Recent advances in high-performance computers facilitate operational
numerical weather prediction by global hydrostatic atmospheric models with
horizontal resolutions of
The Global 7 km mesh nonhydrostatic Model Intercomparison Project for improving TYphoon forecast (TYMIP-G7) is designed to understand and statistically quantify the advantages of high-resolution nonhydrostatic global atmospheric models to improve tropical cyclone (TC) prediction. A total of 137 sets of 5-day simulations using three next-generation nonhydrostatic global models with horizontal resolutions of 7 km and a conventional hydrostatic global model with a horizontal resolution of 20 km were run on the Earth Simulator. The three 7 km mesh nonhydrostatic models are the nonhydrostatic global spectral atmospheric Double Fourier Series Model (DFSM), the Multi-Scale Simulator for the Geoenvironment (MSSG) and the Nonhydrostatic ICosahedral Atmospheric Model (NICAM). The 20 km mesh hydrostatic model is the operational Global Spectral Model (GSM) of the Japan Meteorological Agency.
Compared with the 20 km mesh GSM, the 7 km mesh models reduce systematic errors in the TC track, intensity and wind radii predictions. The benefits of the multi-model ensemble method were confirmed for the 7 km mesh nonhydrostatic global models. While the three 7 km mesh models reproduce the typical axisymmetric mean inner-core structure, including the primary and secondary circulations, the simulated TC structures and their intensities in each case are very different for each model. In addition, the simulated track is not consistently better than that of the 20 km mesh GSM. These results suggest that the development of more sophisticated initialization techniques and model physics is needed to further improve the TC prediction.
Global models provide fundamental information for operational weather
forecasting at daily, weekly and seasonal timescales. Moreover, such models
produce initial and lateral boundary conditions to limited-area models, which
furnish fundamental information for local-scale weather forecasts. Therefore,
operational numerical weather prediction centres have been developing
sophisticated global models with high resolution and accuracy. Because such
models require huge computational resources, their development strongly
depends on advances in high-performance computers. Recent computer progress
has facilitated the reasonable operation of global models with horizontal
resolutions of
Developing high-resolution models with a horizontal grid spacing of
Because developing operational numerical weather prediction models with high accuracy requires huge computational and human resources, the concept of transition of research to operations (R2O) has recently been encouraged. For example, the Hurricane Weather Research and Forecasting Model (Bernardet et al., 2015) and an atmosphere–ocean coupled limited-area model (Ito et al., 2015) have been developed based on R2O in the United States and Japan, respectively. In Japan, two next-generation, nonhydrostatic global atmospheric models have already been developed and used in the research community. These are called the Multi-Scale Simulator for the Geoenvironment (MSSG) and the Nonhydrostatic ICosahedral Atmospheric Model (NICAM). In addition, the Meteorological Research Institute (MRI) of the Japan Meteorological Agency (JMA) has developed a next-generation nonhydrostatic atmospheric model called the nonhydrostatic global spectral atmospheric Double Fourier Series Model (DFSM). To gain knowledge, to develop and improve nonhydrostatic global models and to share them with the research and operational communities are some aims of the present project.
Tropical cyclones (TCs) are characterized by violent winds and torrential
rain. These events cause tremendous damage to human lives, property and
socioeconomic activity via landslides, floods and storm surges. Because an
average of 26 TCs (
Errors in TC track prediction by the JMA operational global atmospheric model
at a given lead time have decreased on an average by half over the past
20 years (JMA, 2014) as the operational model has been upgraded. For example,
TC track prediction error in a 30 h forecast with a 60 km mesh global model
was
Despite the advances in TC track prediction, improvements in TC intensity
predictions by global atmospheric models remain a challenge. One factor that
impedes improvement in the intensity prediction is the lack of horizontal
resolution to capture essential mechanisms of TC intensity changes. TC
intensity and its variation are closely related to the inner-core structure
and convective activity (e.g. Rogers et al., 2013; Wang and Wang, 2014).
Recent studies using a high-resolution, limited-area atmospheric model show
that the use of a horizontal resolution of a few kilometres is necessary to
realistically reproduce the inner-core structure and associated convection
(e.g. Braun and Tao, 2000; Gentry and Lackmann, 2010; Kanada and Wada, 2015).
Fierro et al. (2009) examined the dependence of TC intensity prediction using
horizontal resolutions from 30 to 1 km and pointed out that the predicted TC
intensity became increasingly realistic with resolutions between 15 and
5 km. Therefore, the use of a high-resolution global atmospheric model with
a horizontal resolution of
The primary objectives of the Global 7 km mesh nonhydrostatic Model Intercomparison Project for improving TYphoon forecast (TYMIP-G7) are to understand and statistically quantify the advantages of high-resolution global atmospheric models towards the improvement of TC track and intensity forecasts. The project is conducted as a strategic program of the Earth Simulator of the Japan Agency for Marine-Earth Science and Technology (JAMSTEC). We accomplish this objective via a model intercomparison of three 7 km mesh nonhydrostatic atmospheric models (DFSM, MSSG and NICAM) and a 20 km mesh hydrostatic operational atmospheric model of the JMA (Global Spectral Model; GSM) in various cases. Because a huge amount of data is produced by each model, we developed an effective method to handle and visualize the data. Sharing the knowledge obtained in this project with research and operational communities will facilitate R2O.
In this paper, we describe the specifications of TYMIP-G7 and the set of metrics used to validate the model performances. Some preliminary results concerning the metrics are also shown. This paper comprises six sections. Section 2 describes the common experimental design, including the cases and the output dataset. Section 3 briefly overviews the scientific outcomes of each model and describes the detailed specifications. Section 4 presents the metrics, analysis method and visualization. Preliminary results concerning the advantages of high-resolution models for TC prediction and the simulated TC wind structure are given in Sect. 5. Section 6 is devoted to conclusions and future work.
We imitated JMA operational specifications to conduct 5-day numerical
experiments with the models (DFSM, GSM, MSSG and NICAM). The JMA 6-hourly
global objective analysis data were used for each model to derive
atmospheric initial conditions. The data were provided based on the GSM grid
system, a linear Gaussian grid with a horizontal resolution of 20 km and a
hybrid sigma-pressure vertical coordinate. DFSM and GSM interpolated data
directly onto their model grids, whereas MSSG and NICAM preliminarily
interpolated the data onto common latitude–longitude grids and pressure
levels and then interpolated this to their model grids. A merged satellite
and in situ data global daily sea surface temperature (SST) product with a
horizontal resolution of 0.25
The project was implemented using the Earth Simulator, a supercomputer system operated by JAMSTEC. The Earth Simulator is based on NEC SX-ACE, a distributed-memory, massively parallel vector system with a total of 5120 computational nodes. Each node has one central processing unit, which comprises four processing cores and a 64 GB main memory. The theoretical peak performance of the entire system is 1.3 peta floating-point operations per second.
List of initial times for stage 1 of TYMIP-G7. Typhoon cases in italic font are weaker than a tropical storm and those in bold italic font are extratropical cyclones.
List of initial times for stage 2 of TYMIP-G7. Typhoon cases in italic font are weaker than a tropical storm and those in bold italic font are extratropical cyclones.
Continued.
We conducted the project for two stages: from June 2015 to September 2015 and from October 2015 to March 2016. The first stage addressed TCs from September to October in 2013, during the most active TC season since 1951. We forecasted nine TCs in 52 runs (Table 1). However, we detected some flaws in MSSG and NICAM, and we could not perform some of the numerical experiments. The second stage addressed the life cycle of a TC, e.g. genesis, rapid intensification, recurvature and extratropical transition in addition to the Madden–Julian oscillation (MJO; Madden and Julian, 1972) and the boreal summer intraseasonal oscillation (BSISO; Wang and Rui, 1990; Wang and Xie, 1997). After we improved the detected flaws, we examined 13 TCs in 85 runs (Table 2) in addition to the numerical experiments in the first stage. We analyse the model output obtained in the second stage in this paper.
Model output data for every 1 or 3 h from each experiment (Tables 1 and 2)
were stored for analyses. The components of the output are listed in Table 3.
Even though each model uses its own grid system, the output data were
prepared for a regular latitude–longitude (lat–long) grid system. In TYMIP-G7,
we used GrADS file formats (pairs of 4-byte IEEE 754 floating-point standard
with a big-endian binary file and a control file in text format) that are
common in the atmospheric and oceanic research fields. The domain of the
output data covers the globe, including the western North Pacific Ocean
(100–180
We used three 7 km mesh nonhydrostatic global atmospheric models in TYMIP-G7
(Fig. 1). The DFSM was developed in the MRI of the JMA. The MSSG was
developed at JAMSTEC. NICAM was developed at JAMSTEC, the University of Tokyo
and the RIKEN Advanced Institute for Computational Science. In addition, we
used GSM with a horizontal grid spacing of
Output variables and domains.
GSM (JMA, 2013) is a hydrostatic global spectral atmospheric model using
spherical harmonics. The JMA has used this model operationally to provide
fundamental information for forecasts. The model was put into operation in
1988 with T63L16 resolution (200 km mesh), where “T
The JMA has also used GSM as the principal part of an ensemble prediction system for medium-range weather forecasts. The forecast data are widely provided via the framework of “The observing-system research and predictability experiment Interactive Grand Global Ensemble” (TIGGE) for the research community. TIGGE data have been used for various applications, including TC track prediction (Yamaguchi et al., 2012, 2015) and the MJO (Matsueda and Endo, 2011). In addition, GSM has been used to produce atmospheric reanalysis datasets, i.e. the Japanese 25-year ReAnalysis (JRA-25; Onogi et al., 2007) and the Japanese 55-year ReAnalysis (JRA-55; Kobayashi et al., 2015). MRI global climate models have been developed based on GSM and have been used in climate research, such as global warming projections (e.g. Mizuta et al., 2006; Yukimoto et al., 2011) and stratospheric studies (e.g. Shibata et al., 1999). TC activity in future climates has been intensively studied using various model physics and horizontal resolutions (Murakami and Sugi, 2010; Murakami et al., 2012a, b).
Schematic diagram of the horizontal grid structures of the three models used in TYMIP-G7.
Brief description of the specifications for each global nonhydrostatic model.
The MRI developed DFSM by changing the hydrostatic dynamical core of GSM using spherical harmonics to a nonhydrostatic dynamical core using a double Fourier series (Yoshimura, 2012). DFSM uses the same basis functions of the double Fourier series as Cheong (2000). In DFSM, a fast Fourier transform is used instead of a Legendre transform in the meridional direction. Because the computational cost of the fast Fourier transform is much smaller than that of the Legendre transform, especially at high resolution, DFSM is applicable to finer-resolution simulations. DFSM gives nearly the same results as GSM using the Legendre transform; a comparison of 2-day forecasts using the 60 km resolution model was shown by Yoshimura and Matsumura (2005).
In GSM and DFSM, a two-time-level, semi-implicit, semi-Lagrangian scheme (e.g. Hortal, 2002) is used to facilitate long time steps for computational efficiency. The vertically conservative semi-Lagrangian scheme is used in the advection calculation (Yoshimura and Matsumura, 2003, 2005; Yukimoto et al., 2011), and a correction method similar to that described by Priestley (1993) and Gravel and Staniforth (1994) is used for global conservation in the material transport. To save computational costs, we used a reduced grid (Miyamoto, 2006) in which the number of zonal grid points is decreased, especially at high latitudes (Fig. 1).
Because the DFSM resolution is
Physical packages included in GSM and DFSM are the same as those in the March 2014 version of the operational global atmospheric model of the JMA. A prognostic cumulus parameterization scheme (Randall and Pan, 1993) and other schemes in GSM are used in DFSM without any changes. The physical process is described in detail in the JMA (2013).
MSSG is an atmosphere–ocean coupled nonhydrostatic model aimed at a seamless simulation from global to local scales (Takahashi et al., 2006, 2013). The MSSG comprises atmospheric (MSSG-A) and oceanic (MSSG-O) components. MSSG uses a conventional lat–long grid system for regional simulations and the Yin–Yang grid system (Kageyama and Sato, 2004; Baba et al., 2010), which comprises two overlapping lat–long grids to avoid the polar singularity problem, for global simulations. MSSG has been used in a wide range of applications. A cloud-system-resolving global ocean–atmosphere coupled MSSG successfully simulated an observed MJO propagation (Sasaki et al., 2016). A global atmosphere–ocean coupled experiment with 11 km horizontal resolution with a nested region with 2.7 km horizontal resolution simulated sea surface cooling caused by a TC along its track (Takahashi et al., 2013). High-resolution regional atmospheric simulations have been conducted to investigate the influence of the choice of cloud microphysics scheme and in-cloud turbulence on cloud development (Onishi et al., 2011, 2012). MSSG-O with a 2 km horizontal resolution has been used to investigate the dispersion of radionuclides released from the Fukushima Daiichi nuclear power plant (Choi et al., 2013) and the effect of wind on long-term summer water temperature trends in Tokyo Bay, Japan, with 200 m horizontal resolution (Lu et al., 2015). MSSG-A with a 5 m spatial resolution has been used in building-resolving urban atmosphere simulations to examine the heat environments of streets (Takahashi et al., 2013).
In this study, MSSG-A is exclusively used. Its dynamical core is based on the
nonhydrostatic equations, and it predicts the three wind components, as well as air
density and pressure. Each horizontal computational domain covers
4056
During the first stage of the project, extraordinary increases in precipitable water appeared in the 5-day integrations when the conventional bulk surface flux model of Zhang and Anthes (1982) was used for both land and ocean surfaces. This issue was solved by the use of the COARE 3.0 model (Fairall et al., 1996, 2003) for ocean surface fluxes with Zhang and Anthes (1982) being used only for land surface fluxes. This combination was used for all simulations in the second stage, and we plan to rerun all the simulations in the first stage.
NICAM (Satoh et al., 2008, 2014) was developed as a climate model and can explicitly resolve clouds without any convective parameterization, which is known to be the most ambiguous component in conventional climate models (Randall et al., 2003). From the first appearance of realistic cloud-resolving simulations using a 3.5 km mesh horizontal resolution by Miura et al. (2007a), NICAM has primarily been used to study tropical meteorological systems, such as the MJO (Miura et al., 2007b; Nasuno, 2013; Miyakawa et al., 2014), TC genesis from the MJO in boreal winter (Fudeyasu et al., 2008, 2010a, b), TC genesis from the BSISO in the western North Pacific (Oouchi et al., 2009; Nakano et al., 2015) and BSISO in the northern Indian Ocean (Taniguchi et al., 2010; Yanase et al., 2010). NICAM has also been used for quasi-real-time forecast systems during field observation campaigns to support field observations (Nasuno, 2013). Recent progress with high-performance computing infrastructures, such as the K-computer, a 10-petaflop supercomputer in Japan, facilitates 870 m mesh global simulations (Miyamoto et al., 2013, 2015; Kajikawa et al., 2016). This is the highest resolution to date (10 July 2016). Climate simulations (of 30 years) using a 14 km mesh model (Kodama et al., 2015) and large member (10 240 members) ensemble data assimilations based on an ensemble Kalman filter (Miyoshi et al., 2015) have also been executed.
NICAM uses an icosahedral grid system that covers the globe with a nearly
uniform grid size, avoiding the polar singularity problem. Increased
horizontal resolution is attained by recursively dividing horizontal grids in
half. Therefore, the possible horizontal resolution is discrete and
represented by the “g-level”, which indicates the number of divisions of a
horizontal grid. In this project, the 2014 version of NICAM (called
NICAM.14.3) was used with a horizontal resolution of g-level 10,
corresponding to a 7 km mesh. The vertical level comprises 38 vertical
layers to a top height of 36.7 km with the lowest layer at 80 m. NICAM uses
a fully compressible nonhydrostatic equation system for the dynamics of the
atmosphere. The model uses an icosahedral grid system in the horizontal
direction with the Arakawa A-grid and terrain-following coordinate with the
Lorenz grid in the vertical direction. The equations are discretized using
the flux form of the finite volume method. The numerical scheme guarantees
conservation of total mass and energy. The second-order Runge–Kutta scheme
is primarily used for time integration, whereas the third-order Runge–Kutta
scheme is used in some cases to avoid computational instability. NICAM uses
the split-explicit scheme together with the horizontal explicit and vertical
implicit scheme to avoid the restriction of the Courant–Friedrichs–Lewy
condition for acoustic waves. The NICAM Single-moment Water 6 cloud
microphysics scheme (Tomita, 2008) is used for cloud microphysics without any
convective parameterization. Planetary boundary layer processes are
calculated using the Mellor–Yamada–Nakanishi–Niino level 2 scheme
(Nakanishi and Niino, 2004) implemented and examined by Noda et al. (2010).
Longwave and shortwave radiation transfer is calculated using MstranX
(Sekiguchi and Nakajima, 2008). Land surface processes are computed by the
Minimal Advanced Treatments of Surface Interaction and Runoff (MATSIRO;
Takata et al., 2003). NICAM is coupled with a simple slab ocean model. This
model calculates SST based on the local heat balance between the ocean slab
and the atmosphere, and the other ocean dynamics, such as vertical mixing and
advection, are not considered. The slab has a specific heat capacity
determined by its thickness (15 m). The calculated SST is nudged with a
persistent SST anomaly with an
During the first stage of this project, there were frequent problems of divisions by zero in MATSIRO that had not been experienced in simulations with coarser horizontal resolutions. This issue was fixed before simulations in the second stage, and abnormal cases in the first stage had to be rerun. The fix had a slight impact on the prediction results. During the second stage, however, two cases were still unable to be completed due to numerical instability (Table 2).
Here, we define the following metrics to evaluate the TC forecast
performance:
computational resources for a 5-day forecast on the Earth Simulator
(node hours), TC track (position) error every 6 h of forecast time (km), TC intensity (central pressure) error every 6 h of forecast time
(hPa), averaged radius of surface 50-knot (25 m s averaged radius of surface 30-knot (15 m s
It is important for the operational model that the calculation is completed in less time with smaller computational resources so that we applied metric (1). The metrics (2)–(5) measure the accuracy of the track, intensity and surface wind structure prediction based on the RSMC Tokyo best-track data.
Screen capture of the web application: outgoing longwave radiation at 14 September 2013, 10:00:00 UTC, simulated in experiments initialized at 12 September 2013, 06:00:00 UTC.
We extract TC tracks from the model experiments using the hourly mean sea
level pressure (SLP) data with a horizontal resolution of
The RSMC Tokyo best-track data contain longest and shortest radii of 50-knot and 30-knot wind speeds and their direction. AR50 and AR30 are defined as the average of the longest and shortest radii of the 50-knot and 30-knot wind speeds, respectively. The directions of the longest and shortest radii are defined by eight directions (N, NE, E, SE, S, SW, W and NW) in the best-track data. Therefore, we calculated the radii of the 50-knot and 30-knot wind in the model in each of the eight directions first and then determined the direction of the longest and shortest radii. Then, the radii in those two directions were averaged to obtain AR50 and AR30.
The multi-model ensemble mean (MME) is applied to the three 7 km mesh models (DFSM, MSSG and NICAM). MME is a simple ensemble average derived from a combination of individual models, which reduces the average forecast error relative to the best individual predictions by the individual models. MME also provides additional information about the forecast uncertainty, enhancing forecast confidence (Goerss, 2000; Yamaguchi et al., 2012).
We developed a web application that allows the simultaneous visualization of
multi-model results. Figure 2 shows a screen capture of this application,
which portrays digital globes using Cesium.js (Analytical Graphics, Inc.,
Computational performance is an important metric for an operational numerical weather forecast model. DFSM, MSSG and NICAM models consumed computational resources equivalent to 682, 2330 and 1155 node hours, respectively, for a case on 12 September 2013, 00:00:00 UTC. These quantities did not vary greatly between cases because the computational nodes were occupied in each calculation and the disk I/O was executed from/to the work disk mounted on each computational node. Note that the computational resources required for each model are highly dependent on the model specifications (e.g. the physics scheme, advection scheme, number of vertical layers, vertical resolution and time step) and the degree of optimization for the Earth Simulator.
Errors in the track prediction for GSM, DFSM, MSSG, NICAM and MME (in the second stage). Each grey bar indicates the number of samples at each forecast time (right-vertical axis). Error bars indicate 95 % confidence levels of the central pressure difference between the prediction and the RSMC Tokyo best-track data.
To quantify the advantage of using finer resolutions for TC track prediction, we examined the time series of TC track prediction errors with reference to the RSMC Tokyo best track for the second stage (Fig. 3). TC track predictions by DFSM, MSSG and NICAM performed better than GSM. However, the reduction in the track errors depended on the TC case. That is, the use of finer resolution alone does not always improve TC track prediction. This suggests that improvements in the initial conditions and those of the physical processes in each model are also required to improve track prediction.
We also validated MME using track predictions of the three models with
reference to the RSMC Tokyo best-track data. MME track prediction gave the
smallest track errors for forecast time (FT) of 24–120 h. The reduction
rate of the MME position error from that of GSM was
Errors in the predictions of the central pressure for GSM, DFSM, MSSG, NICAM and MME (in the second stage). Each grey bar indicates the number of samples at each forecast time (right-vertical axis). Error bars indicate 95 % confidence levels of the central pressure difference between the prediction and the RSMC Tokyo best-track data.
Figure 4 shows time series of the average central pressure and the standard
deviation in each model relative to the RSMC Tokyo best-track data for the
second stage. Because the global objective analysis data, which were used as
initial conditions of the numerical experiments, tend to reproduce TC central
pressure shallower than those in RSMC Tokyo best-track data, cases with an
initial bias
Errors in the averaged radius of the 50-knot wind (AR50) for GSM, DFSM, MSSG, NICAM and MME (in the second stage). Each grey bar indicates the number of samples at each forecast time (right-vertical axis). Error bars indicate 95 % confidence levels of the AR50 difference between the prediction and the RSMC Tokyo best-track data.
Accurate predictions of AR50 and AR30 lead to accurate estimations of the
area affected by TCs. Figure 5 shows the validation result of AR50 based on
the RSMC Tokyo best-track data. All models had negative bias of 80–90 km
even at the initial time. This negative bias is partially attributed to the
shallower estimation of the central pressure by
Errors in the averaged radius of the 30-knot wind (AR30) for GSM, DFSM, MSSG, NICAM and MME (in the second stage). Each grey bar indicates the number of samples at each forecast time (right-vertical axis). Error bars indicate 95 % confidence levels of the AR30 difference between the prediction and the RSMC Tokyo best-track data.
Figure 6 shows the validation results of AR30. All models show a negative
bias of more than 200 km at FT
Composite analysis of the radius-height cross section of the
axisymmetric mean radial (shaded) and tangential (contour) wind speed for TCs
at the time of the analysed central pressure between 920 and 940 hPa in the
RSMC Tokyo best-track data. Contour intervals are 5 m s
An accurate prediction of the three-dimensional TC structure can lead to accurate predictions of the intensity, AR30 and AR50. Because there is no high-resolution TC observation that is suitable for the validation of the simulated TC structure, here we made an intercomparison of the TC wind structures simulated by the 7 km models and 20 km mesh GSM. Figure 7 shows a composite of the radius-height section of the azimuthal mean radial and tangential wind speeds for TCs at the time of the RSMC Tokyo best-track central pressure between 920 and 940 hPa, corresponding, in the life cycle, to the mature stage of a TC. A total of 347 snapshots were used for the composite analysis. If the models can simulate the TC structure perfectly, the result should be the same for all models. While all 7 km mesh models reproduced typical axisymmetric mean inner-core structures, such as primary and secondary circulations, the simulated TC structures differed significantly between the 7 km models as expected above. The TCs calculated by DFSM had the highest maximum tangential wind speed and the smallest radius of maximum wind (RMW) of the models. In addition, its primary circulation was the deepest, reaching up to 100 hPa in the vertical direction and the narrowest in the horizontal direction. The depth of the inflow and outflow layers in DFSM was relatively thin and had the strongest radial velocity. The TCs in NICAM and MSSG showed relatively similar structures to each other; however, MSSG had thicker inflow and outflow layers. Differences in the heating and inertial stability in the inner core led to differences in the primary and secondary circulation (Shapiro and Willoughby 1982). Understanding the cause of the differences in the simulated structures in the models will lead to improvements in all the models.
TYMIP-G7 was implemented in two stages from June 2015 through March 2016. The
aim of the project was to statistically quantify and understand the
advantages of high-resolution global atmospheric models to improve 5-day TC
track, intensity and wind radii forecasts. We performed numerical experiments
for multiple TC cases in 137 runs using three 7 km mesh global
nonhydrostatic atmospheric models: DFSM, MSSG and NICAM. We also included a
20 km mesh global hydrostatic atmospheric model, GSM, on the Earth Simulator
of JAMSTEC. We statistically evaluated errors in the TC track, intensity and
wind radii predictions with the following primary results.
The 7 km models statistically improve both the TC intensity and track
predictions, whereas the improvement in the individual TC tracks depends on
the case. The MME is a promising approach to further enhance the TC track and
AR50 predictions. The predicted TC structure differs greatly between the three models
even though they have the same horizontal resolution.
To follow up the above results to further improve TC prediction, we must
answer the following questions.
Why are the TC predictions improved by high-resolution models? What are the factors that cause the differences in the simulated TC structure (such as the radius of the maximum
winds, the eyewall slope, the inflow and outflow layers and the rainbands) in the three
7 km mesh atmospheric global models?
To answer (Q1), an intercomparison of forecasts by the 20 km mesh GSM and the 7 km mesh models (DFSM, MSSG and NICAM) is the first step. Concerning (Q2), the predicted TC structure depends on the physics schemes, such as cloud microphysics, planetary boundary layer and surface flux, as well as the dynamical core of the model. To understand the impacts of the model physics schemes, sensitivity experiments altering the schemes and/or tuning parameters will be required.
In addition, the following topics are suggested for future work:
extended-range forecasts, contributing to TC genesis and MJO/BSISO
forecasts; atmosphere–ocean coupled experiments to examine impacts on TC
intensity and track and MJO/BSISO; further high-resolution experiments to study impacts of better
inner-core representation on TC intensities and tracks; and data assimilation to contribute for validating the models and
understanding the TC processes and model initializations.
These topics are addressed below.
An advantage of global models for TC prediction over limited-area models is the coverage of multi-scale atmospheric phenomena from a mesoscale vortex to synoptic environments. Because TC genesis strongly depends on synoptic environments modulated by the MJO/BSISO, global models should be used for its forecasting. Indeed, Nakano et al. (2015) and Xiang et al. (2015) showed that TC genesis is predictable up to 2 weeks in advance; this great skill in TC genesis forecasting was attributed to its strong ability to forecast BSISO/MJO. We are conducting extended-range (longer than 2 weeks) forecast experiments using the four models in several cases and will investigate the advantage of high-resolution modes.
In the present project, atmosphere-only models were used, except for NICAM, which is coupled with a simple slab ocean model. However, studies have shown that fully coupled atmosphere–ocean processes are essential for especially slow-moving, intense TCs (e.g. Yablonsky and Ginis, 2009). Recently, Zarzycki (2016) reproduced sea surface cooling caused by TCs realistically using a global atmospheric model coupled with a slab ocean model with a simple parameterization of ocean turbulent mixing, which is not considered in NICAM, and demonstrated that the cooling led to significant reduction in TC intensity. These processes affect the TC structure and therefore the track and intensity. In addition, a fully coupled atmosphere–ocean model is better for MJO/BSISO forecasts. MSSG is already capable of coupling MSSG-A with MSSG-O (Sasaki et al., 2016; Takahashi et al., 2013). In addition, NICAM has been coupled with the Center for Climate System Research Ocean COmponent Model (COCO; Hasumi, 2006). Therefore, we will use these coupled global models to examine the impacts of global atmosphere–ocean processes on TC forecasts.
To improve the high-resolution models, the validation of simulated phenomena
using observations is essential. An understanding of the essential processes
and the modelling therefore requires high-resolution spatiotemporal
observations. Recent advances in satellite observations furnish
quantitatively and qualitatively rich observational data. However, the
spatiotemporal resolution is still insufficient for the validation of TC
structures simulated by high-resolution models. Aggressively developing data
assimilation techniques using satellite observations (e.g. Zhang et al.,
2016; Okamoto et al., 2016) is a promising means of obtaining
high-resolution, spatiotemporal, three-dimensional TC structures, including
those at the cloud convection scale (
Access to the initial and boundary data for the models and model outputs can be granted upon request, under a collaborative framework between MRI, JAMSTEC and related institutes or universities.
The authors declare that they have no conflict of interest.
This project was conducted as “The Earth Simulator Strategic Project with Special Support” of JAMSTEC. All numerical experiments were run on the Earth Simulator (NEC SX-ACE). This study was partly supported by HPCI Strategic Programs for Innovative Research (SPIRE) Field 3, the FLAGSHIP 2020 project of the Ministry of Education, Culture, Sports, Science and Technology (MEXT) and KAKENHI 26282111, 26400475 and 15K05292 of the Japan Society for the Promotion of Science (JSPS). The authors thank Mikiko Ikeda, Yuichi Saitoh and Hiromitsu Fuchigami for supporting the experiments on the Earth Simulator. The authors also acknowledge Hideaki Kawai and Eiki Shindo for the fruitful discussions. The schematic diagram of the NICAM grid was provided by Masaki Satoh. Edited by: P. Ullrich Reviewed by: two anonymous referees