Causes of climate change over the historical record

The 19th century began as one of the coldest periods of the last millennium, at least for the Northern Hemisphere (see Masson-Delmotte et al 2013), following a slightly warmer 18th century (figure 1). Some of the coldest observed periods followed in the two years after the powerful eruption of Mount Tambora in 1815 (Raible et al 2016). After that, temperatures began to show a slow rise, interrupted by cooling induced by volcanic eruptions in the 1830s (Brönnimann et al 2019a) and then the Krakatoa eruption in 1883 (see figure 1). Global temperatures rose particularly rapidly over the early 20th century, showing anomalous warming from the 1920s through the 1940s (see figure 1; and Hegerl et al 2018), before plateauing in the 1950s and 60s, and beginning their strong ongoing increase.

Zoom In Zoom Out Reset image size

Much of this temperature variability has been driven by external forcing (figure 1): following a small dip of CO₂ in the Little Ice Age (Schmidt et al 2012, Masson-Delmotte et al 2013), CO₂ began to rise since the beginning of industrialization along with other greenhouse gases. The strongest increase in radiative forcing by greenhouse gases occurred in the recent few decades, an increase that is steadily continuing to date. With the burning of fossil fuels, anthropogenic aerosols began to increase as well, with aerosol forcing estimates peaking globally around 1980, although emissions have continued increasing in South and East Asia while decreasing in Europe and North America since then.

Natural forcing has imposed decadal scale variations on the total forcing (figure 1): while global radiative forcing by solar irradiance variations was quite small, with an increase towards the mid-20th century and a minimum in the 17th and early 19th centuries, episodic volcanic eruptions caused periods of stronger or weaker than average negative forcing. The largest was the eruption of Mount Tambora in 1815, which came shortly after an eruption of unknown origin in 1808 or 1809 (Guevara-Murua et al 2014, Cole-Dai et al 2016, Raible et al 2016). A strongly smoothed version of the total forcing (anthropogenic and natural forcings combined) deviates from the anthropogenic forcing substantially over some periods, most notably, the period around the Tambora eruption, and the mid-20th century. The natural forcing in the latter period is largely due to a hiatus in volcanism. It has been argued that volcanic eruptions only cause short-term cooling. However, climate model simulations show an extended cold period following the 1809/Tambora period, with no single year in model simulations with HadCM3, for example, reaching the average of the 20 years prior to the eruptions up to the 1830s (Schurer et al 2014), when another period of volcanism kept temperature low until the 1840s (Brönnimann et al 2019a). Equally, climate models simulate long-term warming in periods with little volcanic forcing, such as the early 20th century (Hegerl et al 2018). The climate model simulated response to all forcings combined (multimodel mean, concatenated between Coupled Model Intercomparison Project Phase 5 (CMIP5) and Paleoclimate Modelling Intercomparison Project (PMIP) simulations, see methods; figure 1(c)) closely follows the forcing and replicates the observed and reconstructed global temperature estimates largely within uncertainties. Some studies have argued that the response to forcing could account for more of the observed global variability if forcing uncertainty is taken into consideration (Haustein et al 2019).

The observations deviate from the model mean and range during some periods, the first of which is 1900–1910. This was a period of anomalously cold SST conditions developing in the South Atlantic and spreading northward (Hegerl et al 2018). Both long-term homogeneous stations in southern Africa and South America as well as ship data support the anomalously cold conditions during this period, which clearly deserves more attention (see discussion of ocean below). Observations are warmer than models during the peak of the early 20th century warming around 1940, which was particularly pronounced in the Arctic and Atlantic sector (Brönnimann 2009, Wood and Overland 2010, Hegerl et al 2018). The most recent deviation between climate models and observations occurred during the 'hiatus' of warming from around 1998 to 2012 (figure 1(c); Lewandowsky et al 2016, Yan et al 2016, Medhaug et al 2017) which has since ended with global warming rapidly resuming (Hu and Fedorov 2017).

The spatial pattern of observed trends (figure 2) shows that both warming and cooling/flat periods can show distinctly different spatial signatures. A trend towards cool conditions in some regions prior to 1910 shows also relatively cool conditions in data covered parts of the Southern Ocean. The early 20th century warming emerges from this cold period (figure 2(b)), which is equally strong or stronger over ocean and, while it started with strong Arctic and North Atlantic warming (Hegerl et al 2018), it is relatively uniform for the 1910–1950 trend. From 1950 to 1980, observations show a hemispherically asymmetric spatial trend pattern, with more regions warming in the Southern Hemisphere and some oceanic regions of cooling in the Northern Hemisphere (figure 2(c)). From 1980 onwards, a strong warming emerges that is almost global in nature with very strong trends (figure 2(d)). Exceptions are the off-equatorial and tropical regions of the central and eastern Pacific associated with the transition to a negative Interdecadal Pacific Oscillation phase (Zhang et al 1997, Power et al 1999) coincident with the early-2000s hiatus (Kosaka and Xie 2013, England et al 2014), and the high-latitude Southern Ocean (Armour et al 2016, Jones et al 2016).

Zoom In Zoom Out Reset image size

What caused these spatially diverse long-term trends? The similarity between simulated and observed/reconstructed changes in figure 1 suggests a strong role of external forcing. Detection and attribution methods are able to disentangle which of the forcings have played key roles in observed changes, and which are less important. Table 1 summarizes published global and hemispheric scale detection and attribution results from various timelines. Analysis of palaeoclimatic records for hemispheric and global mean data suggests that significant trends in response to greenhouse gas increases can already be detected and attributed by 1900, both across the Northern Hemisphere and in some regions such as Europe (Hegerl et al 2011, PAGES 2k Consortium et al 2013, Schurer et al 2014). This result is based on an analysis that captures the time evolution of hemispherically averaged temperature from the 15th century (table 1) and hence captures both the temperature response to a sustained, small CO₂ drop over parts of the Little Ice Age (Koch et al 2019) and the response to the CO₂ increase following industrialization. Abram et al (2016) also found sustained warming in regional proxy-reconstructions from the early mid-19th century, consistent with climate modelling.

However, volcanism is important over much of the 19th century as well: figure 1(d) illustrates that in models, the warming period up to the eruption of Krakatoa was in large parts a relaxation from a period of heavy volcanism around the Mount Tambora eruption (Brönnimann et al 2019a). From the 50 year trend centred around 1860 onwards (ending around 1885, figure 1(d)), climate models indicate that the warming trend originating from the recovery after heavy volcanism is exceeded by the warming trend caused by greenhouse gas increases. Detection and attribution analyses (table 1) confirm detectable responses to both forcings.

Analyses over the entire instrumental period robustly detect the influence of greenhouse gases when using fingerprints that are derived by averaging across many available climate models. Results based on fingerprints from individual models can vary more, with separate detection of greenhouse gas responses in an analysis simultaneously estimating natural, greenhouse gas and aerosol forcing only in about half of the models (Gillett et al 2013, Jones et al 2013, Ribes and Terray 2013). However, integrating attribution results across model uncertainty in a Bayesian analysis yields a robustly detectable greenhouse gas signal even in the presence of model uncertainty (Schurer et al 2018). The response to other anthropogenic forcings, particularly from aerosols, is less clearly detectable unless prior assumptions exclude very large or negative responses (Schurer et al 2018) and their role in regional anomalies is discussed below (see also table 1; Bindoff et al 2014).

Both for reconstructed palaeoclimate and climate over the instrumental period, the response to natural forcing (solar and volcanic combined) is robust across studies, although the best estimate magnitude is only about 70% of that in climate model simulations (see also figure 3). Scaling factors in the top panel indicate the best fit and uncertainty range of the magnitude of the model simulated pattern to observations a in equation (1). In instrumental data this slightly smaller response to natural forcing may, at least in part, be due to the confounding effect of El Niño events in the later 20th century following eruptions which, when accounted for, brings models and observations in closer agreement in attribution studies (Lehner et al 2016). A strong role for volcanism in decadal temperature variability is confirmed in detection and attribution studies for the last millennium (table 1; see Bindoff et al 2014, Schurer et al 2014) although again the amplitude of detected changes appears smaller in reconstructions than model simulations. Volcanism has also been implicated in long-term hiatus and surge events of global warming (Schurer et al 2015, Neukom et al 2019).

Zoom In Zoom Out Reset image size

Only few studies are available that estimate the role of solar forcing alone. Over the last five centuries, reconstructions support only a moderate magnitude of solar forcing (Schurer et al 2014), as do analyses of the instrumental period based on formal attribution (Stott et al 2003, Benestad and Schmidt 2009) and global time series regression analyses (Folland et al 2018, Lean 2018). Analysis also suggests a role of solar forcing in trends (figure 1(d)), although it is not significant against internal variability in climate models (indicated by the spread of simulations) yet may have slightly influenced trends. The solar influence may be stronger on regional climate where solar forcing may influence modes of climate variability. For instance, during solar minima, there appears to be an increased likelihood of the negative phase of the NAO and increased North Atlantic/Eurasian blocking frequency (Lockwood et al 2010), linked with cold winters in Europe and warm ones in Greenland (e.g. Woollings et al 2010, Ineson et al 2011, Scaife et al 2013, Gray et al 2016). Possible effects have also been found on equatorial Pacific SSTs, sea level pressure in the Gulf of Alaska and the South Pacific, the strength and location of tropical convergence zones, the strength of the Indian monsoon and the location of the descending branch of the Walker circulation, impacting on precipitation (Meehl et al 2009, Gray et al 2010, Bindoff et al 2014, and references therein). These hypothesized effects may either arise through SST influences, or through solar influence on the stratosphere (Gray et al 2010), and remain uncertain.

Figure 3 shows the implications of detection and attribution of greenhouse gas, other anthropogenic, and natural forcings over the instrumental period on causes of the trends over the periods shown in figure 2 (note that the results shown are from Schurer et al (2018) but are qualitatively and quantitatively similar to those in other studies, table 1). The attribution analysis is based on a multimodel mean fingerprint over the instrumental period, with uncertainties enlarged to avoid overconfidence (by increasing the variance of the control simulation by a factor of 2.6, see methods and Schurer et al 2018). It yields a well-constrained greenhouse gas response that is consistent and close in magnitude to the multi-model mean response in climate models. It also shows a detectable response to natural forcing, which is slightly smaller in observations than in climate models (figure 1, yellow). The response to other anthropogenic forcing is more uncertain and depends on prior assumptions (figure 3(a), the informative prior assumes no negative scaling factors and decay at about 3, peaking at 1 while the flat noninformative prior covers a −1 to +3 range).

These attribution results can be interpreted as observation-based estimates of the contribution of forcing to different periods, in a similar way that the IPCC has estimated the greenhouse gas contribution to the recent 60 years (Bindoff et al 2014). This is done by inflating or deflating the multi-model mean forced contribution to a period within the range of the estimated scaling factors. Note that interpreting the results of the long analysis over shorter segments carries additional uncertainties in that errors in the time evolution of the response may impact shorter periods, yet average out over longer periods. Where this occurs, uncertainties over the shorter period may be larger than indicated by the scaling factor uncertainty only.

Results show that the observed cooling from 1870 to 1910 in observations (uncertainty in observed change expressed in grey histogram, Morice et al 2012) occurred despite a small greenhouse forced warming, and appears to be due to a combination of internal climate variability, natural forcing (e.g. Mount Krakatoa eruption) and aerosols. The noticeable contribution by greenhouse gases is consistent with the early detection of greenhouse warming from proxy-based data discussed above. This period shows stronger cooling than simulated, consistent with the above discussed period of anomalously cold SSTs in the very early 20th century. The subsequent period (figure 3(c)) is dominated by the early 20th century global warming trend. The combined response to anthropogenic forcing (purple) is smaller than the observed trend, indicating a role of internal climate variability in the warming to 1950 (see also Hegerl et al 2018). The detection and attribution results further suggests that the plateau in observed trends from 1950 to 1980 occurred despite a net positive anthropogenic forcing, which is a strong greenhouse warming counteracted in large part by very strong aerosol induced trends. The analysis indicates that this net anthropogenic forcing was counteracted by slightly negative natural forcing (e.g. eruption of Mount Agung). Subsequently, greenhouse gases caused a strong warming trend from 1980 to 2012 (when CMIP5 simulations end); with the aerosol influence weakening and largely counteracted (in best estimate) by a slightly positive response to natural forcing, which is consistent with the eruption of El Chichon in 1982, and Mount Pinatubo in 1991 in the first half of the period (see also figure 1(d)).

These results, particularly, the varying contribution of natural forcings to different decades across the industrial periods as well as the early detectable greenhouse gas influence illustrates the difficulty of finding a suitable and 'typical' pre-industrial period (Hawkins et al 2017, Schurer et al 2017): periods are influenced differently by natural forcing, and CO₂ has been rising since 1750, following its enigmatic drop (Koch et al 2019) around the Little Ice Age. Therefore there is no constant and unambiguous preindustrial background temperature. Figure 1 shows that the period 1850–1900 (which is frequently used as proxy for the pre-industrial baseline due to the availability of instrumental observations with some global-scale coverage; Allen et al 2018) is a fairly stable climatic period with only small trends superimposed on the anthropogenic forcing, which, however, by that time had already caused a warming trend.

Model simulations and attribution results suggest that aerosols have been playing a key role in shaping regional climate, over the entire 20th century and before. It has been argued that small aerosol perturbations in early industrial time may have caused substantial impacts in a less polluted atmosphere (e.g. Carslaw et al 2013)—with potential consequences for our best estimate of aerosol radiative forcing (Stevens 2015, Kretzschmar et al 2017, Booth et al 2018). This is, however, difficult to quantify as the natural aerosol loading due to biomass burning and natural sources is uncertain, as is the magnitude of aerosol-cloud interactions and therefore the realistic nature of their representation in models (e.g. Zelinka et al 2014, Wilcox et al 2015, Toll et al 2017).

Long-term aerosol impacts on regional climate are supported by climate modelling (figure 4). European mean surface temperature, similar to global temperature, shows a plateau in warming at the period of strongest European and North American aerosol emissions that is reflected in aerosol only simulations, but not in greenhouse gas or natural only runs (figure 4(a)). Furthermore, the observed daily temperature range over Europe has decreased throughout that time period, although with quite strong variability and data uncertainty. Comparison with surface solar radiation (e.g. Wild et al 2007, Makowski et al 2008), and single-forcing simulations (figure 4(b), Undorf et al 2018b) suggest a contribution from anthropogenic aerosols to this decrease. While the temperature impact of aerosols is expected to have been largest over their emission regions, and downstream thereof (e.g. Shindell et al 2010), the global, heterogeneous patterns of temperature change simulated by models (e.g. Shindell et al 2015, Wang et al 2016) suggest relevant impact elsewhere, too. Aerosol impact on other variables is mediated by the change in temperature, as suggested for Arctic sea ice (e.g. Acosta Navarro et al 2016, Mueller et al 2018), or by temperature gradients, like for the inter-tropical convergence zone (ITCZ, Chang et al 2011, Hwang et al 2013, Undorf et al 2018c; see also figures 8(a), (b)) and the strength and position of the Northern Hemisphere subtropical jet stream and the tropical belt width (e.g. Allen and Ajoku 2016, Undorf et al 2018b).

Zoom In Zoom Out Reset image size

Long-term variability in the monsoons has also been linked to aerosol forcing, in East Asia (Guo et al 2013, Li et al 2016), South Asia (e.g. Bollasina et al 2011, Guo et al 2015), and Australia (Rotstayn et al 2012, Dey et al 2019) as well as over the Sahel (e.g. Rotstayn and Lohmann 2002, Held et al 2005, Ackerley et al 2011, Dong et al 2014). In particular the decrease between the early to mid-20th century and the 1980s and the subsequent recovery in precipitation over the global land surface (Wilcox et al 2013) and global-scale monsoon precipitation, averaged across Asian, African, and American monsoon regions in the Northern Hemisphere, has been attributed to aerosol forcing (Polson et al 2014). The aerosol influence on monsoon precipitation changes in South Asia (figure 4(c)) was found to be driven by a combination of emissions from North America, Europe and South Asia that are all simulated to weaken the monsoon circulation (Undorf et al 2018c). North American and European aerosols, along with natural forcings, are detectable drivers for African monsoon precipitation, but the model simulated changes over this region appear weak compared to observed changes, a source of concern about both modelling future changes and understanding monsoon variability (Biasutti 2013, Polson et al 2014). The dominant mechanism by which the aerosol impact is mediated, on the other hand, seems to be related to shifts of the ITCZ and is as such a well-studied response to aerosol-induced changes of the interhemispheric temperature gradients (Chang et al 2011, Chiang and Friedman 2012, Hwang et al 2013, Undorf et al 2018c; see also figures 8(a), (b)). Related impacts have been suggested on other large-scale atmospheric circulation features like the strength and position of the Northern Hemisphere subtropical jet stream and the tropical belt width (e.g. Allen and Ajoku 2016, Undorf et al 2018b).

Global warming will affect the global water cycle and has probably already done so (IPCC 2013). Diagnostics of atmospheric water content and humidity, while showing clear anthropogenic signals (see Santer et al 2007, Bindoff et al 2014) only go back through the satellite era. There is evidence from the mixed satellite/in situ record that the expected intensification of the hydrological cycle with wet regions getting wetter and dry getting drier is indeed detectable, although only if tracking wet and dry regions with the seasonal cycle and over time (Polson et al 2013, Polson and Hegerl 2017). In situ rainfall stations are able to support a longer time horizon, but are spatially sparse, fairly uncertain and only available over land (Zhang et al 2007, Harris et al 2014). However, over land precipitation is influenced by a complex combination of SSTs, land use influences, land-sea contrast and direct forcing (Greve et al 2014), and future changes over land are hence uncertain and strongly model dependent (IPCC 2013). Furthermore, climate model simulations indicate that over the historical period, precipitation changes over land are moderated by other forcings, particularly shortwave forcings (e.g. Richardson et al 2018). Nevertheless, since the second half of the 20th century, a human induced increase in intense precipitation has been detected, consistent with a moister, warmer atmosphere (Min et al 2011, Zhang et al 2013). Similarly, there is some evidence from in situ data over the second half of the 20th century that the high latitudes are becoming wetter (Min et al 2008) although data uncertainty here is substantial (Hegerl et al 2015). Zonal land precipitation shows a change since the 1920s that is broadly consistent with the expected response to anthropogenic forcing (Zhang et al 2007), although particularly seasonal responses are uncertain and noisy (Sarojini et al 2012, Polson et al 2013). Drought atlases from proxy data and instrumental data support a long-term change in drought frequency with a detectable human influence by the middle of the 20th century (Marvel et al 2019). While it remains to be seen to what extent this reflects precipitation change and to what extent increased evaporation due to warming, it provides powerful evidence that greenhouse gases have influenced aspects of the water cycle early on.

In contrast to changes over land, model-simulated precipitation changes over ocean are fairly robust across models, with a signal of wet regions getting wetter and dry regions getting drier. Many precipitation records over islands go back to the 1920s. While they are too sparse to constrain global precipitation changes, the island stations are able to evaluate the pattern of precipitation change associated with global temperature changes: the so-called precipitation sensitivity diagnoses the precipitation response to warming (for any reason including greenhouse gas induced warming) and is expressed as the change [%] in mean precipitation per degree of global mean warming. It has also been found to be a useful constraint on future changes if applied to extremes (O'Gorman 2012). Island stations support a pattern of precipitation sensitivity that is in fairly good agreement with historical simulations from climate models (figure 5) and shows a stronger correlation with historical simulations over the period since 1930 than with satellite data over the period since 1979 (Polson et al 2016). This both supports the simulated large-scale precipitation response and emphasizes the need for long records for noisy precipitation signals. An amplification of the global hydrologic cycle is also supported when using sea surface salinity as an 'indirect rain gauge' (Schmitt 2008, Terray et al 2012): globally from the 1950s (Durack et al 2012, Skliris et al 2014) and over the Atlantic from the early 20th century onwards (Friedman et al 2017).

Zoom In Zoom Out Reset image size

Forcing is also expected to directly affect rainfall through a change in energy available for evaporation. Such a response leads, at least in models, to a rapid precipitation decrease after volcanic eruptions over land (Iles et al 2013). The response to greenhouse gas induced warming is muted compared to that to aerosols since changes in lapse rate and atmospheric energy budget constraints reduce the precipitation response to warming (Allen and Ingram 2002, Lambert and Allen 2009, Andrews et al 2010, Bala et al 2010, Cao et al 2012, O'Gorman 2012). This is why the global land response to shortwave forcing such as that from anthropogenic and volcanic aerosols may be more detectable over the historical period than that to greenhouse gas increases (Allen and Ingram 2002). Long-term streamflow data are an excellent opportunity to study the water cycle response to forcing. Several large rivers have streamflow records back into the 19th century. These may reflect changes in response to a combination of precipitation and evaporation, possibly including CO₂ induced changes in transpiration (Gedney et al 2006, Piao et al 2007), and can be affected by human influences including irrigation, land use changes, dam construction, and extraction (Gerten et al 2008, Dai et al 2009, Dai 2016). However, they show a detectable and more robustly observed short-term response to volcanic eruptions, with drying on average in the wettest regions of the planet, detectable in the tropics and northern Asia, and detectable wettening in some dry regions such as the southwestern US (Iles and Hegerl 2015).

Long-term changes in atmospheric circulation are best documented for the Northern Hemisphere Atlantic sector. Here, the NAO is the dominant mode of variability at the surface, and related to variations in storm tracks, particularly, in the winter. The NAO is often defined by the pressure difference between the Icelandic low and the Azores High (e.g. Hurrell 1995), has a distinct spatial pattern of sea level pressure (e.g. Hurrell and Van Loon 1997) and has been observed over a long period of time (Hurrell et al 2003). The NAO is closely related to the Northern Annular Mode (Thompson and Wallace 2001) which is more zonal in nature and links to variations in the polar stratospheric vortex. Due to the longer record, we focus here on the NAO over the Atlantic Sector (Hurrell and Deser 2009). While the NAO is fairly white on timescales longer than interannual, it shows some long-term trends over the period of record. After a variable period with no pronounced trend in the second half of the 19th century, the NAO increased and then showed a marked long-term decrease between the early 20th century and the 1970s. This was followed by a strong upward trend peaking in the 1990s and then a downward trend into the so-called hiatus period (e.g. Hurrell 1995, Thompson and Wallace 2001, les and Hegerl 2017). Winters with an anomalously high NAO index tend to be warmer over Eurasia and anomalously cold in eastern North America and parts of Greenland as well as of the North Atlantic (figure 6, central panel middle). The opposite is true for low NAO values (Hurrell and Van Loon 1997, Iles and Hegerl 2017, figure 6, central panel left and right). If this linear relationship holds for trends in the NAO (and aggregates of NAO trends from climate models suggest it does; Deser et al 2017, Iles and Hegerl 2017) then NAO trends cause long-term temperature changes (figure 6). The NAO decrease to the 1970s may have led to dynamically induced boreal winter trends counteracting greenhouse warming to 1970, then strengthening it to the 1990s, and then counteracting it again. Residual trends, after linearly removing the NAO influence, show a more uniform warming pattern (figure 6 bottom; see also Thompson and Wallace 2001). Not shown is the impact of the NAO on precipitation, which would lead to expected rainfall trends of opposite sign over the Mediterranean and Northwest Europe (Deser et al 2017). Note that based on climate model simulations the ocean response to the NAO trend may enhance the response relative to that estimated from the interannual relationship (see next section; see also Deser et al 2017, Iles and Hegerl 2017 for regression/composite based results).

Zoom In Zoom Out Reset image size

The zonal mean Hadley Circulation is the globally dominant circulation feature, and changes in its location and strength could potentially have large impacts, particularly, on rainfall. Interannual variability in the strength of the Hadley Circulation, as well as its relationship to the El Niño-Southern Oscillation is well studied, but trends in its strength are not well established (Nguyen et al 2012) and reanalyses may show wrong trends (Chemke and Polvani 2019). Nevertheless, a robust widening of the tropical belt is found since ca. 1980 (although reanalyses data sets tend to overpredict the widening; see Davis and Davis 2018). Although climate models predict a widening due to greenhouse gas forcing and also a response to hemispherically heterogeneous aerosol forcing (section 3.3), internal variability is high and yet precludes attribution (Staten et al 2018). The widening from 1980 onward started from a southward shifted state: the northern tropical belt shifted southward from the 1940s to 1980 (Brönnimann et al 2015), and similar decadal changes in the edge of the northern tropical belt have also been found in earlier periods in tree-ring based reconstructions (Alfaro-Sánchez et al 2018).

Furthermore, a weakening of the Pacific Walker circulation was suggested at a centennial scale since the mid-19th century, in line with model simulations (Vecchi et al 2006). However, observations from recent decades point to a strengthening. Apart from data issues, this strengthening might have been due to internal variability, which is a dominant factor controlling the strength of the Pacific Walker circulation (Chung et al 2019). The Walker circulation is closely linked to El Niño, the climate mode with largest global influence, which again shows substantial variability in its variance on multidecadal timescales (Wittenberg 2009). Thus, the contribution of changes in circulation to observed long term changes in precipitation and temperature remains uncertain, as is a possible role of forcing in these changes. This remains a research priority.

The biggest potential source of decadal climate variability is the ocean. Even an inert ocean would cause decadal climate variability by integrating weather noise, and ocean dynamics are expected to enhance this variability (Hasselmann 1976, Frankignoul and Hasselmann 1977). For example, in the GFDL model, a warming episode similar to the early 20th century warming occurred in a historical simulation due to ocean overturning variability (Delworth and Knutson 2000). Figure 7 shows another example based on the Max-Planck Institute ocean model forced by century-long reanalysis ERA20C (Poli et al 2016) with an experimental set up similar to Müller et al (2015). Surface temperatures in the North Atlantic are closely associated with a downward ocean surface heat flux (latent plus sensible heat flux) on an inter-annual timescale (figure 7(b)) and upward into the atmosphere heat flux on decadal to multi-decadal timescales (figure 7(c), see also Gulev et al 2013). This clearly indicates a short-term ocean response to atmospheric forcing such as by the NAO and a long-term memory (sub-decadal to multi-decadal) by ocean inertia resulting in heat release back into the atmosphere. In fact, the role of wind-driven forcing, such as that linked to the NAO, for the ocean inertia has been widely been documented (e.g. Delworth and Mann 2000, Eden and Jung 2001, Eden and Willebrand 2001). Sub-decadal to decadal variations appear in the coupled North Atlantic climate system following wind forcing and a damped oscillation (Czaja and Marshall 2001, Eden and Greatbatch 2003). Further multi-decadal variations in the North Atlantic are closely associated with buoyancy-forced deep convection (Bersch et al 2007) and the complex interplay between processes in higher latitudes and the North Atlantic (Jungclaus et al 2005, Polyakov et al 2010).

Zoom In Zoom Out Reset image size

The variations of the NAO, North Atlantic heat fluxes and SST underwent strong multi-decadal variations (figure 7(d)) linked to the trends in the NAO discussed above. Similarly, albeit with a delay of a decade, the SST show a cooling period during the 1960s and 1970s flanked by warming periods 1920s–1930s and 1990s–2000s, respectively. The warming period in the 1920s has led to a mean increase of surface air temperature (>0.5°) within the basin and adjacent continents (e.g. Brönnimann 2009) and has the largest effects in high latitudes (Johannessen et al 2004). Changes in the atmospheric circulation have been suggested as a primary precursor of the warming (e.g. Polyakov et al 2010). In fact, by forcing an ocean model with century-long reanalysis data it has been shown that the NAO-like atmospheric circulation induces anomalous northern heat transport in the North Atlantic and incites an Arctic warming with a delay of about a decade (Müller et al 2015). North Atlantic wind anomalies may have contributed to an increased transport of warm waters into higher latitudes and to Arctic warming (Bengtsson et al 2004), contributing a signal of internal variability to the early 20th century warming. Figures 7(e), (f) underlines the importance of the heat transport for the heat release in the North Atlantic and higher latitudes. Variations of the heat transport and the Atlantic Meridional Overturning Circulation (AMOC; here 26°N and 1000 m depth) reveal the close temporal relationship to the NAO, SST and heat fluxes.

This strong ocean/coupled dynamics hypothesis contrasts with the possibility that some observed SST variations in the Atlantic may be linked to aerosols (e.g. Booth et al 2012, Zhang et al 2013, Murphy et al 2017, Bellomo et al 2018, Haustein et al 2019). Figure 4(d) illustrates strong AMV variability, but a downturn around 1970 that also occurs in response to historical forcing, suggesting that the AMV is not purely an internal mode of climate variability. Single-forcing simulations indicate a role for both anthropogenic aerosol and natural forcing (figure 4(d)), the latter of which has also been suggested to have played a role during the last millennium based on proxy records (Knudsen et al 2014, Wang et al 2017) and model studies (Otterå et al 2010). The connection of the AMV to ocean variability is unclear, with some evidence pointing to a contribution from a forced component of the AMOC (e.g. Tandon and Kushner 2015, Undorf et al 2018a, Watanabe and Tatebe 2019) in response to natural and anthropogenic aerosols (Delworth and Dixon 2006, Cowan and Cai 2013, Menary et al 2013). Determining the contribution by forcing and climate dynamics to decadal ocean variability, particularly in the Atlantic, therefore remains uncertain and requires more attention and, probably, a longer data horizon.

Ocean variability has also been implicated in the recent slowdown period of warming: Periods with decadal warming or cooling due to natural or internal variability that is strong enough to double or counteract present anthropogenic warming are dispersed throughout the historical record (Schurer et al 2015, Fyfe et al 2016). Volcanic forcing is a pacemaker particularly for long periods of fast and slow warming, with cooling due to eruption effects and rapid warming during recovery (Schurer et al 2015). However, hiatus and surge periods occur also due to internal climate variability and show a pattern involving the tropical Pacific oceans (Roberts et al 2015), with possibly a contribution by the Atlantic in observations (Schurer et al 2015).

Sea ice responds to ocean and atmospheric temperatures, with decreases in sea ice not only observed recently, but also during the early 20th century high latitude warming (Titchner and Rayner, in preparation; Walsh et al 2017). For example, large changes in the sea ice conditions around Spitsbergen were reported in the 1920s and attributed to additional warm water being pushed north by the Gulf Stream (Ifft 1922). For the recent period since 1953, signals of greenhouse gas, aerosol and natural forcing induced changes have been detected in the observations (Mueller et al 2018). Notz and Stroeve (2016) have suggested that summer Arctic sea ice is melting proportionally to cumulative carbon emissions, although other studies have proposed a role for internal variability in the recent decline (e.g. Day et al 2012). Knowledge of pre-1950s Arctic conditions could be substantially improved through the digitisation of many decades worth of voyager records along with thousands of logbooks(e.g. García-Herrera et al 2018). These logbooks contain instrumental observations of temperature, pressure and winds, and often include information on sea ice extent back to the 19th century.

The climate variability simulated in models is the yardstick against which attribution of change to causes occurs. Climate variability has also been identified as a potential emergent constraint on climate sensitivity, which is theoretically supported by the fluctuation dissipation theorem (Cox et al 2018). Both make it vital to evaluate if long observed records support the decadal variability simulated in climate models. This question has been addressed in multiple IPCC reports (e.g. Flato et al 2013) and is illustrated here for the case of Northern and Southern hemisphere SST variability in in figure 8, comparing two observational SST datasets, ERSSTv5 (Huang et al 2017) and HadSST3 (Kennedy et al 2011a, 2011b), and 10 CMIP5 models. The observations largely follow the multimodel mean and range in the historical forcing simulations over the 20th century, although some of the deviations particularly in the Southern Hemisphere are quite large, such as during the 1920s (figures 8(a)–(b)). There is also an excursion between models and data during the early 1940s that may be connected to biases related to the second world war (Kennedy 2014, Haustein et al 2019). The standard deviation of the residual SST variability (after subtracting the multimodel mean) is large in observations compared to that of the historical simulations, particularly for the Southern Hemisphere (figure 8(c)).

Zoom In Zoom Out Reset image size

To what extent this discrepancy is due to model error, residual forcing or remaining observational error in SSTs (Chan and Huybers 2019, Haustein et al 2019) is presently unclear (Friedman et al in revision). The difference between variability with multimodel mean removed compared to that from control simulations (figure 8(c)) suggests that residual forcing may have contributed to the variability in observations. However, there is a wide range of simulated variability amplitudes in the CMIP5 models, even on the global scale (Knutson et al 2013, Schurer et al 2013, Sutton et al 2015). We conclude that, evaluating which, if any, of the climate models simulate reliable variability on decadal timescales is difficult—both due to limited sampling in observations, and the difficulty to separate forced response from internally generated variability. This should be a high priority for research.

This review focuses largely on identifying answers to key research questions about the instrumental era from existing literature; involving a broad set of coauthors to cover it well, and it includes new results that are being published. Additionally, web of science searches have been conducted to ensure broad coverage, using the keywords.

'Detection and Attribution, global temperature'; using outcomes from 2012 onwards as IPCC WGI will have captured results prior to that. A further search term was '19th century global temperature change'.

Global mean temperature in all simulations is calculated as a blend of surface air temperature over land and sea-surface temperature over ocean with full coverage. To account for a potential difference in the sensitivity of forcings in the multi-model-means (MMMs) drawing on a different ensemble of climate models, the last millennium MMM is regressed onto the CMIP5 MMM during the period of overlap using a total least squares regression (scaling factor 1.20). The last millennium MMM is then scaled by the regression factor and is re-normalised so that it has the same mean over the shared period 1861–1999 as the CMIP5 MMM, where the CMIP5 MMM is plotted as anomalies since 1961–1990.