The El Niño–Southern Oscillation (ENSO) is the primary mode of interannual and decadal climate variability in the tropics (Glantz, 2001; McPhaden et al., 2006). It also affects climate variations in subtropical and midlatitude regions across both hemispheres through teleconnection of Rossby waves (Diaz et al., 2001; Yeh et al., 2018). Therefore, its prediction is an essential part of global climate prediction on these scales (Latif et al., 1998). Additionally, ENSO is a crucial metric for climate model evaluation, especially for atmosphere–ocean coupling and associated physical feedbacks (Planton et al., 2021). Many studies have reported that Coupled Model Intercomparison Project (CMIP) models have successfully represented the basic features of observed ENSO, such as a recognizable ENSO life cycle and sea surface temperature (SST) pattern over the tropical central and eastern Pacific (Guilyardi et al., 2009, 2020; Lloyd et al., 2009, 2011; Bellenger et al., 2014). However, many model biases are also found in CMIP6 models (Beobide-Arsuaga et al., 2021; Capotondi et al., 2020a; Chen and Jin, 2021), resulting in 30 %–50 % uncertainties in future ENSO projection (Beobide-Arsuaga et al., 2021).

The paradigm based on theory and observations depicts ENSO as a closely coupled oscillation system between the ocean and the atmosphere (Jin, 1997; Latif et al., 1998; McPhaden et al., 1998, 2006; Neelin et al., 1998; Philander, 1989; Wang, 2018; Wang and Picaut, 2013). An El Niño event begins from a westerly wind initiated over the western equatorial Pacific, typically in the spring of the first year (i.e., March, April, and May in year 0; MAM0). The westerly wind drives more warm water towards the east and gradually warms the central Pacific. Around summer, eastward-propagating oceanic Kelvin waves are triggered over the central Pacific, reducing the upwelling at the tropical eastern Pacific, deepening the thermocline, and warming the sea surface temperature (SST). Consequently, the zonal SST gradient and the easterly wind in the tropical Pacific are reduced. Such a retreat of the easterly wind further reduces the SST gradient, causing the so-called Bjerknes feedback between the easterly wind and the SST gradient (Bjerknes, 1969; Cane, 2005). Through this feedback, the warm SST increases and reaches a maximum around the following winter (i.e., December of year 0 and January and February of year 1; DJF+1). Furthermore, following the warm SST anomaly (SSTA), the center of deep convection activity shifts toward the central Pacific, increasing latent heat flux and reducing shortwave heat flux into the ocean surface through deep cloud cover. Such seasonal phase locking of an El Niño event is a crucial characteristic of the observed ENSO.

In contrast to the coupling nature of the atmosphere and ocean found in ENSO observations, the atmospheric feedbacks are found to dominate the modeled ENSO frequency and amplitude in CMIP models (Guilyardi et al., 2009; Lloyd et al., 2009, 2011; Bellenger et al., 2014; Beobide-Arsuaga et al., 2021). It has been noticed that the CMIP models tend to simulate a weaker Bjerknes feedback, namely, a weaker SST warming and westerly wind coupling. Furthermore, the heat flux–SST feedbacks are overemphasized in simulated ENSO dynamics, especially for the shortwave heat flux–SST feedback. Such overemphasized heat flux–SST feedback compensates for the weaker warming from the Bjerknes feedback, producing a seemingly realistic ENSO warming in CMIP models (Bayr et al., 2019). The biased Bjerknes and heat flux feedbacks were later found to be related to the biases of the seasonal variations of Walker circulation (Bayr et al., 2019, 2018). Such complexity resulting from intertwined atmospheric–ocean feedbacks makes it challenging for model developers to improve ENSO simulations without fully understanding how these mechanisms are represented in the coupled models.

Taiwan Earth System Model version 1 (TaiESM1; Lee et al., 2020) is the Earth system model developed at the Research Center for Environmental Changes (RCEC), Academia Sinica. It participates in the CMIP6 intercomparison activity and has been used to study major climate variabilities and regional climate features (Chen and Jin, 2021; Park et al., 2020). While its overall performance of climate mean states and major variations has been evaluated and documented in Wang et al. (2021), in this study, we conducted a more comprehensive investigation of ENSO's fundamental features and statistics in historical TaiESM1 simulations. We noticed that the ENSO amplitude increased significantly compared with the observations, with intense and prolonged warm SST from May to December. Especially over the eastern Pacific, the early onset of warming and sustained warming in September, October, and November (SON0) before peaks in December are the two primary prominent biases. We further analyzed the physical processes associated with ENSO's strong warming biases and found it is due to the biased positive feedback between SST and shortwave surface fluxes over the eastern equatorial Pacific. Such feedback is primarily attributed to the prevailing low clouds overlying the cold tongue region, with large cold biases in SON0. Our results provide the baseline for ENSO performance in TaiESM1 and suggest that the current ENSO biases in TaiESM1 are intertwined with biases of the mean state and seasonal variation of the tropical climate system. The remainder of this study includes the following. Section 2 describes the TaiESM1 and observational dataset used for model evaluation and the methodology for analyzing ENSO. Section 3 documents the basic characteristics of ENSO simulated in TaiESM1. More analysis focuses on seasonal variation of El Niño events in TaiESM1 and associated biases in Sect. 4. The study is concluded with a summary and discussion in Sect. 5.

Based on the Community Earth System Model version 1.2.2 (CESM1.2.2; Hurrell et al., 2013) developed by the National Center for Atmospheric Research and sponsored by the National Science Foundation and the Department of Energy in the United States, TaiESM1 includes several physical schemes developed in-house in RCEC. These designs include convective triggering (Y. C. Wang et al., 2015), radiation parameterization of three-dimensional topography (Lee et al., 2013), an aerosol scheme (Chen et al., 2013), and a cloud fraction scheme based on probability density function (Shiu et al., 2021). The ocean component is the same as CESM1 using the Parallel Ocean Program version 2 (POP2; Smith et al., 2010). An overall evaluation of climate variability in TaiESM1 shows that the simulated ENSO features stronger SST warming and atmospheric teleconnection compared with the base model CESM1 (Wang et al., 2021). In this study, we have conducted a more in-depth analysis of ENSO's fundamental features and statistics and identified the physical processes of ENSO biases within TaiESM1. The historical simulation of TaiESM1 from 1850 to 2014, driven by the forcing designed by CMIP6, is analyzed. The historical run is initiated from the pre-industrial control run of TaiESM1. It utilizes an atmospheric model with a horizontal resolution of 0.9∘ latitude × 1.25∘ longitude and 30 vertical layers. The community land model employed shares the same resolution as the atmospheric model. Additionally, the POP2 ocean model has a resolution of approximately 1.125∘ in longitude and 0.47∘ in latitude.

We evaluated the model's performance using the atmospheric variables from the Collaborative Reanalysis Technical Environment Multireanalysis Ensemble version 2 (MRE2; Potter et al., 2018). The MRE2 is a product of the ensemble average of seven reanalysis products, including CFSR (Saha et al., 2010), ERA-Interim (Dee et al., 2011), MERRA (Rienecker et al., 2011), MERRA-2 (Gelaro et al., 2017), JRA-25 (Onogi et al., 2007), JRA-55 (Kobayashi et al., 2015), and 20CRv2c (Compo et al., 2011). Studies have found that the ensemble average can reduce the errors of individual reanalysis for selected atmospheric variables (Potter et al., 2018). For those variables not provided in MRE2, such as cloud cover, we used the ECMWF reanalysis version 5 (ERA5), the most up-to-date reanalysis produced by ECMWF (Hersbach et al., 2020). While previous studies have identified differences in air–sea feedbacks among reanalysis datasets, the ensemble mean of multiple reanalysis datasets can be used as the best estimate by reducing random errors through averaging (Kumar and Hu, 2012). We also obtained precipitation data from the Global Precipitation Climatology Project (GPCP V2; Adler et al., 2003; Huffman et al., 2009), SST data from the Extended Reconstructed SST version 5 (ERSSTv5; Huang et al., 2017), and subsurface ocean data, such as sea surface height (SSH) and potential subsurface temperature, from the Simple Ocean Data Assimilation version 3.3.2 (SODA v3.3.2; Carton et al., 2018). The observational datasets used for this study span 1980 to 2018, except for ERSST, which covers the period from 1900 to 2018.

In this study, we employ regression and composite analyses as the primary tools to investigate ENSO features. We represent El Niño using indices over crucial regions, including Niño 3 (5∘ N–5∘ S, 150–90∘ W), Niño 3.4 (5∘ N–5∘ S, 170–120∘ W), and Niño 4 (5∘ N–5∘ S, 160∘ E–150∘ W), in the regression analysis. For the observational Niño 3.4 index, we use a base period between 1900 and 2014 from ERSSTv5, following the Niño index calculation of the Climate Prediction Center, NOAA (NOAA Climate Prediction Center, 2000). We also use model data from 1900 to 2014 as the base period for TaiESM1's historic run. To avoid impacts of model bias on longer timescales, such as interdecadal variation, we utilize the full length of available simulation data to obtain the most robust statistics of ENSO feature simulated by TaiESM1.

We choose the composite method instead of the regression map to better identify teleconnection signals associated with the El Niño events. In our preliminary analysis, the regressed maps of El Niño events show patterns similar to the composite events in the tropics, but with much weaker signals in the midlatitudes (not shown). Furthermore, as the ENSO events simulated in TaiESM1 show very symmetric alternations between El Niño and La Niña events, our composite based on El Niño events is followed by La Niña events. To build the El Niño composite, we choose strong events with the Niño 3.4 index larger than 1 standard deviation (i.e., 1.22 ∘C) over the simulation period. The ERSSTv5 dataset includes eight El Niño events in 1982, 1986, 1987, 1991, 1994, 1997, 2002, and 2009. In comparison, TaiESM1 simulated 21 events throughout the entire historical simulation. As the El Niño events simulated by TaiESM1 exhibit a very strong amplitude, most of the composite fields of these events passed the significance test at the 95 % confidence level (not shown). Therefore, we will not denote regions passing significant tests in the composite fields in the following analysis.

In our analysis of TaiESM1's ability to simulate ENSO diversity, we examined its capability to distinguish between eastern Pacific (EP) and central Pacific (CP) El Niño events using the Niño 3–Niño 4 approach (Kug et al., 2009). During the historical period of TaiESM1, we identified 23 CP events and 17 EP events. This higher frequency of CP events in TaiESM1 is consistent with previous findings in CMIP models (Capotondi et al., 2020b; Chen et al., 2017; McPhaden et al., 2011). It is worth noting that the total number of El Niño events exceeds 21 when using the El Niño 3.4 index, as some EP events transition into CP events during the mature phase. In contrast, during the observation period from 1980 to 2014, we found a total of four EP events and four CP events. In the subsequent sections, we will also discuss the composites of EP and CP events in further detail.

In this section, we analyze the seasonal life cycle of the strong ENSO signals simulated in TaiESM1 to understand the intense tropical warming anomaly of El Niño events and link this bias with seasonal mean states. Based on the El Niño events defined in the previous section, we construct the seasonal cycle of El Niño events by plotting a Hovmöller diagram averaged over the Equator between 3∘ S and 3∘ N.

Figure 9 shows the Hovmöller diagram of SST anomalies and 1000 hPa zonal winds along the Equator (3∘ S–3∘ N) based on the strong El Niño event selected from the ERSSTv5 and TaiESM1 historical run. In the observation in May, the warm SSTA occurs over the dateline and the westerly wind anomaly starts to propagate to 135∘ W. In JJA0, the warm SST slowly develops at 180–135∘ W, progressively propagates to the eastern equatorial Pacific with westerly anomalies, and reaches maximum amplitude in November, December, and January (Fig. 9a). In contrast, the warm SST when initiated in May intensifies almost simultaneously in the basin east of 150∘ W in TaiESM1 (Fig. 9b). The warm anomaly quickly reaches 2 ∘C over the entire eastern equatorial Pacific through JJA0 and into September. Such warming seems to be coupled, with one branch quickly extending to the eastern equatorial Pacific after the initiation in May, whereas the westerly anomaly's major branch is well coupled with SSTA over the western Pacific most of the time. The warming continued developing even when the westerly anomalies in the eastern equatorial Pacific weakened in JJA0 and reached a maximum in January as in observations. This early development of a warm SSTA in the eastern equatorial Pacific in May and continuous warming in JJA0 and SON0 are two primary model biases that may contribute to strong El Niño events in TaiESM1.

To understand the heat flux–SST coupling, we examined similar composites of net surface heat flux (Fig. S5a, b) and latent heat fluxes (Fig. S5c, d), which are the two heat fluxes prominent in observational El Niño events. Notably, in the observations, the seasonal variation of surface net heat fluxes in MRE2 is primarily controlled by latent heat fluxes, especially from May to November, as other heat fluxes play a minor role (Fig. S5c, d). However, the spatial patterns of net surface fluxes of TaiESM1 are primarily dominated by shortwave surface fluxes (Fig. S5d) and amplified by latent heat fluxes (Fig. 2b). Figure 10 shows the same Hovmöller diagram but for downwelling surface shortwave flux (color shading) and surface temperature (contours) composited over the El Niño events. Evolutions of shortwave radiation and SST during the life cycle of El Niño events in observations and TaiESM1 are shown in Fig. 10a and b, respectively. In the observations, upward shortwave fluxes of about 10 to 20 W m-2 (i.e., shown as negative in Fig. 10) are seen during the entire El Niño period, which is well collocated with the warm SSTA, due to the increased shortwave reflection of deep convection triggered by warmer SST. Such reflection of shortwave fluxes by deep clouds induces negative feedback (i.e., higher SST, large reflected shortwave radiation) between shortwave radiation flux and SST in the tropics (Fig. 10a). The negative feedback intensified in September, reached the maximum in December, and continued into the following February (Fig. 10a). In contrast to the observations, TaiESM1 produced unrealistically strong negative feedback over the western equatorial Pacific and Intertropical Convergence Zone (ITCZ) regions (Fig. 10b), as well as very strong positive shortwave radiation anomalies near the equatorial eastern Pacific after May in year 0 (Fig. 10a). The increase in shortwave influx starts from the eastern Pacific in May and gradually extends to 125∘ W in November, a feature that was not seen in observations. The westward progression of shortwave radiation increase contributes to erroneously strong SST warming, and its westward extension is seen in TaiESM1 from May to Jan in year 1 (Fig. 10b) instead of the observed simultaneous warming in the eastern Pacific.

To understand the cause of the strong SSTA development from May to November identified in Fig. 9 and its relationship with shortwave radiation fluxes in Fig. 10, we analyzed the seasonal cycle of low clouds (gray contours), vertical velocity (green contours), and SST (color shading) in the observations and TaiESM1 in Fig. 11a and b. Figure 11c shows the differences between TaiESM1 and MRE2. Figure 11d shows the climatological seasonal SST cycle at 100∘ W to show the SST differences where cloud biases are prominent. On the seasonal timescale in observations, the SSTA is closely coupled with anomalies of vertical velocity and low-level clouds. Cold surface temperature is commonly collocated with excessive low-level clouds and subsidence (Fig. 11a). While TaiESM1 also exhibits a clear seasonal variation like the observations in the eastern tropical Pacific (Fig. 11b), the simulated seasonal cycle is rather asymmetric with a colder bias during MAM0 and SON0 and warm biases during JJA0 and DJF+1 (Fig. 11c). Compared with the observations, TaiESM1 warmed approximately 1 month later during MAM0 and cooled deeper over the eastern Pacific in SON0 (Fig. 11c). The warmer MAM0 sea surface in TaiESM1 provides a smaller zonal SST gradient and may lead to the earlier onset of the eastward-propagating westerly anomaly found in MAM0 through the Bjerknes feedback (Fig. 9b). However, this cold surface temperature bias during SON0 provides a cold lower boundary for low stratus clouds to develop, leading to an environment for the strong positive feedback between shortwave fluxes and SST during El Niño events. Such impacts of the cold tongue bias are also found in CESM1 and CESM2, which share the same ocean model (i.e., POP2) as TaiESM1 (Y. C. Wang et al., 2015; Wei et al., 2021).

For a further look into the bias in SON0, Fig. 12 shows the changes in cloud fraction and zonal circulations in SON0 during the composited El Niño events. TaiESM1 successfully reproduced the eastward shift of the upward branch related to deep convection in response to the warm SST during El Niño events (Fig. 12a, b). However, TaiESM1 produced a much stronger response in circulations and cloud cover than the observations (Fig. 12b). Especially over the eastern Pacific, a stronger upward motion anomaly occurs and about 10 %–20 % of low clouds are reduced in TaiESM1, indicating a dramatic reduction of the low-cloud regime (Fig. 11c). As a result, more shortwave heat influxes are allowed into the ocean surface in SON0 when El Niño events occur (Fig. 10b). Consistent with the seasonal variation of SSTA in Fig. 9b, the increased shortwave fluxes thus help to keep the warm SSTA in SON0 over the eastern Pacific after the warm SSTA initiates in MAM0 in TaiESM1 (Fig. 10b).

This study documented ENSO's fundamental statistics and features in TaiESM1, a CMIP6 participant. Compared with the observational dataset, TaiESM1 has captured many prominent observed ENSO features, including a 3–5-year spectrum peak, seasonal phasing, the evolution of warm SST, deepening of the subsurface layer during an El Niño event, and teleconnection patterns in midlatitudes. However, the simulated El Niño signals in TaiESM1 are much stronger and more prolonged than the observed ENSO signals. Such strong signals are shown as intense warm SSTA over the tropical Pacific within the spatial structure of the composite of the El Niño event. In the meantime, in response to the tropical warm SSTA, the teleconnection wave activities associated with El Niño events are much stronger in TaiESM1, significantly affecting temperature and rainfall in high-latitude regions, although with a similar spatial pattern to the observations.

To understand the cause of the warm El Niño biases, we investigated the seasonal cycle of strong El Niño events in TaiESM1. Compared with the observed delay of SST propagation to the central Pacific in May after the wind anomaly's initiation, the simulated SST warming quickly propagates toward the eastern Pacific along with the westerly wind anomaly in TaiESM1. The SST warming continued through November and reached a maximum in December, showing a stronger and more prolonged warm period than the observation. Moreover, in contrast to the negative feedback found in observations, our analysis shows that the strong El Niño warm anomalies in TaiESM1 is due to the strong SST–heat flux positive feedbacks, especially over the eastern Pacific in SON0. Further analysis shows that this biased feedback is due to the response of the spurious low-cloud regime off the west coast of South America simulated in TaiESM1. This biased cloud regime results from the seasonal variation of the cold tongue over the eastern Pacific, consistent with previous studies using the CESM family and CMIP models (Ham and Kug, 2012; Wei et al., 2021). During El Niño events, the stratus clouds over the eastern Pacific gradually diminished due to the warmer SST, allowing solar radiation to warm the ocean surface. This result leads to a positive feedback of downward solar radiation and SST over the eastern Pacific, an opposite sign of the observed relationship. This biased relationship is typical in the CMIP5 and CMIP6 models (Bayr et al., 2019; Beobide-Arsuaga et al., 2021).

In summary, TaiESM1 can reproduce many fundamental features of ENSO. However, it still possesses several biases shared by other CMIP6 models, including the lack of randomness of El Niño events, El Niño magnitudes that are too strong, and early SST warming in the early stage of El Niño events. The strong El Niño strength, our analysis found, mainly results from the biased atmosphere–SST coupling accompanied by biases of the mean state and seasonal cycle of the cold tongue and Walker circulation, which is consistent with previous studies with CMIP5 and/or CMIP6 models (Bayr et al., 2018; Chen et al., 2021). However, to resolve this bias (and others), more detailed analysis, including process-related metrics, and more model experiments, such as atmosphere-only and ocean-only experiments, are required to determine the cause and effects of the observed biases. For TaiESM1, we plan to implement ocean-only experiments with the ocean component POP2, allowing us to quantify the ocean's response to biased winds and radiation fluxes. We will also conduct AMIP-type simulations to investigate the development of westerly wind anomalies under biased SST conditions. Combined with process-oriented diagnosis, these model experiments will help us to determine and better comprehend the causes and effects of these observed biases. Also echoed with other studies of ENSO evaluations of CMIP6, our analysis suggests that a more process-based development strategy focusing on atmosphere–ocean coupling rather than a feature-based evaluation of ENSO is needed to reduce the uncertainty of ENSO simulations and future ENSO projection in climate models.