An analysis of a 5-yr (from 1 January 2009 to 31 December 2013) free run of the superparameterized (SP) Climate Forecast System (CFS) version 2 (CFSv2) (SP-CFS), implemented for the first time at a spectral triangular truncation at wavenumber 62 (T62) atmospheric horizontal resolution, is presented. The SP-CFS simulations are evaluated against observations and traditional convection parameterized CFSv2 simulations at T62 resolution as well as at some higher resolutions. The metrics for evaluating the model performance are chosen in order to mainly address the improvement in systematic biases observed in the CFSv2 documented in earlier studies. While the primary focus of this work is on evaluating the improvement of the simulation of the Indian summer monsoon (ISM) by the SP-CFS model, some results are also presented within the context of the global climate. The SP-CFS significantly reduces the dry bias of precipitation over the Indian subcontinent and better captures the monsoon intraseasonal oscillation (MISO) modes. SP-CFS also improves the northward and eastward propagation of high- and low-frequency modes of ISM. Compared to CFSv2, the SP-CFS model simulates improved convectively coupled equatorial waves; better temperature structures both spatially and vertically, leading to a significantly improved relative distribution of variance for the synoptic disturbances and low-frequency tropical intraseasonal oscillations (ISOs). This analysis of the development of SP-CFS is particularly important as it shows promise for improving the cloud process representation through an SP framework and is able to improve the mean as well as intraseasonal characteristics of CFSv2 within the context of the ISM.
The seminal role played by the “internal dynamics” arising from higher-frequency subseasonal oscillations compared to the slowly varying “external forcing” makes the Indian summer monsoon (ISM) climate (e.g., the seasonal mean or the annual cycle) one of the most challenging systems to simulate and predict (Goswami et al. 2006; Wang et al. 2005; Goswami and Mohan 2001). To be able to simulate the ISM climate realistically, therefore, it is imperative that a climate model simulates the statistics of the subseasonal oscillations (synoptic disturbances and subseasonal oscillations) with fidelity. However, all the tropical subseasonal oscillations are convectively coupled. Therefore, in order to realistically simulate the tropical climate in general, cloud feedbacks should be accurately represented (Arakawa 1975, 2004; Frank 1983; Randall et al. 2000; Majda 2007; Moncrieff et al. 2012; Waliser et al. 2012). Traditionally, because of the relatively coarse horizontal grids, the cloud effects in general circulation models (GCMs) have been represented using parameterizations. The progress made toward improving the representation of the subgrid-scale cloud processes has been unacceptably slow (Randall et al. 2003; Randall 2013). A recent alternative approach called superparameterization (Grabowski and Smolarkiewicz 1999; Khairoutdinov and Randall 2001) has proven to be quite promising (Khairoutdinov et al. 2008; Benedict and Randall 2009; Goswami et al. 2011, 2013; DeMott et al. 2013). A microscale–mesoscale dynamic (MMD) model, which is an extension of superparameterization (SP) for numerical weather prediction (NWP), has also been recently introduced (Majda 2007).
Over the past decade, significant effort has been devoted toward understanding the advantages of the GCMs with superparameterization, also known as multiscale modeling frameworks (MMFs). Prior to this study, the superparameterization has been implemented in the Community Atmospheric Model (CAM, known as SP-CAM), the Community Climate System Model (CCSM, known as SP-CCSM), and in the NASA GISS finite volume (fv) GCM (Tao et al. 2009). Zhu et al. (2009) reported a more realistic stratiform heating profile associated with higher rain rate and vigorous Madden–Julian oscillation (MJO; Madden and Julian 1971) in SP-CAM as compared to CAM3.0 simulations. Benedict and Randall (2009) also showed improved simulation of the space–time structure of MJO by SP-CAM. Goswami et al. (2011, hereinafter G11) analyzed the monsoon intraseasonal oscillation (MISO) in SP-CAM simulations. Although SP-CAM was able to produce a reasonable seasonal mean, it showed significant bias in capturing different modes of MISO. Similar to Luo and Stephens (2006), G11 also found SP-CAM to produce a significant wet bias over the Asian monsoon region. The SP-CCSM simulations improved the simulation of the ISM compared to the conventional CCSM; however, there have been notable biases (DeMott et al. 2013; Goswami et al. 2013), which could possibly stem from the biases in the host CCSM model itself (Meehl et al. 2012; Sabeerali et al. 2013). The simulation of ISM is found to have significantly improved in SP-CCSM compared to SP-CAM (Stan et al. 2010; Goswami et al. 2013), which indicates better fidelity of the SP framework in a coupled general circulation model (CGCM). However, Goswami et al. (2013) also noted several biases such as the absence of an oceanic ITCZ over the Indian Ocean region and improper simulation of different modes of ISM variability in SP-CCSM. Recently, Krishnamurthy et al. (2014) also showed that the SP framework in coupled models produce a more realistic ISM simulation as compared to SP-CAM or compared to a GCM where conventional convection parameterization is used (e.g., CCSM). All of the above studies unanimously indicate that resolving the subgrid-scale cloud processes through the SP framework improves model fidelity. Although there is no certainty that every GCM would respond to the SP framework in the same way as CCSM, it is worthwhile to explore the scope of the SP framework for a model, which simulates a rather realistic ISM compared to CCSM. It has been demonstrated (Sabeerali et al. 2013) that the CFSv2 simulates a reasonably realistic ISM, which is significantly better when compared to CCSM simulations. Having said that, CFSv2 also shows many prominent biases (Chaudhari et al. 2013; Saha et al. 2013; Goswami et al. 2014; Abhik et al. 2015).
The CFSv2 model has been selected, as part of the Monsoon Mission of India, as the operational model for dynamical seasonal prediction over the Indian monsoon region. While the suitability for CFSv2 for seasonal predictions of monsoons comes from the fact that it has one of the highest forecast skill for retrospective forecasts of all-India seasonal-mean rainfall among the models of its class (Zuo et al. 2013), it still suffers from some significant systematic biases such as (i) a cold bias in the tropospheric temperature and a dry bias for ISM rainfall, etc. (Saha et al. 2013; Goswami et al. 2014) and (ii) a significant underestimation of synoptic variance over the Asian monsoon region. These biases are likely to prevent us from any further improvement in seasonal prediction skill. This, in fact, provided us with strong motivation to incorporate the SP into the CFSv2 model. Keeping this in mind, we intend to improve the subgrid-scale cloud processes in CFSv2 through the inclusion of an SP framework and subsequently to evaluate the superparameterized CFSv2 (SP-CFS) to explore the impact of SP on the systematic biases observed in CFSv2.
In this paper, we first examine the boreal summer [i.e., June–September (JJAS)] mean spatial pattern of the precipitation field focusing over the ISM domain. We also investigate the mean features of a few other fields that have important influences on deciding the features of the ISM rainfall. As monsoon intraseasonal oscillations (ISOs) are the building blocks of the seasonal-mean monsoon (Goswami et al. 2006), next, we bring our investigation down to the intraseasonal time scale to examine if the monsoon ISO modes are simulated with fidelity. As the simulation of the tropical ISOs by a model, in general, is intimately linked to its ability to simulate the convectively coupled equatorial waves (CCEWs; Wheeler and Kiladis 1999; Zhang 2005) over the entire tropics, we also examine the improvement of simulation of the CCEWs in the CFSv2 model within an SP framework. The relative shares of the synoptic and ISO variance in the CFSv2-simulated boreal summer climate (Goswami et al. 2014) in the SP-CFS simulations are also explored over the zonal belt spanning 50°S–50°N. The model and methodology are described in section 2 followed by results and discussion in section 3.Conclusions are offered in section 4.
2. Model description, data used, and methodology
The CFSv2 is a fully coupled GCM with the atmosphere component known as the Global Forecast System (GFS), which was configured to use a spectral triangular truncation at wavenumber 62 (T62) horizontal resolution and 64 vertical levels. The superparameterization is a cloud-resolving model (CRM) embedded within each grid of CFS. The CRM is the System for Atmospheric Modeling (SAM), version 6.7.5, described by Khairoutdinov and Randall (2003). The conventional cumulus parameterization and the cloud parameterizations in the SP-CFS are turned off. Each CRM has a periodical two-dimensional domain with 32 columns with 4-km grid spacing, and 60 levels collocated with the first 60 CFS vertical levels. The SAM has an anelastic dynamical core and the model integrations are done using a third-order Adams–Bashforth scheme on a staggered Arakawa C grid. Its prognostic variables are liquid–ice water static energy, and total nonprecipitating (cloud water and ice) and total precipitating water (rain, snow, and graupel), and its diagnostic variables are cloud water, cloud ice, rain, snow, and graupel mixing ratio. SAM uses a 1.5-order subgrid-scale closure (prognostic SGS TKE) or Smagorinsky-type closure scheme (Smagorinsky 1963) to compute the subgrid-scale fluxes. Surface fluxes are calculated based on Monin–Obukhov similarity. Precipitation is computed using bulk microphysics that included the ice microphysics processes. The CRMs are forced uniformly with the CFS large-scale dynamic tendencies applied horizontally and return the vertical profiles of the large-scale tendencies due to cloud processes. Coupling between the CFS and CRM generally follows the methodology used in the SP-CCSM as described by Khairoutdinov et al. (2005). However, unlike the SP-CCSM, which uses the semi-Lagrangian dynamical core and two-level forward-in-time scheme, the CFS uses the leapfrog scheme. The CRM in SP-CCSM starts from the CRM solution at the end of the previous CCSM time step integrating forward in time with its own small time step (usually 20 s) over one CCSM time step. However, the CFS uses a three-level leapfrog scheme; so that the CRM is integrated forward over two CFS time steps. Therefore, the CRM on SP-CFS produces two separate solutions: one for the odd and the other for even CFS time steps. Other than that, the coupling between the CRM and CFS generally follows the methodology used in SP-CCSM. For the cloud–radiation calculations, cloud statistics calculated by the CRMs are passed to the CFS radiation package. The land surface fluxes are computed on the GCM grid scale. Momentum feedback is not allowed between the CRM and GCM. The SP-CFS employs MOM, version 4.0 (Griffies et al. 2003), as the ocean model and the four-level Noah land surface model (Ek et al. 2003) represents the land surface.
Computationally, the SP-CFS demonstrates reasonably good parallel performance using the hybrid message passing interface (MPI)–OpenMP model. However, running the SP-CFS requires considerable computer time. Details about the computational aspects of SP-CFS are provided in section 2c. The SP-CFS has been integrated for a period of 5.5 yr as a free run initialized on 16 May 2008, the initial condition being obtained from the National Centers for Environmental Prediction (NCEP). The first 6 months are considered for the spinup period and the remaining 5 yr (1 January 2009–31 December 2013) are analyzed and presented in this paper. The time tag is just for convenience and not to be confused with predictions.
b. Methodology and observational dataset
In this study, the fidelity of the SP-CFS in simulating ISM and MISOs is compared to that of the conventional CFSv2 and the observations. The CFSv2 is also run with T62 horizontal resolution for the same period as SP-CFS. The analyses presented here are for the ISM season (i.e., JJAS), unless mentioned otherwise. Additionally, 14- and 7-yr climate simulations by the CFSv2 at T126 and T382 resolutions, respectively, are also used to demonstrate whether the higher-resolution version of CFSv2 with conventional parameterization could better capture the CCEWs.
The Tropical Rainfall Measuring Mission (TRMM) 3B42 0.25° × 0.25° resolution daily rainfall dataset (Huffman et al. 2010) is used for the observed precipitation. To better compare the model-simulated rainfall with observations over the Indian subcontinent, a high-resolution gridded rainfall (1° × 1°) dataset is generated based on India Meteorological Department (IMD) rain gauge observations (Rajeevan et al. 2006) to use in the study. The NOAA outgoing longwave radiation (OLR) dataset (Liebmann and Smith 1996) and several NCEP–NCAR reanalysis products (Kalnay et al. 1996) are used as observational benchmarks. All the model outputs and observational–reanalysis datasets have been bilinearly interpolated to 2.5° × 2.5° regular grid resolution. For the analyses carried out in this paper, we have used a smoothed climatology, which is reconstructed from the mean and the first three harmonics of the actual climatology. We have used a Lanchoz filter (Duchon 1979) to perform all the filtering mentioned in the paper. To evaluate the vertical heating structure simulated by the model and comparisons with the reanalysis, the diabatic heating Q1 is calculated following Yanai et al. (1973) as
where Cp is the specific heat of air at constant pressure (~1004 J kg−1 K−1), R is the gas constant for dry air (~287 J kg−1 K−1), p0 is the reference pressure (~105 Pa), ω is the vertical velocity, and θ is the potential temperature of the air parcel.
c. Computational cost
Earlier studies (Khairoutdinov et al. 2005) have shown that SP-CAM takes 2 days of wall-clock time for a 1-yr simulation on 1024 cores of an IBM SP2 computer, and they could achieve this with good parallel performance of the model. The SP-CFS model takes around 0.6 and 4 h of wall-clock time for a 1-day simulation on 1024 cores of an IBM Power6 computer at T62 and T126 resolutions, respectively. The SP-CFS model at T62 resolution is 90 times costlier than CFS at the same resolution, while SP-CFS at T126 is 230 times more expensive than CFS at the same resolution. This computational cost of SP-CFS had compelled us to test the latest version of CFSv2 at a lower resolution (T62) for demonstrating this pioneering experiment about the role of SP on CFSv2. Though there is a hindrance to parallelizing the atmospheric component in the CFS, which is a spectral model, through the usage of parallel hybrid threading, we could integrate a 5-yr simulation of the SP-CFS model at T62 resolution on 557 cores while looking for a scope to improve the parallel performance.
To avoid repetitiveness, for the rest of this documentation the phrase “relative/compared to CFS” will be minimally used while analyzing SP-CFS results.
3. Results and discussion
To demonstrate the impact and improvements of the SP-CFS model, we will compare its performance to the performance of the CFSv2 model with the main focus on the biases typically seen in CFSv2 simulations as documented in previous studies.
To get a sense of the amplitude and variability of the daily rainfall and how it contributes to the seasonal mean, we examine the daily rainfall (Fig. 1) over the core monsoon zone (Rajeevan et al. 2010; Singh et al. 2014). It should, however, be noted that these are continuous simulations and we are neither correlating the simulated rainfall with the observations nor doing any year-to-year comparison. The motivation for Fig. 1 is strictly to see the annual cycle and its interannual variability and to see how the subseasonal spells are contributing to the seasonal mean monsoon rainfall over the monsoon zone. A significant and realistic improvement in the amplitude of the day-to-day variability of rainfall during the summer months is noted in the SP-CFS simulations. This is evident from the JJAS standard deviation values shown in the bottom-right panel of Fig. 1. While the CFS underestimates the daily variability, simulating only 30% of observed amplitude, SP-CFS comes close to the observations by simulating 75% of the observed amplitude of the daily rainfall variability. Also, it is noted that CFS rarely simulated any wet spells over central India whereas SP-CFS simulates many wet spells. However, we still note that these spells are of relatively shorter duration and between the wet spells central India becomes completely dry, which is rarely seen in the observations.
a. Mean field analyses
The Indian summer monsoon is a convective–radiative–dynamical coupled system (e.g., Goswami and Shukla 1984; Wang et al. 2005), which critically depends on the background climatic state. Therefore, realistic simulation of the climatological mean field is necessary for any model. In this section we present an overview of the simulated climatological mean precipitation and winds and a few other simulated fields, which are crucial in making the ISM climate what it is.
1) JJAS mean precipitation and OLR
The JJAS global mean precipitation, OLR, and rainfall over the Indian subcontinent are shown in Fig. 2. The mean JJAS precipitation and OLR are broadly captured by both models. The simulated OLR by SP-CFS agrees with the observations, particularly over the Pacific and Indian Oceans. SP-CFS simulates a better rainfall distribution particularly over the ISM region, which is the domain of focus in this study. The dry precipitation bias seen in CFSv2 simulations has been reduced significantly in SP-CFS simulations. However, there are some pockets of excessive rain in the simulated mean precipitation field shown by both models; perhaps, more of those for SP-CFS. These excessive wet regions in SP-CFS simulations may be attributed to the periodic boundary condition considered for the CRMs. Similar biases have been reported in SP-CCSM simulations (Benedict and Randall 2009; Goswami et al. 2013). Apart from the ISM domain, precipitation improvement by SP-CFS is also noticed over the equatorial African region as well as over some areas in the northern part of South America as compared to CFSv2 (Fig. 2a). The improvement in the mean OLR field is consistent with the improvement seen in the mean precipitation. OLR results over the Indian subcontinent and South America have improved in the SP-CFS simulations, consistent with the reduction of dry bias in mean rainfall. Significant improvement is also seen over the southern region of the eastern Pacific. Over the Sahel region SP-CFS fails to simulate the high OLR.
To evaluate the simulated rainfall distribution over the Indian subcontinent, a comparison of CFS, SP-CFS, and IMD land-only rainfall is made, concentrating on the Indian subcontinent in Figs. 2g–i. As we had mentioned earlier, the CFSv2 simulation shows a major dry bias in simulating the ISM rainfall. This dry bias is reduced in the SP-CFS simulations. The improvement in seasonal-mean rainfall in the SP-CFS has come through the increased number of stronger wet spells observed in Fig. 1. However, Fig. 2 shows a wet bias over the equatorial Indian Ocean, across the foothills of the Himalayas, and along the Myanmar coast. The reduction of the dry precipitation bias over the Indian subcontinent, particularly with respect to IMD rainfall data, is noteworthy. The reduction in the dry precipitation bias is more evident in the annual cycle plot of the climatological ISM rainfall over the central Indian monsoon region (15°–25°N, 75°–90°E; Fig. 3). While reduction of the dry bias in SP-CFS is noteworthy, central India still remains significantly dry compared to the observations. We also note that although the timing of the withdrawal of the monsoon is close to the observations in SP-CFS, the onset is delayed in the model by a couple of weeks (Fig. 3). Further work will be required to improve these remaining biases.
The CRM in the SP framework is expected to improve the cloud types and their distribution. Is the improvement of the day-to-day rainfall variability and the reduction of the mean precipitation dry bias in SP-CFS coming from improved simulations of cloud types and their distribution? Are the right kinds of clouds contributing to the right categories of rainfall? To gain more insight into these questions, instead of plotting simple probability distribution functions (PDFs), we plot the joint distribution of rainfall versus OLR, taking OLR as a proxy for cloud-top height. We counted all events with ≥1 mm day−1 rainfall and sorted them into 2-mm-wide classes (1–3, 3–5 mm day−1, etc.). The events in each class are again sorted based on their corresponding OLR values categorized in 5 W m−2 increments, starting from 100 W m−2. In Fig. 4 we have plotted the actual counts obtained following the method mentioned above for every grid point over the domain 15°S–30°N, 50°–110°E, for the five JJAS seasons. For clarity, the values shown in Fig. 4 are after dividing by a factor of 100; that is, a contour line of 1 is equivalent to 100 events.
We note from the shading information in Figs. 4a,b, over the ISM domain, lighter rain events with OLR values of approximately 270 W m−2 show the maximum count for the observations. A secondary peak exists for the lighter rain events with OLR values of approximately 220 W m−2. These peaks possibly correspond to rain coming from middle, or congestus, clouds (primary peak) and from high, or stratiform. clouds (secondary peak). The abundance of rain from congestus clouds was also reported in Morwal et al. (2015), although their study domain was much smaller and only for the monsoon withdrawal phase. As we go from light to moderate to heavy rainfall events, the twin peaks of the joint distribution gradually fades out and we see a single peak centered about OLR values of about 180–200 W m−2. Figure 4a suggests that CFSv2 simulates too many lighter-intensity rain events for OLR values of about 240–280 W m−2. We note that the bimodal nature of the distribution for light-rain events is missing in the model simulation. Also the fact that CFSv2 simulated most of the events within the OLR range of about 180–290 W m−2 indicates that the model-simulated climate fails to reproduce the spectrum of clouds seen in nature. We note from Fig. 4b that SP-CFS also seems to oversimulate the lighter-intensity rain events. However, the overestimation of the light rain events has been reduced compared to CFSv2 and also the bimodal nature of the light rain distribution is reasonably well captured by the model. In addition, the range of simulated OLR values has improved in the SP-CFS simulations, indicating a better cloud distribution.
We performed the same exercise but broadened our area of interest to the entire tropics within the latitude band 15°S–15°N (Figs. 4c,d). We note that the improvement seen in the rainfall–OLR joint distribution over the ISM domain is actually a subset of the overall improvement seen over the tropics. In fact, the improvement is even more impressive over the tropics. We argue that this improvement in the rainfall–OLR joint distribution is a very important improvement, as seen within the superparameterized framework.
Figure 5 shows the JJAS mean circulation pattern at the 850- and 200-hPa levels. Both the SP-CFS (Figs. 5a,d) and CFSv2 (Figs. 5b,e) models simulate the overall observed circulation patterns (Figs. 5c,f) fairly well. From a detailed view, the SP-CFS-simulated winds better resemble the observations compared to CFSv2, as is evident in the strength of the Southern Hemisphere easterlies and the monsoon wind over the Bay of Bengal (Figs. 5a–c). At the 850-hPa level, when simulating the strength of the monsoon low-level jet (LLJ), and the spread of the LLJ while meeting the Indian subcontinent and the extension of the LLJ to the western Pacific, SP-CFS shows better fidelity compared to CFSv2 simulations. However, the winds over the eastern equatorial Indian Ocean and north of Australia (Fig. 5a) are relatively stronger compared to the observations (Fig. 5c). The cyclonic shear in the westerlies over the Bay of Bengal is better captured in SP-CFS (Fig. 5a) as compared to CFSv2 (Fig. 5b). At 200 hPa, SP-CFS captures the extension of the easterly jet over the African continent, which is missing in the CFSv2-simulated winds (Figs. 5d–f). However, the spread of the strong easterly winds in the SP-CFS simulations is confined over a relatively smaller domain (Fig. 5d) compared to the observations (Fig. 5f).
3) Dynamical and thermodynamical constraints
In this section we analyze a few of the dynamical and thermodynamical fields that play crucial roles in determining the ISM climate. One of the major driving forces behind the maintenance of the ISM length of the rainy season (LRS) is the north–south tropospheric temperature (TT) gradient (ΔTT) over the ISM domain (Goswami and Xavier 2005; Xavier et al. 2007). The abrupt seasonal transition of the ISM is driven by the seasonal warming of the upper troposphere, as indicated by ΔTT. We define ΔTT as the difference in TT (averaged between 200 and 600 hPa) between a northern box (5°–35°N, 40°–100°E) and a southern box (15°S–5°N, 40°–100°E), following Xavier et al. (2007). The ΔTT for the observations and model simulations, for both CFSv2 and SP-CFS, are shown in Fig. 6a. We note that both models simulate the annual cycle of ΔTT fairly well compared to the observations. However, an examination of the annual cycle of the TT for the northern (Fig. 6b) and southern (Fig. 6c) boxes individually reveals that the TT results for both of the boxes are highly negatively biased by about 4 K throughout the year in CFSv2. This implies that, although the CFSv2-simulated ΔTT looks realistic, it is actually not realistic, as the gradient is over a cold background. This is consistent with the observations that while the low-level winds are reasonable (Fig. 5), the cold background does not allow convection to intensify and be sustained (Fig. 1). The SP-CFS-simulated TT, on the other hand, is correlated with the observations over the two boxes individually as well. This improvement in the simulation of the TT in SP-CFS is evident in the improvement in the rainfall annual cycle as seen in Fig. 3.
The tropospheric temperature distribution, particularly the north–south temperature gradient in the troposphere, greatly influences the global circulations (Gill 1980) as a whole. One of the major concerns about the CFSv2-simulated climate is the unrealistically cold TT over the entire globe, which did not show improvement from the earlier CFSv1 (Saha et al. 2013). Against the backdrop of the fact that SP-CFS improves the simulation of the ΔTT for the ISM, a natural extension of our research is to examine the global structure of the simulated TT (Fig. 7). We note in Fig. 7 that SP-CFS has been successful in reducing the cold bias, especially over the tropics. We computed the mean TT as the average temperature of the atmosphere between 300 and 600 hPa for JJAS. Figures 7a and 7b show the TT bias of the CFSv2 and SP-CFS simulations with respect to the NCEP–NCAR temperature field, respectively. It can be seen (Fig. 7a) clearly that CFSv2 simulates a significantly cold (~3°–7°C colder) troposphere with respect to NCEP–NCAR, over the entire globe. This temperature bias is reduced significantly in SP-CFS with a warmest bias of 1°C at a few locations in the tropics and a coldest bias of −4°C over the eastern end of Russia. Figure 7c shows the vertical profile of temperature for the tropics. It is clear that the SP-CFS-simulated temperature profile (green line) is almost identical to the NCEP–NCAR temperature profile (black dotted line). This reveals that the improvement seen in simulating the ΔTT for the ISM is a part of a larger scale improvement in the simulation of TT over the whole tropics. This in fact is very encouraging as ISM is not an isolated phenomenon. This improvement in the TT distribution may be attributed to the realistic simulation of the heat sources as a result of the explicit representation of the cloud processes by the CRMs as revealed by the intraseasonal variability and convectively coupled waves discussed below in section 3b.
After briefly diverting to the global scale to examine the simulation of TT, we bring our investigation back to the ISM domain to analyze the meridional structure (averaged between 65° and 95°E) of the easterly shear (difference between zonal winds at 200 and 850 hPa; Fig. 8a), lower-level specific humidity (averaged between the surface and 850 hPa; Fig. 8b), equivalent potential temperature (θe, averaged between 1000 and 850 hPa; Fig. 8c), and apparent heat source Q1 computed using Eq. (1) (averaged over the region 0°–30°N, 60°–100°E; Fig. 8d).
Jiang et al. (2004) proposed that mean easterly shear is one of main processes responsible for the poleward propagation of the convection anomaly during ISM. SP-CFS shows a stronger easterly shear except in the region of 5°–15°N where the SP-CFS and observations match well (Fig. 8a). The CFSv2-simulated easterly shear is stronger than the observations but not as strong as SP-CFS south of 5°N. In the region 5°–15°N, the CFSv2 simulates weaker than observed easterly shear. North of 15°N, CFSv2 and SP-CFS show equally strong easterly shear (which is stronger than the observations). The moisture–convection feedback mechanism proposed by Jiang et al. (2004) explains the northward propagation near the equator. As per this mechanism, the interaction between mean low-level moisture and the mean flow gives rise to the moisture convergence to the north of the convection. CFSv2 (Fig. 8b) shows a prominent bias in capturing the low-level moisture field, which has notably been improved in the SP-CFS. Consistent with an unrealistic moisture field (Fig. 8b), the CFSv2 simulation underestimates θe, implying weak moist instability in the lower levels (Fig. 8c). SP-CFS shows a much improved and realistic θe, implying there is improvement in moist instability. A recent study by Boos and Hurley (2013) argues that θe can be a good proxy for the thermodynamic state of the near-surface air throughout the entire monsoon domain. So an improved θe indicates that SP-CFS has a better near-surface thermodynamic state leading to a more realistic low-level monsoon circulation, which is consistent with the improved mean 850-hPa winds (Fig. 5a). As large-scale heating plays a dominant role in influencing the convectively coupled circulation, we compute the apparent heat source Q1 (Yanai et al. 1973) to examine the vertical profile of heating (Fig. 8d). Both CFS and SP-CFS (Fig. 8d) show an almost similar and underestimated heating structure compared to the observations. A marginal improvement is seen in the SP-CFS-simulated heating structure particularly in the upper troposphere (above 500 hPa) and also in the double-peaked structure of the heating profile, as seen in the observations. The similarity of the amplitude of the heating profiles simulated by the two models is thought provoking against the backdrop of the underestimated moisture and θe in CFSv2. What is responsible in the CFSv2-simulated ISM climate for generating equal amounts of heat as simulated by SP-CFS despite the fact that CFS has less moisture (Fig. 8b) and is more stable (Fig. 8c)? Why does CFSv2 simulate less ISM rainfall compared to SP-CFS despite having similar heating profiles? As mentioned by earlier studies (Goswami et al. 2013), CRMs play a key role in deciding the heating distribution of the model, and the answers to these questions may reveal more interesting insights into CFSv2 model behavior and the impacts of superparameterization on it.
b. Intraseasonal variability
To simulate a realistic mean monsoon, a model needs to reasonably simulate the internal variability. Most of this internal variability is generated by the low-frequency ISOs. It has been documented that the variability due to the ISOs lies within the period range of 10–90 days (Goswami 2012; Waliser 2006; Wang 2012). To bring out the intraseasonal variability (ISV), we bandpass filter (Duchon 1979) the rainfall anomaly for a 10–90-day window, extract the JJAS results from the filtered data, and then compute the standard deviation. We carried out the same exercise for the two model simulations and the observations. Figure 9 shows the ISV for the SP-CFS and CFSv2 simulations as well as the observations. Since rainfall follows a Poisson distribution (Goswami et al. 2011), the ISV follows a pattern similar to that of the JJAS mean rainfall. Consistent with the dry seasonal mean rainfall bias, CFSv2 shows a weak ISV over the Indian subcontinent. Unrealistically high ISV is also noted over the Arabian Sea region in the CFSv2 simulations. SP-CFS simulations realistically capture the ISV as compared to the observations. Over the Indian subcontinent, simulation of ISV is improved compared to that in the CFSv2 simulations although it is slightly weaker than observed in north-central India. The ISV simulated by SP-CFS along the west coast of India also appears to have improved (Fig. 9a). However, The ISV appears to be stronger compared to the observations over the equatorial Indian Ocean.
1) Indian summer monsoon wavenumber–frequency analysis
To examine the model’s fidelity in simulating the dominant modes of ISV during boreal summer, we compute the space–time spectra. Figures 10a,c,e show the north–south space–time spectra and Figs. 10b,d,f show the east–west space–time spectra, highlighting the dominant modes of oscillations in the daily precipitation fields from TRMM, CFSv2, and SP-CFS, respectively, computed following the methodology of Wheeler and Kiladis (1999). We have computed the signal-to-noise ratio (SNR) of the precipitation by dividing the raw power in the precipitation by an estimate of its red noise background. The red noise background is estimated by passing a 1–2–1 filter through the power repeatedly in both wavenumber and frequency, till the filter saturates. We locate the peaks in the space–time spectra (Fig. 10) using the pair (wavenumber, time period) for ease of recording in Fig. 10.
In Figs. 10a,c,e, the meridionally propagating modes over the ISM domain (between 20°S and 30°N) are shown with the SNR of precipitation being averaged between 60° and 100°E. It should be noted that for the north–south spectra, wavenumber 1 corresponds to the largest wave that fits within the 20°S–30°N latitude band. A dominant northward-propagating mode of about 45-day periodicity and wavenumbers 1 and 2 is found (Fig. 10a) in the observations (Sikka and Gadgil 1980; Yasunari 1979). Some southward propagation is also seen for 45-day period and wavenumber −2 (negative wavenumber is for southward propagation). Both CFSv2 (Fig. 10c) and SP-CFS (Fig. 10e) simulate the dominant northward-propagating mode with a time period of 60 days and underestimated power. However, the locations of the peak SNR values, which are centered around (0, 60) for CFSv2 (Fig. 10c) and at (1, 60) for SP-CFS (Fig. 10e), for the northward-propagating mode suggest that the SP-CFS spectra look slightly better. The power seen in the observations (Fig. 10a) at (−2, 45) is simulated by both of the models at a longer time period of 60 days. In Fig. 10e, the equality of power at (2, 60) and (−2, 60) suggests a stationary bias in the SP-CFS-simulated mode. Some power is seen at (−2, 30) in the SP-CFS simulations (Fig. 10e), which do not have any observational counterpart. SP-CFS simulates higher-than-observed northward power in the high-frequency regime.
In the zonal direction, the space–time spectra are computed for the entire tropics between 15°S and 15°N. From Fig. 10b we note that the observations have their highest SNRs centered on a 45-day period with wavenumbers 1–4, which is the eastward-propagating component of the summer monsoon ISO (Goswami 2012). The westward Rossby response to the eastward-moving low-frequency MISO is seen around (−5, 45–60). The power seen around (3–4, 15–20) represents the eastward-propagating Kelvin waves. Some power is also seen in the 10–20-day periodicity and wavenumbers 5 and 7 in the observations, propagating westward. There is also some power seen in the time period of less than 10 days for wavenumbers 3 and 4. In the CFSv2 simulations (Fig. 10d), the SNR peak is seen around (6, 60) and the model almost fails to capture the monsoon ISO signal. It simulates two peaks around (3, 60) and (1, 60); the first one having a slightly higher magnitude among the two. Power is seen in the wavenumbers −1 and −4 around a time period of 60 days. Notably, there is hardly any power seen in the time period of less than 20 days. This is consistent with the fact that the CFSv2-simulated day-to-day (read high frequency) variability was low in Fig. 1. Overall, Fig. 10d shows that CFSv2 simulates the eastward-propagating component of the monsoon ISO at much shorter zonal scales and fails to simulate the scale selection. Figure 10f suggests that SP-CFS-simulated zonally propagating equatorial waves during boreal summer are much improved. The major peak being around (1, 60) indicates that while the model simulates the planetary scale correctly, the period simulated is still slightly longer than observed. Also the power is overestimated and is too localized around (1, 60). Like the observations (Fig. 10b), power is also seen around (−4, 45–60) (Fig. 10f). However, the power representing the eastward-propagating Kelvin waves is underestimated. The power showing westward propagation in the 10–20-day period is missing in the SP-CFS simulations. The power in the less-than-10-day time period propagating eastward is noteworthy. To summarize, SP-CFS-simulated equatorial waves compare better with observations than those simulated by CFSv2 during boreal summer.
2) Convectively coupled equatorial waves
The improved simulation of the equatorial waves in SP-CFS simulations during boreal summer made us to examine the simulation of the CCEWs for the whole 5-yr period of simulated climate and once again we broaden our study domain from the ISM region to the entire tropics. The tropical intraseasonal variability is known to influence the ISM considerably. Nevertheless, it is always important for any GCM to realistically simulate this variability. To see the simulation of tropical intraseasonal variability, we performed a wavenumber–frequency spectrum analysis on the observed and the simulated OLRs for the tropical belt spanning 15° latitude on either side of the equator. We used 90-day segments with 60-day overlapping to calculate the spectral power. We again followed the methodology of Wheeler and Kiladis (1999) to compute the space–time (ST) spectra.
The intraseasonal variability as represented by the ST spectra for the symmetric and antisymmetric components is shown in Fig. 11. In the observations, for the symmetric component (Fig. 11a), the most prominent feature of the spectra is the large power seen for wavenumbers 1–3 and a period of approximately 45 days, which corresponds to the low-frequency summer monsoon ISO signal. The other features are the eastward-propagating Kelvin waves, the westward-propagating equatorial Rossby (ER) waves and the very high frequency (time period <3 days) inertia–gravity (IG) waves. The most prominent feature of the ST spectra of the antisymmetric component for the observations (Fig. 11b) is the power representing the mixed Rossby–gravity (MRG) waves. The CFSv2 ST spectra for the symmetric component (Fig. 11c) seem to indicate that the model has difficulty reproducing the observed intraseasonal variability. The low-frequency monsoon ISO power is much weaker and shifted toward higher wavenumbers. The Kelvin and IG waves are very feeble. The corresponding antisymmetric components (Fig. 11d) are notably absent in the CFSv2 simulation. On the other hand, the ST spectra for the SP-CFS simulations (Figs. 11e,f) look substantially improved compared to those of the CFSv2 spectra. The low-frequency monsoon ISO time period is closer to 40 days, which is slightly shorter than the observed period. However, the power still is weaker than the observations. The eastward-propagating Kelvin waves and the westward-propagating ER waves are realistically captured. The high-frequency IG waves are also well captured. The westward power seen within the range 3–6 days is stronger than in the observations. The antisymmetric component of MRG waves is also found to have improved in the SP-CFS simulation (Fig. 11f), although the amplitude appears to be still weaker than in the observations.
It is noteworthy that in the previous applications of superparameterization, significant improvement in simulating the low-frequency monsoon ISO has also been reported (Khairoutdinov et al. 2008; Benedict and Randall 2009; DeMott et al. 2011). However, it is worth noting that an SP framework does not always guarantee an improvement in the biases of the parent GCM. For example, the oceanic tropical convergence zone (TCZ) over the Indian Ocean (IO) region was clearly simulated by CAM for the monsoon months, whereas it was found to be absent in the SP-CAM (Khairoutdinov et al. 2005) simulation. Similar deficiency is also noticed in SP-CCSM simulating the oceanic TCZ over the IO region during the ISM season. Therefore, it is particularly important to evaluate the improvement in the representation of intraseasonal variability in the SP-CFS. The present analyses of SP-CFS for CCEW in conjunction with earlier studies using different models (Kim et al. 2009) confirm the importance of the explicit representation of the cloud processes in simulating the tropical waves.
The improvement seen in representations of both high- and low-frequency tropical waves in the SP-CFS with a coarse-resolution atmosphere using T62 has been instrumental in improving the daily rainfall variability over the Indian monsoon region (Fig. 1), in turn, leading to an improvement in predicting the seasonal mean monsoon (Fig. 3). Is it possible that the improvements in the representation of CCEWs could have been achieved by simply improving the horizontal resolution of the atmospheric component of the CFSv2? To investigate this issue, we analyzed the space–time spectra of the daily OLR simulated by CFSv2 with T126 and T382 resolutions in the atmosphere. The corresponding ST spectra are shown in Fig. 12. Comparing Figs. 12a,b with Fig. 11e and Figs. 12c,d with Fig. 11f, it is noted that the CFSv2 simulations do not show improvement in the equatorially trapped waves and the low-frequency monsoon ISO with the increase in resolution. Furthermore, it is noteworthy that even with the relatively coarse resolution (T62), the SP-CFS simulation shows a more realistic spectrum of tropical waves (Figs. 11e,f) than do the high-resolution (T126 and T382) CFSv2 simulations (Fig. 12). It may be noted that the high power found in wavenumber 14 in the observations (Fig. 11a) is not a signal, as mentioned by DeMott et al. (2011). This is because the peak at wavenumber 14 in the observations (Figs. 11a,b) is related to the 14 swaths per day a polar-orbiting satellite travels around the earth (Wheeler and Kiladis 1999) and does not have any meteorological significance.
As we had mentioned while discussing the improvement of the simulated tropospheric temperature in SP-CFS, similarly, the above ST-spectra analyses suggest that the improvement seen in the simulation of the equatorial waves in the ISM domain is part of the overall improvement seen over the entire tropics. The improvement in the low-frequency monsoon ISO signal is noteworthy as it is not only one of the major components of the global climate but is also a challenging one for the GCMs to simulate. This further corroborates the importance of physically based representations of cloud processes in simulations of the tropical atmosphere, and exposes the limitations of conventional convective parameterization.
c. ISM intraseasonal modes
1) Propagation features
The above sections demonstrate the improvements found in simulations of intraseasonal variability by SP-CFS for different scales and modes. What still needs to be explored is whether the propagation of the high- (quasi biweekly) and low-frequency (30–60 day) modes are realistic in SP-CFS as compared to observations. Observations show that the quasi-biweekly 10–20-day mode has a clear westward propagation whereas the low-frequency 30–60-day mode shows a northeastward movement over the ISM domain [for a review, see Goswami (2012)]. A lag–regression technique (Wilks 1995) has been used to demonstrate if the models could simulate these propagation features corresponding to the two dominant modes of MISOs. Using bandpass Lanczos filters (Duchon 1979), the rainfall anomalies corresponding to these MISO modes are isolated. A reference precipitation time series for use in regressions is created by averaging filtered anomalies over central India (18°–28°N, 73°–82°E) (Singh et al. 2014); we call this the central India precipitation index (CIPI). One index each for the two modes is created from the 10–20-day (for the quasi-biweekly mode) and 30–80-day (for the 30–60-day mode) bandpass-filtered rainfall anomalies for JJAS. We then lag regress the filtered anomalies for the entire map onto the corresponding CIPIs. The zonal propagation of the respective modes is shown by averaging the regression coefficients between 10° and 20°N, and north–south propagation is shown by averaging regressed values between 70° and 90°E.
The propagation features of the 10–20-day mode are seen in the observations (Figs. 13a,b) and are simulated by the CFSv2 (Figs. 13c,d) and SP-CFS (Figs. 13e,f). In the observations (Fig. 13a), the 10–20-day mode does not show prominent north–south propagation and shows a stationary maximum centered over 16°–20°N. In the CFSv2 simulations (Fig. 13c), two maxima are seen:one north of 16°N with a smaller spatial extent compared to the observations and the other maximum south of the equator. In fact, one more maxima center is seen south of 28°N between lag −5 and −15 days. Also, the mode shows northward propagation from the maximum in the south to the one in the north with a phase speed of 1° latitude per day. The prominent northward propagation of this 10–20-day mode over the equatorial Indian Ocean indicates improper simulation of the air–sea interaction mechanism in the model over the Indian Ocean, as pointed out by Goswami et al. (2014). In the SP-CFS simulation (Fig. 13e), the maximum is seen over 16°N. The horizontal extent of this maximum is smaller than in the observations; however, it is larger than that in the CFSv2 simulations. Between the equator and 8°N, feeble southward propagation is seen. The north–south propagation features over the Indian Ocean appear to be improved in the SP-CFS simulation. Zonally, the 10–20-day mode shows westward propagation in the observations (Fig. 13b). In CFSv2 (Fig. 13d), the westward propagation feature is reasonably captured. However, the phase speed is slower than in the observations and the amplitude is too strong between 85° and 95°E, which is consistent with the strong ISV seen along the Burma coast in Fig. 9. The SP-CFS (Fig. 13f) simulates the zonal propagation features of the 10–20-day mode reasonably well, namely, the maxima over 120°E, the phase speed, and weakening of the mode south of 100°N and peaking again over the Indian subcontinent. However, the mode looks too regular over the Indian subcontinent compared to the observations. Nevertheless, in the SP-CFS simulations, improvement is evident in capturing the propagation features of the 10–20-day MISO mode.
The propagation features of the 30–60-day mode are seen in the observations (Figs. 14a,b) and simulated by CFSv2 (Figs. 14c,d) and SP-CFS (Figs. 14e,f). In the observations (Figs. 14a,b), the 30–60-day mode shows clear northeastward propagation. CFSv2 captures the northward propagation feature reasonably well but tends to stop sharply at about 20°N. In addition, the mode looks too strong, which is possibly an artifact of the low standard deviation value of the CFSv2-simulated CIPI, and as was the case with the 10–20-day mode (Fig. 14c), suggesting improper simulation of the atmosphere and ocean dynamics over the Indian Ocean. This raises again the issue of dry precipitation bias over the Indian subcontinent despite CFSv2 simulating the northward propagation well, as pointed out by Goswami et al. (2014). They attributed this ISM rainfall dry bias to improper simulation of air–sea interaction over the Indian Ocean. CFSv2 (Fig. 14d) totally fails to capture the eastward propagation of the 30–60-day mode; instead, it simulates a westward propagation. In other words, as suggested by Figs. 13d and 14d, CFSv2 simulates westward propagation over the ISM domain for the whole ISO time period. SP-CFS simulates the northward propagation of the 30–60-day mode (Fig. 14e) reasonably well. Comparing Figs. 14c and 14e, the improvement over the Indian Ocean is noticeable. The southward propagation south of the equator is better simulated in SP-CFS. However, it is overemphasized in SP-CFS and makes the mode appear almost stationary. This is consistent with the equality of the southward and northward power around 60 days in the SP-CFS north–south space–time spectra (Fig. 10e). The eastward propagation (Fig. 14f) looks improved, particularly over the Indian longitude band (70°–100°E). Improvement in capturing the east–west propagation feature is consistent with the analyses presented in Figs. 10 and 11. This is an important improvement indicating a better low-frequency MISO simulation in the SP-CFS.
2) Circulation features
To examine the circulation features, we plotted the active composite of the circulation anomalies for the two simulated modes. To do that, first we isolated the wind anomalies using 10–20- and 30–80-day Lanczos bandpass filters. Then, we defined active monsoon conditions based on CIPI defined in the previous section for the two modes. We have considered the ISM as active when the CIPI value is above 1.2 for at least three consecutive days. The middle date of the active monsoon period has been considered as the peak active date. Then, we compute a composite of the wind anomalies for those active dates for the two MISO modes. The corresponding vorticity is shaded in the background for each circulation pattern.
The active composite of the circulation pattern for the 10–20-day mode is shown in Fig. 15 that includes two additional panels in the middle showing the meridional structure (averaged between 60° and 90°E) of the vorticity magnitude. Comparing the 850-hPa circulation pattern for the observations (Fig. 15a), CFSv2 (Fig. 15c), and SP-CFS (Fig. 15e), we see that the cyclonic circulation over central India is reasonably well simulated by both models. The circulation over the central and south equatorial Indian Ocean (Fig. 15a) is missing in CFSv2. SP-CFS captures this feature but with a smaller spatial extent and slightly stronger vorticity. Over the Bay of Bengal, the CFSv2 simulates unrealistic circulation with strong unrealistic vorticity. The SP-CFS circulation pattern looks improved over the Bay of Bengal region; however, it is still biased. Over the Arabian Sea, both models simulate the observed circulation pattern reasonably well. However, the strength of the SP-CFS-simulated vorticity appears to be closer to the observations. Over the region bounded by 10°–30°N and 100°–120°E, alternate centers of opposite vorticities are seen in the observations (Fig. 15a). CFSv2 fails to simulate this realistically. SP-CFS captures the observed pattern over this region although with stronger vorticity. It is seen that the overall vorticity structure and magnitude has improved in the SP-CFS simulation at the 850-hPa level. This is supported by the meridional structure of the magnitude of the vorticity shown in Fig. 15i. The SP-CFS vorticity (green line in Fig. 15i) agrees well with the observations (black line in Fig. 15i), except south of 5°S. On the other hand, CFSv2-simulated vorticity (red line in Fig. 15i) looks completely out of phase south of 10°N, if not from farther north. At the 200-hPa level, the positions of the observed major vorticity maximum and minimum (Fig. 15b) (viz., the cyclonic vorticity centered over 30°N, 60°E; the cyclonic–anticyclonic vorticity couplet over the Indian subcontinent and adjoining Indian Ocean) are better simulated in SP-CFS (Fig. 15f) compared to CFSv2 (Fig. 15d). Also, the bias in the magnitude of the simulated vorticity is reduced in SP-CFS simulations, as evident in Fig. 15ii. However, the agreement with the observations at 200 hPa (Fig. 15ii), for SP-CFS, is not as good as at 850 hPa (Fig. 15i). The CFSv2-simulated meridional structure of the vorticity is clearly unrealistic at 200 hPa. The cyclonic vorticity over central India at both 850 and 200 hPa (Figs. 15a,b) suggests a barotropic vertical structure of the observed 10–20-day MISO mode. The barotropic character of the 10–20-day MISO mode has also been documented in Goswami et al. (2011). A notable feature of the SP-CFS simulations is the realistic simulation of this barotropicity. As circulation is driven by heating, improvement in the vertical structure of the 10–20-day MISO mode is a direct consequence of a better heating structure for which the CRMs are responsible within the superparameterized framework. It is noteworthy that Goswami et al. (2011, 2013) attributed the failure to simulate the MISO modes realistically to unrealistic simulation of the vertical structure of heating in their attempt to understand the simulation of MISO modes in superparameterized GCMs.
The active composite of the circulation pattern for the 30–60-day mode is shown in Fig. 16. The major circulation feature seen in the observations at 850 hPa (Fig. 16a) is the cross-equatorial southwesterly flow pattern generating a cyclonic vorticity over the Indian monsoon zone. Both CFSv2 (Fig. 16c) and SP-CFS (Fig. 16e) capture this southwesterly flow pattern reasonably well. The CFSv2-simulated vorticity is stronger than the observations over central India, which is evident in Fig. 16i, and is unrealistic south of 10°N, particularly over 60°–100°E. SP-CFS improves the simulation of the magnitude and structure of the vorticity north of the equator. The improvement in vorticity magnitude in SP-CFS is clear in Fig. 16i. However, the large-scale organized spatial structure of the positive vorticity from north of the Arabian Sea to the South China Sea across central India (Fig. 16a) is not realistically captured by the SP-CFS. East of 100°E for the domain shown in Fig. 16, both models show limited ability to capture the circulation and vorticity realistically. At 200 hPa, the most prominent feature observed (Fig. 16b) is the anticyclonic vorticity anomaly over an area to the east of the Tibetan Plateau and the cyclonic vorticity anomaly approximately in the region 10°–15°N, 45°–90°E. As we have noticed in previous CFSv2-simulated circulation patterns, CFSv2 again captures the overall structure but with much stronger magnitude for the vorticity (Fig. 16d). In SP-CFS (Fig. 16f), the model performs poorly in simulating the circulation pattern. The vorticity maximum over the Tibetan Plateau is biased toward a southward position in SP-CFS with unrealistic organization. Although the bias seen in the magnitude of vorticity in the CFSv2 simulation is marginally improved in SP-CFS, the poor large-scale organization neutralizes the impact of this improvement. Figure 16ii depicts that both models capture the baroclinic structure of the circulation over central India.
A close examination of Figs. 15 and 16 shows the major difference between the CFSv2 and SP-CFS simulations has been the improvement in the magnitude of the vorticity in SP-CFS, in the observed major vorticity centers. However, the large-scale organization is relatively poor with many of the smaller-scale vortices compared to the host GCM CFSv2. Whenever SP-CFS captures the large-scale organization reasonably well (e.g., in Figs. 15e and 16e), it shows improvement over CFSv2. A detailed study of the cloud-scale variables in SP-CFS should provide a much-needed better understanding of this model behavior, which is a major goal of our future research.
d. Ratio of the synoptic to ISO variance
Seeing the noticeable improvement in the tropical intraseasonal variability, it will be worth evaluating the synoptic-scale and ISO-scale contributions in the SP-CFS simulation compared to the CFSv2 to address the precipitation biases. To explain the dry precipitation bias in the CFSv2 simulations, Goswami et al. (2014) used a metric to examine the share of the synoptic and the ISO variance in the total monsoon variability. Although their study was focused on the simulation of the ISM, the variance analysis brought to light a clearer picture of the deficiency in the CFSv2 simulations: the CFSv2 model systematically underestimates the synoptic variance relative to the variance of the low-frequency oscillatory modes over the entire globe and even more so over the tropics, which is consistent with the weaker power of the equatorial convectively coupled waves (Figs. 10 and 11). A similar analysis has been carried out for the SP-CFS simulation. It can be seen in Figs. 17a,b that the synoptic variance accounts for 60%–80% of the total daily variance during the summer monsoon season, whereas ISOs can explain 30%–40%, over most regions in the ±30° latitude zone. Unlike the observations, CFSv2 underestimates the share of the synoptic variance (Fig. 17d) and overestimates the ISO variance (Fig. 17e). On the other hand, the SP-CFS simulations look much more realistic compared to the CFSv2 simulations (Figs. 17g,h). Although there are some minor random biases in the simulation of the relative variances on the synoptic and ISO scales, systematic underestimation (overestimation) of synoptic (ISO) variance is not evident in the SP-CFS simulation. The realistic simulation of the relative share of the synoptic and ISO variance is one of the major achievements for SP-CFS and its effect is visible in Fig. 1, leading to a reduced dry bias over the central Indian region.
We present an analysis of a 5-yr free run of the newly developed SP-CFS with the atmosphere at T62 horizontal resolution. This is the first time the NCEP CFSv2 model has been run without conventional cumulus parameterization and with a superparameterization of cloud processes. We start our investigation by analyzing the simulation of the Indian summer monsoon by the SP-CFS. We document that some of the known major biases of CFSv2, like the cold tropospheric temperature bias and underestimation of the synoptic variance, have been reduced significantly in the SP-CFS over the entire tropics. We also demonstrate that SP-CFS better simulates the equatorial convectively coupled waves. Better simulation of convectively coupled equatorial waves within an SP framework has also been reported for SP-CCSM (DeMott et al. 2011). As mentioned in DeMott et al. (2011), it is possible that the CRMs within the SP framework precondition the moisture field (Waite and Khouider 2010), thereby creating favorable conditions for deeper convection to follow. Khouider and Majda (2008) have demonstrated that these sequences of evolution of convection can explain the convectively coupled equatorial waves. This explains the success of SP-CFS over the CFS in simulating the equatorial waves. The CFSv2, which uses the simplified Arakawa–Schubert convection scheme, allows detrainment only through the top of the cloud and hence lacks moisture in the middle and upper troposphere (Pattnaik et al. 2013), which fails to simulate the moisture preconditioning. A better simulation of the equatorial convectively coupled waves and associated variability by the SP-CFS, in turn, has been manifested in the capturing space–time structure and the propagation features of different monsoon modes. The improvement of intraseasonal variability of convection and rainfall is reflected in the improvement of the mean seasonal rainfall, the joint distribution of rainfall and OLR, and the annual cycle of rainfall. Consistent with these improvements, the dry bias of the ISM has also been reduced along with the length of the rainy season and the day-to-day rainfall variability over central India. We have also shown that the improvement in the SP-CFS-simulated mean ISM rainfall is actually a result of systematic improvements to the lower-level moisture, north–south temperature gradient, and moist instability.
The SP-CFS thus shows much more promise in improving the intraseasonal variability of the global model (e.g., CFS) and definitely demands further research. There are still some notable biases that appear to be arising from the CRM. Therefore, tuning the CRM microphysics by choosing the appropriate microphysical constants may be required. For example, an ice-to-snow conversion threshold would be just one of the many choices available to possibly reduce the biases in the SP-CFS simulations. We also intend to eliminate the restriction of the periodic boundary conditions on CRMs and make the CRMs sense conditions in the adjacent GCM grids, hoping to further improve the SP-CFS simulations. However, the computational cost will have to be kept in mind. Although the SP-CFS is still a newer model, it has already demonstrated the importance of resolving the subgrid-scale cloud processes in simulating the global climate and the ISM in particular. A further analysis of SP-CFS simulations with the above-mentioned tuning in the CRM would give us a more robust understanding of the biases and possible ways of mitigating them.
The Indian Institute of Tropical Meteorology (Pune, India) is fully funded by the Ministry of Earth Sciences, Government of India, New Delhi. This work is part of the Ph.D. thesis of the first author carried out at the Indian Institute of Tropical Meteorology. We thank the National Centers for Environmental Prediction for the reanalysis data used in this paper. M. Khairoutdinov was supported by the NSF Science and Technology Center for Multiscale Modeling of Atmospheric Processes (CMMAP), managed by Colorado State University under Cooperative Agreement ATM-0425247. We especially thank Dr. S. Moorti (NCEP) for his help with the CFSv2 model. We also express our sincere thanks to Prof W. Grabowski (NCAR) for his valuable comments in orienting the results. We gratefully acknowledge Dr. A. K. Srivastava of Indian Meteorological Department, Pune, India, for helping with IMD data. The authors gratefully acknowledge and thank the reviewers for their constructive comments and suggestions, which helped to improve the paper.