Effect of Horizontal Resolution on the Simulation of Tropical Cyclones in the Chinese Academy of Sciences FGOALS-f3 Climate System Model

. The effects of horizontal resolution on the simulation of tropical cyclones were studied using the Chinese Academy of Sciences FGOALS-f3 climate system model from the High-Resolution Model Intercomparison Project (HighResMIP) for the Coupled Model Intercomparison Project Phase 6 (CMIP6). Both the low-resolution (about 100 km resolution) FGOALS-20 f3 model (FGOALS-f3-L) and the high-resolution (about 25 km resolution) FGOALS-f3 (FGOALS-f3-H) model were used to achieve the standard Tier 1 experiment required by the HighResMIP. FGOALS-f3-L and FGOALS-f3-H have the same model parameterizations with the exactly the same parameters. The only differences between the two models are the horizontal resolution and the time step. The performance of FGOALS-f3-H and FGOALS-f3-L


Introduction
Tropical cyclones are extreme weather phenomena characterized by intense wind speeds and heavy rainfall. Although 40 tropical cyclones alleviate coastal droughts, they can also cause severe economic losses and significant human casualties (Mendelsohn et al., 2012;Aon, 2018). Against the current background of global climate change, the effective simulation, prediction and projection of global tropical cyclone activity is challenging, but essential for disaster prevention and mitigation (Emanuel, 2017).
The simulation of tropical cyclones in global climate models (GCMs) is challenging in terms of both resolution and 45 physical processes. Tropical cyclone-like structures appeared in early GCMs and Manabe et al. (1970), Bengtsson et al. (1982), Krishnamurti et al. (1989), Broccoli and Manabe (1990), Wu and Lau (1992) and Haarsma et al. (1993) were pioneers in using objective feature-tracking approaches to study simulated tropical cyclones. However, the low resolution and incomplete parameterization of the physical processes in these early GCMs meant that their performance in simulating tropical cyclones was limited. For this reason, statistical methods were used to study on the climatology of tropical cyclones. Camargo et al. 50 (2013) found that the simulation of the frequency of tropical cyclones in the Coupled Model Intercomparison Project 5 (CMIP5) was much lower than in the observations. This was mainly due to the cold biases of the sea surface temperature, which amplified the uncertainty of future projections. Emanuel (2013) designed a downscaling method to reduce the uncertainty in projections of tropical cyclone activity.
The horizontal and vertical resolutions of climate system models increased over the following half-century in line with 55 the complex parameterization of the physical processes. As a result, more refined details (e.g., tropical cyclones and tropical waves) can now be resolved. Regional climate system models with smaller spatial scales and lower computing costs can now be used to simulate tropical cyclones. Knutson et al. (2007) used a high-resolution regional model to simulate tropical cyclone activity in the northern Atlantic Ocean. The structure and interannual variability of the simulated tropical cyclones has a high fidelity with the observations. There has also been a significant increase in the resolution of GCMs. Oouchi et al. (2016) used 60 a 20 km mesh global atmospheric model to simulate tropical cyclone activity in a warming climate and found that the highresolution GCM could not only describe the details of typhoons very well, but also captured the variability of tropical cyclones.
The increase in the horizontal resolution of GCMs has led to significant changes in the simulation of the variability of tropical cyclones. Previous studies showed that there are significant changes in the El Niño-Southern Oscillation (ENSO) as the horizontal resolution of GCMs increases (Philander et al., 1992;Kuntson et al., 1997;Schneider et al., 2003;Masson et al., 2012;Larson et al., 2013;Meehl et al., 2020) and the simulation results were mostly positive. However, these improvements 70 in predicting the ENSO with an increase in horizontal resolution did not lead to improvements in the relationship between the ENSO and tropical cyclones (Matsuura et al., 1999;Bell et al., 2014;Krishnamurthy et al., 2016). There is also a close relationship between the Madden-Julian oscillation (MJO) and tropical cyclones (Liebmann et al., 1994;Hall et al., 2001;Camargo et al., 2008). High-resolution GCMs need to not only give a better description of the structure of tropical cyclones, but should also well simulate the relationship between tropical cyclones and large-scale variabilities (e.g., the MJO and ENSO), 75 which is crucial in reducing the uncertainties in the simulation and prediction of tropical cyclones (Manganello et al., 2012;2016;Zhang et al., 2016;Delworth et al., 2020). As the horizontal resolution increases in the models, some key parameters in the physical parameterizations are tuned to give a better performance (Bacmeister et al., 2013;Roberts et al., 2020), e.g., Lim et al. (2015 found that an increase in the threshold of minimum entrainment led to the increasing TC activity,  found that the constrained convective heating in the convective scheme induced intense grid-scale upward motions 80 and promoted large-scale condensation, which was favorable for the development of a more intense TC. These artificial tuning might introduce more uncertainties in terms of the effects of resolution, giving rise to conclusions that are controversial to the tropical cyclone research community. The impacts of horizontal resolution on the simulation of tropical cyclones were studied using the Chinese Academy of Sciences Flexible Global Ocean-Atmosphere-Land System, Finite-Volume Version 3 (FGOALS-f3) model, which was 85 developed by the State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG), Institute of Atmospheric Physics (IAP). The simulated tropical cyclones in FGOALS-f3 were introduced firstly, then the outputs of FGOALS-f3-L and FGOALS-f3-H were used to reveal the influence of horizontal resolution on these simulations. The latest version of FGOALS-f3 participated in the CMIP6 (Eyring et al. 2016), DECK and MIPs endorsements (Zhou et al. 2016;He et al. 2019He et al. , 2020Haarsma et al. 2016). FAMIL2 is the atmospheric component of the climate system 90 model FGOALS-f3. Li et al. (2019) evaluated the simulation performance of tropical cyclone activity in the latest generation atmospheric general circulation models from the LASG-IAP (FAMIL2) using a coarse resolution with standard AMIP experiments. Although FAMIL2 is able to reproduce the many aspects of the activities of tropical cyclones with a horizontal resolution of 1°, there is still some room for improvement in simulating tropical cyclones, such as the weak intensity of tropical cyclones, fewer tropical cyclones in the peak month in the northern Atlantic and eastern Pacific oceans and inaccurate large-95 scale factors. Therefore, the HighResMIP configuration has been applied for both the low-and high-resolution FGOALS-f3.
Both model versions retained the exact model physics and parameters and the only differences were the horizontal resolutions and model time steps, which better meet the rule of HighResMIP: "The experimental set-up and design of the standard resolution experiments will be exactly the same as for the high-resolution runs". This study aimed to address the following issues: (1) the impacts of horizontal resolutions on the simulation of global tropical cyclones in a climate system model; and 100 (2) the possible physical linkages between the horizontal resolutions and the simulated tropical cyclones?
This paper is organized as follows. Section2 introduces the model, data and methods used in this study. Section 3 shows the performances of simulated tropical cyclones in both FGOALS-f3-L and FGOALS-f3-H. Section 4 discusses the possible reasons for the improvement of the simulation of tropical cyclones with increased horizontal resolutions. Section 5 introduces the physical parameterization and its impact on the simulated TC in GCMs, then discusses the potential value-added effect of 105 MJO and TC due to the increase in horizontal resolution from HighResMIP models. Section 6 provides a summary of the results.

Description of FGOALS-f3
FGOALS-f3 is the latest version of the Chinese Academy of Sciences climate system model and was designed for CMIP6. 110 The FGOALS-f3 model consists of four components: (1) the atmospheric component is the Finite-volume Atmospheric Model Version 2.2 (FAMIL2.2) (Zhou et al., 2015;Bao et al., 2018Bao et al., , 2020Li et al., 2019;He et al., 2019), which is the successor to the atmospheric general circulation model of the Spectral Atmosphere Model of LASG (SAMIL) (Wu et al., 1996;Bao et al., 2010Bao et al., , 2013 Hunke et al. 2008;Hunke & Lipscomb, 2010). These four components are coupled by the Version 7 coupler in the CESM (Craig et al., 2012). Li et al. (2019) introduced the atmospheric component FAMIL2 in detail and carried out some tuning to achieve stability in long-term coupled integrations (defined as FAMIL2.2). 120 As is shown in Table 1, the finite-volume cubed-sphere dynamical core (FV3) (Lin, 2004;Putman and Lin, 2007;Voosen, 2017) is used as the dynamical core in FAMIL2.2, which is the atmospheric component of FGOALS-f3. The University of Washington moist turbulence parameterization (Park & Bretherton, 2009) is also used in FGOALS-f3. This is a non-local, high-order closure scheme and uses the diagnosed turbulent kinetic energy to determine the eddy diffusivity in turbulence. The Resolving Convective Precipitation parameterization (Bao and Li, 2020) is used, which involves calculating the microphysical 125 processes in the cumulus scheme for both deep and shallow convection; six species are considered, similar to the Geophysical Fluid Dynamics Laboratory (GFDL) cloud microphysics scheme (Zhou et al., 2019). The gravity wave drag scheme (Palmer et al., 1986), the cloud fraction diagnosis scheme (Xu & Randall, 1996) and the radiative transmission scheme (Clough et al., 2005) are also considered.
The vertical layers of FGOALS-f3-L and FGOALS-f3-H are both set to 32, whereas the horizontal resolutions of 130 FGOALS-f3-L and FGOALS-f3-H are C96 (approximately 100 km) and C384 (approximately 25 km), respectively ( To maintain the stability of the integration for the dynamical core, the two parameters k_split and n_split included in FV3 are different in FGOALS-f3-L and FGOALS-f3-H. k_split is the number of vertical remapping operations per physical time step in the dynamical integration and n_split is the number small dynamic (acoustic) time steps between the vertical remapping operations, which will affect the stability of the integration when the horizontal resolution of the model is changed. Considering 135 that FGOALS-f3-H requires more frequent vertical remapping, k_split and n_split are set to 6 and 15, respectively (they are 2 and 6, respectively, in FGOALS-f3-L). The time steps of the physical processes are both set to 30 minutes, but the update frequency of radiative transmission and the minimum time step of the microphysics scheme in both FGOALS-f3-L and FGOALS-f3-H are 1 h and 150 s, respectively. Li et al. (2017) tested the computing performance between the FGOALS-f3-L and FGOALS-f3-H using the Supercomputer Tianhe-2 and the results indicated a high computing speed-up and low computing 140 costs when the number of parallel processes was increased.

Tracking algorithms
An objective feature-tracking approach is used to detect the model-generated tropical cyclones based on the 6-h outputs 160 of FGOALS-f3-L and FGOALS-f3-H. According to the tracking scheme (Table 3), the sea-level pressure, warm core (the temperature anomaly averaged between 300 and 500 hPa), 10 m wind and the 850 hPa absolute vorticity are used to diagnose the tropical cyclone activity, which is similar to the method used in the climate system model of the GFDL (Zhao et al. 2009;Chen & Lin 2013;Xiang et al. 2015). Li et al. (2019) used this scheme to evaluate the simulated performance of tropical cyclones in FAMIL2 and showed a consistent performance. The wind speed thresholds between FGOALS-f3-L and FGOALS-165 f3-H are consistent with the relationship between the horizontal resolution of the models and the tropical cyclone detection algorithms .

Global climatology of tropical cyclone track density
The climatology of simulated tropical cyclones is the first step in testing the performance of the model. Zhao et al. (2009)  170 used a GFDL GCM with a 50 km horizontal resolution to simulate global tropical cyclone activity and obtained a negative bias in the number of tropical cyclones in the eastern Pacific, northern Atlantic and southern Indian oceans. These biases also appeared in the low-resolution models participating in the US CLIVAR Working Group on Hurricanes (Walsh et al., 2015). Negative biases between FGOALS-f3-L and IBTrACS appear in the mid-latitudes of the western Pacific and northern Atlantic oceans, but positive biases between the FGOALS-f3-H and IBTrACS also appear in these areas, which means that there are more tropical cyclone events at higher latitudes in FGOALS-f3-H than in IBTrACS and the simulation in FGOALS-f3-L. This phenomenon also exists in the high-resolution GCMs that participated in the European Union Horizon 2020 project PRIMAVERA (Roberts et al., 2020). 185 The negative biases of the intensities of tropical cyclones are also improved when the horizontal resolution is increased from 100 km ( Figure 1a Ocean when the horizontal resolution is increased from 100 to 25 km. Similar to the pattern of track density anomalies ( Figure  190 2), the biases of the wind speed densities between the FGOALS-f3 and IBTrACS are improved when the horizontal resolution is increased from 100 to 25 km, but the positive biases are intensified in the mid-latitudes of the western Pacific and northern Atlantic oceans. Figure 4 shows the pressure-wind pairs for each 6-hourly measurement of tropical cyclones between FGOALS-f3 and IBTrACS. These results indicate that the spread of pressure-wind pairs in FGOALS-f3-L is narrow and there is a severe underestimation of intense tropical cyclone events at the lower surface pressures and higher wind speeds in the western Pacific and northern Atlantic oceans (Figure 4a, 4c), although this bias has been dramatically improved (Figure 4b,   4d).
The increased intensity of tropical cyclones in FGOALS-f3-H favors the apparent negative bias of the tropical cyclone lifetime in FGOALS-f3-L when the horizontal resolution is increased. Figure 5 shows the average lifetime of tropical cyclones from 1991 to 2014. The average lifetime of tropical cyclones in the observations is about 8.5, 7.5, 7.5, 4 and 7.5 days in the 200 western Pacific, northern Atlantic, eastern Pacific, northern Indian and southern Pacific oceans, respectively, and the simulation of the average lifetime of tropical cyclones is increased in these five basins when the horizontal resolution is increased from 100 to 25 km. For example, the simulated average lifetime of tropical cyclones increases from 6 to 7.4 days when the horizontal resolution is increased in the western Pacific Ocean.

Seasonal cycles and the interannual variability of tropical cyclones 205
Evaluating the seasonal cycle and interannual variability of tropical cyclones in GCMs is an efficient way to verify the coordination between tropical cyclone activity and large-scale circulation patterns (Manganello et al., 2012;Camargo et al., 2016;Kuntson et al., 2019;. Robert et al. (2020) found no uniform improvement in the seasonal cycle and interannual variability of tropical cyclones at increased horizontal resolutions, which means that there is a difference in coordination between tropical cyclone activity and large-scale circulation patterns in high-resolution GCMs. Figure 6 shows the seasonal 210 cycle of tropical cyclones between IBTrACS, FGOALS-f3-L and FGOALS-f3-H and shows that FGOALS-f3-L gives a consistent underestimation of the seasonal cycle of tropical cyclones in the northern Atlantic, eastern Pacific, northern Indian and southern Pacific oceans. Neither the single peak in the number of tropical cyclones in the northern Atlantic (peak month September), eastern Pacific (peak month August) and southern Pacific (peak month February) oceans nor the double peak in the northern Indian Ocean (peak months May and November) could be reproduced in FGOALS-f3-L. There are two increases 215 in the simulated number of tropical cyclones in the peak months of the northern Atlantic and eastern Pacific oceans and the characteristics of the seasonal cycle of simulated tropical cyclones are improved in the northern Indian, western Pacific and southern Pacific oceans as the horizontal resolution is increased from 100 to 25 km. Figure 7 shows the interannual correlation of the numbers of tropical cyclones between FGOALS-f3 and IBTrACS. The results show that the correlation coefficient is improved in each basin, which reflects the fact that the interannual variability between the tropical cyclone activity and large-220 scale circulation patterns is harmonious. Figure 8 shows the interannual correlation of the accumulated cyclone energy (ACE; Bell et al., 2000) between FGOALS-f3 and IBTrACS. ACE is a measure used by the National Oceanic and Atmospheric Administration, which means that the energy over the lifetime of a tropical cyclone is calculated for every 6-h period: where Vm is the estimated sustained wind speed in knots. The results (Figure 8) show that the correlation coefficient of the ACE between IBTrACS and the simulation is improved in each basin when the horizontal resolution is increased from 100 to

Horizontal structure of tropical cyclones 230
Previous studies have shown that the horizontal resolution influences the horizontal structure of simulated tropical cyclones (Strachan et al., 2013;Murakami et al., 2012;2013;Roberts et al., 2020). Manganello et al. (2012) compared the horizontal structure of moisture content of TC of GCM between low (T511) and high (T2047) resolutions and found that the refined structure of tropical cyclone liquid was simulated when the horizontal resolution is increased.

Modulation of the tropical cyclone track by the MJO
There is a clear evidence of the connection between the MJO and tropical cyclone activity worldwide (Camargo et al., 245 2009;Klotzbach et al., 2014). Zhang et al. (2013) summarized the connection between the MJO and global tropical cyclone activity and found that the MJO affects the formation and movement of tropical cyclones in each phase. Figure Figure 11) and this provides an accurate background for the formation and propagation of tropical cyclones.
Using the phase of the MJO to produce a composite of the daily track densities of tropical cyclones is an effective way of exploring the coordination between the MJO and tropical cyclone activity in GCMs. The biases of track density anomalies ( Figure 12) are in agreement with the precipitation anomalies in the MJO (Figure 10). The results indicate that FGOALS-f3-260 L cannot reasonably reproduce the pattern of track densities in each phase of the MJO in the northern Indian, western Pacific, eastern Pacific and northern Atlantic oceans. There is considerable improvement in the track densities modulated by the MJO when the horizontal resolution is increased (Figure 13). The improvement in the MJO with an increase in the horizontal resolution plays a crucial role in simulating the variability of tropical cyclones at sub-seasonal to seasonal scales.

Large-scale environmental factors 265
The genesis potential index (GPI; Emanuel et al., 2004) is applied to detect the connection between the genesis of tropical cyclones and large-scale circulation patterns. Camargo et al. (2007) and Walsh et al. (2013) found that the correlation between the GPI and the variation of tropical cyclones in GCMs mainly depends on the horizontal resolution and the similarity between the GPI. The variation in tropical cyclones is increased when the horizontal resolution is increased. The GPI used in this work is defined as: 270 where vort850 is the 850 hPa absolute vorticity (s −1 ), RH is the 600 hPa relative humidity (%), Vm is the maximum potential intensity (Emanuel, 1995) and Vshear is the magnitude of the wind shear between 850 and 200 hPa (m s −1 ). Vm (the maximum potential intensity) is defined here as: where Ck is the exchange coefficient of the enthalpy, Cd is the drag coefficient, Ts is the sea surface temperature and T0 is the mean outflow temperature. CAPE* is the convective available potential energy of the air lifted from saturation at sea-level and CAPE b is the convective available potential energy of the boundary layer air. Figure 14 shows the GPI for the FGOALS-f3-L (Figure 14a), FGOALS-f3-H ( Figure 14b) and ERAI (Figure 14c) simulations. There is a considerable underestimation of the GPI in FGOALS-f3-L, which is more remarkable in the eastern 280 Pacific and northern Atlantic oceans. These biases are consistent with the biases of the simulated tropical cyclones (Figure 2a).
There is more consistency between the GPI in the observations (Figure 14c) and the simulated frequency of tropical cyclones  (Figure 15d) oceans, which is favored by the reduction in the bias of the large-scale circulation patterns.

The physical parameterization and its impact on the simulated TC in GCMs
A Resolving Convective Precipitation (RCP) scheme has been used in both the high and low versions of FGOALS-f3 (Bao and Li, 2020;He et al., 2019;Li et al., 2019). The RCP scheme calculates convective and stratiform precipitation at the grid scale, which has the advantage of both scale-awareness and high computational efficiency. The parameterizations of physical processes in traditional GCMs are very sensitive to the change of resolution. Especially, the processes of convection 295 and clouds is considered as effective resolved, which means the assumptions and equations in the low-resolution condition are not suitable for the high-resolution GCMs. As the result, the model convergence with increasing resolution will be degraded (Sakradzija et al., 2016). Simulated TC are very sensitive to the processes of convection and cloud in GCMs (Zhao et al., 2012).
Actually, the effective tunning for the convection (Lim et al., 2015;Murakami et al., 2012), boundary (Zhang et al., 2017), and microphysics (Chutia et al., 2019) parameterizations will contribute to the improvement of intensity, number, track, and 300 structure of simulated TC. Although the fixed parameterization scheme combined with the fine grid will improve the simulation performance of TC obviously, the effect of resolution will be amplified. According to this study, the GCM with the scaleaware parameterizations still slightly underestimates the intensity of TC at 0.25 o degree resolution (Figure 1). Besides, air-sea exchanges and non-hydrostatic processes are both important to enhance intense of TC (Ma et al., 2017). Emanuel and Sobel, (2013) found that the absence of air-sea coupling can lead to potentially large imbalances in the surface energy budget, which 305 is not conductive to the development of TC.

The potential value-added effect of MJO and TC due to the increase in horizontal resolution from HighResMIP models
The results show a clear improvement in the relationship between the MJO and tropical cyclone activity when the horizontal resolution of FGOALS-f3 is increased from 100 to 25 km. This result indicate that the large-scale background 310 associated to the TC is improved when increased the horizontal resolution of FGOALS-f3. It is worth exploring whether this improved relationship is common to all the GCMs participating in CMIP6. The eight models participating in the HighResMIP Tier1 were selected (Table 4) to calculate the MJO phases and GPI separately. Figure 16 shows the ranking of the models according to the average anomaly correlation coefficient of the MJO from phases 1 to 8 and shows that there is an increase in the anomaly correlation coefficient of the MJO in the high-resolution models (red cylinders) relative to the low-resolution 315 models (blue cylinders). The MJO phases simulated by the models could provide an accurate large-scale background for the generation and development of tropical cyclones (Zhang et al., 2013). Figure 17 shows the anomaly of the composite GPI by MJO phases 4-7 between the multi-model mean of the GCMs and the ERA-Interim dataset and the result indicates that the biases of the GPI in the South China Sea and the western north Pacific Ocean are decreased in the high-resolution models ( Figure 17b) relative to the low-resolution models (Figure 17a) as a result of the improved simulation of the MJO phases when the horizontal resolution of the models is increased. Although there seems to be a significant improvement in the MJO-TC relationship when the horizontal resolutions of the GCMs are increased, it is worth noting that not all the GCMs participate in the HighResMIP follow the rules: "The experimental set-up and design of the standard resolution experiments will be exactly the same as for the high-resolution runs", which mean the specific optimization in the parameterization of physical processes for the GCMs in the high horizontal resolution (Roberts et al., 2020). The consequence is that the effect of horizontal resolution 325 on TC simulation will be overestimated or underestimated (Strachan et al., 2013).

Summary and conclusions
The impacts of horizontal resolution on the simulation of tropical cyclones were studied with the latest version of FGOALS-f3, which participated in CMIP6 HighResMIP (Haarsma et al., 2016). Li et al. (2019)  The improvement in the large-scale environmental factors in FGOALS-f3-H contributes directly to the simulation of tropical cyclones. This study shows that it is worth establishing a high-resolution coupled dynamic prediction system based on FGOALS-f3-H to improve the prediction skill of tropical cyclones on sub-seasonal to seasonal scales (Camp et al., 2018;Murakami et al., 2016). This dataset will be uploaded to the sub-seasonal to seasonal prediction project (Vitart, 360 et al., 2018;Vitart et al., 2017).

Code and data availability
The model output of FGOALS-f3 models for CMIP6 simulations, which is used in this work is uploaded to the Earth System Grid Fedration (ESGF), and the users can access to these outputs freely.

Competing interests
The authors declare that they have no conflict of interest.   (Clough et al., 2005) Gravity wave drag scheme Palmer et al. (1996) Cloud fraction diagnosis scheme Xu & Randall (1996) Convection microphysics scheme Resolving Convective Precipitation (Bao and Li, 2020)