.

Matching ocean and atmospheric initial states
To investigate possible reasons why atmospheric initial condition A produced the best fit to the ocean initial condition of 1st May 2015, we calculated the pattern correlation between SAT and SST between 60°S and 60°N (avoiding regions covered by sea-ice) in the initial conditions. Using this metric three combinations stand out, 2015 ocean with atmosphere Figure 5. (a) Spatial pattern correlation between 1st May SAT and SST anomalies with respect to 1981-2010 mean computed globally between 60°S and 60°N. The grey dots show the pattern correlation for SAT for each year with the SST for the year specified on the xaxis and the black triangle is when the SST and SAT are taken from the same year using ERA5 data. The pink (Atmos. A), purple (Atmos. B) and brown (Atmos. C) dots show the pattern correlation between the SAT anomaly in the atmospheric initial condition and SST anomaly in the ocean initial condition used in the model simulations, with the x-axis specifying the year of the ocean initial condition. (b) The relationship between the 60°S-60°N pattern correlation of the initial conditions (x-axis) and the JJA mean SAT anomalies over the larger European Region (y-axis). The atmospheric initial conditions are indicated in pink, purple and brown on the left for Atmos. A, B and C and the ocean initial conditions are indicated in red, orange. green and blue on the right for 2015, 2014, 1993, and 1984, respectively. A, 1993 ocean with atmosphere B and 2014 ocean with atmosphere A, ( figure 5(a)). Two of the 3 atmosphere/ ocean initial condition combinations with the highest pattern correlation also exhibit the highest pattern correlation between predicted JJA SAT anomalies and observed JJA SAT anomalies in the A-E-A region ( figure 4(b), squares) and in the smaller European box (figures 4(b), 5(b) triangles); only the 2014 ocean with atmosphere A has a weak relationship to the observed JJA SAT anomalies ( figure 4(b)). There is a correlation of 0.48, between the initial condition global pattern correlation (coloured dots from figure 5(a)) and the JJA ensemble mean pattern correlation with ERA5 in the European region (triangles in figure 4(b)) ( figure 5(b)). This suggests that there is a relationship between the similarity of ocean and atmosphere initial conditions and the resulting SAT pattern (similar relationships are also seen with A-E-A region pattern correlation and regression (figure S5)). When performing the same computation with observationbased ERA5 data for 1st May and taking the SAT and SST anomalies from the same year the pattern correlation is always much higher (0.38-0.63) than correlating the SSTs of one year with the SAT patterns from another year (−0.26 to 0.28, figure 5(a)). This further motivates using the pattern correlation to determine the best atmosphere/ocean initial condition combination.
To increase confidence in this metric, we repeated the pattern correlation between all 1st May SAT snapshots (daily means) from the historial+RCP4.5 run with the SST in our 2015 ocean initial condition. Taking a window of 41 years between 1995 and 2035 (to minimise the effect of global warming on the initial states) we found that the 2015 1st May snapshot (Atmos. A) would rank third in this timeseries.

Discussion
There is increasing evidence that the ocean state can be a precursor for the shape of a season to come ( (Baker et al 2018). The reasons for this are not yet understood and are topic of ongoing research, but the way forecasting systems are initialised may be a factor. After initialisation by data assimilation, coupled models tend to adjust; they 'repel' the initialised state and drift towards their own attractor. This is particularly troublesome for the subsurface ocean, which is largely adiabatic and data-assimilation introduces artificial sources and sinks. Therefore, the ocean temperature and salinity anomalies that contain a source of predictability may be present in the initial conditions as a consequence of the addition of these artificial sources and sinks of heat (and salt). However, such anomalies may be damped out by adjustment after initialisation shock and the bias/drift correction that is applied in prediction systems that use data-assimilation or full field initialisation. For these reasons we choose to test a gentler initialisation method where neither ocean nor atmospheric initial conditions have been nudged towards observations by artificial sources and sinks. Anomaly fields are taken from an ocean simulation forced by an atmospheric reanalysis, ensuring as much as possible that they are dynamically consistent with the free-running coupled ocean (i.e. structural biases due to the model configuration will be consistent). The conjecture is that this approach reduces model drift and initialisation shocks therefore extending the 'memory' of the ocean in (re-) forecasts.
To keep the number of ensemble simulations within the allocated HPC resources we focused on the impact of the initial states of ocean and atmosphere/ sea ice on re-forecasts. It is well known that soil moisture can have a large influence on temperatures over Europe (e.g. Queseda et al 2012, Ardilouze et al 2017).
In our experiments we eliminated the impact of soil moisture by holding it constant. Our results do not conclude that soil moisture did not play a role in the 2015 heatwave. The fact that the summer of 2015 had the lowest precipitation on record (Orth et al 2016, figure 1(c)) strongly points to soil moisture aiding the development of the anomalously warm summer of 2015, therefore making soil moisture an ideal candidate for future studies. Furthermore, the sea ice initial conditions used in atmosphere A, B, and C do show a fingerprint of the anthropogenic sea ice decline, with the atmosphere initial condition A having the thinnest sea-ice and bearing closest resemblance to the observed 2015 spring (figure S7). Studies have shown that sea ice can have an impact on temperatures in Europe, especially in winter (e.g. Petoukhov and Semenov 2011) and precipitation in summer (Screen 2013). Preliminary tests indicate that the sea ice initial state is important for the successful re-forecast of summer 2015 (not shown).
Our successful re-forecast of the 2015 heatwave depends on the choice of the initial atmosphere/sea ice condition. The choice of the atmospheric initial conditions is arbitrary and the strong JJA temperature signal signal over central Europe simulated when using the atmosphere state A fortuitous. However, the temperature signal in our re-forecast is clearly statistically significant (30-member ensemble) and larger than the signal obtained in current operational seasonal forecasting systems. Whereas the physical mechanisms that lead to this good result are not yet fully clear our preliminary assessment is that the initial atmosphere should act to preserve the initial ocean anomalies as well as possible, allowing both for the correct coupled ocean-atmosphere evolution of the initial ocean anomalies and for the correct teleconnections with remote SST anomalies.
To better understand the role of local and remote teleconnections in the successful re-forecast we performed two additional ensembles with 2015 ocean initial condition and atmosphere initial condition A. For these experiments temperature and salinity anomalies are prescribed in the North Atlantic and set to zero elsewhere and vice versa ( figure S8). The results suggest a role for both regional (North Atlantic) and remote influence on the development of the warm 2015 JJA anomaly over central Europe ( figure S8). In both experiments we find a warm JJA anomaly over central Europe with about half the amplitude of the signal found with atmosphere A and the 2015 ocean. A candidate for remote teleconnections is ENSO.
The prevalent view on the relation between ENSO and European summer climate is that it is weak. However, recently this has been challenged. Rodríguez-Fonseca et al (2016) argue that such a relation does exist, but it appears to be non-stationary, modulated by the state of the North Atlantic (AMV) and the characteristics of the ENSO signal (pattern and amplitude). Our results suggest that such a link did exist in spring/ summer 2015. Averaged over all ocean initial states, atmosphere initial condition A is still associated with a weaker European heatwave (figure 3). The average over all 4 ocean initial condition SSTs does consist of a positive ENSO signal (figure S1) and atmosphere initial condition A itself consisted of the strongest SAT anomalies over the ENSO-region of all atmospheric initial conditions (figure S2).
We found that a pattern correlation between SST and SAT anomalies in the initial state is a tentative metric for the selection of the atmospheric initial condition. We cannot yet prove that a high pattern correlation to select atmospheric initial conditions systematically improves re-forecasts. Also, it is not yet clear if such an atmospheric state can always be found. Nevertheless, the successful re-forecast using atmosphere initial condition A implies that it is possible to predict the European heatwave of 2015. Preliminary tests using the atmosphere of the UK Met Office's forecasts system did not lead to a better re-forecast than using atmosphere initial condition A (figure S9). Using the initial condition from the forecast still produces a warm anomaly over much of central Europe, however the amplitude of the JJA temperature is only about half of when atmosphere initial condition A is used. This suggests that, when applying the ocean initialisation method studied here, using a model atmosphere analogue may lead to more successful reforecasts than imposing the observed one which is not part of the model's attractor. Although further testing is needed, our results suggest that our approach outlines a promising route for making seasonal predictions. However, to substantiate these results, further investigation is required before this method can be used in operational prediction systems.