The single-column model (SCM) functionality of the Energy Exascale Earth System Model version 1 (E3SMv1) is described in this paper. The E3SM SCM was adopted from the SCM used in the Community Atmosphere Model (CAM) but has evolved significantly since then. We describe changes made to the aerosol specification in the SCM, idealizations, and developments made so that the SCM uses the same dynamical core as the full general circulation model (GCM) component. Based on these changes, we describe and demonstrate the seamless capability to “replay” a GCM column using the SCM. We give an overview of the E3SM case library and briefly describe which cases may serve as useful proxies for replicating and investigate some long-standing biases in the full GCM runs while demonstrating that the E3SM SCM is an efficient tool for both model development and evaluation.
Despite advances in computation allowing for general circulation models (GCMs) to be run with progressively finer resolution with each successive generation, the parameterized physics in the atmospheric components of these GCMs have steadily become more complex. Indeed, while this increase in complexity often leads to better climate simulations due to more realistic and comprehensive processes being accounted for, understanding interactions between these parameterizations and the GCM dynamics can be a daunting task. A tool to help GCM physics development and evaluation is the so-called single-column model (SCM) framework, which is a functionality that exists in many state-of-the-art GCMs. This work was pioneered by Betts and Miller (1986), with the link between SCMs, observations, and GCMs studied more extensively in Randall et al. (1996). An SCM is a mode where a single column of the atmosphere is run in isolation with prescribed atmospheric dynamics. Thus, the SCM will simulate unresolved processes within the atmospheric column, such as clouds, microphysics, turbulence, and radiation while removing the complexity of the dynamics–physics interactions.
SCMs are often the first step in the GCM parameterization development and/or implementation process. This is due to the fact that SCMs can provide a framework for quicker and easier debugging compared to the full GCM counterpart. In addition, depending on the regime of interest being targeted, the SCM simulation can readily be compared against observations or large eddy simulation (LES). This allows for rapid feedback of the parameterization performance in a more process-oriented environment. Park (2014) and Bogenschutz et al. (2012) are examples of how an SCM is used to implement and evaluate new and complex families of parameterizations in the National Center for Atmospheric Research's (NCAR's) Community Atmosphere Model version 5 (CAM5; Neale et al., 2012) SCM. The SCM can also be used as a tool to explore configurations that may not be feasible to do in a full GCM. For example, Bogenschutz et al. (2012) explored CAM, with two different physics packages, with very high LES-like vertical resolution that would have been computationally burdensome to do with a full GCM run. Smalley et al. (2019) use the SCM to construct a novel modeling framework that is forced by reanalysis to simulate a variety of environmental conditions in the subtropics to evaluate their parameterization suite.
SCMs are also useful tools for examining GCM physical parameterization performance at the process level. For instance, Zheng et al. (2017) used CAM's SCM to diagnose the cause of a cloudy planetary boundary layer oscillation, which was found to be the result of coupling issues between the turbulence and microphysics schemes. Zhang and Bretherton (2008) performed SCM studies to show that cloud feedbacks in CAM3 were controlled by unphysical oscillations caused by interactions between convection and resolved-scale processes. The SCM also provides a useful tool to perform perturbed parameter sensitivity studies for complex parameterizations that contain an abundance of tunable parameters, such as those performed by Guo et al. (2015).
The Energy Exascale Earth System Model version 1 (E3SMv1; Golaz et al., 2019) is an Earth system model designed with funding by the Department of Energy (DOE) for research and applications relevant to its mission. While E3SMv1 contains three new components (ocean, sea ice, and river) that have not previously been coupled to an Earth system model, the atmosphere and land components were branched from the Community Earth System Model version 1 (CESM1; Hurrell et al., 2013) but have evolved since (Xie et al., 2018; Rasch et al., 2019). Therefore, E3SMv1 inherited CAM and its associated SCM (Gettelman et al., 2019), but its SCM has also evolved. Those changes will be documented in this paper.
While SCMs have demonstrated that they are valuable tools for parameterization development, testing, and process-oriented evaluation efforts, they are unable to elucidate remote impacts or clarify physical–dynamical interactions where three-dimensional transport effects come into play. In addition, the SCM may replicate the behavior of the full GCM better for certain regimes and conditions than others, though this issue is poorly understood and not well studied. In this paper we will preliminarily demonstrate under what conditions the E3SM-SCM can serve as a useful proxy for GCM performance.
This paper is organized as follows: in Sect.
The E3SM Atmosphere Model version 1 (EAMv1; Xie et al., 2018; Rasch et al., 2019) was originally branched off from NCAR's CAM. Therefore E3SM inherited the CAM SCM (SCAM; Gettelman et al., 2019) and several similarities exist between the two.
Similar to SCAM, surface fluxes in E3SM can be prescribed and is the default setting if this information is available in the case forcing file. Otherwise, the surface fluxes are computed interactively via the land model or the data ocean model, using prescribed sea surface temperatures. The E3SM SCM does not currently support running an interactive ocean model, such as the work presented in Hartung et al. (2018), which may be a useful framework towards understanding parameterization feedbacks and climate sensitivity.
Section
The E3SM uses a prognostic modal aerosol model (Liu et al., 2016). While this represents a sophisticated and modern treatment of aerosol in E3SM, it presents challenges for an SCM and has been noted in Lebassi-Habtezion and Caldwell (2015) (hereafter LHC2015). Chief among these is the fact that E3SM initializes all aerosol mass mixing ratios to zero and results in unrealistically low aerosol concentrations until surface emissions loft sufficient aerosol. This is a process that can take several days to spin up (Schubert et al., 1979) and can significantly impact the simulated results of several SCM cases that are only hours in duration. LHC2015 show that CAM5 simulations are very sensitive to the initialization of aerosol for stratiform boundary layer cloud cases but not for shallow and deep convective cases (because deep and shallow convective microphysics schemes were not tied directly to the aerosol scheme). However, E3SM uses a unified treatment of shallow convection and planetary boundary layer turbulence (Golaz et al., 2002; Bogenschutz et al., 2013). Therefore, the shallow convective clouds are tied to the large-scale microphysics scheme and could lead to more severe impacts and sensitivities for the shallow and deep convective cloud regimes if aerosol is not specified adequately.
We have implemented the three options proposed by LHC2015 to initialize aerosol in the E3SM SCM. The first option is to use prescribed aerosol climatology derived from a 10-year E3SM present-day simulation with climatologically prescribed sea surface temperatures (SSTs). The second option is to specify the droplet and ice concentrations in the microphysics, thus bypassing the aerosol–cloud interaction, and the third option is to use observed aerosol information from the intensive observation period (IOP) forcing file if it is available. Selecting an aerosol specification option is mandatory for the E3SM SCM and a runtime error will result if a user attempts to run the SCM with no aerosol specification. For the scripts provided in the E3SM SCM case library (see Sect.
Many published LES comparison studies involving the simulation of boundary layer clouds include “idealizations”. As an example, the goal of the LES intercomparison study of the Barbados Ocean and Meteorological experiment (BOMEX; Holland and Rasmusson, 1973) was to investigate the role of turbulence dynamics for the shallow cumulus boundary layer (Siebesma et al., 2003) while avoiding the complications of microphysics and radiation. As such, none of the LESs participating in the study included a microphysical parameterization in their simulation. In addition, the radiative heating tendencies for the LES comparison were included in the large-scale forcing. Should an E3SM SCM user wish to evaluate the turbulence and cloud structure of the BOMEX case against the LES intercomparison study of Siebesma et al. (2003), not only would an apples-to-apples comparison not be possible with an out-of-the-box configuration of the inherited SCM, but it would also be scientifically invalid due to the fact that radiative tendencies would be double counted. While implementing these idealization switches into the model code is rather trivial, it is not an obvious task for the typical SCM user who is not familiar with the code and who may not be aware of the idealizations needed to match LES results.
Therefore, with the goal of preventing improper case setups, we have implemented idealization switches into the E3SM SCM code to allow for apples-to-apples comparison with IOP forcings corresponding to the appropriate reference for the particular case (see Sect.
The code required to run the CAM SCM has long been entangled with the Eulerian dynamical core. As a result, the SCM could not be run with CAM's current Finite Volume (FV; Lin and Rood 1997; Neale et al., 2012) dynamical core or E3SM's spectral element (SE; Dennis et al., 2012) dynamical core. This is a problem because while the horizontal advection fields are provided by the IOP forcing files, the dynamical core is still responsible for computing the large-scale vertical transport if this is not prescribed in the forcing file. In the E3SM/CAM SCM the only part of the dynamical core that is exercised is the computation of the large-scale vertical transport. The Eulerian and SE dynamical cores use two different methods for this computation: the former uses a simple Eulerian calculation, while the latter uses a semi-Lagrangian method. Therefore, the inherited SCM was inconsistent with the full GCM with regard to how the large-scale vertical advection was computed. In addition, there are differences in the numerics between the two dynamical cores; whereas the Eulerian core uses a leapfrog numerical scheme, the SE dynamical core uses a third-order five-stage explicit Runge–Kutta (RK) method as described in Dennis et al. (2012). This results in different coupling between the prescribed and computed dynamical forcing with the physics and results in different dynamics and physics time steps between the SCM and the GCM run, which could cause inconsistencies between the two configurations if a particular parameterization scheme is sensitive to time step.
Ideally, we want the SCM to be as close a proxy to the full GCM run as possible. Thus we upgraded the SCM dycore to use the same SE dynamical core used by E3SM. Even though horizontal advection is prescribed in the SCM, the dycore still plays an important role for vertical advection. As described in Dennis et al. (2012), the SE dycore operates on quadrilateral elements whose Gauss–Lobatto–Legendre (GLL) quadrature points form the physics columns targeted by the SCM. Because there are many physics columns within each spectral element, it is impossible to initialize a single physics column when running a SE-dycore SCM. In this context the simplest way to initialize the SCM dycore is to “trick” the model by initializing the dynamical core using a low-resolution global configuration but then only actually use a single physics column for our calculations. We do this using the lowest-resolution configuration supported by E3SM, which contains 96 elements and corresponds to a grid spacing of approximately 7.5
The E3SM SCM library is currently comprised of 25 cases that range from widely used cases of idealized boundary layer cloud regimes of a few hours in duration to unique cases that span the duration of years to a decade (i.e., continuous forcing from Atmospheric Radiation Measurement (ARM) Southern Great Plain (SGP) site; Xie et al., 2004). The list of available forcing files and their references can be found in Tables
E3SM case library (part 1).
E3SM case library (part 2).
The E3SM SCM case library is publicly available on the E3SM SCM GitHub project wiki (
A major advantage of the SCM using the same dynamical core as the full GCM is the ability to easily replay a single GCM column to replicate a specific GCM column with only round-off error differences. This is a powerful tool, where the user generates IOP forcing from a full E3SM run, with the intention to replicate a column of interest in SCM mode. This can be used to help diagnose model crashes due to unstable physics parameterizations or to target and address chronic model biases in an efficient manner. It can also help to fill in the gap for a particular regime or location where there is no forcing provided by the E3SM SCM library. The inherited SCM, which used the Eulerian dynamical core, required additional post-processing of the GCM forcing to be compatible with the replay option (as documented in Gettelman et al., 2019), since forcing terms between the two dynamical cores are somewhat different. Since the E3SM SCM uses the same dynamical core as the full GCM, the method to replay a GCM column is straightforward and accurate.
Though the E3SM SCM replay option is accurate, it cannot provide a fully bit-by-bit representation of a GCM column. This is because the GCM and SCM will only give bit-by-bit answers if they do exactly the same calculations. In GCM mode, the end of the dynamics state is computed via a series of sub-stepped loops. For the SCM, the net effect of these loops must be encapsulated by the end-of-step values minus the beginning-of-step values, divided by the time step. This tendency is then added to the SCM state using forward Euler time stepping. Since the GCM and SCM calculations are not identical, roundoff level differences occur. This issue could in principle be resolved using quadruple precision output but we found the related difficulties associated with this to not be worth a roundoff level gain. Our approximate method has proven suitable for most scientific applications of interest to E3SM users. In Sect.
In this section we will demonstrate that the SCM can serve as a tool to reproduce and explore climatological biases within the E3SM. We will also show an example of when the SCM cannot be used as a proxy for the full model. Finally, we will show an example of using the E3SM replay option. Some major biases in the E3SM include (but are not limited to) an overestimate of clouds in the Arctic, lack of subtropical maritime stratocumulus, lack of high clouds in the TWP warm pool, timing of precipitation in the tropics and midlatitudes, and a lack of precipitation over the Amazon rainforest (Xie et al., 2018; Zhang et al., 2019; Golaz et al., 2019; Rasch et al., 2019). In this section we will attempt to replicate a select number of these biases with the SCM.
Unless otherwise stated, the SCM results presented in this paper use the short-term hindcast approach (Ma et al., 2015). The SCM is initiated every day at 00:00 Z and run for 2 d, with prescribed large-scale forcing and surface turbulent fluxes and without temperature and moisture profiles being nudged to observations. The 24 to 48 h forecasts in each simulation are then combined as a continuous time series. With the hindcast approach, the model is well constrained by the large-scale condition, allowing us to isolate problems related to parameterizations. It also avoids the possible impacts of nudging to the clouds and precipitation (Ghan et al., 2000; Zhang et al., 2014). For example, Randall and Cripe (1999) extensively discussed the nudging method for SCMs and conclude that the impact of nudging on SCM simulation depends on the model biases produced without nudging; thus there is no solid theory on what can be expected from a particular model while using nudging. We will, however, explore the differences between nudging and the short-hindcast mode in one example.
SCMs are a useful tool to explore biases due to the model's physical parameterizations. But are there certain conditions and regimes under which the SCM is a better or worse proxy for the full GCM? While it has been demonstrated many times in the literature (e.g., Golaz et al., 2002; Bogenschutz and Krueger 2013; Suselj et al., 2013) that boundary layer cloud cases (such as DYCOMS for stratocumulus and BOMEX for shallow cumulus, as an example) can serve as a useful surrogate to explore and improve biases in the global model (due to the important cloud-forming processes in these regimes being mostly locally driven), the question of whether precipitation due to deep convective processes can be replicated faithfully in SCMs is less understood. Here we will attempt to replicate E3SM's biases in precipitation, both in the mean state and variability sense, to see when the E3SM SCM may be useful to exploit and investigate these biases.
The diurnal cycle of precipitation, especially over land, is a mode of climate variability that GCMs have long struggled to simulate adequately (Covey et al., 2016; Lee et al., 2007). Over land the late afternoon peak in precipitation is typically associated with the transition of shallow to deep convection, while the nocturnal peak is mostly due to elevated convective systems associated with eastward-propagating mesoscale convective systems. Many studies have attributed the GCMs' inability to represent the diurnal cycle of precipitation to deficiencies in the moist convective parameterizations (Dai and Trenberth 2004; Lee et al., 2008), where model errors over land are associated with unrealistic strong coupling of convection to the surface heating (Lee et al., 2007; Xie et al., 2002). Thus, precipitation peaks in the model tend to occur too early over land during the day, especially in summer.
E3SMv1 strongly exhibits these aforementioned biases (Xie et al., 2019; hereafter XIE2019), especially when focused over the continental United States (CONUS; Fig. 9 of XIE2019). In the central plains of the US, observed precipitation peaks in the late evening time, whereas E3SM precipitation peaks around noon. Can E3SM SCM reproduce this bias and can we use the SCM to implement modifications to the parameterized physics that would help improve this long-standing issue? For this experiment we use version 2 of continuous forcing from the Southern Great Plains (SGP) ARM site (Tang et al., 2019; Xie et al., 2004) that spans from 2004 to 2015; however for this study we only consider the warm season from 1 May through 31 August of each year. Note that multi-year SCM forcing allows us to perform robust statistical analysis rather than relying on a single case study as typically done in the past with SCM runs.
To see if we can improve this bias in the SCM, we implemented a revised convective triggering function, as implemented in XIE2019, that has been shown to greatly improve the diurnal cycle of precipitation in E3SM simulations. This new convective triggering is a combination of two methods, known as dynamic convective available potential energy (dCAPE) and the unrestricted launch level (ULL).
Composite of the total precipitation (convective + large scale) in local time for the E3SM SCM (red curve), E3SM SCM with the convective modifications documented in Xie et al. (2019; blue curve), and observations (black curve) from the Southern Great Plains (SGP) ARM site version 2. Panel
Figure
Due to the fact that the SCM can replicate the behaviors seen in the global model for this situation, we can further use this SCM case to explore the exact reason for this behavior. Figure
Another major bias in E3SM that is characteristic of most GCMs is the lack of precipitation over the Amazon (Fig. 9 of Xie et al., 2018). E3SM has a climatological dry bias upwards of 4 mm d
Figure
Precipitation from the GOAMAZON field campaign for SCM (red curve) and observations (black curve). The GCM results (blue curve) are taken from an E3SM run in the column closest to the GOAMAZON location (3
The reasoning why the GCM Amazon dry bias cannot be replicated in the SCM is likely because this bias is primarily due to a misrepresentation of the large-scale environmental conditions in the GCM rather than by parameterized deficiencies. This is important information for E3SM developers and the analysis team. Figure
Annual cycle of observed versus E3SM simulated environmental states for the GOAMAZON location for large-scale vertical velocity (top two left panels), relative humidity (top two right panels), zonal wind (bottom two left panels) and meridional wind (bottom two right panels).
In addition, it is well understood that for deep-convection precipitation is usually balanced mainly by advective moisture convergence, which is prescribed in these experiments. Therefore, this is a prime example of a situation where the SCM is not a useful tool to help improve GCM biases, but it does suggest that efforts should be spent on improving the large-scale circulation, or remote biases, that are probably responsible for the Amazon precipitation bias. As already mentioned, the SCM can replicate the bias related to the early onset of precipitation (similar to that seen in Fig.
Having comprehensive a priori knowledge on what particular biases and regimes could faithful be replicated within an SCM framework would be invaluable for GCM development and improvement. However, this is currently poorly understood and should be the subject of future work.
Zhang et al. (2019) show that E3SM suffers from an overestimate of Arctic clouds, mostly in the form of too much liquid cloud. Here we use the Mixed-Phase Arctic Cloud Experiment (MPACE; Verlinde et al., 2007) case that sampled clouds over open ocean near Barrow, AK, with the goal to collect observations to advance the understanding of the dynamics and microphysical processes of mixed-phase clouds. This is a 17 d case taking place in October 2004.
The top row of Fig.
Top panel displays the evolution of the vertical cloud structure for cloud fraction (observations on the left and E3SM SCM on the right) for the MPACE case from October 2004. The second through fourth panels on the left column represent the cloud liquid mixing ratio, while the second through fourth panels of the right column represent cloud ice mixing ratio from E3SM SCM simulations). The second row represents simulations using the hindcast method, the third row represents simulations with the Bergeron–Findeisen process set to the default tuning value, and row four represent the simulations where temperature and moisture are nudged to observations. The bottom row displays the evolution of the integrated liquid (left) and ice (right) path for the various configurations mentioned.
As previously mentioned, we chose to perform the majority of experiments in this paper in short-term hindcast mode. However, the E3SM SCM also comes with an option to nudge temperature and moisture to observed values. By default the E3SM SCM uses a nudging timescale of 3 h. It is interesting to note that the solution obtained for the MPACE case is strongly dependent on the technique used to constrain the mean state. The fourth row of Fig.
Figure
Vertical evolution of temperature from observations
If, however, one is using the SCM for the purposes of parameterization development/implementation and wishes to assess their new parameterization in conditions with little to no mean state bias (e.g., to avoid compensating errors), then the nudging method is likely preferable. For instance, the results seen using nudging vs. hindcast for MPACE clouds may suggest that Arctic clouds simulated in E3SM are an artifact of compensating errors. When the observed temperature and moisture profiles are used, we see the model struggles to produce any ice cloud at all and is in conflict with observations. This suggests that E3SM developers may need to reevaluate either the parameterizations and/or tuning choices in order to get a desirable solution when the temperature and moisture most resemble observations. In addition, caution is warranted when using nudging, since constantly nudging to the observed temperature and moisture state inherently breaks the water and energy budget by acting as an artificial source. The sequential splitting techniques that E3SM uses could in theory be obscuring the direct effects of this and could be leading to the artificial reduction in condensate. However, this idea needs to be explored more.
Xie et al. (2018), Golaz et al. (2019), and Zhang et al. (2019) all report substantial bias of high clouds in the TWP. Figure
Climatological E3SMv1 high cloud bias, computed relative to MODIS observations, from Xie et al. (2018). Open red star shows location of the TWP field campaign, while solid yellow star shows the location we use for our E3SM SCM replay experiment.
We wish to replay a column near the location where the bias is most severe. Therefore we choose a location near 5
Once the simulation is over we use scripts provided in the E3SM case library to replay our column of choice. The inputs we need to specify are the E3SM-generated forcing file, initial condition file, the latitude and longitude we wish to simulate, and the desired start date and run duration.
Figure
Temporally averaged profiles of cloud fraction
From Fig.
As an example of this, we run the SCM in replay mode for this column but with the Bergeron efficiency set to 1.0 (blue dashed curve in Fig.
This paper describes the E3SMv1 SCM, including modifications made to it since we adopted it from CAM, and how this configuration can be useful for model development and evaluation. A number of important upgrades were made to E3SM SCM since the inherited version, including the ability for the user to specify how the aerosols are treated to avoid unscientific case setups due to the fact that E3SM initializes all aerosol concentrations to zero. Idealizations have also been implemented and turned on by default, depending on the IOP forcing the user selects, to ensure an apples-to-apples comparison with LES benchmarks that the IOP forcing was meant to replicate. Most importantly, the E3SM SCM is now configured to work with the same dynamical core as the full GCM. This ensures that the SCM runs with the same large-scale vertical advection scheme, time step, and physics–dynamics coupling methods as the full model. It also allows the user to trivially replay a column of the full GCM with the SCM without the need to interpolate initial condition files or forcing files from one dynamical core to the next. The upgrade to the SE dynamical core is also advantageous because E3SM no longer needs to maintain and support the Eulerian dynamical core, which was not used in any other model configuration. We note that the E3SMv1 SCM infrastructure is expected to remain the same for E3SMv2.
The E3SM SCM also has an extensive library of IOP cases that span the traditionally used GCSS boundary layer cloud cases (i.e., BOMEX, DYCOMS, RICO) and standard deep-convection cases (i.e., ARM97, GATE). We also include IOP forcing files from more recent and modern cases, such as GOAMAZON, RACORO, and DYNAMO; many of these are unique to the E3SM SCM. For example, the E3SM can simulate conditions at ARM SGP for 12 continuous years. This allows for robust GCM-like statistics to be generated in a computationally efficient manner. Scripts to run each individual case are available at
We provide some examples of when the E3SM SCM may prove to be a useful proxy for GCM performance. For instance, we are able to successfully replicate the diurnal cycle of precipitation bias in the GCM by using forcing generated at ARM SGP. This bias is mostly due to deficiencies in the triggering mechanism in the convective parameterization that is unable to properly handle elevated convection. By implementing the trigger improvements documented in XIE2019, we are able to reproduce the same improved diurnal cycle of precipitation in the SCM found in global simulations. However, we were unable to replicate the seasonal cycle of dry Amazon bias with the SCM. We conclude that the root cause of the bias is due to improper representation of the large-scale environment rather than a deficiency with the parameterizations.
Using Arctic clouds as an example, we use the SCM to experiment with tunable parameter changes to evaluate the sensitivity of the high-latitude cloud bias. We report positive effects with the tuning of one parameter for this particular regime, but we caution that the SCM cannot inform how a full GCM simulation and radiation balance would be impacted with a modification. We also compare the SCM in hindcast or free-running mode versus a run where the SCM is nudged to observations. By running in hindcast or free-running mode the SCM allows the model biases in temperature and moisture to naturally develop, thus providing a better proxy with the model behavior in the full GCM; it should therefore be used if trying to replicate E3SM behavior. Nudging the SCM to observations may not provide a proxy with the full GCM and the behavior that could deviate significantly from E3SM global runs. This mode is, however, potentially useful if trying to improve or implement a parameterization while avoiding compensating errors. We caution the user on the potential unintended consequences of adding artificial sources that nudging could introduce.
We also demonstrate that the replay mode in the E3SM SCM can faithfully replicate a column of the GCM, though bit-by-bit replication is not possible in the current implementation. This mode is useful when trying to simulate a particular regime or region that the extensive E3SM case library does not cover. In our example, we replicated the high cloud bias in the Tropical Pacific Warm Pool. While the SCM cannot inform us directly whether biases are caused predominately by deficiencies in the model physics or the large-scale flow, it can provide clues about the culprit. This allows model developers to focus their energies more efficiently on a solution.
The E3SM SCM is mature and should be a first step in the model physics development and implementation process. With the extensive case library and the ability to simulate many different regimes, users can gain valuable insights into their development efforts and efficiently fix bugs. The SCM is also an important tool for addressing long-standing biases in the model; its incredible efficiency makes large sets of perturbed parameter tests easy. In addition, model instabilities that may arise in the full GCM can be investigated efficiently using the easy to use SCM replay mode, which is a powerful tool that can faithfully replicate a column of the full GCM.
The model code used in this study is located at
LLNL conceived the study and was the primary developer of the SCM. WL did some of the preliminary development work of the SCM, while YSC provided some crucial code contributions for the replay mode for the SCM. All authors contributed to interpreting the results.
The authors declare that they have no conflict of interest.
This research work was performed under the auspices of the U.S. Department of Energy by LLNL under contract DE-AC52-07NA27344.
This work was part of the CMDV Software Modernization program, funded by the U.S. Department of Energy (DOE) Office of Biological and Environmental Research (grant no. DE-AC52-07NA27344). Yao-Sheng Chen is funded by Scientific Discovery through Advanced Computing (SciDAC) (award no. DE-SC0018650).
This paper was edited by Andrea Stenke and reviewed by two anonymous referees.