The role of external forcing and internal variability in regulating global mean surface temperatures on decadal timescales

Global mean surface temperature (GMST) shows considerable decadal variations superimposed on a pronounced warming trend, with rapid warming during 1920–1945 and 1977–2000 and warming hiatuses during 1946–1976 and 2001–2013. The prevailing view is that internally generated variations associated with the Interdecadal Pacific Oscillation (IPO) dominate decadal variations in GMST, while external forcing from greenhouse gases and anthropogenic aerosols dominate the long-term trend in GMST over the last hundred years. Here we show evidence from observations and climate models that external forcing largely governs decadal GMST variations in the historical record with internally generated variations playing a secondary role, except during those periods of IPO extremes. In particular, the warming hiatus during 1946–1976 started from a negative IPO but was later dominated by the eruption of Mount Agung in 1963, while the subsequent accelerated warming during 1977–2000 was due primarily to increased greenhouse gas forcing. The most recent warming hiatus apparent in observations occurred largely through cooling from a negative IPO extreme that overwhelmed the warming from external forcing. An important implication of this work is that when the phase of the IPO turns positive, as it did in 2014, the combination of external forcing and internal variability should lead to accelerated global warming. This accelerated warming appears to be underway, with record high GMST in 2014, 2015, and 2016.


Introduction
, who demonstrated that the MME of historical simulations can capture the observed forced response quite accurately over the past 120 years. Therefore, we use the difference between the observed GMST (representing the combination of internally generated and externally forced variations) and the MME (taken as representing the externally forced variations) to estimate the internally generated variations. We assume that the externally forced and internally generated variations are additive; an assumption has been discussed by Taylor et al (2012) and Dong and Zhou (2014) and found to work reasonably well for most purposes. Note that CMIP5 models may be overly sensitive to greenhouse gas forcing and thus overestimate the warming trend relative to observations as suggested by the lack of agreement between satellite-observed and simulated radiative signatures in the tropics (e.g. To highlight variations on decadal timescales, a long-term linear trend is first removed, then an 8-year low-pass Lanczos filter (Hamming 1989) using 13 points is applied to the time series to remove inter-annual variations. The long-term linear trend is based on a least-squares linear fit to annual mean fields over the entire period of 1920-2013 for each region. For GMST, the trend over 1920-2013 is 0.63°C per century for GISTEMP and 0.59°C per century for HadCRUT. The purpose of removing the linear trend is to highlight the decadal variations rather than to completely remove externally forced signals, which can vary on decadal time scales. The 8-year low-pass filter is commonly used to remove inter-annual variations in previous studies (e.g. Han et al 2014, Dong et al 2016). We also examined the sensitivity of the results to different choices for low-pass window by using a 13-year low-pass filter (not shown), which indicates the robustness of our results. Note that the reduced effective degrees of freedom due to low-pass filtering has been taken into account when computing the confidence limits using a Monte Carlo technique (Dong et al 2014). To extract the leading decadal modes of global surface temperature, we performed an empirical orthogonal function (EOF) analysis of the decadal smoothed surface temperature datasets.
In our study, we consider the role of the IPO in affecting GMST. Some studies we cite discuss the role of the Pacific Decadal Oscillation (PDO) and others the IPO during global warming hiatuses and periods of accelerated warming. For the purpose of our study we consider these to be the same phenomena even though there are some distinctions (e.g. Newman et al 2016). To explore the relationship between regional SST and GMST, we computed correlations on decadal timescales (figures 1(d) and (e)). For the tropical Indian  2000 1920 1940 1960 1980 2000 1920 1940 1960 1980 2000

Figure 1. Time series (in°C) for annual GMST anomalies (red lines) and detrended 8-year low-pass filtered GMST (blue lines) from (a) HadCRUT, (b) GISTEMP and (c) external forcing from 18 CMIP5 models. Correlation coefficients between SST and GMST for detrended 8 year low-pass anomalies during 1920-2013 from (d) HadISST and HadCRUT, (e) ERSST and GISTEMP, and (f) external forcing. Stippling indicates regions exceeding the 95% statistical significance. Standard deviation of detrended 8-year low-pass filtered SST from (g) HadISST, (h) ERSST and (i) external forcing.
Environ. Res. Lett. 12 (2017) 034011 Ocean, western Pacific Ocean and North Atlantic Ocean, observed decadal GMST variation shows a significant positive correlation with regional SSTs, while the eastern equatorial Pacific SST shows a weaker correlation with GMST. These results  1(f)). However, it is noteworthy that while externally forced variations dominate the observed GMST on decadal timescales, it cannot account for the pattern of decadal variation in SSTs (cf. figures 1(g)-(i)). In particular, we observe that on decadal timescales, the tropical eastern Pacific, north Pacific and north Atlantic exhibit stronger magnitude variations than other basins (figures 1(g) and (h)), most likely due to internal variations associated with the IPO and AMO, respectively, since the pattern is not well reproduced by externally forced variations (figure 1(i)).
Considering internally generated variations alone based on pre-industrial control runs, the tropical Indian and Pacific Ocean SSTs show high correlation with decadal variations in GMST (figure S2). This IPO-like spatial pattern of correlation is robust in the 6 CMIP5 models. Also consistent across all the 6 models is the stronger magnitude in SST variations of the eastern equatorial Pacific relative to the Indian Ocean and western Pacific Ocean ( (figures 3(b), (d), and (f))), which is statistically significant at the 95% level of confidence. These high correlations imply that decadal variations in GMST are strongly influenced by external forcing. However, comparing the EOF1 patterns between observations and externally forced variations reveals a large IPO-like difference in the Pacific, indicating that the IPO also contributes to GMST. Therefore, we suggest that external forcing and IPO-related SST in the eastern equatorial Pacific, which is dominated by internal variability, are the two key factors in affecting the decadal variations of GMST.
We further conducted EOF analysis on preindustrial control runs from CMIP5 models to investigate the dominant modes of internally generated decadal variations in global surface temperature. Common to all the 6 CMIP5 models, a robust IPOlike pattern emerges as the dominant mode of the

d), from observations (HadCRUT and HadISST; left column), external forcing based on 18 CMIP5 models (middle column) and internal variability obtained from the difference between observation and externally forced variations (right column). The numbers at the top right denote the correlation coefficients between the observation and externally forced or internally generated variations on decadal timescales (black lines).
Environ. Res. Lett. 12 (2017) 034011 decadal surface temperature variations, with a positive contribution to GMST (figures S5 and S6). However, the magnitude of IPO-like influence on GMST is much weaker than that of externally forced variations, ranging from 0 to 0.05°C , which corresponds to one standard deviation in PC1 based on 6 CMIP5 models (figures S5 and S6). This is not surprising, given the non-uniform spatial pattern of IPO-like mode.   1940 1960 1980 2000 1920 1940 1960 1980 2000 1920 1940 1960 1980 2000 110W  4(a)). The reason that internal variations appear to play only small role during 1946-1976 may be that the IPO was in cold phase at this time, but there were no significant trends in its time series over this period (figures S7(a) and (b)).  1946-1976 1977-2000 2001-2013 1920-1945 1946-1976 1977-2000 2001-2013 1920-1945 1946-1976 1977-2000 2001-2013 Nat   figure S8(a)). This result indicates that internally generated variations from the IPO dominated decadal cooling at the beginning of this period (red lines in figure S8(a)) with weaker magnitude than the observed GMST cooling. Later, naturally forced variations due to volcanism in 1963 had major impact (Zhang 2016) along with a weaker impact from reduced growth in GHG concentrations before 1960 based on CMIP5 models ( figure S8(b)). Therefore, the change in the IPO to a negative phase in 1946 started the decadal cooling, which was subsequently further accelerated by

Summary and Discussion
The main motivation of the present study is to explore the relative importance of external forcing and internal variability in regulating GMST on decadal timescales and to clarify the relationship between GMST and regional SSTs. We have analyzed observed surface temperature datasets and a wide variety of CMIP5 coupled climate models to systematically quantify the relative contribution of externally forced and internally generated variations during 1920-2013. Our main conclusions are that: The relationship between the El Niño-Southern Oscillation (ENSO) and GMST on interannual time scales has been broadly reported in previous studies (e.g. Mann and Park 1994, Yulaeva and Wallace 1994, Klein et al 1999). In particular, during El Niño the ocean loses heat to the atmosphere, which elevates GMST, while during La Niña the ocean gains heat from the atmosphere causing GMST to cool. The IPO can be interpreted in part as the low frequency (inter-decadal) envelope of higher frequency ENSO variations (Zhang et al 1997). Thus, periods of positive IPO (decades dominated by El Niños) lead to warm GMSTanomalies, while periods of negative IPO (decades dominated by La Niñas) lead to cold GMST anomalies. Thus, the interpretation of the role of the IPO in GMST inextricably involves consideration of the ENSO cycle.

Superimposed on a pronounced long-term cen
In this study, we use RCP4.5 runs from CMIP5 to extend external forcing beyond 2006. To examine the reliability of this procedure, which is commonly used by other investigators (see section 2), we compare the observed radiative forcing with different RCP scenarios for 2005-2015 (figure S9). All four RCP scenarios share a synchronous evolution in radiative forcing, with stronger magnitudes than the observed 1.3 Wm À2 increase in TOA radiative forcing. Previous studies estimated that increasing GHGs have led to an increasing TOA radiative imbalance of order 0-1 W m À2 since 2000 (Trenberth 2009, Trenberth and Fasullo 2013). Allan et al (2014) reconstructed the increase as 0.62 ± 0.43 Wm À2 based on satellite data, atmospheric reanalyses and climate model simulations for 2000-2012 period. The overestimates from CMIP5 RCP scenarios may be because they only consider the effects of external forcing, while observations reflect the combination of both externally forced and internally generated variations. Though the results presented here are insensitive to the choices of which RCP scenarios are used after 2006, if we use the observed TOA forcing rather than RCP forcing, the positive contribution of external forcing would be weaker and the resultant negative effect of internal variability derived from the difference between observation and externally forced variations would also be weaker. However, we would still conclude that internal variability dominates the recent global warming hiatus. A more detailed analysis of the TOA radiative forcing during most recent decade needs further study.