An Unprecedented Set of High-Resolution Earth System Simulations for Understanding Multiscale Interactions in Climate Variability and Change
Abstract
We present an unprecedented set of high-resolution climate simulations, consisting of a 500-year pre-industrial control simulation and a 250-year historical and future climate simulation from 1850 to 2100. A high-resolution configuration of the Community Earth System Model version 1.3 (CESM1.3) is used for the simulations with a nominal horizontal resolution of 0.25° for the atmosphere and land models and 0.1° for the ocean and sea-ice models. At these resolutions, the model permits tropical cyclones and ocean mesoscale eddies, allowing interactions between these synoptic and mesoscale phenomena with large-scale circulations. An overview of the results from these simulations is provided with a focus on model drift, mean climate, internal modes of variability, representation of the historical and future climates, and extreme events. Comparisons are made to solutions from an identical set of simulations using the standard resolution (nominal 1°) CESM1.3 and to available observations for the historical period to address some key scientific questions concerning the impact and benefit of increasing model horizontal resolution in climate simulations. An emerging prominent feature of the high-resolution pre-industrial simulation is the intermittent occurrence of polynyas in the Weddell Sea and its interaction with an Interdecadal Pacific Oscillation. Overall, high-resolution simulations show significant improvements in representing global mean temperature changes, seasonal cycle of sea-surface temperature and mixed layer depth, extreme events and in relationships between extreme events and climate modes.
Key Points
- An unprecedented set of multi-century high-resolution Community Earth System Model (CESM) simulations is described
- High-resolution CESM simulations reveal a potential role of Southern Ocean polynyas in multidecadal climate variability
- High-resolution CESM exhibits significantly improved simulations of extreme events, such as tropical cyclones and atmospheric rivers
Plain Language Summary
Although the current generation of climate models has demonstrated high fidelity in simulating and projecting global temperature change, these models show large uncertainties when it comes to questions concerning how rising global temperatures will impact local weather conditions. This is because the resolution (~100 km) at which the majority of climate models simulate the climate is not fine enough to resolve these small-scale regional features. Conducting long-term (multi-centuries) high-resolution (~10 km) climate simulations has been a great challenge for the research community due to the extremely high computational demands. Through international collaboration, we are able to address this challenge by delivering an unprecedented set of multi-century high-resolution climate simulations using the Community Earth System Model (CESM), capable of directly representing tropical cyclones and extreme rainfall events. In this paper, we give an overall assessment of the value and benefit of the high-resolution CESM climate simulations by making a direct comparison to an identical set of low-resolution CESM simulations. We showcase some of the major improvements of the high-resolution CESM in simulating global mean temperature changes, seasonal cycle of sea-surface temperature, and extreme events, such as tropical cyclones and relationships between tropical cyclones and El Niño-Southern Oscillation.
1 Introduction
Climate variations and change can impact the likelihood and intensity of extreme events, including tropical cyclones, heat waves, winter storms, droughts, floods, and coastal sea-level rise. Such events have profound effects on the well-being of humans, agriculture, energy use, industrial activity, marine ecosystems, and coastal sustainability, with economic impacts in the billions of dollars. There is a pressing need to develop more advanced modeling frameworks for high-resolution multiscale Earth system predictions to improve our ability to simulate and predict future changes in extreme events. The International Laboratory for High-Resolution Earth System Prediction (iHESP) was established to address this pressing need. This trilateral collaboration among the Qingdao National Laboratory for Marine Science and Technology (QNLM), Texas A&M University (TAMU), and the U.S. National Center for Atmospheric Research (NCAR) was established to pursue transformational efforts in high-resolution Earth system modeling and prediction through combining and maximizing modeling expertise and computing resource.
One of the major research themes during the initial phase of iHESP (2018–2023) is to assess and quantify the role of mesoscale ocean features, including fronts and eddies, and their interactions with the atmosphere and sea-ice in climate variability, predictability, and prediction by carrying out an unprecedented set of present and future climate simulations at mesoscale-eddy-resolving/-permitting and tropical cyclone (TC)-permitting resolutions in the ocean and atmosphere, respectively. Completing such an extraordinary set of climate simulations requires (1) a state-of-the-art Earth system model, (2) a team of modeling experts and software engineers, and (3) a dedicated High-Performance Computing (HPC) resource. The iHESP collaboration is able to bring these requirements together, creating a unique opportunity to carry through this ambitious research plan.
The Earth system model used for iHESP research is the Community Earth System Model version1.3 (CESM1.3) developed at NCAR in collaboration with the broader climate modeling community. CESM has a long history of performing simulations in support of the Coupled Model Intercomparison Project (CMIP) that are used in the Intergovernmental Panel for Climate Change (IPCC) reports. An early version of the model, CESM1.1, was configured and run successfully for 100 years at a horizontal resolution of 0.25° for the atmosphere and land models and nominal 0.1° for the ocean and sea-ice models to simulate present-day climate (Small et al., 2014). The current high-resolution CESM (hereafter HR) is based on an earlier version of CESM1.3 described by Meehl, Yang, et al. (2019) which used an updated version of the atmospheric model—compared to that of CESM1.1—at the same 0.25° horizontal resolution used by Small et al. (2014), but with nominal 1° ocean and sea-ice models. The version used by iHESP in this study replaced the 1° ocean and sea-ice model described in Meehl, Yang, et al. (2019) by the nominal 0.1° horizontal resolution ocean and sea-ice models of Small et al. (2014) with some additional modifications and improvements.
As part of the iHESP collaboration, the HR simulations are being performed on the Sunway TaihuLight HPC in Wuxi, China. This system has a peak performance of 125.4 petaflops and is currently ranked the third fastest HPC system in the world. This Chinese Sunway-based HPC has a different architecture from the intel-based HPC on which CESM code was developed. Therefore, before these HR simulations could be run, the CESM1.3 code had to be first ported and then optimized on the Sunway HPC. This porting and optimization effort led by the QNLM team was technically challenging and took nearly a year to complete. The details of this effort are described in Zhang et al. (2020).
This paper focuses on the scientific aspects of two long HR simulations just completed: (1) a 500-year preindustrial control simulation (PI-CTRL) and (2) a 250-year historical and future transient climate simulation (HF-TNST) for the 1850–2100 period. These simulations are unprecedented in terms of model resolution and integration length. Previous simulations at comparable model resolutions were rather short, that is, on the order of 100 years, as discussed in Small et al. (2014) for CESM and in Roberts, Jackson, et al. (2020) for the recent High-Resolution Model Intercomparison Project (HighResMIP) simulations. Our long high-resolution simulations allow us to perform a more robust evaluation of the benefits of high-resolution climate simulations and their potential shortcomings. By comparing these HR simulations to their low-resolution counterparts (hereafter LR) and available observations, we attempt to address the following questions: How does the simulation of natural variability and response to anthropogenic forcing of the climate system depend on model resolution? In particular, how do multi-scale interactions arising from coupling the better resolved atmosphere and ocean models influence the low-frequency, large-scale behavior of the Earth system, as well as their impact on extreme events? To what extent can the knowledge about climate variability and change gained from low-resolution modeling studies carry over to the TC-permitting and ocean eddy-resolving/-permitting regime? We believe this unique set of climate simulations will be highly valuable for the climate research community to advance our understanding of multiscale interactions within the Earth system and their impact on future climate projections.
The manuscript is organized as follows. Section 2 gives a brief description of the HR model and the simulation design. Section 3 describes the results of PI-CTRL with an emphasis on energy balance and stability of the simulated climate system, including model drift, as well as intrinsic multidecadal variability. Section 4 focuses on the analysis of the historical period of HF-TNST, in which we compare HR and LR simulations with available observations. The comparative analysis includes major modes of climate variability and extreme events, as well as relationships between them. Section 5 examines simulated future climate response, including extreme events, to anthropogenic forcing in HR and LR simulations. Finally, in section 6, we provide a summary along with a few concluding remarks and discuss lessons learned from this unique set of CESM simulations.
2 Model Configurations and Simulations
2.1 CESM1.3
The CESM version used for iHESP research is not the latest version CESM2.0 (Danabasoglu et al., 2020), but based on an earlier version CESM1.3. The primary reason for using an older version, instead of the latest CESM2, is that the former has been run and evaluated thoroughly for century-long climate simulations in a high-resolution configuration (Small et al., 2014), while the latter has not. Indeed, as of now, there is no high-resolution configuration of CESM2 available. Additionally, even if there were one, CESM2 is computationally much more expensive primarily because of many additional physics parameterizations included in its new atmospheric model version. The particular version of CESM1.3 (CESM1.3-beta17_sehires38) used in this study is based on an earlier version, CESM1.3-beta17_sehires20, described in Meehl, Yang, et al. (2019), which was developed specifically for supporting a high-resolution CESM version with a 0.25° atmosphere and a standard-resolution nominal 1° ocean model. The CESM1.3 component models are the Community Atmosphere Model version 5 (CAM5; Neale et al., 2012), the Parallel Ocean Program version 2 (POP2; Danabasoglu et al., 2012; Smith et al., 2010), the Community Ice Code version 4 (CICE4; Hunke & Lipscomb, 2008), and the Community Land Model version 4 (CLM4; Lawrence et al., 2011).
A detailed description of the progression from CESM1.1 to CESM1.3-beta17 is provided in Meehl, Yang, et al. (2019). The most significant changes and improvements were made in the atmospheric model. A major change from CESM1.1 to CESM1.2 (also used in CESM1.3) involved moving from an Eulerian to a Lagrangian vertical advection scheme within the Spectral Element dynamical core (SE-dycore). Additional changes from CESM1.2 to CESM1.3 included the following most noteworthy aspects: (1) a microphysics rearrangement, (2) a change in the radiation code (Rapid Radiative Transfer Model for General Circulation Models, RRTMG), (3) updates to the heterogeneous freezing code and the gravity wave scheme (McFarlane, 1987; Richter et al., 2010), and (4) changes to dust tuning and soil erodibility. As shown by Meehl, Yang, et al. (2019), these changes in CESM1.3 led to a better positioning of the Southern Hemisphere (SH) jet and improved high and low cloud simulations that agree better with the available observations.
Although a high-resolution version of CESM1.1 was configured and successfully integrated for more than a century under a perpetual present-day (year 2000) climate forcing (Small et al., 2014), a similar configuration for CESM1.3 was not available before the iHESP project. Therefore, the first concerted effort of iHESP was directed at configuring and testing a high-resolution version of CESM1.3, leading to the current version of CESM1.3-beta17_sehires38 tag with a 0.25° resolution in CAM5 and CLM4 and a nominal 0.1° resolution in POP2 and CICE4. There are 30 vertical levels in the atmosphere with a model top at 3 hPa and the atmospheric model parameter settings remain unchanged from those in CESM1.3-beta17_sehires20 described in Meehl, Yang, et al. (2019), except for a couple of bug fixes, including (1) a bug fix in the radiation code that omitted diffusivity angle calculations for key longwave radiation bands, ultimately causing unrealistic atmospheric temperatures and cloud formation, and (2) a fix in logic errors when computing snow water path and snow cloud fraction. A minor change to the processor decomposition for the SE dycore is also included in CESM1.3-beta17_sehires38. The ocean and sea-ice models are essentially the same as those used in Small et al. (2014). The ocean model has 62 levels in the vertical with a maximum depth of 6,000 m. Both the horizontal and vertical grids are identical to those used in Small et al. (2014). Compared to CESM1.3-beta17_sehires20, a couple of changes were made to POP2. First, a more efficient elliptic solver for the barotropic mode was back-ported from CESM2.0 to CESM1.3 to increase the computational efficiency of high-resolution simulations when using large processor counts (e.g., Hu et al., 2015; Huang et al., 2016). The development of this new solver was led by the Chinese team in collaboration with NCAR, representing a contribution of the Chinese climate modeling community to the CESM development. Second, a change of ocean coupling frequency was made from 1 hr to 30 min to alleviate coupling instabilities between the ocean and sea-ice models. In the sea-ice model, an older penetrative shortwave calculation method was replaced with a newer delta-Eddington shortwave computation of Briegleb and Light (2007). Although this latter method had been the default shortwave computation for the standard resolution simulations since Community Climate System Model version 4 (CCSM4) and CESM1, it was not used in the high-resolution CESM1.1 simulation (Small et al., 2014). Unlike the standard resolution version, the high-resolution POP2 in CESM1.3-beta17_sehires38 does not include mesoscale and submesoscale parameterizations (same as in Small et al., 2014). Nor does it include the overflow parameterization (Danabasoglu et al., 2010) (also same as in Small et al., 2014). Finally, the new coupler developed for CESM2.0–Common Infrastructure for Modeling the Earth (CIME) was backported to CESM1.3-beta17_sehires38.
After a preindustrial (year 1850 conditions) control simulation was configured, CESM1.3-beta17_sehires38 was first ported and optimized on Stampede2 at the Texas Advanced Computing Center (TACC) for testing and tuning. Several decades of test runs were made, during which melting snow grain radius and the melt onset temperature parameters were adjusted in CICE4 to ensure that sea-ice thickness and extent were within the observed estimates and that the top-of-atmosphere (TOA) radiation imbalance was small (~−0.05 W m−2 averaged over year 6 to 20). This finalized CESM1.3-beta17_sehires38 tag was then sent to the Chinese team for porting and optimization on the Sunway TaihuLight HPC. The porting and optimization effort on the Sunway HPC was a major software engineering undertaking that took nearly a year to complete—details of this effort are described in Zhang et al. (2020). The outcome was a highly parallelized and efficient Sunway version of CESM1.3-beta17_sehires38 that can achieve a performance speed of close to five simulation years per calendar day on 61,600 cores with no model output. With high-frequency model outputs, including daily mean and 6-hourly variables, the model performance decreases to less than three simulation years per day (~40% decrease). The Sunway version of the model used for this study is available via GitHub at https://github.com/ihesp/CESM_SW. It should be noted that this model version uses the SE-dycore in CAM5 which was employed for both the high- and low-resolution simulations conducted on Sunway.
2.2 Preindustrial Control (PI-CTRL) and Historical and Future Transient (HF-TNST) Climate Simulations
The experimental design of our PI-CTRL and HF-TNST climate simulations follows the prescriptions for the Coupled Model Intercomparison Project phase 5 (CMIP5) experiments (Eyring et al., 2016). The PI-CTRL ocean component was initialized with the January-mean climatological potential temperature and salinity from the World Ocean Atlas (WOA; Locarnini et al., 2013; Zweng et al., 2013) and at state of rest. The other component models were initialized from restart files of previously performed simulations. The climate forcings were set to the 1850 conditions and kept constant throughout the entire 500-year simulation. The first 20 years of HR PI-CTRL were carried out on Stampede 2 at TACC, and the remaining 480 years were completed on Sunway TaihuLight. The first 150 years are a period of rapid adjustment of the upper ocean, and also during this period of the simulation, a hardware issue was identified on Sunway TaihuLight (see Zhang et al., 2020, for more discussion). Therefore, the analysis presented here covers the last 350 years of PI-CTRL. It is worth noting that high-frequency model outputs, such as 6-hourly atmospheric variables required for tracking TCs and atmospheric rivers (ARs), were not saved until year 338, because turning on high-frequency model outputs significantly slows down the model performance. Therefore, only years 338–500 of HR PI-CTRL contain high-frequency model outputs.
The HF-TNST simulation was branched from year 250 of PI-CTRL. It uses historical forcings from 1850 to 2005 and representative concentration pathway 8.5 (RCP8.5) forcings from 2006 to 2100 (Lamarque et al., 2010, 2011; Meinshausen et al., 2011) in accordance with CMIP5 experimental protocol. Ozone concentrations were calculated using a high-top coupled chemistry-climate model (CESM1; Whole Atmosphere Community Climate Model [WACCM]; Marsh et al., 2013) with specified ozone depleting substances. High-frequency model output was turned on from 1877 onwards, giving us 221 years of 6-hourly atmospheric variables to track TCs and ARs during the historical and future periods.
Although the iHESP HF-TNST simulation uses the same forcing data sets as in CESM1 Large Ensemble simulations (CESM1-LENS; Kay et al., 2015), a clean comparison of the respective LR and HR simulations is not possible because two different code versions were used. Specifically, while CESM1-LENS is based on CESM1.1 with the finite-volume dynamical core (FV-dycore) of CAM5, iHESP HF-TNST uses CESM1.3 with the CAM5 SE-dycore. Thus, to allow a more rigorous comparison, an identical set of LR (nominal 1°) CESM1.3 simulations with the same CAM5 SE-dycore was also conducted on Sunway TaihuLight. To achieve an acceptable TOA balance, we retuned some model parameters in LR CAM5 SE-dycore, including cldfrc_rhminl, minimum relative humidity for low stable clouds (0.87 in LR vs. 0.92 in HR), micro_mg_dcs, autoconversion size threshold for ice to snow (5.25 × 10−6 m in LR vs. 6.0 × 10−6 m in HR), and dust_emis_fact, multiplier to activate/aerosolize dust from the land surface (0.95 in LR vs. 1.05 in HR). We also readjusted parameters in CICE4, including xndt_dyn, the number of dynamic timesteps per thermodynamic timestep (1.0 in LR vs. 2.0 in HR), dt_mlt_in, the melt temperature onset value for changes in the snow grain radius (1.0°C in LR vs. 0.5°C in HR), fsnowrdg, the fraction of snow left on the sea ice after ridging (0.5 in LR vs. 1.0 in HR), r_ice, the number of standard deviations away from the mean inherent optical properties of bare ice (0.0 in LR vs. 1.5 in HR), r_pnd, the number of standard deviations away from the mean inherent optical properties of melt ponds (0.0 in LR vs. 1.5 in HR), and rsnw_melt_in, the maximum melting snow grain radius (1.0 × 103 μm in LR vs. 0.5 × 103 μm in HR). Additionally, POP2 in LR CESM1.3 includes both mesoscale and submesoscale parameterizations, as well as the overflow parameterization, as in the standard resolution CESM1.3-beta17 described by Meehl, Yang, et al. (2019). All the comparative analyses shown in this paper will be based on the same set of HR and LR simulations. However, we note that the use of CAM-SE in LR is non-standard in CESM experiments, and this LR CESM configuration is relatively untested and untuned compared to the CAM-FV version of LR CESM. The choice was made to keep consistency between HR and LR in order to isolate the effects of resolution, but at the cost of a somewhat degraded LR simulation compared to standard CESM1 using CAM-FV.
This set of HR and LR simulations only presents the initial phase of iHESP's modeling efforts. The second phase will include a 10-member ensemble of HR HF-TNST simulations from 1920 to 2100, and the third phase will include an ensemble of HR decadal prediction experiments. As such, the iHESP modeling project is truly a “big data” project for the current generation of computing. Completing such an unprecedented set of high-resolution climate simulations requires substantial computational, storage, and human resources. The 480-year segment of the HR PI-CTRL took ∼200 days to complete using 61,600 cores of Sunway TaihuLight and produced more than 360 TB of raw model output. The 250-year HR HF-TNST took ~150 days to complete using 41,400 cores on the same machine, producing close to 300 TB of raw model output. When combined, just these two HR simulations during the initial phase of the iHESP project exhausted a total of over 560 million core hours of Sunway TaihuLight HPC with a total archived data volume exceeding 660 TB.
Some of the data sets from the iHESP CESM simulations are already publicly available via iHESP website at http://ihesp.tamu.edu and QNLM website at http://ihesp.qnlm.ac. The entire dataset will be released by the end of 2020. The model output variables follow the requirements of HighResMIP CMIP6 (Haarsma et al., 2016), which includes a large number of high-frequency fields, including daily mean and 6-hourly outputs, that are necessary for studying extremes, such as TCs and ARs, on societally relevant time scales. We will show some results on these extremes in the following sections.
3 Preindustrial Control Simulations
3.1 Global Energy Balance, Temperature Evolution, and Long-Term Variability
As mentioned previously, both HR and LR were tuned to obtain TOA energy imbalances as small as practically possible before long PI-CTRL simulations were conducted to prevent large temperature drifts in their ocean components. Figure 1 shows the net TOA radiative imbalance, QTOA, timeseries in HR and LR for the last 350 years of PI-CTRL with the first 150 years regarded as model spin-up. In both simulations, QTOA is rather stable, computed as −0.217 W m−2 and +0.033 W m−2 in HR and LR, respectively, for the last 350 years. These QTOA values are acceptably small, considering that the previous CESM1.1 simulations by Small et al. (2014) had QTOA value of above +0.3 W m−2 in HR and +0.5 W m−2 in LR, respectively. We note that tuning high-resolution models is much more challenging than their low-resolution counterparts because of much more intense computing demands.
Figure 1b shows the ocean global-mean potential temperature timeseries from HR and LR. As in many other coupled control simulations (e.g., Danabasoglu et al., 2020), any TOA imbalances are largely reflected in these time series because the other component models cannot store or lose heat in the long term. Thus, while the global-mean ocean temperature shows a cooling trend of about −0.063°C per century in HR, LR simulation warms by about 0.01°C per century. Unfortunately, these global averages are rather deceiving because they do not reflect any drifts in either lateral or vertical redistribution of heat in the ocean as demonstrated by the vertical-profile time series of horizontal-mean global potential temperature in Figures 1c and 1d for HR and LR, respectively. These are anomalies from the January-mean climatological WOA initial conditions. As can be seen, the most significant secular changes take place below 200 m in both HR and LR. The maximum cooling trend of ~ −0.08°C per century in HR is found between 1,000 and 2,000 m, whereas the maximum warming trend of ~ +0.04°C per century in LR is located between 500 and 1,000 m, below which there is a weak cooling trend. Compared to the observed January-mean climatological temperature, the HR has a warm bias in the upper 1,000 m with a maximum value of 0.3–0.5°C between 50 and 250 m and a deep cold bias of ~0.2–0.3°C, giving rise to a positive bias of vertical temperature gradient. In contrast, the LR has a more severe cold bias approaching to −1°C in the upper ocean and a stronger warm bias of +0.6°C between 500 and 1,000 m, leading to a more pronounced negative bias of vertical temperature gradient. These differences in vertical temperature gradient between HR and LR can have an important implication for meridional ocean heat transport as will be discussed later.
Another noticeable difference between HR and LR is the more pronounced multi-decadal variation in the upper 300 m ocean temperature in HR, which is seemingly absent in LR (Figures 1c and 1d). This multi-decadal variation is clearly seen in the global-average sea-surface temperature (SST) time-series with an estimated period of ~40 years (Figure 2a). The regression of SST anomalies onto the global-mean SST time-series reveals a pattern that bears a close resemblance to the simulated Interdecadal Pacific Oscillation (IPO) defined by Meehl and Hu (2006) (Figures 2c and 2d). Therefore, the pronounced multi-decadal variation in the upper-ocean temperature in HR is largely attributed to a super-charged IPO in the model.
Although a detailed analysis for the cause of the super-charged IPO in HR PI-CTRL is beyond the scope of this paper, some preliminary analysis suggests a link between Southern Ocean sea-ice variability and IPO. Superimposed onto the global-mean SST time-series in Figure 2a is a SH sea-ice extent anomaly time-series (SHice). It is evident that the two time-series are negatively correlated, such that a decrease in sea-ice extent in the Southern Ocean will lead to a warming in the tropical Pacific. Such a negative correlation between IPO and Southern Ocean sea-ice is seemingly consistent with some previous studies where it is argued that tropical Pacific SST anomalies can remotely force variability in the Southern Ocean via stationary Rossby waves excited by convective heating anomalies in the tropics (e.g., Meehl, Arblaster, et al., 2019). Indeed, regressions of Z500* anomalies, which are defined as Z500 departures from the zonal mean, onto the global mean SST (Figure 2c) and the IPO principal component (PC) time-series (Figure 2d) display a stationary wave pattern emanating from the tropics. However, a lead-lag correlation analysis between the global mean SST and Southern Ocean sea-ice extent time-series indicates that the latter actually leads the former by ~4 years (Figure 2b), suggesting that Southern Ocean sea-ice variability may not simply be responding to tropical Pacific changes, but rather play a more active role in the energetic IPO cycle.
The multidecadal Southern Ocean sea-ice variability in HR appears to have a close association with open ocean polynyas in the Weddell gyre. Figure 3 shows composites of September-October-November (SON) sea-ice concentration simulated in HR PI-CTRL for years when the annual Southern Ocean sea-ice extent time-series (Figure 2a) is above one standard deviation (Figure 3a, defined as non-polynya years) and below one standard deviation (Figure 3b, defined as polynya years), respectively. Contrasting these two panels clearly shows the presence of a polynya in the Atlantic sector of the Southern Ocean, particularly within the Weddell Gyre in the polynya-year composite. Consistent with other recent studies (e.g., Dufour et al., 2017; Kurtakoti et al., 2018; Weijer et al., 2017), the Weddell polynya is a prominent feature in HR, but does not appear in LR, suggesting that eddy-permitting model resolution may be critical to simulate these polynyas, perhaps through better representation of ocean mesoscale dynamics and small-scale steep topography, such as Maud Rise (Kurtakoti et al., 2018).
When Weddell polynyas occur, the ocean releases heat into the atmosphere, causing warming of the Antarctic atmosphere (Kaufman et al., 2020; Weijer et al., 2017). Such warming can then lead to a reduction in atmospheric heat transport into the Antarctic region during polynya years (Kaufman et al., 2020), which in turn can cause a reduction in the heat transport out of the tropics (Kang et al., 2008, 2009). The reduced southward heat transport from the tropics implies more heat is transported northward across the equator. This requires an anomalous Hadley Cell that has a northward (southward) cross-equatorial flow aloft (near the surface). This anomalous Hadley Cell is consistent with a southward shift of the Inter-Tropical Convergence Zone (ITCZ) (Kang et al., 2008, 2009). Figure 2e shows the regression of precipitation anomalies onto the negative of the SH sea-ice extent time series, that is, −(SHice). During polynya years (decrease in the SH sea-ice extent), there is not only an increase of local precipitation in the Weddell Sea region, but also a decrease of precipitation along the ITCZ and an increase of precipitation just south of the ITCZ, particularly in the western tropical Pacific. This precipitation pattern in the tropics is consistent with a southward shift of the ITCZ. Since the vicinity of the ITCZ marks the weakest trade wind regime, the southward shift of the ITCZ will bring the weak trade regime near the equator, which will trigger a Bjerknes feedback and lead to the development of an IPO-like warming along the equator (Pausata et al., 2015). We, therefore, hypothesize that this interaction between the tropics and Antarctic region may play a key role in the supercharged IPO in HR.
We note that such a polynya-IPO relationship was not found in either the previous HR simulation discussed in Small et al. (2014) or the CESM1.3 simulations with a LR ocean reported by Meehl, Yang, et al. (2019). The fact that the simulations by Meehl, Yang, et al. (2019) were based on the same HR CAM5 but with LR POP2 suggests that the intermittent occurrence of polynyas in the Weddell Sea and its interaction with IPO are likely attributable to the resolution increase in the ocean and sea ice models. On the other hand, the HR CESM1.1 simulation reported by Small et al. (2014) used a present-day (year 2000) climate forcing, which resulted in a warmer mean climate than that in PI-CTRL, and was integrated for only 100 years. Both of these factors may contribute to the absence of a similar polynya-IPO relationship in Small et al. (2014). In fact, Figures 2a and 2b show a much weakened polynya-IPO relationship in HR HF-TNST, particularly after year 2000 when the climate becomes significantly warmer than that during the preindustrial period. The long periodicity (~40–50 years) of the super-charged IPO in PI-CTRL can also make it difficult to identify the polynya-IPO relationship in the short simulation by Small et al. (2014). Therefore, it remains unknown whether the polynya-IPO relationship is unique to this HR CESM1.3.
3.2 Meridional Overturning Circulation and Antarctic Circumpolar Current
The ocean meridional overturning circulation (MOC) plays a key role in Earth's energy budget by redistributing heat, and because of the long inherent time scales of the ocean, it contributes to natural, low-frequency variations of the climate system. The Atlantic component of the MOC (AMOC), in particular, is believed to be critical in sustaining Atlantic Multidecadal Variability (AMV; Sutton & Hodson, 2005) and associated climate impacts over land (Zhang et al., 2019). The MOC streamfunction reduces the complex, three-dimensional ocean circulation to a simplified representation of flow that highlights the large-scale meridional and vertical flows that balance the high-latitude sinking of dense, deep-water masses such as the North Atlantic Deep Water (NADW) and Antarctic Bottom Water (AABW). These meridional flows represent an oceanic link connecting the North Atlantic and the Southern Oceans, although the strength of that connection and the time scales over which it operates remain poorly understood (Buckley & Marshall, 2016; Johnson et al., 2019).
Changing model resolution impacts the mean MOC structure as well as MOC variability characteristics, including the long-term drift (Figures 4a–4d). The biggest difference in mean MOC (black contours) is seen in the strength of the (negative) AABW cell which is considerably enhanced in HR compared to LR (over 16 Sv at 30°S, compared to over 4 Sv). This is associated with a greater penetration of AABW into the Northern Hemisphere (NH), including the North Atlantic. While the maximum AMOC strength (at roughly 35°N) is about 4 Sv weaker in HR, both resolutions exhibit a (positive) NADW cell of just over 16 Sv extending from 30°S to 30°N. The somewhat shallower NADW cell in HR appears to be related to the stronger AABW flow. As a result of the stronger AABW cell and comparable NADW cell in HR, the net southward flow into the Southern Ocean (at depths between about 1 and 3.5 km) is greatly enhanced in HR (~28 Sv compared to ~16 Sv at 30°S). In both HR and LR, the AMOC drift is dominated by a shallowing of the southward NADW flow, with HR also exhibiting a weakening of net AMOC strength over the course of the simulation (Figures 4b and 4d; color fill). There is a large positive trend in the abyssal MOC in LR which indicates a steady decline in AABW production and a concomitant reduction in the strength of southward mid-depth flow into the Southern Ocean (Figure 4c). This MOC drift likely contributes to the deep warming trend seen in LR (Figure 1d), with relatively warm NADW filling up more of the abyssal ocean. HR shows opposite trends: a long-term strengthening of the (negative) AABW cell (Figure 4a) and a steady cooling in the deep ocean (Figure 1c).
The characteristics of intrinsic, low-frequency (15-year cutoff with a Lanczos filter) MOC variability (after removal of the linear trend) appear quite sensitive to model resolution (Figures 4e–4h), with HR showing a pronounced increase in Southern Ocean (SO) variance related to the intensity of the AABW cell and the strength/depth of southward flow around 30°S. A band of enhanced global MOC variance in HR, likely related to large-amplitude, multidecadal variations in SH sea ice extent (Figure 2 and see below), extends from the SO to roughly 20°N (Figure 4e), and it appears to dominate AMOC variability south of about 5°N (Figure 4f). The multidecadal variability of the NADW cell is, however, reduced in HR compared to LR, at least in the extratropical North Atlantic (NA). In LR, high NADW variability appears largely confined to latitudes north of 35°N, due to a diabatic short-circuiting of the AMOC around that latitude, so that HR AMOC variance is greater than LR in the subtropical NA despite being lower in the extratropics. At both resolutions, the northern extratropical variance of AMOC is lower than that of the global MOC (compare Figures 4e and 4f and 4g and 4h). This is attributable to low-frequency variance in the extratropical Pacific meridional overturning circulation (PMOC) which is more pronounced in HR, although the mean PMOC is very weak (<4 Sv) in both resolutions (not shown). The relatively high low-frequency PMOC variance in HR is concentrated between 30 and 40°N which suggests that it could be related to shifts in the strength and position of the Kuroshio (Ma et al., 2016; Qiu & Chen, 2005) (to be further investigated).
Figure 4 indicates the locations of representative circulation indices that reflect large-amplitude multidecadal variations in the strength of the NADW cell (MOCNA) and the AABW cell (MOCSO). Simultaneous regressions (not shown) confirm that these overturning indices are associated with coherent large-scale circulation changes in their respective hemispheres. Both indices serve to quantify the high-latitude deep water formation activity that contributes to slow variations in MOC at lower latitudes (note that because AABW production gives rise to counterclockwise flow in the westward facing meridional plane, an increase in MOCSO corresponds to reduced AABW production and vice versa). The spectral characteristics of MOC (focusing on these two key indices) are very different in HR compared to LR (Figures 5a and 5b). The HR simulation exhibits a spectral peak in MOCSO at about a 40-year period that would clearly appear to be related to the strong variability in SH sea ice extent (SHice) at this time scale (Figures 2 and 5a). While LR also shows significant multidecadal SHice power (peaking at around a 60-year period), there is no corresponding spectral peak in MOCSO (Figure 5b), presumably because the SH sea ice variability in that simulation is not governed by open-ocean polynya formation. The NADW cell in HR shows maximum variance at relatively short decadal time scales (peaking at around a 20-year period), while the AMOC in LR shows significant variance concentrated between about 40- to 100-year periods.
The causal relationships between the low-frequency MOC variations in both hemispheres, SH sea-ice extent, and AMV are explored in a preliminary way using lag correlations (Figures 5c–5f). This analysis underscores the stronger link between MOCSO and SHice in HR than in LR (Figures 5c and 5d), with an expansion of the SH sea ice extent in HR leading by about 5 years an increase in MOCSO (decrease in AABW production). Increased MOCSO corresponds to enhanced upwelling of relatively warm circumpolar deep water at latitudes poleward of roughly 50°S (Figure 4). In both HR and LR, a maximum in MOCSO leads to a reduction in SHice, but the anticorrelation is slightly stronger (and more delayed) in HR. It is noteworthy that there is a stronger interhemispheric linkage between AABW/NADW production in HR than in LR. This manifests as a weak, but significant, positive correlation between MOCSO and MOCNA in HR with the former leading by about a decade (Figures 5c and 5e). We hypothesize that this teleconnection is mediated through the atmosphere (e.g., the atmospheric response to SH polynyas seen in Figure 2), although more work is needed to fully understand the relevant mechanisms. In both simulations, enhanced NADW production and spinup of MOCNA leads, by a few years, to anomalous basin-scale warming of the North Atlantic (i.e., positive AMV; defined here as the detrended and low-pass filtered average SST over the domain 0–60°N, 75–7.5°W). It is interesting that both HR and LR also show a significant, positive relationship between MOCSO and AMV when the former leads by about 17 years (Figures 5c and 5d). This connection is considerably stronger in HR, and the AMV spectrum in HR (with a significant 40-year peak, but weak variability around 20-year time scales where MOCNA is most energetic) implies that SO processes play more of a driving role than local NA processes. This is unexpected and merits further investigation.
AMV is generally believed to be strongly linked to AMOC variability driven by NADW production (see Zhang et al., 2019, for a recent review), and both HR and LR do show that MOCNA leads a warming of the North Atlantic (Figures 5e and 5f) with lag correlation values (~0.5) that are on par with those seen in previous control simulations using CCSM4 (Danabasoglu et al., 2012). In HR, the SST regression onto MOCNA (Figure 6b) exhibits a canonical horseshoe pattern of positive SST in the North Atlantic (maximum in the subpolar gyre region), negative SST anomalies in the South Atlantic (as well as the Indian Ocean and maritime continent regions), and positive anomalies in the northeast Pacific. This SST response in HR bears a strong resemblance to that noted for the observed AMV (Ting et al., 2011; Zhang et al., 2019). However, the SST regression onto the AMV index itself in HR is quite different (Figure 6a), showing a strong and highly significant simultaneous warming of the eastern Pacific (as well as weaker warming in the South Atlantic and Indian Ocean) that is not consistent with the observed AMV, although it bears a strong resemblance to the global IPO mode highlighted above (Figure 2c). Indeed, the global SST in HR (Figure 2a) also shows a significant spectral peak at 30–40 year time scales (not shown, but similar to the peaks for SHice, MOCSO, and AMV in Figure 5a) as well as a high simultaneous correlation with AMV, implying that AMV in HR is dominated by this global mode associated with Weddell Sea polynyas. This strong SO influence on AMV in HR is the likely explanation for the large-amplitude SST regression signals over the SO in the Weddell Sea region (Figures 6a–6d), which are only apparent in the austral winter regression (at much lower amplitude) in LR (Figure 7c). This discrepancy in HR between the AMOC-related AMV (Figure 6b) and the IPO-related AMV (Figure 6a) suggests that the traditional basin-scale AMV index is of limited usefulness for understanding intrinsic modes of North Atlantic variability in this simulation insofar as it represents a superposition of signals with distinct geographic origins and time scales. While the AMOC-related SST signals in HR appear realistic, the expected atmospheric teleconnections—summertime warming of the eastern United States and eastern Europe, enhanced summertime precipitation over Northern Europe and the African Sahel, and reduced summertime precipitation over the southwest United States (see Zhang et al., 2019, and references therein)—are very weak if present at all (Figures 6d and 6f). The AMV in LR is very anemic, whether computed as a regression on an SST index or an AMOC index (Figures 7a and 7b), and there is less discrepancy in LR between these two methods of examining AMV. Despite the weak AMV in LR, there is no clear indication that AMV-related teleconnections are better represented in HR. This is perhaps due to the confounding influence of SO processes in HR, as discussed above.
HR simulates a much stronger SST response to (normalized) AMOC variations than LR (cf. Figures 6b and 7b), and yet low-frequency MOCNA variability in LR is considerably higher than in HR (Figures 4f and 4h). This implies that the surface temperature response to AMOC variations on a per Sverdrup basis is much greater in HR than LR. The heat transport regression shown in Figure 8 confirms that the heat transport efficiency of the AMOC in HR is roughly double that of LR, which explains how weaker AMOC variations can produce a stronger surface response in HR. This analysis also indicates that there is a stronger and faster meridional coherence of heat transport anomalies in HR, with a discernible influence of AMOC-related variations extending into the SH near lag 0. A plausible explanation for the lower heat transport efficiency (and weak AMV) in LR is the large surface cold bias in that simulation (Figure 1d). This may account for the differences between AMV in LR and that analyzed in previous low-resolution versions of the CESM model (e.g., Danabasoglu et al., 2012).
The Antarctic Circumpolar Current (ACC) is the only current system that circles the Earth within a confined latitudinal belt. Its transport and path are impacted by several factors that include the ocean bathymetry, mesoscale eddies, surface buoyancy fluxes, and zonal wind stress strength and location of its maximum. These features and processes—essentially all impacting the meridional density and sea-surface height (SSH) gradients—are expected to be better resolved and represented in high-resolution ocean models in comparison to their low-resolution counterparts. For example, there is evidence that eddy statistics, SSH variability, and SST variance are better represented in high-resolution simulations (e.g., McClean et al., 2011; Small et al., 2014). Further, better representation of the SST front across the ACC gives rise to enhanced air-sea heat loss on its equatorward side, which in turn affects mixed layer depth (MLD) (Lee et al., 2011; Small et al., 2020). However, high-resolution alone does not guarantee a good fit to observed mean SSH as discussed in McClean et al. (2011) where they show large differences in SSH between fully coupled and forced (hindcast) ocean simulations with the same high-resolution.
Figure 9a presents the time series of the ACC transport at Drake Passage from HR and LR simulations, showing that the mean transport in HR (137.43 Sv) is 9.32 Sv lower than in LR (146.75 Sv), computed over the last 200 years of the simulation. Such reductions in this transport when going from low- to high-resolution have been previously reported (Chassignet et al., 2020; Roberts et al., 2019; Small et al., 2014). As a specific example, a recent pair of CESM2.0 forced ocean simulations show that the low-resolution simulation has a much larger ACC transport than its high-resolution version (165 vs. 120 Sv, respectively) (Chassignet et al., 2020). The observational estimates for the modern-day transport range from 134 ± 15 Sv (Whitworth & Peterson, 1985) to 173 ± 11 Sv (Donohue et al., 2016), indicating that both LR and HR ACC transports are within these observational ranges, albeit the HR value is close to the lower limit of the observed estimates. Interestingly, both model simulations exhibit stronger-than-observed zonal wind stress at these Southern Ocean latitudes (Small et al., 2014), implying that the ACC transports could be even lower with more realistic zonal wind stresses in our simulations. Furthermore, comparisons of SSH distributions from HR and LR to available observational estimates (not shown) reveal a mixed picture, meaning that the representation of SSH is not necessarily better in HR than in LR.
In both HR and LR, the ACC timeseries show rather rich variability on interannual to centennial time scales with peak-to-peak changes of order 10–15 Sv. While there is little evidence of a trend in the HR ACC transport, the LR ACC experiences a drop of about 5 Sv between years 300 and 325. In Figure 9b, the lead-lag correlations between the ACC transport and SHice which is closely related to polynya variability in HR, are presented to investigate whether there is a causal link between the ACC transport and polynyas. The figure shows statistically significant, but relatively weak, correlations when the SHice leads by about 10 years and, also, when the ACC transport leads by about 25–30 years.
3.3 Relationship Between SST and Surface Turbulent Heat Fluxes
Previous studies showed that increasing ocean model resolution enhances SST variability, which is predominantly attributed to a larger upper-ocean heat flux convergence induced by the resolved mesoscale ocean eddies (e.g., Kirtman et al., 2012; Small et al., 2014). Both observational and model-based results indicate that the eddy-induced SST fluctuations can be of sufficient magnitude and persistence to have an impact on air-sea exchange of heat and moisture via latent and sensible turbulent heat fluxes (THF), particularly over eddy-energetic regions coinciding with extratropical SST fronts, such as the ACC and the seaward extensions of western boundary currents (Bishop et al., 2017; Ma et al., 2016; Putrasahan et al., 2017; Small et al., 2019, 2020; Wu et al., 2006). This ocean mesoscale eddy-atmosphere (OMEA) interaction contrasts with the traditional view in climate dynamics that SST variability is predominantly driven by the atmosphere, where stochastic THF forcing drives a red spectrum SST response (e.g., Barsugli & Battisti, 1998; Hasselmann, 1976). There is increasing evidence that ocean-eddy-driven SST variability can influence variability in both the ocean and the atmosphere via coupling with THF and exert influence on weather and climate (e.g., Kirtman et al., 2017; Ma et al., 2015, 2017; O'Reilly et al., 2016; Siqueira & Kirtman, 2016).
In this subsection, we contrast the variance, power spectral density, and cross-spectral statistics computed using monthly mean SST and THF anomalies from the HR and LR PI-CTRL simulations. Here, THF is defined as positive out of the ocean, and SST and THF anomalies are defined as fluctuations about a long-term mean, a linear trend, and a seasonal cycle composed of annual and semiannual components. SST and THF are both analyzed on the atmospheric model grid from simulation years 1 to 500 (150 to 500) in LR (HR), and the spectral estimates are computed as described in Laurindo et al. (2019). Our objective is to quantify a set of spatial and temporal scales at which resolved SST variability can induce a response in THF.
Figure 10a shows the ratio between the variance of the monthly SST anomalies in HR and LR. In agreement with previous studies, the SST variance in HR is larger than that in LR throughout most of the extratropics, particularly along eddy-rich oceanic regions such as the ACC, the seaward extension of western boundary currents (where the variance in HR can exceed that in LR by a factor of 10), and the interior of the subtropical gyres of all oceanic basins except the North Atlantic (where the variance ratio is close to one). Major upwelling zones also show enhanced SST variability, including the California and Peru-Chile Current Systems in the Pacific and the Guinea and Angola Domes in the Atlantic. Correspondingly, the THF variance ratio (Figure 10b) also reveals an enhanced THF variability in HR over energetic ocean current systems, consistent with a THF response to the increased SST variability (Small et al., 2019). A notable exception to the correspondence between SST and THF is observed within the tropics, where the SST variance in LR is found to exceed that in HR by a factor of up to three while showing a correspondent impact in the THF variance only in the tropical Atlantic. A preliminary analysis indicates that the excessive tropical SST variance in LR occurs over interannual time-scales and longer, although it does not seem to be linked to low-frequency climate models such as the IPO or ENSO as variability related to the indexes of these modes is actually more energetic in HR than in LR (cf. discussion in section 4.2). Although the cause for the larger tropical SST variance in LR is still unclear, we hypothesize that it can be partially attributed to the difference in tropical MLD between HR and LR. MLD in LR is up to 20% shallower than that in HR (cf. discussion in section 4.1 and Figures 14 and 15).
To evaluate how the excessive SST and THF variance in HR relative to LR is distributed across different spatial scales, we computed the power spectral density of SST and THF as a function of zonal wavenumber (k) for both simulations, also contrasting them with the corresponding estimates from the monthly mean J-OFURO3 observational product (OBS) (Kubota et al., 2002; Tomita et al., 2019) for the period of 1988–2013 at 0.25° resolution. Figure 10c exemplifies the results obtained for a zonal section over the Agulhas Return Current (black dashed lines in Figures 10a and 10b), where it is well known that strong OMEA interaction occurs. Similar results apply to the seaward extension of western boundary currents, such as the Gulf Stream and Kuroshio extension. The SST and THF power spectra resolved by all datasets show a shallow plateau for scales larger than ~2,000 km where the power remains relatively constant. At smaller scales, the power begins to diverge between HR and LR. Specifically, the power in LR decreases at an approximate k − 3 rate for scales less than ~2,000 km, while in both HR and observations, the power remains nearly constant until about ~500 km, when it progressively decreases toward smaller scales at a k−4 rate. This indicates that increasing model resolution can affect variability at not only mesoscales (<100 km) but also subbasin scales (~1,000 km).
Figures 10d and 10e show the cross-spectral statistics magnitude-squared coherency (
and absolute phase factor (|θab|) computed between SST and THF as a function of k.
varies from 0 to 1 and gives an indication of how well-matched SST and THF are at a given k, while |θab| shows the corresponding phase relationship, varying from 0° (in-phase) to 180° (opposite-phase). The SST/THF coherency in HR and observations both sharply increase to values larger than 0.9 over scales smaller than ~2,000 km and show an in-phase relationship, although the behavior shown by these datasets diverges for scales smaller than ~500 km in the sense that the coherency in observations decreases to ~0.5 at scales of about 100 km while in HR it decreases at a slower rate until reaching ~0.8 at 50 km. A similar (albeit much more modest) behavior can also be observed in LR, with the SST/THF coherency increasing from near-zero at ~2,500 km to the peak value of ~0.4 between 500 and 1,000 km while keeping a nearly in-phase relationship. The coherencies then drop toward smaller scales until reaching near-zero values at ~250 km. This suggests that OMEA coupling is much stronger in HR than in LR.
To analyze the temporal distribution of the enhanced SST and THF variance in HR and observations, Figure 10f shows the power spectral density of both quantities as a function of frequency (ω) for the same zonal section over the Agulhas Return Current, computed using spatially high-pass filtered data to isolate variability with scales smaller than about 1,000 km. While the SST power resolved by HR exceeds that in observations by about 50% throughout the analyzed frequency range, the THF power in observations is about 60% larger than in HR. In contrast, SST and THF power spectral density in HR and observations both exceed that in LR by two orders of magnitude. It is particularly interesting to note that the SST power spectra display a red noise structure in HR, LR, and observation, as predicted by the theory (Barsugli & Battisti, 1998; Hasselmann, 1976). However, the THF power spectra show a different structure: The LR spectrum shows nearly equal amplitude for all frequencies, indicative of a white spectrum, while the HR spectrum shows a red noise structure, which is consistent with the observations. This indicates that in HR and observations SST can have a strong influence on THF, confirming the presence of strong OMEA feedback.
Regarding the shape of the SST power spectra in Figure 10f, HR and observations indicate that the power remains relatively constant from the longest resolvable periods in each dataset (about 170 and 13 years, respectively) until about 500 days, while in LR, the power decreases by about a factor of five within the same frequency range. Over periods between 60 and 500 days, the SST power in all datasets decreases toward higher frequencies at an approximate ω−5/3 rate. The shape of the THF power spectra in HR and observations is very similar to that of SST, while in LR, the THF spectra show approximately the same power at all frequencies. Figures 10g and 10h show the corresponding SST/THF
and |θab| estimates, revealing that both quantities in HR and observations remain closely matched (>0.8 coherency) and approximately in-phase within the entire analyzed frequency range. Such close relationships are not reproduced in LR, where maximum coherencies of ~0.4 are reached only over periods longer than about 4 × 104 days (>100 years).
These results suggest that the resolved ocean variability can influence the air-sea coupling variability over a surprisingly wide range of spatial scales (from 50 to 2,000 km) and temporal scales (from 2 months to decadal and potentially longer). Research is currently ongoing to characterize the scales where the atmosphere- and ocean-driven air-sea coupling regimes prevail across the global ocean and to determine the physical conditions in the ocean and in the atmosphere that give rise to the observed scale dependency of air-sea interactions.
4 Historical Simulations
4.1 Global Mean Surface Temperature, Mean Climate State, and Seasonal Cycle
Global-mean surface (2-m) air temperature is a key metric for evaluating climate model performance in simulating historical and future climate change. Figure 11a presents a comparison of simulated and observed timeseries of global mean surface air temperatures (SATs) and SSTs, showing that both temperatures are in much closer agreement with the observations in HR than in LR. In particular, the simulated SST in HR falls in-between the two long-term observed SST datasets, HadISST2 (Titchner & Rayner, 2014) and ERSSTv5 (Huang et al., 2017), where the former is warmer by approximately 0.3 to 0.4°C than the latter. The simulated HR SAT shows an even more impressive agreement with the observed surface temperature, GISTEMPv4 (Lenssen et al., 2019). In contrast, the global mean SST and SAT in LR are both about 1.0°C colder than those in HR and observations. The global mean SST (SAT) averaged over the 1870–2019 (1880–2019) period is 18.26°C (14.04°C) for HadISST2 (GISTEMPv4), 17.89°C for ERSSTv5, 18.13°C (14.04°C) for HR, and 17.03°C (12.44°C) for LR, respectively. The Root-Mean-Square Error (RMSE) of HR SST (SAT) is 0.18°C (0.16°C) relative to HadISST2 (GISTEMPv4) and 0.27 to ERSSTv5, while the corresponding RMSE of LR SST (SAT) is 1.23°C (1.61°C) to HadISST2 (GISTEMPv4) and 0.87°C to ERSSTv5, respectively. The colder SST in LR cannot be simply explained by the difference in net surface heat flux into the ocean between HR and LR, because the globally averaged net surface heat flux into the ocean in HR has a value of −0.06 W m−2 compared to +0.26 W m−2 in LR for the period of 1877–2018 (because surface heat fluxes were output from 1877 onwards), indicating that more heat is pumped into the ocean in LR than in HR. Therefore, the warmer SST in HR can only be attributed to differences in oceanic processes between HR and LR.
We hypothesize that the warmer SST in HR is likely caused by an enhanced upward ocean vertical heat transport (OVHT) in the upper ocean. The OVHT consists of vertical turbulent heat transport and vertical heat transport by mean currents and eddies. In both HR and LR, the vertical turbulent heat transport is parameterized by the K-Profile Parameterization (KPP, Large et al., 1994). However, the eddy vertical heat transport is computed differently in HR and LR. In HR, it is explicitly computed, whereas in LR, it is implicitly computed via mesoscale-eddy parameterization (GM, Gent & Mcwilliams, 1990) and submesoscale-eddy parameterization (Fox-Kemper et al., 2008). As pointed out by Griffies et al. (2015), “mesoscale eddies act to transport heat upward in a manner that partially compensates (or offsets) for the downward heat transport from the time-mean currents.” Therefore, any differences in explicit versus parameterized eddy vertical heat transport may have a significant impact on the net OVHT. Griffies et al. (2015) show that the net upward vertical heat transport by mean currents and eddies (including the parameterized submesoscale eddy heat transport in their models) within the top 50 m increased to ~1.6 PW in 0.1° GFDL CM2.6 model that explicitly resolves mesoscale eddies from ~0.2 PW in 1.0° GFDL CM2.0 model that does not resolve eddies and computes mesoscale eddy heat transport using GM parameterization (see Figure 12 of Griffies et al., 2015). It is this increase in upward OVHT due to explicit rather than parameterized eddy heat transport that is conjectured to be a key contributing factor for the warmer SST in HR than in LR.
Figure 11b displays the timeseries of the simulated and observed global-mean surface air temperature anomalies relative to the 30-year mean temperature for the base period of 1951–1980. Despite the significant cold bias in LR, the simulated time evolution of the global-mean surface air temperature anomalies in HR and LR both agree remarkably well with the long record of observed surface air temperature. The projected global SAT in LR and HR also tracks closely with each other. However, this does not mean that there are no major regional differences in the projected SATs. We will discuss these regional differences in section 5.1.
Although the global-mean SST in HR matches well with observations in terms of mean difference and RMSE measures, there are still significant regional biases. For example, SSTs along the Kuroshio and Gulf Stream extension as well as Southern Ocean fronts are warmer than observed (Figure 12b). The severe warm bias in excess of 3°C along the Gulf Stream extension is found in the Northern Recirculation Gyre region in both HR and LR (Figure 12a), suggesting that both simulations suffer from an overshooting Gulf Stream, which leads to the problem of the North Atlantic Current being too zonal in the model. However, compared to LR, the SST bias in HR, particularly the cold bias along the Gulf Stream, is significantly reduced (Figure 12c). Nevertheless, increasing ocean resolution from 1° to 0.1° in HR does not fundamentally improve the simulation of the Gulf Stream and North Atlantic Current system. This finding is consistent with a recent ocean model intercomparison study, which shows that in many models similarly increasing horizontal resolution does not remove the model biases in the representation of the Gulf Stream and North Atlantic Current (Chassignet et al., 2020). Aside from the warm biases, cold biases prevail over the tropical Atlantic and eastern tropical Pacific in HR with an amplitude up to 1°C. The cold bias is particularly pronounced over the tropical Atlantic and potentially can have a negative impact on Atlantic TC simulations (e.g., Hsu et al., 2019). A direct comparison between SSTs in HR and LR clearly shows that the SST is warmer in HR than in LR almost everywhere except along eastern boundary upwelling regimes, such as the California Current, Peru-Chile, and Benguela upwelling systems, where warm biases in LR are reduced in HR (Figure 12c). The mechanism of the warm bias reduction along the eastern boundary regimes has been discussed by Small et al. (2014, 2015), who demonstrate that improved simulation of the atmospheric low-level coastal jets due to the increase in atmospheric model resolution is the primary cause for the bias reduction. Another distinct feature from the HR-LR SST difference map is the fact that strong relative warming tends to occur along strong eddying regions, including the Kuroshio and Gulf Stream extension, the Brazil/Malvinas Current, the Agulhas Current and the ACC, consistent with some high-resolution model simulations (e.g., Gutjahr et al., 2019). This finding reinforces the hypothesis that the warmer SSTs in HR may be partially related to the difference between explicit and parameterized ocean eddy heat fluxes in HR and LR.
Annual-mean precipitation biases in HR and LR are compared in Figures 12d and 12f, computed relative to the GPCPv2.3 (Adler et al., 2018) climatology for the 1979–2018 period. The global mean precipitation is 1.95 mm d−1 for HR, 1.85 mm d−1 for LR, and 1.72 mm d−1 in GPCPv2.3. Precipitation RMSE is 6.53 mm d−1 for HR and 4.39 mm d−1 for LR relative to GPCPv2.3. Therefore, overall precipitation biases are exacerbated by the resolution increase. The overall bias pattern, however, is similar between HR and LR (Figures 12d and 12f), and both display a double ITCZ bias except that this bias is reduced over the eastern tropical Pacific in HR, which is consistent with the previous CESM1.1 simulation results by Small et al. (2014). Also similar to the finding of Small et al. (2014), the northern ITCZ precipitation amount in HR is much too high, indicative of an overactive Hadley Cell in HR. There is also some indication that the position of the ITCZ is shifted northward in HR, particularly over the tropical Atlantic (Figure 12f). Another visible difference in the annual-mean precipitation between HR and LR is the increase of precipitation along the major oceanic frontal zones, such as the Kuroshio and Gulf Stream extension, the Brazil/Malvinas Current, the Agulhas Current, and the ACC, where warmer SSTs are seen in HR (Figures 12c and 12f), consistent with the finding of Kirtman et al. (2012). Over some continental areas, such as the Amazon basin, precipitation is considerably increased in HR, acting to reduce the regional dry bias in LR. Other areas where precipitation bias is reduced in HR are the western United States and central and western China where a wet bias exists in both HR and LR. Overall, the global mean precipitation in HR is increased by a relatively small amount (~5%) compared to LR, despite HR being considerably warmer than LR (Figure 12c). However, the partitioning of the parameterized convective precipitation and resolved large-scale precipitation shows a significant difference between HR and LR. In LR, about 67% of the total precipitation is attributed to the convective precipitation, while this number drops to only 49% in HR. Therefore, there is an increase in the resolved large-scale precipitation as a result of model resolution increase from 33% in LR to 51% in HR (Figure 13). These findings are in line with some recent studies showing that increasing atmospheric model resolution can lead to an increase in resolved large-scale precipitation, contributing to an increase in precipitation extremes (Kooperman et al., 2018; O'Brien et al., 2016; Rauscher et al., 2016). Rauscher et al. (2016) argue that the intensified large-scale precipitation as grid spacing decreases is driven by resolution-dependent increase in updraft strength. This difference in the convective versus large-scale precipitation partitioning has implications for the simulation of extreme events in HR and LR, as will be discussed later.
In addition to the difference in the mean climate state, there is a marked difference between the seasonal cycle of SST in HR and LR as measured by the peak-to-peak value of the seasonal cycle (Figure 14). This peak-to-peak value is obtained by first computing the monthly climatology of SST at each ocean grid point over the 1870–2018 period and then identifying the warmest and coldest SST and the values before and after the warmest and coldest month in the climatology and finally taking the difference between the average of the three warm month SSTs and the three cold month SSTs. It is evident that in many parts of the global ocean, particularly the Southern Ocean, the seasonal SST variation is weakened by more than 50% due to the increase in model horizontal resolution. The globally averaged peak-to-peak SST seasonal variation is 2.96°C for HR and 3.19°C for LR, respectively. The weaker SST seasonal cycle in HR is in a better agreement with the observations, especially in the Southern Ocean (Figures 14d and 14e). The globally averaged RMSE of the peak-to-peak SST seasonal variation in HR and LR relative to HadISST2 is 0.78°C and 0.82°C, respectively. We also computed the SST seasonal cycle amplitude using a harmonic analysis, and the results are consistent with the analysis shown here. This improved SST seasonal cycle in HR may be partially related to the improvement in the model MLD which will be discussed next.
MLD, as an indicator of the extent of ocean vertical mixing, is an important variable in climate studies, acting to modulate air-sea coupling and heat uptake, and influencing the seasonal cycle of fields like SST. We compare simulated and observed MLD using the definition described in Large et al. (1997) that identifies the shallowest depth at which the local density gradient is as strong as the largest bulk density gradient (the value of the largest density gradient from the surface to some depth). Although other MLD definitions exist (e.g., see Holte & Talley, 2009, for a method and review), we find this density gradient method a useful approach to compare models and observations across seasons, noting that Whitt et al. (2019) applied the same method to Argo data which we use as an observational benchmark.
Figure 15 shows a comparison of MLD among HR, LR, and Argo-based estimates. The period used for the simulated and observed MLD is 2006–2020 and 2004–2017, respectively. Slightly different periods considered in the comparison are due to the fact that MLD was not saved properly for HR before 2006 and Argo data are available from 2004 to 2017. In summer (defined as July-August-September [JAS] and January-February-March [JFM] mean for NH and SH, respectively, representing the lag between solar forcing and ocean response), the MLD is generally shallow (a few meters to 100 m) and in the midlatitudes is mainly governed by the seasonal and diurnal cycles of solar insolation and the strength of wind (and wave) forcing. Some of the deepest MLDs in summer are in the Southern Ocean underneath the atmosphere storm track, which is much stronger than the NH summer storm tracks. HR tends to capture these deeper mixed layers more realistically than LR (Figure 15 left), possibly due to an enhanced storm track because of the increase in atmosphere resolution. The deeper MLD gives rise to a smaller SST seasonal cycle.
More significant differences between HR and LR occur in the winter season. The winter MLD (JFM and JAS for NH and SH, respectively) in mid-to-high latitudes is driven by a complex interaction of surface buoyancy loss, strong winds and waves, and is influenced by advection by ocean currents and eddies as well as the presence of SST gradients. North of the sub-Antarctic Front in the Southern Ocean deep mixed layers form in response to strong surface heat loss. The surface heat loss is generally better represented in models with more realistic SST fronts, such as HR. Further, a strong surface fresh bias in LR is reduced (but not removed) in HR, due to changes in pathways of salinity transport (improvements in Agulhas retroflection, etc.). As a consequence, a very shallow mixed layer in the sub-Antarctic zone is found in LR, while HR captures the main features seen in observations (Figure 15, right). Another distinct feature in HR is the very deep MLDs along the southern edge of the Weddell Gyre (Figure 15e), which is associated with the polynyas. Other deep MLDs are seen around the north and east rim of the North Atlantic subpolar gyre including the Labrador Sea and Greenland-Iceland-Norwegian (GIN) seas during boreal winter. LR tends to have too deep MLD in a too-broad area (Figure 15 right), whereas HR has a more confined region of less deep MLDs which are in reasonable agreement with observations. This is particularly the case in the Labrador Sea, where ocean eddies are critical in setting the location of deep mixing (deep mixing only occurs away from strong eddies which act to restratify the water column), giving rise to the deepest mixing only in a small patch in the southwest Labrador Sea. This is seen in observations and HR, but not in LR.
Simulated annual-mean sea-ice concentrations in HR and LR are shown in Figure 16 along with the observed sea-ice concentration from the National Snow and Ice Data Center (NSIDC; Cavalieri et al., 1996) in the NH and SH for the period of 1979–2018. Overall, HR tends to underestimate the sea-ice concentration more than LR does in both hemispheres. This is particularly evident along the southern Weddell gyre where a tongue of low sea-ice concentration region is reminiscent of polynyas shown in Figure 3b. This is consistent with the finding that HR has an overactive polynya in the Weddell gyre (the polynya was rarely observed in the period of 1976–2015). In contrast, LR overestimates the sea-ice concentration along the southern Weddell gyre, consistent with a lack of polynyas in LR. The underestimation of sea-ice in HR is further confirmed by the sea-ice extent seasonal cycle shown in the bottom panels of Figure 16. In the NH, the amplitude of the sea-ice extent seasonal cycle is well simulated by both HR and LR, but HR has a negative mean bias that corresponds to a 20% underestimation of the sea-ice extent. In the SH, however, HR shows not only a negative mean bias but also an underestimation of the amplitude of the seasonal cycle. Overall, the agreement between LR and observations is better. However, larger sea-ice extent in LR is likely due to colder surface temperatures in comparison to both HR and observations. Nevertheless, there is a significant overestimation of sea-ice extent by nearly 50% during austral summer in LR and neither LR nor HR is able to properly simulate the rapid rate of decline of the sea-ice extent during the SH melting season. The differences in the NH and SH sea-ice simulations between HR and LR may have an important implication for the degree of polar amplification simulated by the two models in response to projected anthropogenic forcing, which will be further discussed in section 5.1.
4.2 Modes of Climate Variability
The three most dominant modes of climate variability on interannual to multidecadal time scales based on available long historical SST records are El Niño-Southern Oscillation (ENSO), Pacific Decadal Oscillation (PDO), and AMV (Tung et al., 2019). We examine the fidelity of HR and LR HF-TNST in reproducing these observed modes of climate variability.
ENSO is the most dominant mode of climate variability on interannual time scales and is an important gauge for a climate model. As shown in Small et al. (2014), increasing CESM1.1 horizontal resolution leads to a more realistic ENSO amplitude compared to the standard resolution CESM1.1 simulation that displayed an over-energetic ENSO cycle. The high-resolution CESM1.3 simulation confirms that ENSO amplitude in HR HF-TNST is realistically simulated. From Figure 17, it is evident that the amplitude of Nino3.4 in HR is comparable to that of ERSST and HadISST2. The standard deviation for the detrended Nino3.4 during the 1920–2019 period is 0.75°C for HR, 0.63°C for LR, 0.75°C for ERSST, and 0.73°C for HadISST2. We note that the Nino3.4 indexes shown in Figure 17 are not detrended, so that secular changes in the simulations can be compared to the observations. Spectral analysis shown in Figure 17 further confirms the close agreement in Nino3.4 amplitude between observations and HR. It also reveals that the simulated ENSO in HR has two dominant peaks, one centered around 3 years and the other over the range of 5–10 years. The observed Nino3.4 SST spectrum shows multiple spectral peaks between 2 and 10 years, one between 2 and 3 years, the second between 3 and 4 years, and the third between 5 and 6 years. Therefore, the simulated ENSO cycle in HR tends to have a longer period than the observed. A visual comparison between the Nino3.4 SST anomalies in HR and observations also indicates that the simulated El Niño in HR has a tendency to persist too long (Figure 17), which is confirmed by auto-correlation analyses (not shown). Also, it is worth noting that the warming trend of ~0.77°C per century in HR that is in an excellent agreement with the warming trend of 0.78°C per century in ERSST Nino3.4 index, but considerably greater than the trend of 0.12°C per century in HadISST2. The uncertainty in the observed equatorial Pacific SST trend has long been debated in the literature (e.g., Deser et al., 2010).
In section 3, it was shown that a super-charged IPO dominates global mean SST variability at multidecadal time scales in HR PI-CTRL (Figure 2). This is reflected in the Nino3.4 SST spectrum in HR PI-CTRL (Figure 17e), where there is a stronger-than-observed spectral peak, albeit statistically insignificant, at about 40 years. Such a multidecadal spectral peak is absent from HR HF-TNST where the low-frequency Nino3.4 SST variation is dominated by a warming trend (Figure 17c), indicating a much weaker IPO in HR HF-TNST. However, the Southern Ocean sea-ice variability still displays a prominent multidecadal oscillation during the historical period of 1850–2018 in HR HF-TNST (Figure 2a), indicating a similar multidecadal Weddell polynya variability as in HR PI-CTRL. But unlike HR PI-CTRL, the correlation between the global mean SST and Southern Ocean sea-ice variability is much weaker in HR HF-TNST and is below the statistical significance level. Furthermore, the maximum correlation no longer occurs when the Southern Ocean sea-ice leads the global-average SST, but occurs at lag = 0 (Figure 2b), suggesting that Weddell polynyas no longer play an active role in multidecadal global SST variability in HR HF-TNST. The absence of a strong IPO with a connection to Southern Ocean sea-ice variability in HR HF-TNST may be attributed to the external climate forcing that produces secular changes in global SST (Figure 11), disrupting the super-charged IPO cycle of HR PI-CTRL. It is also interesting to note that with the presence of the multidecadal peak in HR PI-CTRL, the ENSO spectral peak at 5–6 years is weaker than that in HF-TNST, raising the possibility that some of the IPO energy may be merged into the stronger and broader ENSO peak between 5 and 10 years in HR HF-TNST through nonlinear interactions between ENSO and IPO. However, we caution that spectral analysis is sensitive to time-series record length and the ENSO spectral characteristics in Figure 17 can be subject to uncertainties (e.g., Wittenberg, 2009). Further studies are clearly required to understand the relationship among ENSO, IPO, and tropical Pacific warming trend.
In contrast to the finding of Small et al. (2014), ENSO in LR HF-TNST shows weaker amplitude than that in HR and in observations (Figure 17). The Nino3.4 spectrum in LR has a broad and weak peak between 3 and 5 years and another weak peak around 10 years. Compared to the spectral peaks in HR HF-TNST where the low-frequency peak (5–10 years) dominates over the high-frequency peak (3 years), the low-frequency peak (10 years) in LR HF-TNST is only slightly stronger than the 3- to 5-year peak and barely passes the statistical significance, indicating a much weaker low-frequency ENSO variability in LR. The Nino3.4 SST in LR also shows a warming trend of 0.73°C per century, similar to that in HR and ERSSTv5, but as in HR HF-TNST, no multidecadal peaks are found. Interestingly, in LR PI-CTRL, the Nino3.4 spectrum shows a stronger single spectral peak at around 5 years, which is close to the observed spectral peak (Figure 17). Nevertheless, we conclude that increasing CESM horizontal resolution does not lead to fundamental improvements (or differences) in ENSO simulations. This conclusion deviates from Small et al. (2014) but is consistent with a recent study by Caldwell et al. (2019), which used the Energy Exascale Earth System Model version 1 (E3SMv1) with a similar SE-dycore CAM-based atmospheric component model but a different ocean component model. We speculate that this difference between our study and Small et al. (2014) may be attributable to the differences in the atmospheric component model between CESM1.1 and CESM1.3 (see section 2.1) as well as to the simulation lengths in both studies that may be too short to robustly document ENSO variability characteristics (e.g., Wittenberg, 2009).
Bjerknes feedback is key to ENSO. Its strength can have a significant impact on ENSO's characteristics. There are three essential elements to Bjerknes feedback: (1) feedback between SST and zonal wind stress, (2) feedback between zonal wind stress and thermocline, and (3) feedback between thermocline and SST (Keenlyside & Latif, 2007). We analyze and compare these three feedback processes in HR, LR, and observations. Figure 18 shows that HR and LR share more similarities between them than with observations. A major difference between HR and LR is that in the former the region of maximum zonal wind-SST feedback (element 1) is moved further eastward with a higher strength compared to that in the latter. Compared with the observations, the feedback between thermocline and SST (element 3) is not only underestimated in strength by both HR and LR but also too confined in the eastern equatorial Pacific. In contrast, the feedback between zonal winds and thermocline (element 2) in HR and LR does not extend further enough into the eastern equatorial Pacific. These structural differences in Bjerknes feedback are likely to be responsible for systematic biases, such as westward shift of maximum SST anomalies (not shown), in simulated ENSO in HR and LR.
PDO is predominantly an ocean mixed layer response to surface heat flux forcing induced by variability of the Pacific-North-American (PNA) circulation pattern, which is the most dominant pattern of atmospheric circulation variability over the North Pacific sector (Battisti et al., 2020; Wills et al., 2018), although there is also a component that is linked to ENSO variability (Newman et al., 2016). Figure 19 (left) shows the PNA captured by the leading EOF of December-January-February (DJF) sea-level pressure (SLP) anomalies over the North Pacific sector (120°E to 120°W, 20°N to 85°N) in observations, HR, and LR, respectively, and the corresponding February-March-April (FMA) SST regressions onto the respective leading Principal Components. As expected, the resultant SST regression patterns closely resemble the PDO pattern (Mantua et al., 1997; Newman et al., 2016). The 2-month lag used for SST regression is to take account of the fact that PNA tends to lead PDO by about 2 months (Newman et al., 2016). There is obviously an overall agreement between observed and simulated PDO/PNA in both HR and LR. The position and shape of the Aleutian Low anomaly that characterizes the PNA are particularly well simulated in HR, although the amplitude and explained variance are slightly higher than the observed values. In LR the shape of the Aleutian Low anomaly is more elongated, and its amplitude is weaker compared to observations, but the explained variance is comparable to that of HR. The explained variance is ~53% in HR, ~43% in LR, and ~44% in the observations, respectively, for the period of 1920–2018. The regressed SSTs also display a better agreement between HR and the observations, both of which show a negative SST anomaly extending along the northern flank of the Kuroshio Extension where mesoscale eddy-induced SST variability is strong (Jing et al., 2019; Ma et al., 2015) and a narrow coastal warming zone extending along the coast of Alaska and Canada. These features are less well represented in LR.
In contrast to PNA and PDO, large discrepancies between HR and LR are found in representing the North Pacific Oscillation (NPO)—the second EOF of winter SLP variability in the North Pacific (see Linkin & Nigam, 2008, and references therein)—and the corresponding SST response. Figure 19 (right) compares the observed and simulated NPO and the associated SST response in HR and LR to those of observations where the NPO is characterized by a north-south dipole of winter SLP anomalies centered over the PNA pattern represented by EOF1. The associated surface heat flux anomaly (not shown) has a tripolar structure that drives an SST response. In particular, the southern lobe of the NPO can exert an influence on the Pacific trade winds that in turn generate surface heat flux anomalies and trigger wind-evaporation-SST (WES) feedback (Amaya et al., 2017; Chang et al., 1997; Xie & Philander, 1994) in the subtropics of the North Pacific. The resultant SST response is termed as the Pacific Meridional Mode (PMM; Chiang & Vimont, 2004), which has been shown to have an impact on ENSO (Chang et al., 2007; Chiang & Vimont, 2004; Di Lorenzo et al., 2015). From Figure 19, it is evident that a dipole-like SLP anomaly that resembles the observed NPO is present in HR with the exception that the center of the southern lobe in the model is shifted somewhat eastward and its strength is overestimated. The associated SST response in HR is also in a reasonable agreement with the observation in the North Pacific. In contrast, the NPO SLP anomaly in LR is very weak and does not resemble the observed NPO well. The corresponding SST response also bears little resemblance to the observed SST regression pattern. Since the NPO and related PMM variability are important in linking extratropical North Pacific variability to ENSO (Chang et al., 2007; Chiang & Vimont, 2004; Di Lorenzo et al., 2015), we conclude that increasing model resolution does lead to improvements in simulating North Pacific climate variability as a whole. We further conjecture that the weak NPO in LR may contribute to the weak ENSO in LR, while the stronger-than-observed NPO variability may be partially blamed for the enhanced low-frequency ENSO variability in HR. Some recent studies show that Kuroshio SST front and eddies can have an impact on North Pacific storm track variability (Foussard et al., 2019; Kuwano-Yoshida & Minobe, 2017; Ma et al., 2015, 2017; O'Reilly & Czaja, 2015). To what extent OMEA interaction along the Kuroshio Extension region contributes to the difference in PNA and NPO between HR and LR requires further in-depth analyses.
Long observed temperature records reveal strong multidecadal variability concentrated in the North Atlantic at time scales of roughly 50–90 years (Delworth & Mann, 2000; Tung & Zhou, 2013). The characteristics of observed AMV are sensitive to the method used to isolate it from the externally forced warming which dominates the global temperature record beginning in the early to mid-twentieth century (Frankignoul et al., 2017; Ting et al., 2009; Wu et al., 2011). For a preliminary look at AMV in HF-TNST simulations, we adopt the method of Trenberth and Shea (2006) and examine a relative North Atlantic warming index (AMV*; defined as the area-averaged SST in the domain 75–7.5°W, 0–60°N minus globally averaged SST over the domain 60°S to 60°N). The resulting AMV* timeseries from HR and LR are compared to ERSSTv5 observations in Figure 20d. Both simulations exhibit anemic AMV* compared to observations, but the amplitude discrepancy is improved by roughly 33% in HR (low-pass-filtered AMV* standard deviations in ERSSTv5, HR, and LR are 0.14°C, 0.08°C, and 0.05°C, respectively). Furthermore, there is a striking correspondence with the observed phasing of AMV* in HR HF-TNST (but not LR HF-TNST) starting around 1940 that hints at a role for external forcing in setting the timing of AMV* phase transitions in the late twentieth century, as has been argued in some studies (e.g., Booth et al., 2012). The relative contributions of internal versus externally forced variability in HR HF-TNST remain impossible to ascertain, but a planned ensemble of HR HF-TNST simulations should clarify the relevant mechanisms.
The global SST regression onto the observed (low-pass filtered) AMV* index reveals a familiar horseshoe pattern of warming over the North Atlantic with maximum amplitude in the subpolar gyre region (Figure 20c). The HR HF-TNST regression has a similar pattern and amplitude in the North Atlantic, but it also shows significant signals in the tropical and western Pacific that are only weakly present in the observed data (Figure 20a). This could be related to the use of the Trenberth and Shea (2006) method, which can give spurious signals in the Pacific (Frankignoul et al., 2017). The regression pattern in LR HF-TNST is much less realistic than in HR HF-TNST, with very weak amplitude in the subpolar Atlantic and excessive variability in the Greenland and Norwegian Seas. Overall, the representation of AMV appears much improved in HR over LR in terms of amplitude, time scale, and pattern.
A quantitative comparison of the model AMOC profiles with the profile based on the RAPID data at 26.5°N is provided in the left panel of Figure 21. The HR and LR profiles represent time means for the 1986–2005 period, and the RAPID profile is for the April 2004 to September 2018 mean. The use of the 1986–2005 period, rather than the same RAPID period, from the simulations is to avoid the influence of the RCP8.5 forcing on the simulated AMOC, which starts from 2006. In the upper 1,000 m, both HR and LR transports are in very good agreement with the RAPID profile with both model transports of about 18 Sv within the observational range of 17 ± 3.3 Sv where the range represents one standard deviation (Frajka-Williams et al., 2019; Smeed et al., 2018). Below about 2-km depth, LR profile is in better agreement with the observations. In particular, the penetration depth of the NADW as measured by the depth of the zero-crossing is only slightly deeper than 3,000 m in HR, but about 4,000 m in LR, which is much closer to a penetration depth of about 4,500 m in RAPID. The deeper penetration of NADW in LR is due to the overflow parameterization used for the Nordic Sea overflows (Danabasoglu et al., 2010), which is not employed in HR. As such, HR suffers from this shallow bias common to many coarse and high-resolution models without proper representations of the overflows, especially in level coordinates (e.g., Danabasoglu et al., 2014; Gutjahr et al., 2019; Roberts, Jackson, et al., 2020). The signature of the AABW—the negative transport below about 4,500-m depth in RAPID—is rather weak at this latitude. In LR, AABW is not really present when zonally integrated as done here. In contrast, a maximum AABW transport of about 4 Sv in HR appears to be too strong in comparison with the observations. The Atlantic Ocean time-mean meridional heat transport distributions from HR and LR for the same 20-year period as above are presented in Figure 21 (right). For comparison, the figure also includes the implied transport estimates from Large and Yeager (2009) calculated using the Coordinated Ocean-ice Reference Experiments phase II (CORE-II) inter-annual fluxes and observed SSTs and sea-ice for the 1984–2006 period as well as the direct estimates with their uncertainty ranges from Bryden and Imawaki (2001) and the estimate from the RAPID data (Johns et al., 2011) for the April 2004 to February 2017 period. The maximum heat transport of 1.2 PW around 26°N in HR is in remarkably good agreement with the RAPID estimate. In contrast, the LR maximum heat transport of ~1.05 PW is near the lower range of these observational estimates. While south of the equator, both LR and HR heat transports are very similar to each other, both near the observational mean transports, HR heat transport north of the equator is larger than in LR. Indeed, HR transport is even larger than the implied transport range between about 30° and 60°N. Because the AMOC transports are very similar between HR and LR, the larger heat transport in HR is due to its vertical temperature structure with a warm (cold) bias in the northward (southward) flowing upper (lower) branch of the NADW cell (Figure 1c). The opposite and larger bias structure in LR is responsible for its lower heat transport (Figure 1d).
We end this subsection with a brief comparison between observed and simulated intraseasonal variance in HR and LR. Of particular interest is the Madden-Julian Oscillation (MJO), which is the dominant mode of intraseasonal (30–90 days) variability within the tropical atmosphere. Figure 22 shows the Wheeler-Kiladis zonal wave number-frequency power spectra (Wheeler & Kiladis, 1999) of equatorially symmetric daily outgoing longwave radiation (OLR) from observations, HR and LR. The observations clearly reveal a strong variance within the 30–90 day frequency band and eastward zonal wave numbers 1–2, denoting the MJO. The simulated MJO variance in both HR and LR is a significantly weaker than observed, despite of some apparent similarity in frequency and wave number characteristics. The simulated MJO in HR is about 30% weaker than the observed MJO and is only slightly stronger than that in LR, indicating that increasing horizontal resolution alone does not lead to significant improvements in MJO simulations. Recent modeling studies suggest that MJO simulations can be sensitive to physics parameterizations in atmospheric models. CESM2, which includes many new changes to atmospheric physics parameterizations relative to CESM1, shows a more realistic MJO simulation than CESM1 (Danabasoglu et al., 2020). Hannah et al. (2020) also report a significantly improved MJO simulation by using Super-Parameterized Energy Exascale Earth System Model (SP-E3SM) that embeds a 2-D cloud resolving model (CRM) within each grid column of the atmospheric component model to replace a conventional convective parameterization. Both HR and LR also underestimate the variance of the observed equatorial Kelvin waves, although the Rossby wave variance is more realistically simulated in HR than in LR (Figure 22).
4.3 Extreme Events
Extreme events often depend strongly on local physical conditions and small-scale processes, such as convection and interaction with steep topography, which are not well resolved by coarse resolution climate models. Therefore, we expect that increasing model resolution will lead to improvement in the simulation of extreme events. Here, we consider two phenomena: (1) TCs during warm seasons and (2) ARs during cold seasons, both of which are well known for their ability to produce extreme precipitation. Figure 23 shows global TC tracks simulated by HR and LR during 1877–2018 (because 6-hr variables were output from 1877 onwards as noted in section 2.2) and their comparison to observations. The international best track archive for climate stewardship (IBTrACS) dataset (Knapp et al., 2010) for the period of 1950–2018 was employed for the observed TC tracks. An observed TC is defined as having a 1-min maximum sustained wind speed of 34 kt (17.5 m s−1) or higher. Simulated TCs are tracked using TempestExtremes algorithm (Ullrich & Zarzycki, 2017; Zarzycki & Ullrich, 2017) with six-hourly model output data.
The observed, global average annual number of TCs for the 1950–2018 period is about 82 per year, and the corresponding values are 112 and 25 per year, respectively, in HR and LR HF-TNST. Figure 23 shows that the annual number of TCs in LR is severely underestimated in all basins, except for the South Tropical Atlantic (STA). The North Tropical Atlantic (NTA) is particularly challenging, where the number of TCs in LR is less than 1 per year, compared to 12 per year in observations. This large negative bias in LR is consistent with previous studies (Camargo, 2013; Tory et al., 2013). Even in HR, the annual NTA TC number is only about half of that in the observations. A recent study by Roberts et al. (2020b) suggests that this persistent low bias is a common problem to many high-resolution climate models. In addition to the underestimation of TC numbers, there is a clear bias in NTA TC development in HR, where strong TCs, instead of forming in the deep tropics, tend to occur in the higher latitudes. In terms of the annual mean TC number, TCs in the Western North Pacific (WNP) in HR show the closest agreement with the observations, while the numbers in the other basins, that is, the Eastern North Pacific (ENP), North Indian Ocean (NIO), South Indian Ocean (SIO), South Tropical Pacific (STP), and perhaps most notably STA, all show an overestimation in HR. TC strength is also underestimated by both HR and LR, but much less severely in HR than in LR. While LR simulates only a few category-1 TCs and no TCs stronger than category-1, HR is able to produce a small number of category-4 TCs but nearly no category-5 TCs. In spite of these issues, it is unequivocal that HR improves overall TC representation over LR. In a recent HighResMIP study, it is shown that CESM is one of the models whose TC simulations are highly sensitive to model resolution (Roberts et al., 2020b). In fact, the CESM HighResMIP contribution has the largest number of TCs among all the HighResMIP models; the corresponding low-resolution CESM contribution is among those that produced the least number of TCs, based on the same TempestExtremes tracking algorithm. As noted by Roberts et al. (2020a, 2020b), TC identification and tracking can be sensitive to various tracking algorithms, but we do not think that these general results will be impacted.
It is well known that TCs are modulated by modes of climate variability. Among the most studied is the effect of ENSO on TCs (see Lin et al., 2020, for a review and references therein). Figure 24 shows regressions of TC track density (annual total TC numbers passing through each 4° × 4° box) anomalies onto NINO3.4 SST index from 1950 to 2018 in observations and in HR and LR HF-TNST. During the warm phase of ENSO, TC activities are suppressed in NTA due to the increase of the vertical wind shear (Lin et al., 2020), but increased in WNP. This feature is well captured by HR. In contrast, LR is unable to show any significant TC response signal in the NTA and ENP, because the number of TCs in these regions is severely underestimated. Both models seem to demonstrate some skill in simulating TC reduction in SIO during El Niño and an equatorward shift of TC activities in STP.
AMV is also known to have an impact on Atlantic TCs. Figure 24 also shows regressions of TC track density anomalies onto the AMV* index (Figure 20d) from 1950 to 2018 in observations and from 1877 to 2018 in HR and LR HF-TNST. As can be seen, in both observations and HR, there is an increase of Atlantic TC activities during the positive phase of AMV (Chylek & Lesins, 2008; Goldenberg et al., 2001), although the amplitude of TC response in HR is considerably weaker than that in observations, possibly due to the combined factors that both AMV variability and NTA TCs are underestimated in HR. As expected, an even weaker response in NTA TCs to AMV is found in LR because of even weaker TC activities in NTA and the absence of multi-decadal SST variation in the North Atlantic in LR (Figure 20d). Outside of the North Atlantic sector, the observed TC response to AMV shows an overall decrease in TC activity in all the basins. Some hint of this decrease is seen in HR, particularly in the North Tropical Pacific, but in LR, the response shows no coherent structure. We caution that these analyses are based on short records. Considering the multidecadal time scale of AMV, the observed record from 1950 to 2018 barely covers one complete cycle of AMV.
ARs are synoptic-scale features characterized by long and narrow corridors of intense lateral water vapor transport in the lower troposphere from the tropics to extratropics. Occurrence of ARs is often associated with extratropical cyclones during cold season, although the relationship between the two is highly complex and only a fraction of extratropical cyclones is found to coexist with ARs (see Payne et al., 2020, for a review and references therein). The synoptic nature of ARs raises a question of whether they are adequately represented in climate models. A few previous studies have attempted to address this question using atmosphere-only models (e.g., Hagos et al., 2015). Here we compare some simple AR statistics from HR, LR, and observations.
Observed ARs are derived from ERA-5 analysis dataset (Hersbach et al., 2020) for the period between 1979 and 2005, based on tracking long and narrow closed areas of daily mean integrated vapor transport (IVT) anomalies of greater than a threshold IVT value of 250 kg m−1 s−1 for all seasons over the entire globe. The anomalies are defined as departures from the climatological mean over the analysis period. Closed contours of the threshold value are defined as ARs' outer edge and a closed area must be longer than 2,000 km and narrower than 1,000 km to be classified as an AR (Gimeno et al., 2014; Zhu & Newell, 1998). The same time period and tracking algorithm are applied to HR and LR to identify simulated ARs. We are fully aware of the fact that AR tracking can be sensitive to various tracking algorithms and the simple IVT-magnitude threshold method used here has its limitations in tracking ARs globally (e.g., Shields et al., 2018; Xu et al., 2020). However, the threshold method will be sufficient for our present intent of simply illustrating some of the major AR differences between HR and LR in contrast to observations. More detailed AR tracking sensitivity studies are planned in the future.
Figure 25 (left) shows the mean IVT value carried by all the detected ARs over the globe in the observations, HR, and LR, respectively. It is evident that the mean IVT is much more realistically simulated in HR than LR. Compared with the observations, the IVT in HR has similar spatial structures and amplitudes in both the NH and SH, except that the IVT strength is slightly underestimated by about 12% in the North Pacific, but overestimated by about 15% in other basins in HR. In contrast, IVT values in LR are severely underestimated in all the AR active regions with a threefold underestimation in the Northwestern Pacific. Consistent with the higher IVT, mean precipitation concurrent with ARs in HR is also much higher than in LR and closer to the observed precipitation (Figure 25 middle). In AR active regions, the mean precipitation in LR is less than one third of the value in HR. In the figure, the observed precipitation is taken from the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Climate Data Record (PERSIANN-CDR; Ashouri et al., 2015) data set that provides daily rainfall estimates at a spatial resolution of 0.25° in the latitude band 60°S to 60°N from 1983 to 2005. ARs are well known for their impact on precipitation along the West Coast of North America, causing major flood events in the region. Figure 25 (right) shows a zoom-in of precipitation along the West Coast of North America in observations, HR, and LR, respectively. Clearly, high precipitation along the coast caused by topographic lift is much more realistically represented in HR than in LR. The improved simulation in HR is likely to come from both the improved IVT and better resolved topography in HR (Figure 25 right). However, a more detailed analysis of ARs and related precipitation in HR and LR and the potential cause for the differences warrant future studies. Here, we simply conclude that simulations of both ARs and the associated precipitation are significantly improved by the increase in model resolution. Much of AR-related precipitation is caused by resolved large-scale precipitation (not shown). Therefore, the improvement of AR-related precipitation is closely linked to the increase in the resolved large-scale precipitation in HR as previously discussed.
5 Future Climate Simulations
5.1 Polar Amplification and MOC Response
The climate response to anthropogenic forcing is expected to be characterized by enhanced warming over the poles (polar amplification) as a result of albedo feedbacks associated with sea-ice loss (Hall, 2004), but significant uncertainty remains regarding the magnitude of such an effect and to what extent other processes (such as AMOC decline) might offset the warming at high latitudes. Figure 26 shows the time evolution of zonal-mean SAT as a function of latitude and time in HR and LR HF-TNST and their relative difference. Although both simulations reveal a clear polar amplification, particularly in the Arctic region, that appears to be related to a steady decline in sea-ice area, HR shows a much stronger warming (~25% greater than LR; see Figure 26c) over the Arctic cap. We can surmise that this is linked to the more rapid loss of Arctic sea-ice in HR, which is presumably related to the lower overall sea ice coverage in HR compared to LR early in the 20th century (Figures 16 and 26). Furthermore, there is roughly 10% more warming in HR at all latitudes north of 20°S by the end of the 21st century. There appears to be a different sensitivity to resolution in the SH, with HR showing up to ~40% less warming poleward of 50°S. A caveat here is that there are large internal variations in these single-member projections, so some of these differences between the LR and HR simulations may not be significant.
A map of future minus historical surface temperature change (Figure 27) reveals that the enhanced warming in HR at high northern latitudes is primarily associated with stronger warming over historically ice-covered ocean regions: the Arctic Ocean, Baffin Bay, and parts of the Labrador Sea. A notable exception is the Barents Sea, which warms much less in HR than LR, probably because of large differences in present-day sea-ice coverage in this region (Figure 16). At lower latitudes in the NH, enhanced warming in HR is found over the North Pacific and Bay of Okhotsk, as well as across a broad swath of the Eurasian continent. In the SH, HR exhibits systematically less warming over the entire Southern Ocean, with large reductions in warming relative to LR along the Antarctic continent in the east Pacific and Atlantic sectors. This difference may be possibly ascribed to the difference in the Weddell polynyas between HR and LR. Both simulations show a relative lack of warming in the subpolar North Atlantic. The North Atlantic “warming hole” has been associated with long-term forced AMOC decline (e.g., Drijfhout et al., 2012; Rahmstorf et al., 2015), although the relative role of AMOC compared to other possible drivers remains a topic of ongoing debate (Keil et al., 2020).
In response to RCP8.5 forcing, the AMOC in HR HF-TNST (quantified using the MOCNA index; see section 3.2 and Figure 28) shows a significant weakening trend at a rate of nearly 8 Sv per century over the period of 2000 to 2100 (Figure 28a). The decreasing AMOC trend in LR HF-TNST over 2000–2100 is substantially larger (~13 Sv/century). Figures 29e and 29f confirm that the MOCNA index is indeed representative of large-scale AMOC decline in both simulations and that the magnitude of 21st century decline is substantially larger in LR than HR over the entire Atlantic basin. This weaker AMOC decline in HR stands in contrast to the results of a recent multi-model analysis of CMIP6 coupled HighResMIP experiments (Roberts, Jackson, et al., 2020) in which higher ocean resolution was generally found to be associated with greater externally-forced AMOC decline. The CESM1.3 contributions to that study, however, were an outlier insofar as the eddy-resolving simulation showed less AMOC decline from 1950 to 2050 than the eddy-parameterized simulation. Thus, the analysis here corroborates the CESM results shown in Roberts, Jackson, et al. (2020) while also strengthening the counterargument that higher resolution does not necessarily result in greater projected AMOC weakening as a result of anthropogenic forcing.
In both HR and LR, there is a significant correlation between AMOC strength (MOCNA) and a warming hole index (WHI; defined as the difference between surface temperatures averaged over the subpolar North Atlantic, 40° to 15°W and 45° to 60°N, and over the NH) during the historical period of 1850–2010 when the former leads the latter by a few years (Figure 28). The connection is slightly stronger in HR (r ~ 0.6 compared to r ~ 0.5 after low-pass filtering). This implies that a causal link exists between intrinsic, multidecadal AMOC variability and a relative warming of the subpolar Atlantic, which is in line with the analysis of the AMOC and SST relationship in PI control simulations shown above (Figures 6b and 7b). Furthermore, the larger regression slope between WHI and MOCNA in HR over the historical period (not shown, but evident by visual inspection of Figure 28) would appear to be related to the higher heat transport efficiency of AMOC in HR (Figure 8), which results in relatively large WHI change per Sv AMOC change in HR compared to LR. The warming hole deepens over the 1850–2100 period in both simulations, with pronounced reductions in WHI over the 21st century. In both simulations, strong WHI deepening starts circa 2010, in tandem with a sharp increase in the rate of AMOC decline. Consistent with the linkage seen in the historical period, multidecadal fluctuations in the rate of WHI deepening appear to be linked to changes in the rate of AMOC decline in both simulations. By the end of the 21st century, the magnitude of WHI is comparable in HR and LR (Figure 27), albeit slightly deeper in LR, despite much stronger AMOC reduction in LR. This is again connected to the higher heat transport efficiency in HR and can be explained by the fact that the AMOC-related decline in Atlantic meridional ocean heat transport over the 21st century is roughly equal in both simulations (maximizing at about −0.35PW at ~30°N; not shown). The comparable Atlantic heat transport reduction represents a larger percentage reduction in LR than in HR, given the lower historical heat transport values in LR (Figure 20). In short, this resolution comparison highlights that the magnitude of ocean heat transport change is more relevant than the magnitude of AMOC decline for understanding patterns of projected surface temperature change.
Both simulations also show a reduction in the strength of the AABW cell, which in this case is more pronounced in HR than LR (Figures 29a and 29b). While both HR and LR exhibit a general weakening of the overturning circulation in both hemispheres, presumably related to a weakening of high latitude surface buoyancy forcing, the stronger reduction in AABW in HR could be part of the explanation for weaker AMOC decline in that simulation. This follows from the relative strong interhemispheric MOC linkages seen in the HR control simulation (Figures 5c and 5e). The weakening of the (counterclockwise) AABW results in a slowdown of poleward transport of warm water (most easily seen in the Pacific sector MOC change plots; Figures 29c and 29d) and a corresponding steady reduction in ocean heat transport toward the Southern Ocean in both simulations (not shown). The southward ocean heat transport reduction is greater in HR than LR (by about 0.1–0.2 PW between 30°S and 60°S), which may partially explain the reduced Southern Ocean warming seen in HR (Figures 26 and 27).
5.2 Tropical Cyclones and Atmospheric Rivers in Future Climate
There are considerable uncertainties and debates about potential changes in tropical cyclone activity in response to anthropogenic forcing (see Knutson et al., 2019, 2020, for a review and references therein). These uncertainties are partly due to the fact that the current generation of climate models does not explicitly permit or resolve TCs due to their coarse resolution and, thus, remains deficient in simulating strong TCs. Therefore, they cannot be directly used to assess potential changes in TC intensity for future climate. Until very recently, no multi-modeling effort has been carried out to study TC response to future climate change using TC-permitting coupled climate models (Roberts et al., 2020b). However, Roberts et al. (2020b) is based on the HighResMIP protocol that covers a 100-year period spanning 1950–2050. Here, we compare TC response to anthropogenic forcing between HR and LR HF TNST for a longer (224 years) period of 1877–2100.
Figure 30 shows the difference in TC track densities between a future period (2006–2100) and the 1877–2005 historical period from HR and LR. It is evident that TC response to the RCP8.5 climate forcing is much more muted in LR than in HR, particularly in the NTA where there is no discernible TC response in LR. There are, however, some consistencies in the simulated TC response between HR and LR in the SIO and the ENP, despite the weaker amplitude in LR. In particular, there is an indication of reduced TC activities in the future climate over the SIO, which is consistent with previous CMIP5 studies (Bell et al., 2018; Gleixner et al., 2013; Knutson et al., 2020; Tory et al., 2013), and there is an indication of decreased TC activity in the future climate over the ENP consistent with the HighResMIP results reported by Roberts et al. (2020b). In HR, there is a hint of a poleward shift of TC activities under the future climate in the NH with a particularly clear signal over the WNP. This finding is consistent with Altman et al. (2018), Kossin et al. (2014), Sharmila and Walsh (2018), and Roberts et al. (2020b). Compared to the response in the 100-year HighResMIP simulations shown in Roberts et al. (2020b), the enhanced TC activities north of 15°N in the western Pacific extending to the Northern Indian Ocean become much more pronounced in HR compared to LR (Figure 30). This regional increase in future TC activities deserves further investigations. As in Roberts et al. (2020b), the TC response in NTA is generally weaker than in other TC basins, which is likely related to the negative model bias in TC count. However, HR does hint a decrease in TC activity in the Atlantic hurricane Main Development Region (MDR) and an increase to the north of the MDR, again consistent with a poleward shift of TC activities. Globally, the total number of TCs in HR decreases from an average of 110 per year during the historic period of 1877–2205 to 100 per year during the future period of 2006–2100, amounting to a decrease of roughly 10%. In contrast, the total number of TCs in LR decreases from an average of 25 per year during the historic period to 17 per year during the future period, which is a decrease of more than 30%. Probability Density Function (PDF) of TC 10-m wind speed analysis indicates a shift toward high wind speeds during the future period in both HR and LR (not shown), indicating a possible TC intensity increase in the future (Emanuel, 2017; Knutson et al., 2020). Overall, the TC response to the RCP8.5 climate forcing in HR and LR shows a generally consistent picture to that presented in Roberts et al. (2020b).
ARs are expected to become more active and carry more moisture as climate continues to warm which causes atmospheric moisture content to increase (Espinoza et al., 2018; Payne et al., 2020, and references therein). Although moisture increase is an important factor, storm-track changes and jet stream shifts can also play a role in future AR response to warming. As a consequence of anticipated future AR changes, areas with high terrain, such as the West Coast of North America, are expected to be most susceptible to increases in AR-related precipitation extremes. Despite the potential impact of future AR changes, high-resolution global climate model simulations, particularly coupled model simulations, capable of representing AR processes remain limited (Payne et al., 2020). Here, we take advantage of our HR simulations to investigate how ARs will be changed with a warmer climate.
Figure 31 (left) compares the future change in ARs between HR and LR by differencing mean AR IVT over the future period of 2006–2100 and the historic period of 1877–2005. Although IVT increases globally in both HR and LR, the increase in HR is more substantial: The globally averaged IVT difference between the future and historic period in HR is 6.56 kg m−1 s−1 that is nearly double of the value of 3.34 kg m−1 s−1 in LR (Figure 31 left), indicating that more ARs in the future are projected in HR than in LR. One potential reason for the more ARs is that the moisture content in the atmosphere is increasing more rapidly in HR than in LR. Figure 32 shows the differences in vertically integrated water vapor (IWV) between HR and LR averaged over the historical period of 1877–2005 and the future period of 2006–2100, respectively. It is evident that the difference becomes larger for the future period (by a factor of ~2.56), indicating that IWV is increasing at a faster rate in HR (~18.8%) than in LR (~8.7%) in response to climate forcing. It is also interesting to note that the largest IWV difference between HR and LR tends to occur along the edges of the tropical moisture pool, suggesting a more poleward expansion of the moisture pool in HR than in LR. These areas of large IWV increases along the vicinity of the Kuroshio Extension, Gulf Stream Extension, South Pacific Convergence Zone, and South Atlantic Convergence Zone coincide with the active AR regions. However, we note that higher moisture content alone does not necessarily translate to higher moisture transport, as storm-track and jet stream differences between HR and LR may also play a role in the AR differences between HR and LR.
Along with the increase in ARs, it is no surprise that AR-related precipitation is also increased during the future period compared to the historic period in both HR and LR (Figure 31 center). However, the amount of AR-related precipitation increase averaged over the globe in HR is about 36% higher than that in LR. In many areas the difference is more than a factor of 2. Worth particular note is the significant increase of AR-related precipitation in HR over the West Coast of North America, British Isles and Norwegian West Coast in West Europe, Chilean Coast in South America, and Southeastern Coast of China (Espinoza et al., 2018; Payne et al., 2020). As an example, Figure 31 (right) shows the difference of projected rainfall changes for HR and LR along the West Coast of North America. Although both models project that the highest future rainfall increase occurs off the coast of British Columbia, the amount of the projected rainfall increase in LR is lower than that in HR by a factor of 2 or more. In HR the projected rainfall increase also reveals more rich spatial structures that are closely linked to the local orographic features compared to LR. Since AR-induced precipitation is closely related to the orographic lift effect, the stronger precipitation response and the rich spatial rainfall variation in HR are at least partially attributable to the better resolved mountainous terrains in the high-resolution model (Figure 31 right). In summary, increasing model resolution not only leads to a significant increase in projected AR activity and AR-related precipitation but more importantly results in more detailed regional changes in rainfall response to climate forcing over AR landfalling areas.
6 Summary and Concluding Remarks
We have presented some preliminary results from a set of unprecedentedly long high-resolution coupled global climate simulations at 0.25° atmosphere and land and 0.1° ocean and sea-ice horizontal resolutions. These simulations consist of (1) a 500-year PI-CTRL run and (2) a 250-year HF-TNST from 1850 to 2100 branched from PI-CTRL at year 250. Previous attempts at similar model resolutions were limited to relatively short (less than 200 years) integrations (e.g., Small et al., 2014; Roberts, Jackson, et al., 2020). These long simulations allow for a more robust evaluation of effects of model resolution on the simulated mean climate, internal modes of variability at seasonal-to-multidecadal time scales, and model fidelity in simulating observations and future climate change.
Results from the 500-year PI-CTRL show that the model climate system in both HR and LR attains acceptably small TOA heat flux imbalances of −0.217 W m−2 in HR and +0.033 W m−2 after 150 years of integration. The larger TOA imbalance in HR is reflected in a stronger cooling trend in the deep ocean in HR compared to a weaker warming trend in LR. Despite the strong trend, the global-average vertical ocean temperature gradient shows a smaller positive bias (warmer [colder]-than-observed in the upper [deep] ocean) in HR compared to the larger negative bias in LR. This difference in vertical ocean temperature biases is likely a contributing factor for the stronger meridional heat transport in HR which leads to differences in ocean-driven intrinsic and externally-forced climate variations.
One feature that stands out as a major difference between HR and LR PI-CTRL is the presence of polynyas within the Weddell Gyre in HR. These polynyas are a prominent feature in HR, occurring every 40–50 years, but notably absent in LR, consistent with other previous studies (e.g., Kurtakoti et al., 2018; Weijer et al., 2017). A new finding in this study is that the multidecadal Weddell polynya variability is not only closely correlated with an IPO-like pattern in SST but also leads the IPO by about 4 years. This Weddell polynya-IPO mode is found to dominate multidecadal variability of global-mean SST in HR. We hypothesize that this multidecadal mode is a result of interactions between the Southern Ocean sea-ice and the tropical Pacific ocean-atmosphere coupled systems. If this hypothesis holds, it would imply an active role for the Weddell polynyas in the global climate system (Latif et al., 2013), which may be underrepresented in low-resolution climate models. However, whether this mode is an artifact of our HR configuration or has relevance to reality requires additional scrutiny. Unfortunately, the lack of any long-term observations in the Southern Ocean, particularly regarding polynyas in the Weddell Gyre, poses a great challenge to address this issue.
The Weddell Gyre polynyas also impact AMOC in HR PI-CTRL. Specifically, we have seen not only a significant increase in the strength of the bottom overturning circulation associated with the AABW but also a significant contribution to AMOC variability from SH origins, both of which are likely closely linked to the Weddell polynyas. If these results are robust, then it will raise some interesting questions about the potential role of the Southern Ocean in decadal climate variability and predictability (Latif et al., 2013).
Over major eddying regions, such as the ACC and the seaward extensions of western boundary currents, HR simulates much more realistically OMEA feedbacks between SST and THF than LR. These improved OMEA feedbacks are expected to have an impact on simulations of extratropical modes of variability. However, such impacts are not explored in this study and require further investigations.
Comparison of HR and LR HF-TNST to available observations during the historical period reveals a number of improvements in HR due to model resolution increase, which can be briefly summarized as follows:
- Simulated global-mean SST and 2-m air-temperature evaluations in HR both show an excellent agreement with the historical observations, while a cold bias of ~1.5°C is present in these fields in LR. However, anomalous 2-m air-temperature in LR (referenced to the 30-year mean between 1951 and 1980) shows equally good agreement with observed temperature, suggesting that the cold bias in LR does not hamper its ability to simulate surface temperature response to climate forcing.
- SST seasonal cycle in HR is weaker than that in LR and is in a better agreement with the observations, particularly in the Southern Ocean. The more realistic SST seasonal cycle in HR is at least partially attributed to the improved representation of MLD, which is too shallow in LR, contributing to a stronger SST seasonal cycle.
- The simulated MLD agrees much better with Argo-based observations in HR than LR, particularly in the Southern Ocean and in the North Atlantic, as well as along the Kuroshio extension region. The improved MLD in HR has impacts on not only the seasonal cycle of SST but also persistence of SST anomalies that can exert an influence on climate variability and predictability.
- North Atlantic Ocean meridional heat transport in HR is stronger than that in LR by up to 20% and shows an excellent agreement with the observed estimate at 26°N from the RAPID array. This improvement in ocean heat transport is likely related to the reduced model biases in ocean vertical temperature profile in HR, as simulated AMOC at 26°N has similar strengths and profiles in both HR and LR, and both show good agreement with the observed volume transport estimate.
- Multi-decadal SST variations associated with AMV are enhanced and in a better agreement with observations in HR than in LR, with the latter exhibiting a much too weak multi-decadal SST variability in the North Atlantic.
- TC simulations in HR, in terms of both number and strength, are much improved over LR, even though there are still many noted discrepancies in HR in comparison to available observations. TCs' expected relationships with ENSO and AMV are also better simulated in HR than LR, particularly over the NTA.
- There is a drastic improvement in simulations of ARs and related precipitation in HR compared to LR. The simulated AR IVT in HR is substantially stronger than that in LR and compares much better with the observations. AR-related precipitation is also more realistically simulated in HR than in LR. In particular, precipitation over land areas where frequent landfalling ARs occur, causing extreme precipitation events, is vastly improved in HR.
Not everything is improved by enhancing model resolution, and indeed, there are aspects of the climate in HR that are worse than in LR. For example, sea-ice extent and sea-ice concentration in both hemispheres in HR are underestimated compared to LR and are less realistic. The ACC transport at Drake Passage in HR is at the lower limit of the observed estimates, while the value in LR is well within the range of observed estimates, pointing to a low ACC transport bias in HR. The rainfall bias along the ITCZ seems to be exacerbated by the resolution increase, resulting in a severe positive rainfall bias along the ITCZ in HR.
There are also phenomena that show relatively low sensitivity and/or mixed response to model resolution increase. For example, in comparison to observations, ENSO in the historical period of HR HF-TNST shows better agreement in amplitude but worse agreement in frequency compared to LR. PNA/PDO representations in HR and LR are comparable and both show good agreement with observations, but NPO and related SST response are more realistically simulated in HR than in LR. The better simulated NPO may be partially responsible for the stronger ENSO in HR, as NPO can exert an influence on ENSO (Chiang & Vimont, 2004).
Future climate response to RCP8.5 forcing in HR and LR also displays differences. Particularly, there are significant differences in regional climate response to RCP8.5 forcing between HR and LR. The Arctic amplification is significantly stronger (by ~25%) in HR than in LR, whereas warming in the Southern Ocean is projected to occur at a much slower rate (by ~40%) in HR than in LR. TC response to RCP8.5 forcing in HR shows a more significant and more coherent spatial structure than that in LR, indicating a reduced TC activity in SIO and ENP and a poleward shift, particularly in WNP. Finally, projected ARs under the RCP8.5 forcing scenario have a much enhanced increase (by 45% or more) in HR than in LR. AR-related precipitation shows a corresponding increase in HR, particularly in areas of frequent landfalling ARs, such as the West Coast of North America, where the increase in precipitation in HR can reach more than twofold compared to LR. This suggests that future hydroclimate response to climate change may be significantly underestimated by the current generation of low-resolution climate models utilized in IPCC reports (Roberts et al., 2018).
We conclude by providing some comments to the questions that were posed in the introduction:
- How does the simulation of natural variability and response to anthropogenic forcing of the climate system depend on model resolution? Although basic structures of modes of climate variability and global-mean temperature response to anthropogenic forcing are not fundamentally altered by increasing model resolution, we provide evidence for many important regional differences between HR and LR. Overall, HR is capable of providing more credible regional-scale information that is not available in LR. A good example is the much-improved precipitation simulation related to landfalling ARs over many coastal regions in HR.
- How do multi-scale interactions arising from coupling the better resolved atmosphere and ocean models influence the low-frequency, large-scale behavior of the Earth system, as well as their impacts on extreme events? We see clear evidence of improved representation of multi-scale coupling between atmosphere and ocean in HR. We also see increases in resolved large-scale precipitation and in variances of SST and THF, as well as ocean circulations across broad time scales in HR. In particular, there is intriguing evidence that multi-decadal SST variations associated with AMV are enhanced and more realistic in HR, and the observed TC-AMV relationship is better captured by HR. However, future studies are required to examine the extent to which the improved multi-scale interactions in HR contribute to enhanced realism and fidelity of low-frequency, large-scale climate variability and extreme events.
- To what extent can the knowledge about climate variability and change gained from low resolution modeling studies carry over to the TC-permitting and ocean eddy-resolving regime? Although many of the findings about climate variability and change from low-resolution modeling remain applicable to the high-resolution CESM simulations, there are new features that appear to emerge from eddy-resolving regime and are notably absent in LR. A particular example is open ocean polynyas in the Southern Ocean that stand out as a prominent feature in HR on multi-decadal time scales, but are not seen in low-resolution CESM simulations. The polynyas in HR are further shown to have a potentially remote impact on the tropics at multi-decadal time scales, raising some interesting questions about the role of the Southern Ocean in decadal climate variability and predictability. However, additional studies are required to validate this new finding and its relevance to reality.
Having stated the above, it should be noted that this LR-HR intercomparison study does not solely reflect horizontal resolution differences. Changes in physics parameterizations (e.g., overflows, mesoscale and submesoscale eddies, and river runoff), parameter settings (i.e., TOA tunings, see section 2.2 for details), and even vertical grid (POP2 in HR uses 62 levels whereas 60 levels are used in LR) between HR and LR could also be playing a nontrivial role in the differences. Some of these changes are unavoidable from physics considerations. For example, at 0.1° horizontal resolution ocean models are capable of resolving mesoscale eddies except at polar and subpolar latitudes and should not employ mesoscale eddy parameterizations, such as GM parameterization, to avoid double-counting of mesoscale eddy effects. However, this raises the issue of whether parameterized eddy fluxes in LR models accurately represent eddy effects and to what extent differences between HR and LR models can be attributed to differences between explicit and parameterized eddy fluxes. Griffies et al. (2015) show that within the near-surface layer the explicit mesoscale eddy vertical heat transport in the 0.1° CM2.6 model is considerably larger than the parameterized mesoscale eddy vertical heat transport in the 1.0° CM2.0 model, leading to an increase in the net upward OVHT when mesoscale eddies are explicitly represented. We hypothesize that this increase in the upward OVHT may be responsible for the warmer SST in HR than in LR (see section 4.1). A recent LR-HR ocean model intercomparison study by Chassignet et al. (2020) also noted potential errors in parameterizations of vertical eddy heat transport in LR ocean models that may cause the warming difference below 700 m between HR and LR models. There is clearly a need for future studies addressing the impact of uncertainties in parameterized eddy heat transports in climate model biases.
High-resolution climate simulations are computationally very demanding and energy-consuming, raising concerns about the carbon footprint of this type of simulations. The effort at iHESP attempts to address some of these concerns by porting and optimizing the CESM code on a Sunway HPC that has much lower energy consumption than more conventional CPU-based HPCs (Zhang et al., 2020). Therefore, aside from the fact that this is the first time such long and complete climate simulations have ever been conducted at this high resolution, it is equally important to recognize that these unprecedented CESM simulations were successfully carried out on a greener and more energy efficient HPC system. With the success of CESM on Sunway TaihuLight System, we are planning the next phase of high-resolution CESM simulations at iHESP, which will include an ensemble of 5–10 HR HF-TNST runs from 1920 to 2100 and an ensemble of high-resolution decadal climate prediction simulations. The ensemble of HR HF-TNST runs will allow us to validate the robustness of the results presented here, which are based on only one simulation. The decadal prediction ensemble will allow us to further explore the impact of model resolution on predictable dynamics and model predictive skills. As open science and strong international collaborations are the founding principle of iHESP, all the products from iHESP, including model codes and simulation datasets, are being and/or will be made available to the broader community. We believe this unprecedented set of high-resolution CESM simulations can provide an invaluable basis for a careful evaluation of the benefits of global high-resolution climate simulations, and as such, they will be a useful guide for planning of future high-resolution climate experiments.
As a final note, we comment on a few lessons learned from our first experience with conducting a large set of high-resolution climate simulations. Compared to low-resolution simulations, achieving a small TOA imbalance in high-resolution settings is a more formidable challenge due to the combined factor of the exceedingly high computational cost and the need for many iterations of order 30–50-year simulations with small adjustments in model parameters. The high-resolution simulations presented in this study have not been extensively tuned, resulting in a somewhat large, but still acceptable, TOA heat flux imbalance compared to its low-resolution counterpart. It is certainly possible to reduce the TOA heat flux imbalance by further turning HR model parameters. However, it is less certain whether a further reduction in TOA heat flux imbalance will lead to a more realistic climate simulation, as based on our LR modeling experience small model biases do not necessarily correspond to small TOA heat flux imbalance. This reiterates the point made above that the results of this study are not solely due to horizontal resolution differences but may be caused by model parameter differences. The high computational cost also makes it very difficult to evaluate equilibrium climate sensitivity (ECS) of high-resolution climate models. Instead, transient climate response (TCR) experiments are a more viable approach for evaluating differences in climate sensitivity between high- and low-resolution climate models. We have completed a pair of 80-year HR and LR TCR simulations and the results of these experiments will be presented in a future study. The HR PI-CTRL shows a significant model drift in the first 150 years of the simulation, which unfortunately could be caused by some hardware issues on the Sunway HPC. Nevertheless, the simulation after 150 years reveals more stable and well-behaved results, as shown in this study. Therefore, based on this modeling experience, we would conclude that a 30–50-year model spin-up as suggested by the HighResMIP protocol (Haarsma et al., 2016) is likely insufficient as a starting point for HR climate model simulations. We would recommend a minimum of 150 to 200 years of model spin-up as a future strategy for initializing HR climate model simulations.
Finally, there is certainly a need to examine the impact of physics parameterizations and improve these parameterized physics in high-resolution CESM. Particularly, some of the parameterized ocean physics in the standard resolution CESM are not included in the current high-resolution version. These include submesoscale eddy parameterization (Fox-Kemper et al., 2008, 2011) and overflow parameterization (Danabasoglu et al., 2010). At 0.1°, the ocean model does not resolve either submesoscale eddies or the details of overflows and their physics. Based on a previous overflow parameterization sensitivity study by Yeager and Danabasoglu (2012), we venture that that the shallower NADW cell (and stronger AABW cell) in the Atlantic in HR compared to LR could be attributable to the absence of overflows. We further submit that including overflow parameterization in HR would lead to a further reduction in AMOC variance, because according to Yeager and Danabasoglu (2012), overflows tend to damp low-frequency AMOC variability. On the other hand, submesoscale eddy parameterization has been shown to have a significant impact on near-surface ocean heat budget (e.g., Griffies et al., 2015) and submesoscale eddy induced vertical heat flux can potentially enhance turbulent heat exchange between the ocean and atmosphere (Su et al., 2018). Future studies need to examine the impact of such missing physics by including their parameterizations in the model.
Acknowledgments
This research is completed through the International Laboratory for High Resolution Earth System Prediction (iHESP)—a collaboration among the Qingdao National Laboratory for Marine Science and Technology (QNLM), Texas A&M University (TAMU), and the U.S. National Center for Atmospheric Research (NCAR). NCAR is a major facility sponsored by the US NSF under Cooperative Agreement 1852977. The majority of the simulations were completed on Sunway TaihuLight HPC in Wuxi, China, but initial tuning of high-resolution CESM was performed on HPC systems at the Texas Advanced Computing Center and many simulation analyses were carried out on the HPC systems at the Texas A&M High Performance Research Computing. We knowledge Alper Altuntas for assistance with analyzing simulation datasets. We thank Malcolm Roberts, Anne-Marie Treguier, Magdalena Balmaseda, Arne Biastoch, Bernadette Sloyan, Mark Taylor, Guoxiong Wu, and Colin Zarzycki for their scientific advice and critical evaluation of the iHESP project. We are also grateful to two anonymous reviewers for helpful comments.
Open Research
Data Availability Statement
The data used in this work are available from both iHESP data portal (https://ihesp.tamu.edu/products/ihesp-products/data-release/PI_control/index.html) and QNLM data portal (http://ihesp.qnlm.ac). The Sunway version of the CESM code is available at ZENODO via https://doi.org/10.5281/zenodo.3637771.
