Local correlation between sea surface temperature (SST) and rainfall is weak or even negative in summer over the Indo–western Pacific warm pool, a fact often taken as indicative of weak ocean feedback on the atmosphere. An Atmospheric Model Intercomparison Project (AMIP) simulation forced by monthly varying SSTs derived from a parallel coupled general circulation model (CGCM) run is used to evaluate AMIP skills in simulating interannual variability of rainfall. Local correlation of rainfall variability between AMIP and CGCM simulations is used as a direct metric of AMIP skill. This “perfect model” approach sidesteps the issue of model biases that complicates the traditional skill metric based on the correlation between AMIP and observations. Despite weak local SST–rainfall correlation, the AMIP–CGCM rainfall correlation exceeds a 95% significance level over most of the Indo–western Pacific warm pool, indicating the importance of remote (e.g., El Niño in the equatorial Pacific) rather than local SST forcing. Indeed, the AMIP successfully reproduces large-scale modes of rainfall variability over the Indo–western Pacific warm pool. Compared to the northwest Pacific east of the Philippines, the AMIP–CGCM rainfall correlation is low from the Bay of Bengal through the South China Sea, limited by internal variability of the atmosphere that is damped in CGCM by negative feedback from the ocean. Implications for evaluating AMIP skill in simulating observations are discussed.
Summer is the major rainy season for the Indo–northwest Pacific (NWP) region, including South, Southeast, and East Asia, home to more than three billion people. In boreal summer, some of the warmest water in the world occupies the Indo–northwest Pacific Oceans. Driven by the differential diabatic heating between the Asian continent and the Indo–northwest Pacific Oceans, the southwest monsoons prevail from the Arabian Sea through the South China Sea up to 140°E, while the easterly trade winds prevail farther to the east. Atmospheric convection is active over this Indo–northwest Pacific warm pool, with regional centers off the west coast of India, in the Bay of Bengal and the South China Sea (Fig. 1). Over East Asia, there is a northeastward-slanted rainband called mei-yu in China, changma in Korea, and baiu in Japan. Summer rainfall often accounts for more than half of the annual total for the Asian monsoon region. Hereafter, seasons refer to those of the Northern Hemisphere.
Asian summer rainfall displays large variability on interannual time scales, and the predictive understanding of summer climate anomalies (including modes of variability) is a great challenge for the climate research community to meet urgent societal needs of mitigating climate change (Xie et al. 2015). In the tropics, sea surface temperature (SST) variability is an important driver for interannual variability in rainfall (Deser et al. 2010; Kumar et al. 2013), both locally and remotely.
There are two conflicting arguments about the local SST influence on the atmosphere in the Indo–western Pacific warm pool region during summer. On one hand, since the waters of the Indo–western Pacific warm pool are warmer than the convective threshold (currently at 27°C, Fig. 1), some argue that even a small change in SST can induce large changes in atmospheric convection (Gadgil et al. 1984; Sud et al. 1999; Johnson and Xie 2010). This idea motivated a large international field experiment called the Tropical Ocean and Global Atmosphere Coupled Ocean–Atmosphere Response Experiment (TOGA COARE) in 1992–93, which studied the role of the warm pool in climate (Webster and Lukas 1992). On the other hand, an opposing camp argues that SST variability there is passive and has a limited atmospheric influence by citing weak local correlations between SST and rainfall (Trenberth and Shea 2005; Wang et al. 2005; Wu and Kirtman 2007; our Fig. 2a). Here, the local SST–rainfall correlation is used as a metric for ocean influence on the atmosphere. The correlation is positive when SST forces rainfall as in the equatorial Pacific where ocean dynamical effects on SST variability are strong. Kumar et al. (2013) showed that the prediction skill of seasonal-mean rainfall tends to be high over the regions where the local SST–rainfall correlation is positive and low where the SST–rainfall correlation is either small or slightly negative. The weak or negative SST–rainfall correlation questions the atmospheric influence of SST variability over the Indo–northwest Pacific region during summer. The implied lack of ocean influence on the atmosphere in the region would limit the skill in simulating and predicting interannual variability in rainfall.
Large rainfall variability can also be induced by remote SST forcing [e.g., El Niño–Southern Oscillation (ENSO) in the equatorial Pacific]. During the El Niño–developing summer, the weakened Walker circulation induces coherent rainfall anomalies over the Indo–northwest Pacific. During a post–El Niño summer, SST anomalies over the tropical Indian Ocean induce tropospheric Kelvin waves to suppress convection over the tropical northwest Pacific region (Xie et al. 2009).
Atmospheric Model Intercomparison Project (AMIP) experiments have been widely used to study atmospheric variability by specifying monthly SST variations derived from observations or coupled general circulation models (CGCMs) in atmospheric general circulation models (AGCMs). Such AMIP simulations have successfully reproduced the Southern Oscillation, revealing the essential role of equatorial SST variability. The lack of explicit atmosphere–ocean coupling in AGCMs may cause inconsistency in surface energy fluxes, limiting the simulation skill especially when atmospheric internal variability is important in driving SST changes (e.g., Barsugli and Battisti 1998; Kumar and Hoerling 1998; Wang et al. 2005). Over the Indo–NWP region, SST is affected by the atmospheric fluctuations of weather on intraseasonal time scales (Qiu and Kawamura 2012; McPhaden 1999). In AGCMs, SST is specified, biasing the SST–rainfall relationship and limiting the skill in simulating atmospheric variability. In AMIP, positive SST–rainfall correlations are found over the western Pacific during summer while the correlations are weak or even negative in observations (Wang et al. 2005). While the biases in simulating local SST–rainfall correlations might imply low skills of AMIP (e.g., Zhou et al. 2009), many studies showed that when forced by observed SST variations, AMIP can simulate large-scale variability of the atmosphere over the Indo–NWP region (e.g., Li et al. 2008; Xie et al. 2009; Huang et al. 2010; Chen and Schneider 2014; Song and Zhou 2014a,b). Specifically, these studies pointed to Indian Ocean SST as a driver for the NWP anomalous anticyclonic circulation (AAC) in post–El Niño summers.
The present study aims to resolve the contradiction regarding AMIP skills in simulating local SST–rainfall correlation and large-scale variability. Specifically, we compare a pair of AMIP and CGCM simulations with an identical atmospheric model and identical monthly SST variations. This is different from most AMIP studies, which use SST variations from observations or different models (e.g., Wang et al. 2005). Our study is motivated by the question of whether the AMIP simulation can reproduce atmospheric variability in CGCM over the Indo–NWP region. We show that the AMIP is highly skilled in simulating summer rainfall variability over the NWP region. The local SST–rainfall correlation is not a reliable metric to measure the atmospheric influence of SST variability in the region that is subject to strong remote forcing from ENSO SST anomalies.
In the following, section 2 describes the datasets, models, and experiment design. Section 3 presents the relationship between rainfall and local SST in summer. Section 4 compares rainfall interannual variability between AMIP and CGCM. Section 5 discusses the biases of AMIP in simulating rainfall interannual variability. Section 6 is a summary and discussion.
2. Datasets and methods
We use the Extended Reconstructed Sea Surface Temperature, version 3b (ERSST.v3b), dataset (Smith et al. 2008), Climate Prediction Center (CPC) Merged Analysis of Precipitation (CMAP) (Xie and Arkin 1996), and the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (1948–2016) surface winds and sea level pressure (SLP) datasets (Kalnay et al. 1996). We have regridded these datasets onto a common 2.5° latitude × 2.5° longitude grid for a 38-yr period from 1979 to 2016 (limited by CMAP’s satellite rainfall estimates).
The present study focuses on variability on interannual time scales. Monthly anomalies are derived relative to the climatological mean over the whole period (1979–2016) after removing the linear trend. To reduce the effect of intraseasonal variability, we perform a 3-month running average. An 11-yr running mean is then applied (separately for each calendar month) to remove decadal and longer variations. Our conclusions remain unchanged when using the unfiltered data.
b. The CESM1 Large Ensemble
We make use of two simulations from the Community Earth System Model Large Ensemble (CESM-LE) project conducted by CESM1-Community Atmosphere Model, version 5 (CAM5) with biogeochemistry (BGC), at approximately 1° horizontal resolution in all model components. CESM1 (CAM5) is a global coupled climate model and consists of coupled atmosphere–ocean–land–sea ice component models. See Kay et al. (2015) for a full description of the CESM-LE. The CGCM simulation we use is the multicentury 1850 control simulation with constant preindustrial forcing. Monthly SST and sea ice for the AMIP simulation are specified from the above CGCM simulation starting at year 401. The radiative forcing is the same between the AMIP and CGCM simulations. We use a 100-yr period from year 401 to 500, to assess the effects of ocean–atmosphere coupling. Table 1 summarizes the experiments conducted.
SST and sea ice are updated at monthly intervals following the AMIP protocol of the WCRP Working Group on Numerical Experimentation (WGNE). This facilitates comparison with conventional AMIP runs forced with observed SST and sea ice. The correlation with observations is commonly used as a metric of skill but it is affected by model biases. Our study takes a different, “perfect model” approach and evaluates the “true” AMIP skill based on correlation with the mother coupled simulation.
The coupled model updates SST and sea ice daily. Our preliminary analysis shows that on the submonthly time scales, SST and rainfall are phase shifted by nearly 90° over the Indo–NWP warm pool in the coupled model. Even with updating SST and sea ice daily, AMIP-type runs will have difficulty simulating this high-frequency SST–rainfall relationship. While the scope of this study is limited to evaluating a perfect-model AMIP with monthly SST, it may be desirable to conduct AMIP-like runs in the future that update ocean surface boundary conditions at the same frequency as the coupled model does. It would be interesting to investigate the effects of high-frequency coupling.
To study atmospheric internal variability, we use another atmosphere-only simulation (f.e11.F1850C5CNTVSST.f09_f09.002) from the CESM-LE project, forced by the monthly climatologies of SST and sea ice from the CGCM simulation, referred to as the “ClimSST” simulation. Because of the absence of SST variability beyond the monthly time scale, interannual variability over ocean primarily results from atmospheric internal dynamics.
For comparison with the above perfect-model AMIP, we analyze one additional set of AMIP protocol (f.e11.FAMIPC5CN.f09_f09.historical.toga.ensXX, XX = 01, 02, 03, 04, 05, 06, 07, 08, 09, 10) carried out with the same AGCM (CAM5) from the CESM-LE project. The AMIP experiments are forced by monthly observed SST and sea ice evolution for 1880–2005 (AMIP-Obs) with different initial conditions in each run. Ideally, we would like to use one single AMIP-Obs run of 100 yr to compare with the above 100-yr perfect-model AMIP run. This is, of course, limited by the availability of both CMAP (1979–2016) and AMIP-Obs experiments: 27-yr period (1979–2005) in common. As an option, we combine 4 out of 10 runs in AMIP-Obs experiments for a longer period of 108 yr. Our conclusions remain unchanged when using one single run or combinations of any 4 out of 10 runs for AMIP-Obs experiments.
3. Local SST–rainfall correlation
Figure 2a shows the local pointwise correlation between seasonal-mean SST and rainfall anomalies for June–August (JJA) calculated based on ERSST.v3b and CMAP precipitation for the period of 1979–2016. Significantly positive local SST–rainfall correlations are found in most of the tropics (e.g., the western Indian Ocean, the equatorial southeastern Indian Ocean, and the equatorial central and eastern Pacific) (Trenberth and Shea 2005; Wang et al. 2005; Wu et al. 2006). Often used as a metric for ocean influence on the atmosphere, the positive local SST–rainfall correlation suggests that SST forces rainfall in these regions. For example, a large rainfall increase often occurs over the warmed equatorial central and eastern Pacific Ocean during the ENSO event, where it is usually cold and dry. SSTs over the Indo-Pacific warm pool exceed 28°C in the Bay of Bengal and 29°C in the Maritime Continent, South China Sea, and NWP (Figs. 1 and 2a), above the threshold for convection. Local SST–rainfall correlations are weak in the warm pool region and even negative in NWP (Fig. 2a), which suggests that SST variability there is passive and the local effect of SST on rainfall is not very clear.
In AGCMs, this local SST–rainfall correlation can be different from that in observations and CGCMs due to the lack of ocean coupling. Figures 2b and 2c compare the local pointwise SST–rainfall correlations in JJA from the CGCM and its parallel AMIP simulations. The monthly SSTs in AMIP are prescribed from a 100-yr segment of the CGCM simulation. Consistent with observations and previous studies (Trenberth and Shea 2005; Wang et al. 2005; Wu et al. 2006), local SST–rainfall correlation is positive over most of the tropical oceans in both CGCM and AMIP simulations, with values exceeding 0.8 over the equatorial central and eastern Pacific and southeastern Indian Ocean. The correlation is weak over the Indo–NWP region. CGCM shows significantly negative SST–rainfall correlation, lower than −0.4 from the Bay of Bangel to the east of Philippines. However, the correlation becomes insignificant over most of the Indo–NWP region in the AMIP simulation, despite negative and marginally significant correlation east of the Philippines and positive and marginally significant correlation near Taiwan.
The local negative SST–rainfall correlation over the Indo–NWP region implies atmospheric forcing of SST variability and hence low skill of AGCMs in simulating and predicting rainfall variability (Trenberth and Shea 2005; Wang et al. 2005; Wu et al. 2006; Wu et al. 2009; Chen et al. 2012; Kumar et al. 2013; He et al. 2017). This local view, however, does not distinguish effects from remote SST variability. For example, rainfall variability over the Indo–NWP region can result from slow ocean variability over a remote region via the atmospheric bridge mechanism (Klein et al. 1999; Alexander et al. 2002). If such remote forcing dominates, local SST–rainfall correlation is not a valid metric for skill in simulating and predicting rainfall variability.
4. AMIP–CGCM comparison
Figure 3 compares JJA seasonal-mean rainfall climatologies of AMIP (Fig. 3b) and CGCM (Fig. 3a) simulations and their differences (AMIP minus CGCM, Figs. 3c,d). The AMIP rainfall climatology closely resembles that of CGCM, with large rainfall in the intertropical convergence zone (ITCZ), South Pacific convergence zone (SPCZ), and from the north Indian Ocean to the NWP. Rainfall is weak in the southeast Pacific (Figs. 3a,b). Over the Indo–NWP region, AMIP overestimates rainfall by 5%–10% from the north Indian Ocean through NWP and 10% over south India and the Maritime Continent.
a. Local rainfall correlation between CGCM and AMIP
We examine the local correlation of JJA seasonal-mean rainfall itself between the CGCM and AMIP simulations (Fig. 4) as a direct metric for AMIP skill in replicating interannual rainfall variations in the CGCM run. In general, AMIP is highly skilled in simulating summer rainfall variability in most of the tropics. Specifically, the local rainfall correlation exceeds 0.9 over the equatorial central and eastern Pacific and southeastern Indian Ocean, and 0.8 over the western Indian Ocean (Wu and Kirtman 2005). AMIP–CGCM correlation is high (above 0.6) east of the Philippines, and marginally significant from the Bay of Bengal to the South China Sea despite negative correlation with local SST (Fig. 2). This indicates that local correlation between SST and rainfall is not a good metric of AMIP skill. Over the Asian continent, local rainfall correlation is generally low as noted previously (e.g., Zhou et al. 2009). In addition, the cold bias of the model background SST over the western Pacific may spuriously reduce the local SST forcing of convective variability (Song and Zhou 2014b; Zou and Zhou 2013).
b. Coupled variability
We perform an empirical orthogonal function (EOF) analysis of JJA rainfall anomalies in CGCM simulation to extract spatial modes of rainfall variability. The domain of analysis is the Indo–western Pacific region (10°S–25°N, 40°–160°E, shown as the red box in Fig. 8c). Figure 5 shows anomalies of JJA rainfall, 850-mb winds (1 mb = 1 hPa), and surface temperature regressed onto the leading principal components (PCs) of rainfall anomalies in CGCM simulation. The first three leading modes explain about 60% of the total rainfall variance.
The first two EOF modes are related to ENSO, corresponding to the developing and post–El Niño summers, respectively. The first EOF mode captures 37.9% of the JJA rainfall total variance in the analysis domain (Figs. 5a,d), and the PC is highly correlated with the succeeding December–February (DJF) Niño-3.4 index at 0.78. During the El Niño–developing summer, SST warms up in the equatorial central and eastern Pacific Oceans (Fig. 5d), with westerly wind anomalies in the western-central equatorial Pacific (Fig. 5a) as a result of the weakened Walker circulation. An Indian Ocean dipole (IOD) mode develops with an east–west dipole pattern of SST and rainfall anomalies with easterly wind anomalies along the equator (Figs. 5a,d). The maximum negative SST signal appears along the Java and Sumatra coasts, suggesting that the upwelling effect is important for the IOD development (Fig. 5d). This indicates a positive relationship between rainfall and local SST in the tropical Pacific and Indian Ocean.
Over the north Indian Ocean to the NWP region, rainfall decreases, and the intensified westerly wind anomalies (Fig. 5a) cool the underlying ocean (Fig. 5d). This causes an apparent negative SST–rainfall correlation locally. The positive rainfall anomalies, however, are caused by remote SST effects in the tropical Pacific, rather than the local SST anomalies. AMIP experiments indeed confirm the remote forcing of Indo–NWP rainfall anomalies by equatorial Pacific SST anomalies of a developing El Niño (e.g., Xie and Zhou 2017).
The second EOF mode (Figs. 5b,e) captures 15.2% of the total variance and is the Indo–western Pacific Ocean capacitor (IPOC) mode, which often occurs during post–El Niño summers (Xie et al. 2016). Indeed, the correlation between PC2 and the preceding DJF Niño-3.4 index amounts to 0.67. During the post–El Niño summer, the most prominent feature is a warming tropical Indian Ocean and South China Sea and an AAC over the NWP region, resulting from positive feedback of the Indo–NWP ocean–atmosphere interaction that prolongs the response to El Niño (Du et al. 2009; Kosaka et al. 2013). The north Indian Ocean is kept warm by the anomalous easterlies on the south flank of the NWP AAC, which oppose the prevailing southwest monsoon and reduce surface evaporation (Du et al. 2009). Meanwhile, the north Indian Ocean warming anchors the AAC via an atmospheric Kelvin wave adjustment (Xie et al. 2009). The westward extension of the AAC acts to reduce rainfall there, resulting in a negative correlation between rainfall and local SST anomalies. Here, the effect of Indian Ocean SST on NWP rainfall is also nonlocal. In South and Southeast Asia, El Niño–related surface temperature anomalies are larger in post–El Niño summers (EOF2) than in El Niño–developing summers (EOF1; Figs. 5e vs 5d).
The third EOF mode (Figs. 5c,f) captures 6.9% of the total variance with a significant spectral peak at an 8-yr frequency (not shown). It shows a weak IOD-like pattern and a strong negative center near the New Guinea islands. Meehl et al. (2006) noted a similar 8-yr mode in an earlier version of CESM.
To evaluate the AMIP simulation, we project the AMIP rainfall anomalies onto the EOF patterns of the CGCM in each summer (Fig. 6). The resultant PCs of AMIP are highly correlated with PCs of CGCM at 0.94, 0.8, and 0.87, respectively. Alternatively, we have performed EOF analysis on rainfall variability in the same domain using the AMIP simulation (Fig. 7). AMIP reproduces the leading modes. Spatial correlation between AMIP–CGCM pairs of leading modes reaches 0.95, 0.9, and 0.77 for EOFs 1–3, respectively.
This suggests that rainfall variability on the interannual time scale from the north Indian Ocean to tropical northwestern Pacific is largely induced by remote SST forcing (e.g., El Niño in the equatorial Pacific) rather than local SST. Forced by monthly varying SSTs from CGCM, AMIP is able to simulate interannual rainfall variations despite lacking ocean coupling. Initialized CGCMs generally show high skills in predicting EOFs 1–2 one or more months in advance (Ma et al. 2017).
5. AMIP biases
Rainfall variance shares a similar spatial pattern in AMIP and CGCM simulations that resembles the climatology (Figs. 8a,b). Upon a closer look, the rainfall variance is much larger in AMIP, by a factor of 2–3 from the north Indian Ocean to NWP (Figs. 8c,d). Assuming that rainfall variability in AMIP is the sum of the SST-forced component that is identical to that in CGCM and random atmospheric internal variability, the weak rainfall variance in CGCM explains low local correlations with the AMIP simulation from the north Indian Ocean to the South China Sea (Fig. 4) because of a low signal-to-noise ratio.
We perform an EOF analysis on AMIP minus CGCM rainfall difference in the same region as in section 4. Figure 9c shows rainfall and 850-mb wind anomalies regressed onto the leading PC for AMIP–CGCM difference. The first EOF mode captures 18.7% of total variance in the domain. There are strong positive rainfall anomalies from the north Indian Ocean to NWP and weak negative rainfall anomalies in the equatorial and south Indian Ocean. Low-level winds show a C-shaped pattern, southeasterly south and strong southwesterly north of the equator. An anomalous cyclone develops above the Philippines, with southwesterly anomalies from the north Indian Ocean to NWP. This pattern agrees remarkably well with Gill’s solution with a heating displaced north of the equator (Fig. 3 in Gill 1980) and low-level mean circulation in summer (Fig. 1).
From the north Indian Ocean to NWP, increased convection is associated with the strengthened southwesterly monsoon winds. In CGCM, the reduced surface insolation and intensified winds would lower SST through wind–evaporation–SST (WES) feedback (Xie and Philander 1994). The reduced SST would decrease rainfall and westerly winds in return. This implies a negative ocean feedback on atmospheric internal variability. In AMIP, this internal variability is spuriously strong in the absence of this negative ocean feedback. This explains why rainfall interannual variance increases over the north Indian Ocean to NWP in AMIP.
Because AMIP and CGCM share the same monthly SST, the enhanced rainfall variance in AMIP is likely to result from intrinsic dynamics of the atmosphere. We have examined to what degree the AMIP–CGCM difference might be related to CGCM variability. We define a north Indian Ocean precipitation index PNIO as the domain average in (10°S–20°N, 80°–120°E). AMIP–CGCM difference in PNIO is not well correlated with PNIO or any PC1 of CGCM (Table 2).
To test further the hypothesis that the AMIP–CGCM difference is indeed due to atmospheric internal variability, we examine an AGCM simulation (ClimSST) that is forced by repeating monthly SST and sea ice climatologies of CGCM. Interannual variability in ClimSST is internal to the atmosphere. Figures 9b and 9d show rainfall variance and the leading mode (explains 23.3% of total variance) for a 100-yr period in ClimSST. Both the variance and leading mode of rainfall in ClimSST closely resemble those of the AMIP–CGCM difference, which is large over the tropical Indo–western Pacific.
We have also examined the simulation with the Geophysical Fluid Dynamics Laboratory Atmospheric Model, version 3 (GFDL AM3) forced by observed monthly climatological SST. The leading mode is very similar to that of CESM (not shown). This indicates that the distribution of atmospheric internal variability is robust, which can induce substantial differences in rainfall variability between AMIP and CGCM simulations.
6. Summary and discussion
We have evaluated the skill of AMIP in simulating rainfall variability over the Indo–western Pacific region during boreal summer on interannual time scales. The AMIP uses the atmospheric component of the CGCM (CESM1-CAM5) and is forced by the monthly SST and sea ice variations derived from a parallel CGCM simulation. Overall, AMIP shows high skill in simulating rainfall interannual variability over tropical oceans. The local correlation of rainfall variability between AMIP and CGCM simulations is high (above 0.6) east of the Philippines while positive and marginally significant from the Bay of Bengal to the South China Sea.
Local correlation between SST and rainfall is weak or even negative over the Indo–NWP, challenging the view that SST variability in the warm pool is disproportionately effective in affecting atmospheric convection. The local negative SST–rainfall correlation, however, does not necessarily indicate a negative ocean feedback on the atmosphere over the Indo–NWP region. Rather it reflects the fact that large-scale variability of rainfall is largely determined by remote rather than local SST forcing. Indeed, Indo–NWP SST anomalies anchor coherent atmospheric anomalies in post–El Niño summers, but the atmospheric adjustments are nonlocal. Likewise, during the El Niño–developing summer, SST anomalies in the remote Pacific induce rainfall anomalies over the Indo–NWP by slowing down the Walker circulation. AMIP shows high skills in reproducing both the spatial pattern and temporal evolution of large-scale variability of rainfall, as represented by the leading three EOF modes that account for about 60% of the total rainfall variance over the Indo–NWP. The local correlation of rainfall variability between AMIP and CGCM simulations are significantly improved using rainfall reconstructions derived from the leading three EOF modes (Fig. 10). Thus, the local SST–rainfall correlation is not a reliable metric for ocean–atmospheric coupling. Instead, the AMIP–CGCM pair with the identical atmospheric model and SST is necessary to measure AMIP skill, using metrics such as the AMIP–CGCM correlations as illustrated in this study.
AMIP systematically overestimates rainfall variability over the Indo–NWP region compared to CGCM by a factor of 2–3. This bias is largely due to atmospheric internal variability. We identify a coherent mode of atmospheric internal variability, which in the positive phase, features increased rainfall and intensified southwesterly monsoon winds over the Indo–NWP. This atmospheric mode would be damped through cloud radiative effect and WES feedback if SST is allowed to vary. The spurious atmospheric variability limits the skill of AMIP, especially over the South China Sea. In the tropics, atmospheric internal variability is often assumed to be weak because of strong ocean–atmospheric coupling. It remains to be studied why atmospheric internal variability is high from the north Indian Ocean to South China Sea, and why it is organized in large-scale coherent modes as in Fig. 9.
The AMIP protocol was originally developed to evaluate AGCM skills in simulating observations, and model errors are always a concern. Figure 11 shows local correlation for JJA seasonal-mean rainfall anomalies between observations and an AMIP simulation (AMIP-Obs) forced by monthly observed SST and sea ice. Rainfall correlation is positive east of the Philippines (exceeding 95% significance level), in support of the idea that rainfall variability there is largely induced by remote rather than local SST anomalies. The smaller correlation coefficients for AMIP-Obs than that in “perfect model” AMIP (Fig. 4) may be due to model errors in the AMIP-Obs and too-strong ENSO amplitudes in the CGCM. Our AMIP–CGCM comparison sidesteps model errors in the traditional AMIP-observation comparison. We identify spurious internal variability of the atmosphere as an intrinsic factor that limits the AMIP skill as measured by local rainfall correlation in the AGCM–CGCM pair. Internal atmospheric variability, however, features spatiotemporal patterns that are distinct from the coupled ocean–atmospheric modes; the EOF1 of AMIP–CGCM difference explains only 18.7% of the variance and is marginally correlated with the AMIP EOFs (r = 0.49, −0.39, and −0.3 for EOF1–3, respectively). Thus, our results suggest that it is more appropriate to compare AMIP simulations and observations based on leading EOFs and PCs than on point-to-point local correlations (section 4, Figs. 5–7). Multimember AMIP ensemble simulations are also effective in suppressing the spurious atmospheric internal variability. Needless to say, improving models is important to improve their skills.
We wish to thank Isaac Held for useful discussions. This work was supported by the National Key Research and Development Program of China (2016YFA0601804), U.S. National Science Foundation (1637450), National Basic Research Program of China (2012CB955600), and China Scholarship Council (201506330007).