Numerical experiments are conducted to investigate the differences between summer precipitation over continental East Asia simulated by the Community Atmosphere Model, version 5 (CAM5), and superparameterized CAM5 (SPCAM5, a multiscale modeling framework). The results show that SPCAM5 effectively alleviates several original biases. Overestimates of precipitation on the eastern periphery of the Tibetan Plateau are reduced from CAM5 to SPCAM5 as a result of decreases in both the average hourly precipitation frequency and mean hourly intensity. Underestimates along the coastal regions in southern China are improved following a corresponding increase in mean hourly intensity and a decrease in average hourly precipitation frequency. The frequency–intesnsity relationship is also more realistic in SPCAM5. For western China, overestimated frequency values (in CAM5) of both weak-to-moderate (0–20 mm day−1) and heavy (20–50 mm day−1) intensity ranges are reduced in SPCAM5. For southern China, overestimates of frequency values (in CAM5) in the weak-to-moderate range are also reduced, whereas underestimates in the intense ranges are enhanced. In terms of diurnal variability, SPCAM5 generally exhibits a delay in the afternoon peak time and greater diurnal amplitude. The possible physical reasons for the variations in the precipitation between the models are further investigated. It is found that the change in deep convection intensity is a primary factor governing the shift in the precipitation simulations. SPCAM5 better simulates an intermediate transition stage from shallow to deep convection, which helps the deep convection to grow more fully to a greater magnitude, thus delaying the peak time and increasing the precipitation maxima.
Precipitation is a result of processes associated with water vapor condensation, latent heat release, and cloud occurrence, which fundamentally influence the water balance and radiative forcing. The simulation of precipitation in atmospheric general circulation models (AGCMs) is a major metric to assess model performance (Randall et al. 1991; Trenberth et al. 2003; Dai and Trenberth 2004). However, state-of-the-art climate models have long been unable to satisfactorily reproduce the observed precipitation features (e.g., Dai 2006; Mehran et al. 2014).
A particularly important reason for this inadequate precipitation performance lies in the representation of cloud processes, which influence precipitation via the hydrological process. The representation of clouds in AGCMs is a long-standing problem (Randall et al. 2003; Randall 2013) because of the intrinsic multiscale nature of clouds, which are connected not only with large-scale circulations (e.g., Zhang et al. 1996; Bony et al. 1997; Cess et al. 2001; Yu et al. 2004; Zhang et al. 2014a) that an AGCM is able to resolve but also with mesoscale organizations (e.g., Donner 1993; Mapes and Neale 2011) and small-scale turbulent and convective motions (e.g., Xu and Randall 2001; Stevens 2002; Grabowski et al. 2006) that an AGCM has to parameterize.
For precipitating cloud systems, one of the core efforts of parameterization is to represent the cumulus activity in the atmosphere and its interactions with large-scale circulations (Riehl and Malkus 1958; Yanai et al. 1973; Arakawa and Schubert 1974; Emanuel 1991). Cumulus convection plays a central role in most of the interactions between the dynamical and hydrological processes, the radiative and dynamical–hydrological processes, and the atmosphere and oceans (Arakawa 2004). Nevertheless, the conventional cumulus parameterization approach introduces large uncertainties. Different components in the cumulus parameterization can yield quite different results from model to model. For example, Xie and Zhang (2000) showed that the deficiencies in the convection-triggering function are responsible for the large thermal biases found in the single-column simulations of the National Center for Atmospheric Research (NCAR) Community Climate Model, version 3 (CCM3). They revealed the model’s sensitivity by using a new triggering function based on the observations. Del Genio and Wu (2010) examined several entrainment-rate parameterization schemes based on the results from a cloud-resolving model (CRM) and showed that only one scheme well reproduced the entrainment profiles from the explicit simulation. In addition, GCMs with conventional moist convection parameterizations usually fail to reproduce the right timing of afternoon convective precipitation peaks over land (Yang and Slingo 2001; Bechtold et al. 2004; Dai 2006). GCMs often simulate peaks that are earlier compared with the observations. Studies (e.g., Guichard et al. 2004; Chaboureau et al. 2004; Wang et al. 2007) have shown that these early peaks are related to a lack of the transition from shallow to deep convection because of the crude triggering criteria and unconstrained entrainment rates. In contrast, explicit simulations of cumulus convection by cloud-resolving or large-eddy models usually indicate a gradual moistening of the free troposphere and an increase in the cloud-top height, which are usually absent in models with highly parameterized cumulus physics.
As an alternative approach to modeling the subgrid-scale convective activity, the superparameterized (SP) GCM, otherwise known as the multiscale modeling framework (MMF), has been proposed to better simulate the multiscale nature of clouds. This approach, which was first proposed by Grabowski and Smolarkiewicz (1999) and Grabowski (2001) and was referred to as cloud resolving convection parameterization (CRCP), replaces the cumulus and stratiform cloud parameterizations with explicit simulations that are provided by an embedded two-dimensional (2D) CRM in each grid column of a large-scale host model. Khairoutdinov and Randall (2001) extended this approach further by integrating a 2D version of a three-dimensional (3D) CRM (Khairoutdinov and Kogan 1999; Khairoutdinov and Randall 2003) into a realistic GCM, the NCAR CCM3. This constituted the first prototype of the SP Community Atmosphere Model (CAM), which has been widely used and modified since its development (e.g., Khairoutdinov et al. 2005, 2008; DeMott et al. 2010; Marchand et al. 2009; Wang et al. 2011; DeMott et al. 2011; Xu and Cheng 2013; more references are available online at http://www.cmmap.org/research/pubs-mmf.html). Although MMF also has its own problems (e.g., it tends to underestimate marine stratocumulus clouds because the embedded CRM is still too coarse to explicitly resolve the large turbulent eddies), it is generally considered to improve the simulations associated with deep convection (e.g., the diurnal cycle of rainfall and tropical variability like the Madden–Julian oscillation; Khairoutdinov et al. 2005). The superparameterization has been considered as a parallel to conventional parameterization in the future development of model physics (Randall et al. 2003; Arakawa et al. 2011; Randall 2013).
On the other hand, summer precipitation over East Asia has long been difficult to simulate well, largely because of the influences of complex orographic features, land–sea distributions [see Fig. 1 for an overview of the complex surface features across East Asia, including the vast Tibetan Plateau in the west, the Sichuan basin (27°–32°N, 103°–108°E) on its lee side, the eastern plain, and numerous hills in southern China], and monsoon systems (e.g., Yu et al. 2000; Zhou and Li 2002; Chen and Frauenfeld 2014). In addition, the simulation is especially sensitive to the choice of convection schemes (e.g., Chen et al. 2010b). East Asian precipitation systems comprise phenomena at various spatial and temporal scales (e.g., Tao et al. 2003; Chen et al. 2010a; Luo et al. 2014; Chen et al. 2014). The rainbands associated with the East Asian summer monsoon have significant variability at different time scales [see Ding and Chan (2005) for a review]. Moreover, summer precipitation over continental China exhibits significant diurnal variability (Yu et al. 2007), and the frequency and intensity patterns of precipitation also vary (Zhou et al. 2008). This variety of features makes the East Asian summer precipitation regime an ideal test bed for assessing the model performance and investigating the sensitivity of different types of parameterizations.
Current state-of-the-art AGCMs have considerable biases in the simulation of precipitation over continental East Asia. Figure 2 plots the bias of summer [June–August (JJA)] precipitation rates averaged from 23 models participating in phase 5 of the Coupled Model Intercomparison Project (CMIP5) against the Tropical Rainfall Measuring Mission (TRMM) data (years 1998–2005). The models tend to overestimate the precipitation at the southern and eastern edges of the Tibetan Plateau while underestimating the precipitation in southern China. Such common and stubborn model biases severely hamper the model performance and motivate us to study them in relation to model sensitivity. Because the MMF can explicitly simulate features associated with deep convection on various spatial and temporal scales, improvements should be seen in simulated summer precipitation over continental East Asia. Nevertheless, few studies available in the literature have assessed the performance of MMF in its simulations of East Asian precipitation, constituting a general motivation of the present study.
In this paper, numerical experiments are conducted to investigate the differences between CAM5 and a superparameterized CAM5 (SPCAM5). The climatological mean state, frequency–intensity structures, diurnal variations, and reasons governing the changes in precipitation are all assessed in detail. This study will likely benefit both the observational and modeling communities. It can help us understand how we can benefit from the use of SP-type GCM in the simulations of East Asian summer precipitation. It will also improve our understanding of how to better simulate the precipitation characteristics over East Asia in the context of a global model.
The remainder of this paper is organized as follows. Section 2 describes the model, data, and methods used in this study. Section 3 presents an overview of the climatological mean state. Section 4 investigates the differences between simulated frequency–intensity structures. Section 5 documents the simulations of diurnal variations. Section 6 further explores the causes of precipitation variations between the models. Section 7 provides a summary and discussion of the results.
2. Models, data, and method
a. Models and experimental setup
The CAM5.2 model and its SP version are used in this study. As stated in the introduction, SPCAM5 is derived from the model used in Khairoutdinov and Randall (2001), Khairoutdinov et al. (2005), and Khairoutdinov et al. (2008). Wang et al. (2011) further upgraded the SP to the framework of Community Earth System Model (CESM) with CAM5 to study the aerosol effect on climate. The model code (version spcam2_0-cesm1_1_1) can be downloaded from the NCAR code repository online and was run locally at the National Meteorological Information Center (NMIC) of the China Meteorological Administration (CMA).
CAM5 is a component of CESM developed at NCAR with many external collaborators. The model uses a default finite-volume (FV) dynamical core (Lin 2004) with a hybrid pressure–sigma vertical coordinate (Simmons and Burridge 1981) that has 30 levels with a top at 2.255 hPa. Almost all of the parameterizations in CAM5 have been updated from the CAM4 version, except the deep convection scheme (Zhang and McFarlane 1995; Richter and Rasch 2008; Neale et al. 2008). CAM5 features (i) a new shallow convection scheme and a new moist turbulence scheme (Bretherton and Park 2009) developed by the University of Washington (Park and Bretherton 2009), (ii) a two-moment cloud microphysics scheme (Morrison and Gettelman 2008; Gettelman et al. 2008) and a cloud macrophysics scheme (Park et al. 2014) for the parameterizations of clouds, and (iii) the open-source Rapid Radiative Transfer Model for GCMs (RRTMG) package developed by Atmospheric and Environmental Research (AER) is used as the radiation module (Mlawer et al. 1997; Iacono et al. 2008). A new modal aerosol module is also implemented in CAM5 (Liu et al. 2012).
When SP is active, it replaces the convective and stratiform cloud parameterizations, as well as the boundary layer scheme. Hence, unlike CAM5, SPCAM5 no longer produces large-scale and convective precipitation separately, which are generated by grid- and subgrid-scale moist physics (stratiform cloud microphysics and moist convection), respectively. In other words, the precipitation is thought to be fully resolved in this framework.
The large-scale host model has a resolution of 1.9° × 2.5° and 30 vertical levels. The model’s physics time step is 30 min; it is internally split for the dynamical core to achieve computational stability. The embedded 2D CRM has a resolution of 4 km in the west–east direction and 28 vertical levels, and the integration time step is 20 s. Khairoutdinov et al. (2008) noted that the horizontal direction of the CRM (east–west or south–north) appeared to alter the climatology of precipitation in the western Pacific for the summer months. Nevertheless, the sensitivity to the orientation of the 2D CRM is beyond the scope of the present study and may require a further study. At 4-km resolution, processes associated with deep convection can be resolved, but the processes associated with shallow cumulus or stratocumulus clouds can only be crudely represented (Weisman et al. 1997; Petch et al. 2002). The cloud microphysics scheme used by the CRM is a two-moment scheme that was proposed by Morrison et al. (2005).
Because of the heavy computational burden of running the SPCAM5, 5-yr integrations were performed for the two models with a prescribed sea surface temperature dataset from 2000 to 2004 [i.e., the Atmospheric Model Intercomparison Project (AMIP)-type simulation]. Although a 5-yr integration may be relatively short for climate simulations, we believe that it is suitable to check the differences between the models because our focus is on those fast atmospheric processes. Actually, most model systematic differences associated with moist processes are apparent in the ensembles of short-term runs (e.g., Xie et al. 2012; Wan et al. 2014; Zhang et al. 2015). A 5-yr climatology is indeed different from a longer (e.g., 30 yr) climatology, but their differences are much smaller than differences between two selected models in this study. We have checked the differences between a 5-yr precipitation climatology and a 30-yr one generated by CAM5 (datasets used in a previous study; Zhang et al. 2014b) and confirmed this statement. In addition to the default monthly mean output at the global scale, the models are sampled at a 1-h interval at a regional domain (5°–60°N, 75°–135°E).
b. Data and method
Two precipitation datasets are used in this study. A high-resolution (1-hourly, 0.1° × 0.1°) gauge–satellite merged precipitation dataset (years 2008–12 because this dataset begins only from 2008) is used, which combines the Climate Prediction Center morphing technique (CMORPH) dataset with hourly gauged rainfall data from approximately 30 000 automatic weather stations (Pan et al. 2012). This dataset is referred to as OBSCMO in this paper. The gauged rainfall data from stations are processed following strict quality control procedures. The CMORPH data are used as background values, which are further modified according to the results from the weather stations. This dataset provides a better quality than the original CMORPH dataset. The TRMM 3B42 (Huffman et al. 2007) dataset (years 2000–04) is also used in this study. It combines fine-scale (3-hourly and 0.25° resolutions) precipitation estimates from multiple satellites with gauge analyses where feasible. Because the horizontal resolution of models is approximately 2°, both datasets are uniformly binned to a 2° × 2° grid to facilitate the analysis. Although the time periods of the two datasets are different, they generally show comparable climatological results and can be used to evaluate the model performance.
The hourly precipitation frequency and mean hourly intensity are calculated to reveal details related to the simulated precipitation (section 3b). For each interval in the diurnal cycle (twenty-four 1-h or eight 3-h intervals), the occurrence of precipitation within a given grid box is determined when the hourly or 3-hourly precipitation rate (also referred to as amount hereafter; the unit is scaled to millimeters per day throughout this paper) is greater than 0.5 mm day−1. Hence, the hourly precipitation frequency at each hour of the day is defined as a ratio between the total number of precipitating events occurring at that hour of the day and the total number of days (e.g., 460 = 5 yr × 92 days). The mean hourly intensity at a given hour is the cumulative hourly precipitation rate divided by the number of total precipitating times. The formulations are as follows:
where HFRi and MHIi are the hourly precipitation frequency and mean hourly intensity at hour i of the day, respectively. The variable Ni is the number of total precipitating times, and Pi is the cumulative hourly precipitation rate. The climatological mean precipitation rate at a given hour should be the product of hourly precipitation frequency and mean hourly intensity, if we ignore the 0.5 mm day−1 difference in the statistics.
To reveal the frequency–intensity structure of precipitation (section 4), the precipitation frequency is calculated (binned) against the hourly or 3-hourly rainfall rate at a 1 mm day−1 interval beginning from 0.5 mm day−1. For a given grid box and interval of the actual precipitation rate (say 0.5–1.5 mm day−1), the number of precipitating times (within this interval) are first counted and are then divided by the number of entire hourly or 3-hourly times in statistics (11 040 or 3680, respectively). The relation curve is then averaged for each region of interest.
It should be mentioned that although all four datasets can be scaled to the same unit (mm day−1), they actually represent different characteristics. For hourly datasets (OBSCMO, CAM5, and SPCAM5), the precipitation rate at time n means the average precipitation amount during time n − 1 and n. For 3-hourly datasets (TRMM), the precipitation rate at time means the average precipitation amount during time n − 1.5 and n + 1.5. Thus, the 3-hourly data will smooth the hourly variations when compared with the hourly datasets. For hourly datasets, we have created their 3-hourly counterparts (mimicking the approach of TRMM) and repeated the frequency and intensity calculations conducted in this study (section 3b and section 4). The conclusions of this paper still apply even with the smoothing. Nevertheless, because the focus of this study is the model difference and the TRMM dataset just provides auxiliary help to better constrain the observation, we still present the original results directly obtained from the hourly datasets. The smoothed results are not shown in the paper. For interested readers, the results can be acquired via e-mail.
With reference to the diurnal cycle of precipitation, the diurnal amplitude A is computed as follows:
where Pmax is the hourly or 3-hourly precipitation maximum and is the daily mean value. Thus, the diurnal precipitation variation is normalized as
The variable N(h) is the normalized precipitation value, and P(h) is the hourly or 3-hourly precipitation rate.
The apparent heat source Q1 and apparent moisture sink Q2 are calculated to quantify the subgrid-scale processes involved in temperature and moisture budgets, respectively. Following Yanai et al. (1973), they are calculated using the residual method as follows:
where cp denotes the specific heat constant of dry air, T denotes the temperature, q denotes the specific humidity, ω denotes the pressure vertical speed, p denotes the pressure, R denotes the air constant of dry air, ∂/∂t denotes the time derivation, V ⋅ ∇ denotes the horizontal advection, ∂/∂p denotes the vertical advection, and L denotes the latent heat of evaporation or condensation. In this paper, the vertical transects of Q1 and Q2 are scaled using cp to achieve a unit of kelvin per day. The vertically integrated Q1 and Q2 are not scaled by cp, and the unit is watts per square meter. The sign convention in Eqs. (5) and (6) implies that positive values of Q1denote heating and positive values of Q2 denote drying.
3. Climatological mean states
a. Precipitation rate
A global view (limited between 50°S and 50°N because of TRMM data) of the simulated mean summer (JJA) precipitation rate is presented before we address East Asia. In general, the two models studied here (Figs. 3b,c) simulate similar rainbands across the globe as compared with TRMM (Fig. 3a). The spatial correlation coefficient with TRMM slightly increases from CAM5 (0.81) to SPCAM5 (0.84). By examining the differences between the models (Fig. 3d), differences over several major regions become clear, including (i) central Africa, where SPCAM5 produces more precipitation amount (Fig. 3d) and mitigates the original negative errors between CAM5 and the TRMM data (Fig. 3e); (ii) the northwestern Indian Ocean, where SPCAM5 generally produces less precipitation (Fig. 3d) and reduces the original positive errors between CAM5 and the TRMM data (Figs. 3e,f); (iii) Indo-China and the northwestern Pacific Ocean, where the large negative biases between CAM5 and the TRMM data (Fig. 3e) are replaced with large positive errors (Fig. 3f) because SPCAM5 simulates much more precipitation in these regions; (iv) the central and eastern tropical Pacific Ocean, where the original positive errors over a broad area (Fig. 3e) are reduced to a narrow band in general but where some regions are dominated by larger positive errors; and (v) East Asia, which is investigated in detail below.
Figures 4a,b present the mean summer (JJA) precipitation rates over East Asia, derived from the OBSCMO and TRMM, respectively. The two datasets generally show similar spatial patterns in the mean precipitation rates. The major continental rainbands are located to the south of 35°N, especially on the southern edge of the Tibetan Plateau, over the coastal regions in southern China, and along the Yangtze River valley (the white line around 30°N over eastern China in Fig. 1). The precipitation maxima over these regions are usually associated with several major regional ambient features, including the orographic forcing (Tibetan Plateau), southwesterly winds (southern China), and the subtropical mei-yu, baiu, or changma front (Yangtze River). In addition to their general consistency, the two datasets have some differences. Overall, the OBSCMO precipitation values to the south of 30°N are smaller than those in the TRMM, especially on the southern edge of the Tibetan Plateau. The OBSCMO precipitation maxima over southern China are also slightly weaker than those in the TRMM. Nevertheless, the general picture over East Asia can be easily grasped and confirmed with the aid of the two datasets, which helps us further investigate the model performance.
The results of the CAM5 simulation are shown in Fig. 4c. The precipitation maxima in CAM5 extend from the southern edge to the eastern periphery (west of the Sichuan basin) of the Tibetan Plateau, as well as farther toward northern China, identifying a southwest–northeast rainband. The major deficiency lies in that CAM5 severely overestimates the precipitation at the southern edge and on the eastern periphery of the Tibetan Plateau. In the coastal regions of southern China, the model simulates a dry region, with overall values lower than 5 mm day−1. The problems apparent in the CAM5 simulation of East Asian precipitation are quite similar to those apparent in the CMIP5 model ensemble (Fig. 2). As shown in Fig. 4d, which describes the differences between CAM5 and TRMM, large positive errors are observed along the regions where the model rainband is located. To the south of this rainband, rainfall values in CAM5 are much weaker. The model generally produces a northern-wet–southern-dry pattern.
The results from the SPCAM5 simulation are shown in Fig. 4e, and the difference between the two models is presented in Fig. 4f. These figures illustrate that SPCAM5 effectively alleviates several original biases. SPCAM5 simulates the rainband in southern China well and produces much more precipitation along the northeastern side of the Bay of Bengal, correcting the original dry biases in these regions. Meanwhile, SPCAM5 significantly reduces the precipitation values identified on the eastern periphery of the Tibetan Plateau, where the original artificial rainfall center in CAM5 is located (Fig. 4b). Nevertheless, SPCAM5 did not improve the simulations on the southern edge of the Tibetan Plateau.
b. Average hourly precipitation frequency and mean hourly intensity
The climatological mean of precipitation amount can be further decomposed into two metrics: (i) average hourly precipitation frequency that measures the percentage of hourly precipitation occurring over time and (ii) mean hourly intensity that measures the mean hourly precipitation amount when precipitation occurs. As shown by Eqs. (1) and (2) in section 2b, the hourly precipitation frequency and mean hourly intensity are first calculated for each individual hour. Thus, the results shown here take a daily average of the hourly or 3-hourly climatology.
Many studies have shown that frequency and intensity simulations are important tools that reveal model biases (Chen et al. 1996; Dai and Trenberth 2004; Dai 2006; Li et al. 2015) because different combinations of frequency and intensity could lead to similar climatologies of precipitation. A reasonable simulation of precipitation relies on a correct combination of both frequency and intensity. An overall view of the results (Fig. 5) suggests that CAM5 suffers from the common “low-intensity, high-frequency” problem as many other numerical weather and climate models (e.g., Chen et al. 1996; Osborn and Hulme 1998; Sun et al. 2006; Dai 2006). Although our precipitation rate samples are at hourly or 3-hourly intervals, while daily precipitation data were mostly used in those earlier studies (e.g., Osborn and Hulme 1998; Sun et al. 2006; Dai 2006), this problem generally persists despite the different time scales.
In terms of average hourly precipitation frequency, the two datasets generally present similar images (Figs. 5a,b). The regions with the most frequent precipitation are located to the south of the Tibetan Plateau. On the lee side of the Tibetan Plateau (to the east of 103°E), the frequency values are generally lower than 50%. Overall, CAM5 is found to overestimate the frequency values downstream of the plateau, while SPCAM5 reduces that error. A comparison of the two models’ simulations identifies three regions where these differences are emphasized: a western box (27°–35°N, 103°–108°E) covering the artificial rainfall center on the eastern periphery of the Tibetan Plateau in western China, a southern box (23°–27°N, 108°–118°E) covering the coastal regions in southern China, and an eastern box (28°–35°N, 110°–120°E) covering the Yangtze River valley in eastern China. The precipitation frequency values are reduced in all three regions from CAM5 to SPCAM5, thus offsetting the original positive biases.
With regard to the mean hourly intensity, the two observational datasets (Figs. 3e,f) show similar distributions, but OBSCMO has relatively smaller values than TRMM at the major rainbands, especially on the southern edge of the Tibetan Plateau. This results in the lower total precipitation amount in the OBSCMO dataset (Fig. 2). However, the differences between the models and the observational data are much larger than the discrepancies between the different datasets. As shown in Fig. 5g, CAM5 not only underestimates the magnitude of rainfall intensity over the regions where observed rainbands are located (eastern and southern China) but also produces two artificial intensity maxima centers at the southern edge and eastern periphery of the Tibetan Plateau, respectively. Mean hourly intensity values over southern and eastern China are mostly enhanced from CAM5 to SPCAM5, which brings the values closer to the observed values. In contrast with the general increases, the intensity values in the western box are reduced from CAM5 to SPCAM5, again offsetting the original positive bias. Generally, the simulated mean hourly intensity in SPCAM5 is more realistic than that in CAM5. The remaining problem lies at the southern edge of the Tibetan Plateau, where the mean hourly intensity values are enhanced, thus further departing from the observed values. Meanwhile, the intensity values over eastern and northern China are also greater than that in the observation. The analysis above demonstrates that the changes in simulated precipitation rates are ascribed to contributions from both average hourly precipitation frequency and mean hourly intensity.
4. Frequency–intensity relationship
To examine the changes in average hourly precipitation frequency and mean hourly intensity in more detail, this section will focus on the relationship between frequency and intensity. This will help us better understand the changes of frequency at different precipitation categories (from light drizzle to extreme precipitation events). The western and southern boxes are chosen here because variations of precipitation amount in these two regimes are more evident (Fig. 4f) and are mainly governed by the continental convection processes (as will be shown in section 6).
As outlined in section 2b, precipitation frequency is calculated (binned) by the actual hourly or 3-hourly precipitation rate with a 1 mm day−1 interval beginning from 0.5 mm day−1. To better show the results in a single graph, the relationship curve is first plotted on a logarithmic frequency and intensity axis (see e.g., Fig. 6a). Next, the actual intensity axis is separated into four ranges to better indicate the results within different precipitation categories (see e.g., Figs. 6b–e). The ranges are defined as follows: (i) first category, with weak-to-moderate precipitation (0–20 mm day−1); (ii) second category, with heavy precipitation (20–50 mm day−1); (iii) third category, with heavy-to-extreme precipitation (50–100 mm day−1); and (iv) fourth category, with very extreme precipitation (greater than 100 mm day−1).
The western box is analyzed first. The simulations for this region are characterized by decreases in both average hourly precipitation frequency and mean hourly intensity (Fig. 5) from CAM5 to SPCAM5. An overall view in Fig. 6a shows that CAM5 tends to severely overestimate the hourly frequency values across the range of intensity, while SPCAM5 generally reduces values.
In the weak-to-moderate range (Fig. 6b), the two datasets show similar variations with the increase of intensity. The frequency value in OBSCMO (TRMM) begins at approximately 11% (9%) with around 1 mm day−1 intensity, then gradually decreases with an increase in intensity, and finally ends at approximately 0.45% with 20 mm day−1 intensity. A key difference between the datasets is that OBSCMO has higher frequency values between 1 and 2 mm day−1; that is, more light-rain amounts are reported in the OBSCMO dataset. In CAM5, above the intensity value of 3 mm day−1, precipitation frequency values are greater than those in two observational datasets, which results in precipitation frequency overestimates in this range. Moving from CAM5 to SPCAM5, the frequency values at all intensity ranges are reduced. Below the intensity value of 6 mm day−1, SPCAM5 even simulates a slightly lower frequency than the observed datasets; in contrast, above this threshold, the frequency values in SPCAM5 are much closer to the observed values and lower than the values reported by CAM5.
In the heavy rainfall category (Fig. 6c), CAM5 still overestimates frequency values, whereas SPCAM5 restricts the values to a lower range. The averaged frequency values for the four data sources within this range are 0.32% (OBSCMO), 0.35% (TRMM), 0.39% (SPCAM5), and 0.45% (CAM5). In the heavy-to-extreme (Fig. 6d) and very extreme (Fig. 6e) categories, SPCAM5 still produces lower frequency values than CAM5. The results in Fig. 6 suggest that in the western region, the reduction in average hourly precipitation frequency and mean hourly intensity from CAM5 to SPCAM5 is caused by decreased frequency in both the weak-to-moderate and intense (including heavy and extreme) precipitation intensity categories.
The southern box is characterized by decreases in average hourly precipitation frequency but increases in mean hourly intensity (Fig. 5) from CAM5 to SPCAM5. As shown in Fig. 7a, OBSCMO, TRMM, and SPCAM5 all show lower frequency values in the first category but greater values in the other three ranges. The change in the frequency–intensity relationship in the weak-to-moderate category is quite different from those in the other three categories. In the weak-to-moderate class (Fig. 7b), SPCAM5 largely alleviates the high frequency values reported in CAM5 for all intensity intervals and brings the simulated values closer to the observed results. The averaged frequency values in this range are 1.9% for OBSCMO, 1.6% for TRMM, 2.0% for SPCAM5, and 3.1% for CAM5.
Although SPCAM5 reduces the weak-to-moderate precipitation frequency as it did in the western box, it simulates increased frequencies in the other three categories. For the heavy category (Fig. 7c) and the heavy-to-extreme category (Fig. 7d), OBSCMO, TRMM, and SPCAM5 all have greater frequency values than CAM5. Moreover, above the intensity value of approximately 75 mm day−1 in the third category and in the entire fourth category, CAM5 simulates very little precipitation, thus revealing the difficulty in simulating extreme precipitation values. In contrast, SPCAM5 is able to reproduce the precipitation in these ranges. The results in Fig. 7 suggest that compared with CAM5, SPCAM5 restrains the weak-to-moderate precipitation frequency estimates but enhances the more intense precipitation frequency estimates. Meanwhile, SPCAM5 is also able to simulate those extreme values more clearly, although these constitute only a minor portion in the samples (e.g., the third category). As a result, SPCAM5 simulates greater precipitation amount and higher mean hourly intensity but lower average hourly precipitation frequency in southern China.
5. Diurnal variations
This section investigates the diurnal cycle of precipitation. The diurnal timing of precipitation is important because the associated clouds strongly interact with both shortwave and longwave radiation to modulate the energy balance. Therefore, the diurnal cycle is not only observationally important to understand the nature of precipitation but also critical as a basic metric to assess the simulated precipitation.
Continental diurnal variation is tightly coupled with solar heating in the surface and atmospheric boundary layers and is thus stronger in summertime. Observational evidence has shown that the diurnal maxima of continental deep convection and associated precipitation occur frequently in the late afternoon or early evening (Dai et al. 1999; Dai 2001; Yang and Slingo 2001; Nesbitt and Zipser 2003). Specifically for East Asia, previous studies (e.g., Yu et al. 2007; Yuan et al. 2012b) have reported observed spatial features for diurnal peak of precipitation (also shown in Figs. 8a,b). Along a zonal band averaged within 28°–35°N from the Tibetan Plateau to its lee side, four distinct regimes with different diurnal variations are documented, including late-afternoon and midnight peaks over the Tibetan Plateau, midnight to early-morning peaks in western China, double peaks in the late afternoon and early morning in eastern China, and early-morning peaks over the East China Sea. In addition, southern (to the south of 27°N) and northeastern China (40°–50°N, 110°–130°E) have afternoon precipitation peaks.
Because of the relatively coarse horizontal resolution used in this study, the models are not expected to reproduce the detailed and fine regional features that might be seen from higher-resolution models (e.g., Sato et al. 2009; Dirmeyer et al. 2012; Yuan et al. 2013), where the topography is better resolved. However, the models should be able to simulate some representative and significant features, including prominent continental afternoon precipitation peaks, which are usually a proxy for isolated deep convection that is driven from the surface, and contrasts between the plateau and the plain and the land and sea, which usually reflect the roles played by the heterogeneous underlying surface.
Figure 8 shows the diurnal peak time for mean summer precipitation over East Asia. The observed features have been described above. The two observational datasets generally reveal similar patterns in the diurnal peaks. The major differences are primarily located over the Tibetan Plateau, where rain gauges capture the diurnal variation in the valleys and satellites capture variation over the mountains [see discussions in Chen et al. (2012) and Yuan et al. (2012a)]; the eastern plain (28°–35°N, 110°–120°E), where the TRMM dataset tends to miss the early-morning peak; and southwestern China, where the nocturnal precipitation peaks are underestimated by TRMM. Smoothing the hourly OBSCMO dataset to a 3-hourly dataset generally does not change these differences.
Comparing the two model simulations and the observations reveals several distinct differences in diurnal variation. The notable regions are identified by blue boxes. Over the Tibetan Plateau (Reg1), both models are able to simulate the afternoon and midnight precipitation peaks. However, for both models, the areas with midnight peaks are smaller than those in the observations. Meanwhile, in CAM5, the afternoon peaks occur mostly during 1300–1500 local solar time (LST) and 1500–1700 LST, which is earlier than the observed late-afternoon peaks (1500–1700 and 1700–1900 LST). SPCAM5 delays the simulated peak times over the plateau, with most values occurring within 1500–1700 LST and some values falling between 1700 and 1900 LST. A similar delay in the afternoon peak simulations can also be found along the coastal regions in southern China (Reg2) and in northeastern China (Reg3), where peak values of 1300–1500 LST simulated by CAM5 are mostly delayed to 1500–1700 LST in SPCAM5. Because the afternoon precipitation peaks are usually associated with deep convection, these results may suggest that SPCAM5 better represents the deep convection processes. Meanwhile, over the eastern plain (Reg4), CAM5 simulates peak times of approximately 0900–1100 and 1300–1500 LST, unlike the double peaks in the late afternoon and early morning reported in the observational datasets (Fig. 8a). SPCAM5 partially simulates the early-morning peaks in some locations but still fails to adequately adjust for the afternoon peaks to a proper timing (still earlier).
The diurnal amplitude is another important metric used to examine diurnal variability. Figure 9 shows the spatial distributions of the normalized amplitude values (scaled by daily mean values). Large diurnal amplitudes are located along the coastal regions in southern, eastern, and northwestern China, as well as over the Tibetan Plateau (Figs. 9a,b). Generally, both models are able to reproduce the spatial distributions of diurnal amplitude. Compared with CAM5, SPCAM5 tends to simulate greater magnitudes, especially over the Tibetan Plateau and the coastal regions in southern and eastern China. Meanwhile, it can be found that the diurnal amplitude minima coincide with the location of each model’s primary rainbands (e.g., the southern edge and eastern periphery of the Tibetan Plateau). This phenomenon is less apparent in the observational dataset. This is because models are found to simulate heavy precipitation throughout the diurnal cycle (figure not shown); therefore, the diurnal variation is relatively flat. This feature contributes to the “rainband” that forms in the climatological mean state.
To better illustrate these differences, a zonal band averaged within 28°–35°N is selected to describe the diurnal–longitudinal variations of the normalized precipitation rate (Fig. 10). The two observational datasets present similar features, except the TRMM has stronger absolute values at both the peak and trough. Furthermore, the propagation speed of the precipitation signal (positive values) over time is slower in OBSCMO than in TRMM. Figure 10 also highlights an eastward-delayed diurnal phase between 100° and 110°E (marked by black lines). In CAM5, the absolute values at the peak and trough are relatively weaker than the observed values, and the eastward-delayed phase is not evident. In SPCAM5, the absolute values at both the peak and trough become larger than those in CAM5, especially over the Tibetan Plateau. Additionally, SPCAM5 more evidently simulates the eastward-delayed diurnal phase (the black lines). The results in this section suggest that, overall, SPCAM5 better simulates the diurnal variability of precipitation over continental East Asia. In the next section, the changes from CAM5 to SPCAM5 are further explored.
6. Sources of changes
This section investigates the factors that govern the changes in precipitation simulations from CAM5 to SPCAM5. The first notable change from CAM5 to SPCAM5, as presented in the previous sections, is that SPCAM5 not only reduces excessive precipitation amounts on the eastern periphery of the Tibetan Plateau but also enhances precipitation amounts in southern China. Both changes enable SPCAM5 to simulate a more realistic precipitation climatology. An analysis of the precipitation types using the CAM5 results shows that the subgrid convective precipitation constitutes a dominant portion in both regions. Although the fully resolved precipitation in SPCAM5 does not allow for a direct comparison of precipitation associated with convective processes, some other fields associated with convective processes should exhibit evident differences.
Figures 11a,b compare the convective available potential energy (CAPE) derived from the two models. Over continental East Asia, SPCAM5 simulates much stronger CAPE values than CAM5 does. For instance, along the coastal regions in southern China, CAM5 simulates a low CAPE region, corresponding to the area with underestimated precipitation amounts (Fig. 4). In SPCAM5, this region is characterized by larger CAPE values and enhancement of precipitation amounts (Fig. 4f). Nevertheless, an evident exception is the eastern periphery of the Tibetan Plateau, where CAM5 simulates a prominent maximum CAPE center that corresponds with the artificial rainfall center. This change is in tune with the decreased precipitation amount from CAM5 to SPCAM5.
The values of Q1 and Q2 are further examined to quantify the subgrid-scale processes involved in temperature and moisture budgets. The value of Q1 consists of the heating due to radiation, the release of latent heat by net condensation, and the vertical convergence of the vertical eddy transport of sensible heat. The value of Q2 represents a moisture sink caused by the net condensation and vertical divergence of the vertical eddy transport of moisture (Yanai et al. 1973). As shown in the results (Figs. 11c–f), SPCAM5 reduces the magnitudes of Q1 and Q2 on the eastern periphery of the Tibetan Plateau but largely enhances those magnitudes along the coastal regions in southern China. The vertical structures of Q1 and Q2 are further compared in Fig. 12, which shows a vertical transect along 105°E, crossing the artificial rainfall center on the eastern periphery of the Tibetan Plateau. In CAM5 (Figs. 12a,c), a strong heating and condensation center is located between 25° and 40°N, extending from the surface to 200–300 hPa. SPCAM5 shows lower Q1 and Q2 values in this region. The suppressed convective processes inhibit excessive precipitation on the eastern periphery of the Tibetan Plateau.
Other notable changes are revealed by the diurnal variability metric. For regions where afternoon precipitation peaks dominate, SPCAM5 simulates later peak times and increased diurnal amplitudes. The coastal region in southern China serves as an example of such cases. Over this region, the delayed precipitation peak time allows the deep convection to grow more fully. As a result, the SPCAM5 model simulates lower hourly rainfall frequency but greater mean hourly intensity (Fig. 5), thereby producing more realistic frequency–intensity structures along with the enhancement of total precipitation amounts.
Figure 13 highlights several physical factors in a selected region (23°–27°N, 112°–118°E) over southern China. For precipitation (Fig. 13a), the peak time changes from about 1400 (CAM5) to about 1600 LST (SPCAM5), and the precipitation amounts increase. SPCAM5 also produces a corresponding increase in CAPE magnitudes over CAM5 (Fig. 13b). A key difference is that in CAM5, the precipitation peak and CAPE peak occur at almost the same time (~1400 LST), while in SPCAM5, the precipitation peak (1500–1600 LST) occurs 1–2 h later than the CAPE peak (1300–1400 LST). This is because the deep convection parameterization in CAM5 directly relates the strength of convection to CAPE (Arakawa and Schubert 1974; Zhang and McFarlane 1995). However, the results from numerous cloud-resolving simulations and some field observations (e.g., Chaboureau et al. 2004; Guichard et al. 2004; Khairoutdinov and Randall 2006; Kuang and Bretherton 2006; Zhang and Klein 2010; Del Genio and Wu 2010) usually indicate that a transition from shallow to deep convection exists under the presence of substantial CAPE, and the precipitation gradually increases toward its maxima with a gradual moistening of the free troposphere and an increase in cloud-top height.
To illustrate our idea that SPCAM5 more successfully simulates the continuous transition stage from shallow to deep convection, the budget fields are further compared for the two models. Although the budget fields shown are averaged for the entire summer months and not specifically composited for rainy cases, the results generally reveal the distinctive differences between the models.
Figures 13c,d compare the heating rates (the radiative heating is removed, and it is referred to as Q1 − Qrad), and the differences (SPCAM5 minus CAM5) are shown in Fig. 13e. Both two models simulate the boundary layer heating during 0800–1400 LST. However, in CAM5, the maximum heating in the upper troposphere occurs at around 1400 LST, weaker in magnitude and earlier in phase than that in SPCAM5. The positive differences between the models gradually develop from the lower to upper troposphere over time (Fig. 13e).
Figures 13f–h compare the moisture sink field. Both two models simulate the lower-tropospheric moistening during 0800–1500 LST. However, in SPCAM5, a more obvious drying (net condensation) signal occurs below the upper-level moistening during 1100–1700 LST, when the precipitation amount in SPCAM5 increases from approximately 10 to approximately 12 mm day−1. The drying signal further extends to upper levels over time. In CAM5, the surface drying signal is weaker, and the extension to upper levels is blocked by surface moistening (1700 LST in Fig. 13f), which suggests that the convective process is relatively more discrete in CAM5 than in SPCAM5.
As suggested in previous CRM studies (e.g., Bechtold et al. 2004), drying in the subcloud layer can be considered a crude proxy for shallow cumuli. That is, SPCAM5 better simulates the intermediate continuous transition stage, thereby allowing the convection to develop more fully, delaying the precipitation peak time and increasing the diurnal amplitude. The more abundant shallow cumuli in SPCAM5 can also be identified from the cloud condensate field (Figs. 13i–k), especially between 1100–1700 LST, when precipitation gradually increases. In terms of large-scale vertical motions (Figs. 13l–n), SPCAM5 shows much stronger rising motions during the precipitating period, which indicates that the atmosphere in SPCAM5 is more convective. This correlates with the more abundant unstable energy and heating magnitude generated during the diurnal progression.
7. Summary and discussion
This study compares the CAM5 and SPCAM5 simulations of summer precipitation over continental East Asia. An analysis of precipitation changes based on the climatological mean state, frequency–intensity relationship and diurnal variability is presented. Possible physical explanations for the precipitation changes are further explored. The major conclusions are summarized below.
In terms of the climatological mean state, SPCAM5 enhances precipitation amounts along the coastal regions in southern China and reduces overestimates of precipitation on the eastern periphery of the Tibetan Plateau. These two changes improve the original dry and wet biases generated by CAM5. On the eastern periphery of the Tibetan Plateau, the artificial overestimates of precipitation are reduced by SPCAM5 as a result of the reductions in both average hourly precipitation frequency and mean hourly intensity. Conversely, in southern China, the precipitation amounts are enhanced by a corresponding increase in mean hourly intensity and a decrease in average hourly precipitation frequency from CAM5 to SPCAM5.
In terms of the frequency–intensity relationship, SPCAM5 reproduces more realistic frequency–intensity structures in both western and southern China. For the western region, the frequency values in the weak-to-moderate and heavy precipitation intensity categories are reduced from CAM5 to SPCAM5 and align more closely with the observational datasets. For the southern region, the frequency values in the weak-to-moderate category are reduced in SPCAM5, but the frequency values in the intense precipitation category are increased, thus offsetting the original biases generated in CAM5. SPCAM5 can also reproduce the occurrences of some precipitation extremes that are missed by CAM5.
In terms of the diurnal variability, SPCAM5 generally simulates a later afternoon precipitation peak time than CAM5 and increased diurnal amplitude, especially over the Tibetan Plateau and the coastal regions in southern China. Meanwhile, an eastward-delayed diurnal phase between 100° and 108°E also becomes more evident in SPCAM5.
The reasons for the precipitation changes from CAM5 to SPCAM5 are investigated in relation to unstable energy and budget fields. SPCAM5 generally produces more unstable energy than CAM5, except on the eastern periphery of the Tibetan Plateau. The changes in mean precipitation amounts on the map generally correspond to changes in the CAPE and Q1 and Q2 fields, which reflects the fact that the change in deep convection intensity might be a primary reason for the changes in precipitation simulations.
As a representative region in which isolated deep convection frequently occurs, an area in southern China is further selected to examine differences in simulated diurnal variation. Unlike in CAM5, where the diurnal evolution of precipitation almost coincides with the development of CAPE, the simulated peak time in SPCAM5 occurs approximately 1–2 h later than the peak time of CAPE. The positive anomalies in subgrid-scale heating rate between SPCAM5 and CAM5 gradually develop from the lower to upper troposphere over time. The difference in moisture sink field suggests that a shallow cumuli regime becomes more evident in SPCAM5. The results reflect that a continuous and intermediate transition stage from shallow to deep convection becomes more evident in SPCAM5, helping the deep convection to grow more fully to a higher magnitude, which results in a delay of the precipitation peak time and an increase in the precipitation maxima.
b. Discussion and concluding remarks
This study describes several major differences between two models’ simulations of precipitation over continental East Asia. From the authors’ perspective, the root of these differences is found in the MMF feature of the SPCAM5 model. MMFs (e.g., SPCAM5; Tao et al. 2009) are a bridge between conventional GCMs (e.g., CAM5) and global cloud-resolving models (GCRMs; e.g., Satoh et al. 2008). Because GCRMs are still too expensive to use regularly in climate simulations, MMFs are an ideal tool for studying the multiscale modeling of the atmosphere from the climate aspects. Compared with a conventional GCM, an MMF is not only a global model but also a process model, and it enables the parameterizations of microphysics, turbulence, and radiation to directly operate on the CRM’s grid. The improvements revealed in this study are likely associated with changes in how clouds and precipitation are treated between the models.
Generally, an embedded CRM (although 2D) is better than a parameterized column model in simulating local-scale process rates associated with clouds and precipitation (e.g., Xu et al. 2002; Guichard et al. 2004), and it is usually used as a benchmark (e.g., Bechtold et al. 2000; Raymond 2007) to evaluate the parameterization. When CRMs are embedded in the GCM’s grid box and forced by large-scale dynamics, the resultant models have two dynamical systems and can include the scales of atmospheric motion on various orders. The results in this study, particularly those for southern China, are a typical representative of daytime convection processes over the continent. The changes from different aspects (e.g., climatological mean, frequency–intensity structure, and diurnal cycle) are interrelated. The changes in the diurnal cycle of precipitation, which are characterized by a delayed phase time, an increase in the precipitation maximum, and a gradual and continuous transition stage from shallow to deep convection, directly contribute to decreased average hourly precipitation frequency, increased mean hourly intensity, and enhanced total precipitation amounts, as seen in the climatological mean state.
It is well known that most climate models usually precipitate too frequently at weak intensity. The models tend to overestimate the frequency of light rain while underestimating the frequency of heavy precipitation over land. Using daily precipitation data, DeMott et al. (2007) discussed that SPCAM shows advantages over CAM in simulating heavy rainfall rates. Via correlation analysis, they showed that in the CAM, there is little or no lag between boundary layer energy buildup and rainfall, whereas the SPCAM successfully simulates the observed increase (decrease) in boundary layer height prior to (following) a rain event. This conclusion is generally in accordance with our finding that the SPCAM5 model better simulates a gradual transition stage from shallow to deep convection as revealed from the budget fields.
In nature, the atmospheric instability is usually accumulated before the intense convection starts. This is why most continental precipitation (especially that driven from the surface) usually occurs in the late afternoon. However, in CAM5, the convection and associated precipitation over land is more like an instantaneous response to the thermal forcing (Fig. 13a). This is because the convection scheme in CAM5 instantly removes CAPE once CAPE is generated. Therefore, the convection process is relatively more discrete with a premature onset (e.g., the earlier afternoon peak time with smaller amount), which is a common bias in models with parameterized physics (e.g., Guichard et al. 2004). Therefore, the convection and precipitation could be frequently activated but with relatively weak magnitude, leading to the “low-intensity, high-frequency” problem in the sense of climatology.
By contrast, SPCAM5 shows a delay in the timing of intense convection, accompanied by a more evident surface drying, which indicates the occurrence of more abundant shallow cumuli. The behavior of SPCAM5 is more similar to those derived from observations or cloud-resolving simulations. Yanai et al. (1973) concluded that shallower, nonprecipitating cumulus clouds support the growth of deep, precipitating cumulus towers. Studies from explicit simulations all show that convection develops from dry, shallow to deep cumuli. The existence of shallow cumulus clouds inhibits the quick dissipation of the unstable energy (CAPE) and maintains a gradual development of convection. The convection can grow to a more intense magnitude. Thus, precipitation occurs less frequently but with increased precipitation amount. Therefore, the fundamentally different diurnal mode of convection changes the frequency and intensity structure.
A reasonable simulation of the diurnal cycle depends on simulating a succession of regimes, from dry to moist nonprecipitating to precipitating convection (Chaboureau et al. 2004; Guichard et al. 2004). For models with parameterized physics, Grabowski et al. (2006) suggested that a possible route to improve the early onset of deep convection is to use a time-evolving cumulus entrainment rate as convection evolves from shallow to deep. Del Genio and Wu (2010) also suggested that the inferred entrainment rate from a CRM weakens with increasing time of day as convection deepens. Further, Grabowski et al. (2006) also noticed that although 2D CRMs generally reproduce similar results to 3D CRMs, they tend to simulate too rapid a transition from shallow to deep convection. This might partly explain why SPCAM5 still simulates an early-afternoon peak time compared with the observational datasets.
The comparison between CAM5 and SPCAM5 presented here provides new evidence that can help improve the precipitation simulations over continental East Asia in the context of global models. Considering that the global atmosphere model is gradually moving toward a unified formulation of large-scale GCMs and local-scale CRMs (e.g., Arakawa and Konor 2009; Arakawa et al. 2011; Arakawa and Wu 2013), comparing the results between conventional GCMs and MMFs will shed light on the scientific merits of multiscale atmospheric modeling. More experiments and analyses will be conducted to understand the multiscale dynamical and physical processes associated with the simulations of cloud and precipitation over East Asia.
This research was supported by the National Natural Science Foundation of China (Grants 41505066, 41375004, and 41221064) and the Basic Scientific Research and Operation Foundation of Chinese Academy Meteorological Sciences (Grants 2015Z002, 2015Y005, and 2014R013). The authors are grateful for the comments from three anonymous reviewers and the editor, which helped improve the original manuscript.