Introduction

Tropical cyclones (TCs) are the most catastrophic natural disasters with heavy precipitation, strong wind gusts, and storm surges. 25–30 TCs occur every year over the western North Pacific, accounting for ~30% of global TCs1,2,3. Furthermore, more than 60% of TCs make landfall in Asian countries, causing tremendous economic losses and casualties4. An average of three TCs affect Korean Peninsula every year5,6 but recent records suggest its possible increase. In 2019, seven TCs affected South Korea, including three TCs in September, which was a record-high number since 19597. In addition, the northward shift in location of the maximum intensity over the western North Pacific can make a condition that is more vulnerable to TC in the Korean Peninsula8,9,10,11,12.

Lee et al.13 have also shown that the frequency of TCs affecting Korea and Japan has increased in early autumn due to favorable large-scale environments. Warm sea surface temperature (SST) over the western North Pacific has a positive effect on TC activity, especially in mid-latitudes, which is partly related to the Pacific Decadal Oscillation14,15,16. A more significant warming trend in mid-latitudes than in other regions in recent decades13,17 suggests that damage from TC might increase in the Korean Peninsula. However, few studies have quantified the influence of human-induced warming on TCs affecting the Korean Peninsula.

Climate change contributes to TC activity in the mid-latitude, which can be divided into two effects: Dynamic and thermodynamic. Dynamic effect is important because they are directly involved in the vortex of the TC. SST is a representative thermodynamic effect that can significantly influence TC activity. Many previous studies have shown that SST can affect the destructiveness and intensification of TCs18,19,20,21. SST is also closely associated with TC-induced precipitation. Water vapor holding capacity of the atmosphere exponentially increases as the temperature increases from the Clausius–Clapeyron relationship (~7% increase per 1 K)22,23,24. The resulting moisture increase becomes a major driver that intensifies extreme precipitation such as TCs rather than mean precipitation23,25.

In terms of modeling, previous studies have suggested that the performance of simulated TC intensity improves as the horizontal resolution increases26,27,28. In particular, a convection-permitting model (CPM) simulation is critically required to simulate the inner-core dynamics of TC realistically29. Furthermore, a grid length of 3 km or less is required to simulate the TC structure reasonably30. However, TC attribution studies using CPM simulations for the western North Pacific were rarely done. Thus, the aim of this study is to investigate how human activities could affect the intensity and precipitation of TCs landfalling in the Korean Peninsula using CPM simulations.

Results

Evaluation of the CPM simulations

The Weather Research and Forecasting (WRF) model was used to simulate TC activity. CPM simulation was conducted over a single domain with a horizontal resolution of 3 km. Spectral nudging was applied only in the 10th sigma level to the top of the model to reduce differences in TC tracks from observations. ALL experiment represents the current climate simulation while the NAT experiment is based on counterfactural conditions with human-induced warming removed. Here, anthropogenic SST warming pattern (delta SST) and associated lateral boundary conditions for temperature and relative humidity were estimated from differences between Coupled Model Intercomparison Project Phase 6 (CMIP6) historical (natural plus anthropogenic forcing) and historicalNat simulations (natural forcing only). To consider inter-model uncertainties, we obtained four delta SST patterns (delta SST clusters) using an empirical orthogonal function (EOF) analysis. More detailed information about model configuration and experimental design is provided in the Method section.

This study focused on extremely strong TCs, which belong to Saffir-Simpson Categories 4 and 5 (Supplementary Table 1). In addition, all TC cases considered underwent rapid intensification, having an increase of maximum wind speed by at least 15.4 m s−1 in 24 h31. In particular, TC SANBA and CHABA were destructive cases with lifetime maximum wind speed approaching 80 m s−1 (Fig. 1a, b). The WRF CPM simulations captured observed TC intensities reasonably well although some underestimation remained for TC SANBA and CHABA. This is consistent with the systematic model underestimation of the rapidly intensified TCs32,33. For relatively weak TCs (MAYSAK and HAISHEN), CPM simulations reproduced observed TC intensities realistically, showing good agreement with JTWC best track data (Fig. 1c, d). All TC cases made landfall on the southeastern Korean Peninsula and caused tremendous damage.

Fig. 1: Simulated maximum wind speed and minimum sea-level pressure.
figure 1

a–d Solid and dashed lines represent maximum wind speed (m s−1) and minimum sea level pressure (dashed line, hPa), respectively for four TCs (SANBA, CHABA, MAYSAK, and HAISHEN). Black and red colors indicate Joint Typhoon Warning Center (JTWC) best track data and ALL experiments for each TC case, respectively.

Full size image

We also attempted to evaluate simulated precipitation using the Tropical Rainfall Measuring Mission (TRMM) and Integrated Multi-satellitE Retrievals for Global precipitation mission (IMERG) data. Figure 2e–h represents the average precipitation within a 500 km radius of the TC center at each simulation time. Since the TRMM data are available up to 2019, it was not used for the 2020 TC cases (MAYSAK and HAISHEN). The WRF model tends to simulate larger precipitation than the IMERG but its precipitation is similar to that from the TRMM in the case of SANBA and CHABA. The WRF model skill is evaluated for extreme precipitation by comparing 6-hourly precipitation distributions (Fig. 2a–d). The WRF model reproduces extreme precipitation of the IMERG while the TRMM exhibits much weaker extreme precipitation than the IMERG. Note that differences between the satellite-based observations are substantial in both average and extreme precipitation, which might be due to their differences in temporal and spatial resolution, satellite databases, algorithms, and sensors34. Considering the large observational uncertainty, our evaluation indicates that the WRF CPM simulations have a reasonable skill for TC-induced extreme precipitation. Further, to reduce possible influences of model biases, anthropogenic influences are assessed by comparing differences between ALL and NAT experiments.

Fig. 2: Simulated TC-induced precipitation in comparison with satellite-based observations.
figure 2

a–d Probability density function (PDF) of 6-hourly precipitation in ALL experiment, IMERG, and TRMM data (logarithmic scale on the vertical axis). Precipitation <10 mm per 6 h was not considered in the PDF estimation. e–h TC precipitation in each simulation time. Red, black, and gray solid lines indicate ALL experiment, IMERG, and TRMM data, respectively.

Full size image

All CPM experiments regardless of ALL or NAT produced TC tracks very similar to JTWC best track data, reflecting influences of the employed spectral nudging (Fig. 3). Similar spectral nudging techniques were also implemented in previous studies to keep TC tracks as observed33,35. It should be noted that the influence of global warming on TC tracks could not be evaluated with this setting. This study rather focused on understanding the dynamic and thermodynamic influences of global warming on TC intensity and precipitation by controlling TC tracks as observed. See below for a further discussion.

Fig. 3: JTWC best track data and simulated TC tracks.
figure 3

a–d Joint Typhoon Warning Center (JTWC) best tracks (black line) and simulated tracks of four TC cases (colored line) in ALL and all NAT experiments. NAT experiments consist of five ensemble members based on different ‘delta-SST’ estimates (MME, C1, C2, C3, and C4).

Full size image

Comparison between current and counterfactual experiments

The maximum wind speed and minimum sea-level pressure were analyzed as indicators representing TC intensity. In the ALL experiments, mean intensities of TC were simulated stronger than those of NAT experiments except for a few cases (Fig. 4b). The difference in lifetime maximum intensity was much more significant than the mean intensity (Fig. 4a). This result indicates that anthropogenic forcing could affect the lifetime maximum intensity more than the overall intensity. However, the difference in TC intensity between delta-SST clusters (MME, C1, C2, C3, and C4) was insignificant. In conclusion, ALL and NAT experiments show that the TC intensity is largely amplified when human-induced global warming is added.

Fig. 4: Difference in TC intensity between ALL and NAT experiments.
figure 4

Differences between ALL and NAT experiments in minimum sea level pressure (hPa) and maximum wind speed (m s−1) in a lifetime maximum intensity and b mean intensity. Blue, red, green, and purple circle represent tropical cyclone SANBA, CHABA, MAYSAK, and HAISHEN, respectively. Delta-SST cluster number (from C1 to C4) and MME are indicated for each mark in NAT experiments.

Full size image

Figure 5a–d illustrates probability density function (PDF) distributions of TC precipitation, indicating an integrated area (grid points) of TC rainfall during its lifetime for a given precipitation intensity. TC precipitation is defined as precipitation within a 500 km radius grid points from the TC’s center at 6-hour intervals, which is aggregated to construct PDF distributions. This criterion has been widely used in previous works36,37,38, covering primary wind circulation and adjacent spiral rainband of TCs (80–400 km radius from the TC center)39,40. Using a radius smaller than 500 km may carry the risk of not capturing all of the TC precipitation. Furthermore, Prat and Nelson39 found little change in TC precipitation when adjusting the radius to 500–1000 km. Grid points with precipitation of more than 10 mm were only considered. PDF differences in light to moderate precipitation (<100 mm) between ALL and NAT experiments remained small since there were relatively broad areas in all TC cases (Fig. 5a–d). However, a clear difference was detected in the heavy precipitation area. ALL experiments simulated increased PDFs (i.e. more grid points) above 150 mm by 16–37% compared to NAT experiments in all four TCs and all delta SST clusters (Bar graphs in Fig. 5a–d). We also calculated extreme precipitation (90, 99, 99.9 percentiles) from each experiment, using the same size of space-time samples (Table 1). Consistent with the PDF results, the precipitation intensity in the ALL experiments was larger than NAT results for all four TCs, more strongly at 99.9 percentile, supporting the intensification of extreme precipitation.

Fig. 5: Probability density function and spatial distribution of TC precipitation.
figure 5

a–d Probability density function (PDF) of 6-hourly precipitation in ALL and all NAT experiments (logarithmic scale on the vertical axis). Precipitation <10 mm per 6 h was not considered in the PDF estimation. Bar graphs indicate percentage of grid points having precipitation above 150 mm. e–h TC precipitation (mm per 6 h) within a 500 km radius in the ALL experiment averaged throughout the simulation period and i–l its difference between ALL and NAT_MME experiments.

Full size image
Table 1 Extreme precipitation at different percentiles.
Full size table

Spatial distributions of precipitation are analyzed using 6 hourly precipitation amounts averaged throughout the TC simulation period (Fig. 5e–l). Heavy precipitation of TC in ALL experiments commonly developed around the TC’s inner core regions (Fig. 5e–h). Large differences in precipitation between ALL and NAT were also seen near the TC center (Fig. 5i–l). This could explain why heavy precipitation areas are significantly reduced after removing global warming influences. These results are in line with the higher sensitivity of extreme precipitation to global warming than mean precipitation23,25. All NAT experiments simulated less precipitation than ALL experiments near the TC center, especially in delta SST cluster 4 (NAT_C4) experiments which exhibit stronger SST warming (see Method), representing robust expansion of TC extreme precipitation area due to human influences (Supplementary Fig. 1).

Mechanisms driving precipitation intensification

Vertical motion (omega) and water vapor mixing ratio were investigated to understand detailed processes of TC precipitation intensification in a warmer climate (Fig. 6). Here, 6 hourly snapshots of vertical motion and water vapor mixing ratio throughout the simulation period are composited for each TC. Strong ascending motion (decreased omega) was observed near the eyewall and along spiral rainbands (Fig. 6a for TC SANBA). The difference in omega between ALL and NAT experiments tends to be larger around the TC core region, which is commonly found in other TC cases (Supplementary Fig. 2). Generally, more water vapor resides around the TC core region and water vapor decreases as it moves away from the TC center (Fig. 6c). However, the difference in water vapor between ALL and NAT is found to be relatively uniform in all areas within a 500 km radius (Fig. 6d), a common characteristic shared by other TC cases (Supplementary Fig. 3). This is well contrasted with the omega response that is stronger near the TC core. Considering these different spatial scales, the relationship between precipitation and omega (water vapor) was calculated at a radius of 200 (500) km.

Fig. 6: Distribution of vertical motion and low-level water vapor.
figure 6

Omega (hPa s−1) at 500 hPa within a 500 km radius of TC SANBA in a ALL and b difference between ALL and NAT_MME experiment. Water vapor mixing ratio (g kg−1) at 850 hPa within a 500 km radius of TC SANBA in c ALL and d difference between ALL and NAT_MME experiment. Note that 6 hourly snapshots of omega and water vapor mixing ratio are composited throughout the simulation period.

Full size image

Scatter diagrams between the precipitation difference between ALL and NAT and the corresponding difference in omega and water vapor are displayed in Fig. 7. Note that here we consider grid points with precipitation of more than 10 mm, as done in Fig. 5a–d. For individual TC cases, ALL experiments simulated stronger upward motion than NAT experiments except for the CHABA case (Fig. 7a). When considering all TC cases together, as the difference in upward motion is larger, the difference in precipitation becomes stronger with a significant inter-simulation correlation coefficient (r = −0.92). This indicates that the increased SST under human influences in ALL provides favorable environmental conditions in which more water vapor could evaporate. The increased water vapor could increase moist static energy in the lower troposphere, enhancing atmospheric instability and stimulating convection around the TC core region. When comparing delta SST cluster results, there were no distinct differences in precipitation or upward motion relations (see below). Compared to a strong relation between precipitation and upward motion relation, a weak inter-simulation positive correlation (r = 0.34) is found between precipitation and water vapor averaged within a 500 km radius (Fig. 7e). In line with the Clausius-Clapeyron relationship, ALL experiments exhibit more water vapor than NAT experiments in all TC cases. Among four delta SST clusters, the largest reduction in water vapor is simulated in NAT_C4 runs which are characterized by the warmest delta SST. As a result, the weakest precipitation is seen in NAT_C4 experiments except for the SANBA case, resulting in the largest ALL-NAT difference in moisture.

Fig. 7: Relationship between precipitation and vertical motion and water vapor.
figure 7

Relationships between precipitation and omega at 500 hPa averaged within a 200 km radius in a all simulation period, b first, c second, and d third period. e–h Same as in a–d but for precipitation and water vapor mixing ratio averaged within a 500 km radius. The correlation coefficient (cc) and p-value are provided. Blue, red, green, and purple circles represent TC SANBA, CHABA, MAYSAK, and HAISHEN, respectively. Delta-SST cluster numbers (MME, C1, C2, C3, and C4) are indicated in NAT experiments.

Full size image

We note that the increase in simulated water vapor and upward motion in the WRF model cannot be directly compared to the increase in precipitation upon corresponding grids. Precipitation in the WRF model is cumulative value calculated at each 6-hourly time-step, whereas water vapor and upward motion are snapshots at every 6 h. This is the major reason why the extreme precipitation and the upward motion are not matched systematically (compare Fig. 5e–h with Supplementary Fig. 2). Therefore, we performed area averaging to identify the relationship between precipitation, water vapor, and upward motion, as they cannot be compared on a one-to-one grid basis (Fig. 7). Furthermore, if the area-averaged environmental conditions are favorable for TC-induced precipitation, the area of extreme precipitation will also expand.

To further check which stage of TC lifetime is affected more by human influences, the simulation period has been divided into three sections based on an equal time interval. The first period represents that the TC intensified and was relatively close to the equator. The second period includes the time when the TC reaches its maximum intensity, and the third period is when the TC made landfall on the Korean Peninsula and weakened. Three periods divided for four TCs TC are shown along their tracks in Supplementary Fig. 4. There is no substantial difference in the inter-simulation relationship between omega and precipitation during divided periods (Fig. 7b–d). However, the relationship between precipitation and water vapor varies depending on the TC period (Fig. 7f–h). Water vapor is not associated with precipitation in the first or second period when TC is located over the low-latitude ocean. However, water vapor and precipitation exhibit a significant inter-simulation correlation in the third period. This suggests that the influence of moisture increase becomes important when TCs approach the Korean Peninsula, contributing to TC intensification more efficiently. In addition, NAT_C4 runs with noticeable differences in temperature and water vapor display the largest difference in precipitation from ALL experiment across all three periods. In conclusion, our results based on CPM simulations indicate that anthropogenic global warming is likely to increase precipitation of TC affecting South Korea by 4–12% (averaged within the 200 km radius) through stronger ascending motion and increased water vapor.

Discussion

In this study, CPM simulation was conducted to determine how human activities could affect the precipitation and intensity of four TCs that recently struck South Korea. Ensemble CPM experiments were conducted for the current climate (ALL) and counterfactual conditions without human influences (NAT). Historical simulations with and without human influences from nine CMIP6 models were used to estimate counterfactural background conditions in SST, atmospheric temperature, and relative humidity. To consider uncertainties in SST change patterns and associated atmospheric lateral boundary conditions, four delta SST clusters were used as additional ensemble members. Overall, more intense TCs were simulated in ALL experiments than in NAT experiments. The difference between ALL and NAT experiments was more noticeable in TC lifetime maximum intensity than mean intensity. There was no substantial ALL-NAT difference in relatively weak precipitation regions, while a significant difference was found in extreme precipitation regions. Importantly, TC-driven extreme precipitation (grid points with 6-hourly precipitation above 150 mm) was found to be expanded by 16–37% due to global warming influence. Furthermore, stronger upward motion near the TC core and more water vapor over a larger area were identified as major causes of increased extreme precipitation. It was also revealed that the influence of water vapor increase became significant in mid-latitudes, potentially playing an important role in maintaining TC intensity. Our results demonstrate that global warming caused by human influences can increase probabilities of more-intense TC with heavier precipitation affecting South Korea. Continuous warming in the future is expected to lead to further increases in risks of heavy precipitation, floods, wind gusts, and storm surges by stronger TCs.

Although the present study was focused on TCs affecting the Korean Peninsula, the findings of this study have important implications for other regions. One study32 has applied an event attribution approach with a horizontal resolution of 3 or 4.5 km to identify human influences with three hurricane cases over the Atlantic. More precipitation of TC was simulated in the current climate compared with the pre-industrial condition around the TC’s inner core regions, similar to our results. However, human influences were not detected in the intensity of TC. Another study41 has conducted an event attribution analysis for hurricane Harvey in 2017 where linear trends of SST and tropospheric variables were removed. Current climate conditions simulated more intense TC with increased precipitation compared to detrended simulations. Although tropospheric environments were found to be more important factors for TC intensity and precipitation than SST conditions, the role of water vapor and upward motion near the TC core region was not examined. Furthermore, the present study found that increased water vapor due to global warming contributes to TC precipitation more in the mid-latitudes, providing useful insights into detailed physical mechanisms. It is speculated that two external factors may enhance the water vapor–precpitation relation when TC approaches mid-latitude lands. One is the increased topographic effect due to the TC’s landfall, which can intensify convection and associated water vapor flux33,42. The other is the possible influence of a mid-latitude upper tropospheric jet streak, which can interact TC and enhance convection43,44,45. However, there is a limitation in directly verifying why moisture in the third period had a greater impact since this study only considered TCs affecting the Korean Peninsula. Various TC cases that make landfall in Asian countries with different latitudes need to be examined to explore whether and how the role of moisture availability changes during the TC phases.

Spectral nudging is an alternative technique that forces a regional model not only at its lateral boundaries but also in the interior46,47. Thus, it helps to keep the simulated track of TCs very similar to observations and enables attribution analyses of TC events, focusing on thermodynamic effects. For example, Mori et al.48 conducted hindcast simulations using the WRF model and successfully reproduced realistic intensity and track of typhoon Haiyan when applying spectral nudging. Many other studies have applied spectral nudging to reproduce the track of TCs similar to observations33,35,49,50,51. We have conducted additional model simulations to identify how much the track and intensity of TC would be changed when spectral nudging is not applied. Except for the TC SANBA case, other TCs were not realistically simulated in terms of TC track when spectral nudging was not applied (Supplementary Fig. 5). Diverse drawbacks were found depending on TC cases, including direction of TC movement, translation speed, and landfall location. In terms of TC intensity, experiments that applied spectral nudging showed overall more realistic simulations than those without spectral nudging (Supplementary Fig. 6). Our analysis indicates that no use of spectral nudging can affect TC tracks and intensity substantially, depending on TC cases, and make it difficult to compare TC simulations with the observations. In addition, it should be noted that since we maintained the TC tracks using spectral nudging, it is difficult to consider the influence of anthropogenic warming on TC translation speed. If a warmer climate makes a slowdown of TC translation as suggested by some previous studies52,53, the potential for damage caused by landfalling TCs could increase.

TCs affecting East Asian countries are expected to become more intense as SST over the western North Pacific substantially increases, especially in mid-latitudes, than in other regions54,55,56. Nevertheless, the understanding of associated physical processes remains uncertain. In this respect, our study focused on quantifying thermodynamic effects, keeping simulated TC tracks as observed by applying a spectral nudging technique. For more comprehensive assessments, the full dynamic and thermodynamic effects associated with possible changes in TC tracks need to be carefully considered in future studies, including possible influences of the changes in vertical wind shear and thermal stratification (e.g., Garner et al.57). In addition, CPM-based studies are needed for future TCs under different CO2 pathways58.

Methods

Data

TC location (longitudinal and latitudinal position) and intensity (sustained maximum wind speed and minimum sea level pressure) were obtained from the Joint Typhoon Warning Center (JTWC) best track dataset (available at https://www.metoc.navy.mil/jtwc/jtwc.html?best-tracks). Integrated Multi-satellitE Retrievals for the Global precipitation mission (IMERG) Final Run and Tropical Rainfall Measuring Mission (TRMM) datasets were used for evaluation of TC-induced precipitation59,60.

Model configuration

Weather Research and Forecasting (WRF) model version 4.0 was employed for TC simulation. It is widely used to forecast and reproduce diverse atmospheric phenomena, including TC activity. ERA5 reanalysis of the European Centre for Medium-Range Weather Forecasts (ECMWF) was utilized as an initial and lateral boundary forcing61. It covers the global atmosphere with a 30 km horizontal resolution. Daily Optimum Interpolation Sea Surface Temperature (OISST) version 2.0 from the National Oceanic and Atmospheric Administration (NOAA) was used for the ALL experiment. The WRF model has a single domain with a 3 km horizontal resolution (1601 × 2001, latitudinal and longitudinal grid points, respectively). All experiments were performed with the convection-permitting setting, which could simulate intense TC vortex structures realistically compared to the coarse horizontal resolution simulation. The domain contains 50 vertical levels and uses a sigma vertical coordinate from the surface up to the top of the atmosphere at 50 hPa. The WRF model utilized the WSM 6 (WRF single-moment 6-class) microphysics scheme62, the Rapid Radiative Transfer Model for General circulation model (RRTMG) short and long wave radiation schemes63, and Yonsei University (YSU) planetary boundary layer scheme64,65. We turned off the cumulus scheme because these simulations used a convection-permitting horizontal resolution.

The intensity and precipitation of TC are extremely sensitive to the TC track. It is difficult to identify the anthropogenic warming effect if TC tracks are different between experiments. Therefore, the spectral nudging technique46 was applied to simulate TC tracks realistically33. Spectral nudging was not applied near the surface (10 lowest sigma levels, below about 800 hPa) because TC dynamics could be suppressed66. The spectral nudging coefficient was increased linearly from the 10th sigma level to the top of the model (0 at the 10th sigma level and maximum at the top of the model).

Experimental design

TC cases were selected among TCs landfalling on the Korean Peninsula with extreme strength from 2011 to 2020. Lifetime maximum intensity and maximum wind speed near the Korean Peninsula (33–40oN and 123–131oE) should exceed 60 and 35 m s−1, respectively. Four TC cases were chosen according to these criteria (SANBA in 2012, CHABA in 2016, MAYSAK and HAISHEN in 2020). The initial time of simulation was determined when TC was intensified from tropical depression (TD, maximum wind speed \(<\) 17 m s−1) to a tropical storm (TS, 17 m s−1 \(\le\) maximum wind speed \(<\) 32 m s−1). The last time of simulation was determined when TC was weakened to TS. More detailed information on the TC simulation period is described in Supplementary Table 1. To determine the simulated TC track, we used the minimum sea-level pressure as a variable. The initial location of TC was identified as the minimum sea level pressure near the best track data. The subsequent TC location was defined as the minimum sea-level pressure within a 500 km searching radius from the previous TC center.

Multimodel datasets of the Coupled Model Intercomparison Project phase 6 (CMIP667) were used to compare SST and atmospheric temperature and relative humidity between recent climate and counterfactual conditions. Nine CMIP6 models were used and results for all ensemble members were averaged for each model. Models used in this study and the number of ensemble members are summarized in Supplementary Table 2. Historical experiments (natural plus anthropogenic forcing, ALL6) and historicalNat experiments (natural forcing only, NAT6) were compared to estimate anthropogenic warming (‘delta’) patterns for 2011–2020. For ALL6, historical simulations for 2011–2014 are combined with corresponding SSP (Shared Socioeconomic Pathway) 2-4.5 scenarios for 2015–2020. We obtained differences between ALL6 and NAT6 experiments (ALL6-NAT6) during the recent 10 years (2011–2020) to determine the anthropogenic warming effect on SST and atmospheric temperature and relative humidity. We did not adjust horizontal wind fields in the lateral boundaries to reduce possible perturbations to the simulated TC track as previous studies identified that the difference in horizontal winds does not affect the intensity of TCs much32. We also note that wind shear condition is generally more important for the genesis of TCs than for its mature phase (e.g., Liu and Chan68) and that our study considers extremely strong TC events which usually occur under weak wind shear conditions32. In the NAT experiment, the anthropogenic warming signal was removed from the ALL boundary conditions for SST, atmospheric temperature and relative humidity. Greenhouse gas concentrations were also adjusted to pre-industrial levels: CO2, CH4, and N2O at 284.3 ppm, 808.2, and 273.0 ppb, respectively69. This pseudo-global warming approach has been widely used to examine anthropogenic climate change influences in numerical experiments49,70,71,72,73,74.

Ensemble clusters were organized through an empirical orthogonal function (EOF) analysis of delta SST. CMIP6 models exhibit large differences in their SST responses to external forcings such as greenhouse gases and anthropogenic aerosols. Therefore, using a single pattern from multi-model mean (MME) or a particular model is not appropriate to consider the possible ranges of delta SST patterns (ALL-NAT). One can estimate delta SST patterns from 9 individual models but some models can share similar patterns than others due to the model dependency. Given this situation, one way to obtain representative cases of delta SST (ALL–NAT pattern) among 9 models is to apply an EOF analysis. Results from the EOF analysis indicate that there are four groups of models having similar delta SST patterns, and we have obtained corresponding delta SST from four clusters. Principal component 1 (PC1, x-axis of Fig. 8f) represents a change in the whole domain. Principal component 2 (PC2, y-axis of Fig. 8f) indicates whether the stronger warming region is low-latitude or mid-latitude. Nine CMIP6 models were classified into four clusters (Fig. 8f): NAT_C1, NAT_C2, NAT_C3, and NAT_C4. Together with all model averages (NAT_MME), we have five delta SST patterns which are used for NAT ensemble simulations for each TC case (Fig. 8a–e). Associated delta patterns of atmospheric lateral boundary conditions are obtained for temperature and relative humidity.

Fig. 8: Difference in SST between ALL6 and NAT6 and delta-SST clusters.
figure 8

Differences between ALL6 and NAT6 in sea surface temperature (SST) in a MME, b Cluster 1, c Cluster 2, d Cluster 3, and e Cluster 4. f Scatter plot of EOF first and second principal component (PC1 and PC2, respectively) of delta SST patterns (ALL-NAT) from 9 CMIP6 models and corresponding ensemble clusters.

Full size image

Expanding the ensemble members requires computational resources and time, and one needs to find a better way of ensemble construction that can better consider uncertainty factors. As we discussed above, the influence of atmospheric circulation on TC tracks is very difficult to handle and we apply a spectral nudging method to keep simulated TC tracks similar to the observed. This is equivalent to the storyline approach in which we keep dynamics as observed and evaluate thermodynamic influences75,76. In this setting, the internal variability influence will be much weaker than the influence of delta-SST77. In this respect, Wang et al.78 performed 100 ensemble experiments using perturbed initial conditions and found that the magnitude of dynamic anomalies was almost identical across the ensemble members due to the influence of spectral nudging. Other studies showed that applying spectral nudging reduces internal variability79, which could lead to a decrease in model ensemble spread in TC simulation80. These indicate that different delta SSTs can have a greater impact than different initial conditions based on internal variability under our experimental setting based on a storyline approach. As a simple check of the internal variability influence, we have repeated the ALL experiment by adjusting the initial conditions for the SANBA case (Supplementary Fig. 7). ALL_6 and ALL_12 represent experiments with initial conditions from 6 and 12 h later than the original ALL experiment, respectively. Results show that the simulated TC tracks remain essentially the same as the original simulation, indicating the weak influence of internal variability. The TC intensity and precipitation also remained unchanged, representing the insignificant impact of different initial conditions due to the use of spectral nudging.