Introduction

The East Asian summer monsoon (EASM) precipitation features large interannual variability producing floods and droughts that have wide-ranging impacts on billions of people.1,2,3,4,5,6 Current state-of-the-art climate models generally project increased precipitation-related extremes in monsoon regions under global warming.7,8 However, these models tend to perform poorly in simulating present-day monsoon precipitation over East Asia,9,10,11,12,13,14 limiting our confidence for the future projections of the East Asian monsoon climate.15

Compared to atmosphere-only models driven by observed sea surface temperatures (SST), atmosphere-ocean coupled models can produce better monsoon precipitation.16,17 Therefore, the lack of air-sea coupling has been blamed for the poor performance of atmosphere-only models. This is supported in theory by the fact that atmospheric forcing with considerable random internal variability can exert a strong influence on the oceans on interannual timescale.18,19,20,21,22 Previous studies16,17,19,23,24 have confirmed the importance of coupling for monsoon simulation by comparing coupled model results with their counterpart atmosphere-only model results. However, the better monsoon precipitation in coupled models is often accompanied by biases in SSTs.23,24,25,26

The better performance in coupled models can be attributed to two reasons, i.e., the damped internal variability of the atmosphere27 and the compensation of biases between atmosphere and ocean.28 Quantifying the contributions of the two processes is important to understand to what degree the limited model skill is caused by experimental design problem (i.e., whether or not including coupling) or by model biases. Here, we explore the physical basis for why air-sea coupling improves monsoon precipitation simulations over East Asia through analysis of simulations from 18 pairs of atmosphere-only and the corresponding coupled models from the Coupled Model Intercomparison Project phase 5 (i.e., CMIP5)29 (see methods). We show that the improved monsoon precipitation, in terms of not only climatological mean but also interannual features, is largely due to compensation from SST biases that originate from the biased atmospheric forcing evident in the atmosphere-only models. Analysis of surface energy budgets quantitatively demonstrates how biases in the atmosphere-only models largely determine the benefits from the coupled models. This knowledge is fundamental to understanding how monsoon precipitation and SSTs in climate models can be simultaneously improved for more credible projections of the future changes in monsoon precipitation.7,30

Results

Monsoon precipitation in atmosphere-only and coupled models

The ensemble mean summer precipitation from the atmosphere-only models exhibits very large deficiencies in reproducing the observed monsoon precipitation features, but the model biases are considerably reduced in the air-sea coupled simulations (Fig. 1). In East Asia, large precipitation amounts are observed over the South China Sea and western Pacific (Fig. 1a). The Meiyu/Changma/Baiu rain-belt (referred to as the Meiyu front precipitation hereafter) extends from eastern China to Japan. In the atmosphere-only models, the Meiyu front precipitation (Fig. 1c) is too weak. Meanwhile, precipitation is overestimated by more than 2.4 mm day−1 over the northwestern Pacific (NWP; white solid box in Fig. 1a) and underestimated by more than 4.8 mm day−1 over the tropical western Pacific. Air-sea coupling improves the simulated precipitation broadly over the monsoon region (Fig. 1e vs. Fig. 1c) but produces evident cold SST biases over NWP (shading in Fig. 1e).

Fig. 1
figure 1

Climatological mean (left) and interannual (right) patterns of SST and precipitation during JJA (1984–2005) in observations and simulations. a Climatological mean precipitation in observations; b SST (shading) and precipitation (dots) regressed against the standardized EASM index in observations; c Ensemble mean biases of climatological mean precipitation (dots) in the atmosphere-only models; d Ensemble mean biases of SST (shading) and precipitation (dots) regressed against the standardized EASM index in the atmosphere-only models; e Climatological mean SST (shading) and precipitation (dots) for the coupled model ensemble mean minus the atmosphere-only model ensemble mean (i.e., coupling effect); f same as d but for the coupled model ensemble mean minus the atmosphere-only model ensemble mean. In a, the boxes denote the regions of NWP (white solid), Meiyu front (white dashed), the Indian Ocean (red solid), the western Pacific (red dashed), and the southern tropical Indian Ocean (gold solid).

Full size image

At interannual timescale (Fig. 1b), a strong monsoon (see methods) features suppressed precipitation over NWP (blue dots) and enhanced Meiyu front precipitation (red dots). Positive SST anomalies are found over the Indian Ocean, the South China Sea, and NWP (shading in Fig. 1b), consistent with previous studies.31,32,33,34 In NWP, the negative precipitation anomaly corresponds well to the positive SST anomaly, suggesting that the SST and precipitation anomalies are related through local processes. In the atmosphere-only models, the precipitation anomalies of a strong monsoon are negatively biased over the Arabian Sea, Bay of Bengal and NWP, but positively biased over the tropical western Pacific (Fig. 1d). Coupling reverses the precipitation biases, bringing the interannual pattern to a closer agreement with the observation (Fig. 1f vs. Fig. 1d).

Bias compensation improves monsoon climatology in coupled models

In the atmosphere-only models, the climatological mean precipitation biases over the oceans (Fig. 1c) are closely related to the biases in the net gain of surface energy (Supplementary Fig. 1; see methods). For example, the large wet biases in NWP is corresponding to large negative biases in net surface energy gain. The underestimated surface energy gain over NWP might explain the colder SST in the coupled simulations relative to the atmosphere-only simulations and improve the precipitation in NWP in the coupled simulations17 (Fig. 1e). Similar bias patterns can also be found in many individual models (supplementary Figs 2 and 3). These results imply that SST biases associated with the precipitation and surface energy biases evident in the atmosphere-only simulations may provide a negative feedback to the atmosphere. Therefore, precipitation over NWP is biased in the atmosphere-only simulations forced by observed SST, while the precipitation biases are reduced in the coupled simulations due to the compensation between biases in SST and biases in the simulated precipitation at given SST conditions.23,24 Some previous studies have also revealed the transfer of biases from atmosphere to ocean but SST biases then further amplify atmosphere biases via positive feedback.35,36

To quantitatively demonstrate the chain of processes that relates precipitation biases in the atmosphere-only and the coupled models, Fig. 2 shows the relationships among four key variables. These include the biases of surface energy gain and precipitation in the atmosphere-only models, as well as the SST and precipitation changes due to coupling, averaged over NWP across the 18 models. In the atmosphere-only models, the biases of the surface energy gain and precipitation are negatively correlated with each other (Fig. 2a). Most models underestimate the surface energy gain over NWP. Consistently, most coupled models have colder SST and lower precipitation relative to the atmosphere-only models, with the SST (Fig. 2b) and precipitation (Fig. 2d) changes significantly correlated with the surface energy biases in the atmosphere-only models. Hence, the hypothesized chain of processes relating the precipitation biases in the atmosphere-only models with the precipitation changes due to coupling instigated by the surface energy biases is supported by the correlations in Fig. 2a (−0.84), 2b (0.54), 2d (0.88), and 2e (−0.89), which are all statistically significant at the 95% confidence level. With the negative energy biases in most atmosphere-only models, cold SST changes dominate in the corresponding coupled models. However, the correlations between the SST changes with the precipitation biases in the atmosphere-only models (Fig. 2c) and with the precipitation changes in the coupled models (Fig. 2f) are smaller than those associated with the surface energy biases and not statistically significant. This suggests that some processes relating SST with precipitation may be non-local. Actually, the SST change over the southern tropical Indian Ocean (gold solid box in Fig. 1a) plays a considerable role in precipitation over NWP (Supplementary Fig. 4). When taking into account the remote effect from the southern tropical Indian Ocean, the correlations between the SST changes with the precipitation biases in the atmosphere-only models (Supplementary Fig. 4c) and with the precipitation changes in the coupled models (Supplementary Fig. 4f) are both statistically significant. Overall, models with larger precipitation errors in the atmosphere-only simulations benefit more from coupling and models with smaller errors benefit less (Fig. 2e and Supplementary Fig. 4e) because of the negative feedback.

Fig. 2
figure 2

Scatter plots between any two of four variables listed in the diagonal panels based on the climatological mean results averaged over NWP (see Fig. 1a) from the 18 CMIP5 models. The four variables include the biases of surface energy gain (sum of shortwave and latent heat fluxes; see methods) and precipitation in the atmosphere-only models, as well as SST and precipitation changes due to coupling. In each panel the correlation coefficient between the two variables is given at the top-right corner, with purple indicating statistical significance at the 95% confidence level.

Full size image

To illustrate the impacts of coupling at a broader spatial scale, the 18 models are categorized into three groups (each with six members) based on the surface energy gain over NWP in the atmosphere-only models. In the atmosphere-only models with low surface energy gain, the precipitation bias pattern (dots in Fig. 3a) is overall similar to the ensemble mean results (Fig. 1c) but the magnitude of bias is larger with precipitation overestimated by more than 4.8 mm day−1 over NWP. The precipitation biases are broadly reduced in the corresponding coupled models but cold SST biases ensued over NWP and the western Pacific (Fig. 3b) from the energy biases in the atmosphere-only models (shading in Fig. 3a). The colder SST over NWP influences precipitation locally through changes in evaporation and sea level pressure and remotely through changes in circulations (vectors in Fig. 3b). Compared to the low-energy atmosphere-only models, the high-energy models can generally simulate precipitation better (Fig. 3c). Notably, the precipitation biases in the Meiyu front and across the oceans are reduced. The increased surface energy gain reduces the cold SST biases over NWP in the high-energy coupled models (Fig. 3d). Hence, the benefits from coupling are less evident in the high-energy models than in the low-energy models with respect to precipitation and circulation patterns (Fig. 3b vs. 3d).

Fig. 3
figure 3

Influences of atmosphere-only model biases on the climatological mean patterns simulated in coupled models. a Atmosphere-only model biases of surface energy flux (shading), precipitation (dots), and 850hPa wind (vectors; units in m s−1) and b coupling-induced changes (i.e., coupled model results minus atmosphere-only model results) of SST (shading), precipitation (dots), and 850hPa wind (vectors; units in m s−1) during JJA (1984–2005) in the low-energy models; c, d are the same as a, b but for results in the high-energy models. Only vectors with wind speed larger than 1.8 m s−1 are shown.

Full size image

In general, the atmosphere-only models with large biases in surface energy gain over NWP fail to reproduce the southwesterly at 850-hPa over the Meiyu front and evidently underestimate the associated precipitation, while the models with higher skill over NWP can perform better over the Meiyu front (Supplementary Fig. 5). We further show the relationships among the mean precipitation biases over the Meiyu front region (white dashed box in Fig. 1a) in the atmosphere-only models and the coupling-induced precipitation changes over the Meiyu front region and over NWP (Supplementary Fig. 6). We can see that nearly all the models underestimate the Meiyu front precipitation (Supplementary Fig. 6a), although some of them can well simulate the precipitation and surface energy gain over NWP (Fig. 2a). Besides the biases from NWP, the relatively coarse resolution of atmosphere models might also play a role in the too weak Meiyu front partly because low-resolution models cannot realistically resolve the topographic effects.37 Nevertheless, the coupling-induced precipitation changes over the Meiyu front region are significantly correlated with that over NWP (Supplementary Fig. 6c), indicating the important role of the negative air-sea feedback over NWP in the mean-state Meiyu front precipitation simulation.

Bias compensation also improves monsoon interannual features in coupled models

At interannual timescale, the biases of surface energy gain in the atmosphere-only models are linked to the biases of SST in the coupled models as well. The 18 atmosphere-only models simulate very different interannual surface energy patterns. The inter-model differences are mainly dominated by a contrast of surface energy gain between the Indian Ocean (red solid box in Fig. 1a) and the western Pacific (red dashed box in Fig. 1a) (see methods), indicating a strong teleconnection between these two regions.14,17

Here we examine the relationships among the same four variables shown in Fig. 2 but based on the regressed interannual patterns averaged over the Indian Ocean and the western Pacific, respectively (Supplementary Fig. 7). For both regions, the relationships among the different fields are generally consistent with those of the climatological means over NWP (Fig. 2), indicating that the chain of local processes relating the climatological precipitation biases with the effects of coupling on the climatological precipitation in NWP works in a similar way at interannual scale over the Indian Ocean and the western Pacific. However, because of the strong teleconnection between the two ocean basins, the relationships based on the Indian Ocean–western Pacific contrasts are more evident (Fig. 4) than those in each basin. Notably, the correlations between any two of the four variables are all statistically significant at the 95% confidence level but this is not the case for the individual ocean basin. Models with larger errors in the atmosphere-only mode benefit more from coupling.

Fig. 4
figure 4

Scatter plots between any two of four variables listed in the diagonal panels based on the Indian Ocean minus the western Pacific averages (see Fig. 1a for the region information) from the 18 CMIP5 models. The four variables and numbers in each panel have the same meaning as those in Fig. 2 except that the variables are for the regressed interannual patterns.

Full size image

Based on the contrast of surface energy gain between the Indian Ocean and the western Pacific in the atmosphere-only models, the 18 models are divided into three groups (see methods). For the first group, also called the eastern-type, the atmosphere-only models generally overestimate the gradient of surface energy gain from the Indian Ocean to the western Pacific (shading in Supplementary Fig. 8a). Consistently, the corresponding coupled models produce a SST pattern with overestimated gradient from west to east (shading in Supplementary Fig. 8b), resulting in modified precipitation and circulation patterns that compensate for the biases in the atmosphere-only models. Compared to the eastern-type atmosphere-only models, the third group or the western-type models produce more surface energy gain over the Indian Ocean and less over the western Pacific (shading in Fig. 5a). The energy contrast reduces the west-east SST gradient (shading in Fig. 5b) that further influences the Indian Ocean–western Pacific teleconnection (vectors and dots in Fig. 5b) in the coupled simulations. The difference between the eastern-type models and western-type models suggests that both local and remote impacts of air-sea coupling are closely associated with the biases in the atmosphere-only models.

Fig. 5
figure 5

Influences of atmosphere-only model biases on the interannual patterns simulated in coupled models. a Atmosphere-only model simulations of the regressed interannual patterns of surface energy flux (shading), precipitation (dots), and 850hPa wind (vectors; units in m s−1) and b coupling-induced changes of the regressed interannual patterns of SST (shading), precipitation (dots), and 850hPa wind (vectors; units in m s−1) during JJA (1984–2005) based on the western-type model minus eastern-type model results. Only vectors with wind speed larger than 0.4 m s−1 are shown.

Full size image

In East Asia, the responses of the coupling-induced precipitation changes are most remarkable over south China and the adjacent oceans (dots in Fig. 5b). To find out whether the precipitation response over this particular region is due to local or remote SST changes, we examine the spatial distribution of correlation coefficient between the coupling-induced precipitation change over this region and the coupling-induced SST change at each grid point (Supplementary Fig. 9). We can see that the precipitation change is weakly correlated with the local SST change but significantly negatively (positively) correlated with the SST change over the western Pacific (eastern Indian Ocean). Previous studies16,17,38 have shown that colder SST over the western Pacific may enhance the western Pacific anticyclone, which further affects the precipitation over East Asia.

Discussion

The lack of coupling can cause inconsistencies at the air-sea interface, which limit the skill of atmosphere-only models in simulating monsoon precipitation over East Asia.18,19,20,21,22,30 Biases in the atmosphere-only models may be partly related to random atmospheric internal variability. Coupling to ocean models weakens the atmospheric internal variability through negative air-sea feedbacks,27 which may partly explain the improved skill in the coupled models. Meanwhile, air-sea coupling can improve monsoon simulation via bias compensation between atmosphere and ocean, which is also related to the negative air-sea feedbacks. If the random internal variability is the main reason for the low skill of atmosphere-only models, the SST biases in coupled simulations should only have a weak relationship with the atmosphere-only model biases. However, the significant relationships revealed here (Figs 2–5) from the CMIP5 multi-model ensemble demonstrate the importance of the deterministic biases in the atmosphere models in explaining the coupled model SST biases that compensate for the precipitation biases evident in the atmosphere-only models. While the chain of processes involved in the relationships is local (e.g., over NWP), some improvements in monsoon precipitation in the coupled models can also be achieved through remote effects (e.g., Meiyu front precipitation and the Indian Ocean–western Pacific precipitation contrast).

Here, our quantitative analyses demonstrate that the improved monsoon precipitation in coupled models is largely due to bias compensation from SSTs originating from biases in the atmospheric forcing. Models with larger atmosphere model errors benefit more from coupling and models with smaller errors benefit less. For the climatological mean, models that better simulate the surface energy gain over NWP benefit less from coupling. On interannual scale, models that better simulate the surface energy contrast between the Indian Ocean and the western Pacific benefit less from coupling. Therefore, reducing the biases in atmosphere models is important not only for improving monsoon precipitation in atmosphere-only models, but also for simultaneously improving monsoon precipitation and SSTs in coupled models, which is crucial for operational climate prediction and understanding future changes in monsoon precipitation. To reduce the atmosphere model biases, we need to improve the representations of the cloud and precipitation processes,39,40 which might be accomplished by reducing the uncertainty in the parameterizations of subgrid atmospheric processes associated with turbulence and deep convection.41,42

Methods

Models

The AMIP (Atmospheric Model Intercomparison Project) experiments from 18 atmosphere-only models and the historical experiments from the corresponding 18 coupled models in CMIP529 are examined. The model names and host institutions are given in Supplementary Table 1. The seasonal-mean precipitation, SST, surface sensible and latent heat fluxes, net shortwave and longwave radiation at the surface, and 850hPa winds during boreal summer from 1984 to 2005 are used in each model. All model simulations are interpolated to a lat/lon mesh with 2.5° × 2.5° horizontal grid spacing.

Bias of surface energy flux in atmosphere-only models

Surface energy fluxes consist of four physical components, i.e., shortwave radiation, longwave radiation, sensible heat flux, and latent heat flux. Oceanic mixed-layer heat budget analysis26 showed that the SST biases in the CMIP5 coupled models are primarily caused by the biases in downward shortwave radiation and upward latent heat flux due to atmospheric processes. Therefore, in this study, the net shortwave radiation flux and latent heat flux are used to calculate the biases of surface energy gain in the atmosphere-only models. We find that including the longwave radiation flux and sensible heat flux has only a very small impact on the results (Supplementary Figs 10 and 11).

Definition of the EASM index

The EASM index is defined as the difference between the 850hPa zonal wind averaged in (22.5°–32.5°N, 110°–140°E) and in (5°–15°N, 90°–130°E), which can represent the dominant mode of EASM.5

EASM interannual pattern

The EASM interannual patterns for different variables are derived by regressing the summer mean time series at each grid point against the standardized EASM index for observations and simulations, respectively. The ensemble mean of the simulated EASM patterns is derived by averaging the patterns in each model with equal weights.

In this study, we regress the simulated variables against the simulated EASM index for comparison between the atmosphere-only and the coupled models. Similar conclusions can be drawn when the results in the atmosphere-only models are regressed against the observed EASM index (Supplementary Fig. 12).

The EASM interannual patterns in observation and in the atmosphere-only models are denoted as Obso and Modelm, respectively. The subscripts “o (m)” denotes that the EASM pattern is regressed against the EASM index in observation (model). Thus, the difference between the two patterns, referred to as Δ (Fig. 1d), can be given by:

$$\Delta = {\mathrm{Model}}_m - {\mathrm{Obs}}_o = ({\mathrm{Model}}_m - {\mathrm{Obs}}_m) + ({\mathrm{Obs}}_m - {\mathrm{Obs}}_o).$$
(1)

The first term on the right side represents the difference between the simulated and observed patterns (including that for SST) when both model and observation are regressed against the simulated EASM index. Thus, the simulated and observed surface energy gain can be directly compared as they are associating with the same SST pattern. The second term represents the difference of observations regressed against the simulated and observed EASM index. We found that the total difference Δ is largely contributed by the first term (Modelm − Obsm) for the surface energy gain and precipitation (Supplementary Fig. 13).

Eastern-type and western-type models

Empirical orthogonal function (EOF) analyses are performed to quantify the inter-model differences in the EASM interannual pattern for the surface energy gain (Supplementary Fig. 14). The three dominant EOF eigenvectors (Supplementary Fig. 15) respectively explain 22, 20, and 11% of the total inter-model differences. The first mode shows a contrast of surface energy gain between the Indian Ocean (red solid box in Fig. 1a) and the western Pacific (red dashed box in Fig. 1a). The second (third) mode is mainly dominated by the energy change in the western Pacific (Indian Ocean). Based on the contrast of the surface energy gain between the Indian Ocean and the western Pacific in the atmosphere-only models, the 18 models are categorized into three groups (each with six members), with the first and third groups referred to as the eastern-type and western-type models, respectively.