Impacts of hemispheric solar geoengineering on tropical cyclone frequency

Solar geoengineering refers to a range of proposed methods for counteracting global warming by artificially reducing sunlight at Earth’s surface. The most widely known solar geoengineering proposal is stratospheric aerosol injection (SAI), which has impacts analogous to those from volcanic eruptions. Observations following major volcanic eruptions indicate that aerosol enhancements confined to a single hemisphere effectively modulate North Atlantic tropical cyclone (TC) activity in the following years. Here we investigate the effects of both single-hemisphere and global SAI scenarios on North Atlantic TC activity using the HadGEM2-ES general circulation model and various TC identification methods. We show that a robust result from all of the methods is that SAI applied to the southern hemisphere would enhance TC frequency relative to a global SAI application, and vice versa for SAI in the northern hemisphere. Our results reemphasise concerns regarding regional geoengineering and should motivate policymakers to regulate large-scale unilateral geoengineering deployments.

. Bold font indicates the values that are not significant at the 5% level (i.e. > 0.05).
To determine whether TC frequency changes are statistically significant with respect to HIST, we employ a Wilcoxon rank sum test (WRST) [8], which is similar to a Student's t-test but without the underlying assumption of normally distributed data. We apply the WRST to raw TC counts and compare the HIST period with various subsets of the RCP4. 5 Figure 2 shows scatter plots of the observations and reanalysis data in Fig. 7 with linear fits and Pearson correlation coefficients (r). It is clear that there are positive relationships between TC frequency and relative SST (r = 0.59) and between TC frequency and precipitation (r = 0.55). Additionally, there is a negative relationship between TC frequency and wind shear (r = -0.53). Consequently, due to the transitive principle, there are similarly strong relationships between precipitation, wind shear, and relative SST. Supplementary Figure 3 shows the fitted values (i.e. Λ i ) from the statistical models plotted alongside the HURDAT TC frequencies. All of the models seem to capture the oscillating trend with peaks (troughs) in 1950 and 2010 (1900 and 1980). There seems to be little sign of overdispersion (i.e. the observations are generally within +/-2σ) suggesting that the Poisson model is appropriate for this data [9].

Suppleme ntary Note 5. Is the tempe rature response similar to El Niño?
There is no clear consensus on the impact of stratospheric aerosol on El Niño Southern Oscillation (ENSO) response, with modelling suggesting either no response [17], or SST patterns that project onto El Niño [18] or La Niña [19]. However, a growing body of evidence suggests that volcanic eruptions that preferentially load the northern hemisphere stratosphere with aerosol, lead to SST anomalies that resemble those of naturally occurring El Niño [20].
Here we use the frequently used Niño3.4 metric [21] to assess whether our simulations give SST patterns that project onto ENSO features. La Niña conditions. However, the results above show that HadGEM2-ES does exhibit a credible ENSO performance in terms of the frequency, duration and spatial distribution particularly when compared to the biases in other global coupled atmosphere ocean models [23,24]. We therefore examine the response in these variables in our geoengineering simulations. Interestingly, ERA-I and other reanalyses are presently unable to reproduce observed TC intensities due to their own model-resolution issues [13,14], which suggests that even if we were to calculate HURDAT MSWSs as 6-hourly means, ERA-I would still not reproduce the observed TC intensity distribution.
Supplementary Figure  previous studies [6,15]. The G4 and G4NH scenarios also appear to show a shift toward more intense storms, suggesting that solar geoengineering would do little to modulate increases in storm intensity despite effectively modulating global warming (e.g. Fig. 3). However, Supplementary Fig. 9 does not account for changes to TC frequency. Supplementary Figure   10 shows changes to TC frequency based on their maximum-intensity relative to the intensity distribution in the HIST era. For instance, a positive value in the first column (< α-HIST = 5 %) would indicate an increase in the number of storms that achieve a maximum-intensity that is less than the lowest 5% maximum-intensity value of the HIST distribution (i.e. indicating an increase in weak storms). It is clear that the reductions in overall storm frequency in the RCP4.5 and G4NH scenarios (Fig. 4) are mainly due to reductions in weak storms (< α-HIST = 75 %). The frequency of the most intense storms increases in all of the scenarios (> α-HIST = 95 %). G4SH exhibits increases in storm activity at all intensities, whereas G4 exhibits little Suppleme ntary Figure 10. TC frequency changes grouped by intensity relative to historical intensities. TC frequency changes between various percentiles of the HIST maximum-intensity distribution for two TRACK configurations: a, TRACK at (6,5.5,4) and b, TRACK at (4.5,3.5,4) conformity between the two TRACK configurations. Note that we inject aerosol at a constant rate in the SAI simulations and the temperature rise due to enhanced greenhouse gas concentrations is not completely offset by SAI (Fig. 3a). If we had instead injected at such a rate as to stabilise global-mean temperature the changes to storm frequency and intensity in the SAI simulations might have been different.
A tendency for an overall reduction in North Atlantic TC frequency alongside an increase in the incidence of the most intense storms under global warming (RCP4.5) is a common result from previous research [6,15,16] and has also been identified here ( Supplementary Figs 9 and   10). However, the inability of HadGEM2-ES to produce storms with observed intensities in the North Atlantic basin (Supplementary Fig. S8) reduces confidence in the model's ability to predict changes to the intensity distribution under global warming and SAI. In light of this, we have utilised a statistical-downscaling model based on the HadGEM2-ES climatologies to study changes to the intensity distribution (Fig. 9). The disparities between the results of the explicit storm tracking and the downscaling method should galvanise further research into storm intensity changes under SAI.

RCP4.5 simulations compared to HIST for the statistical-dynamical downscaling model?
We perform the same statistical analyses as in Supplementary Section 3to note, a Wilcoxon rank sum test and a Student's t-testto assess whether TC frequency changes are significant in the results of the statistical-dynamical downscaling simulations. Supplementary Tables 3-5 show the p-values when comparing the 2020-2070 and 1950-2000 periods using a WRST (top line) and t-test (bottom line). Supplementary Table 3 compares TC frequencies (Fig. 9a), Supplementary Table 4 compares hurricane frequencies (Fig. 9b), and Supplementary Table 5 compares major hurricanes (Fig. 9c).
TC frequencies in the G4 and G4NH scenarios are not significantly different to HIST whereas all other differences are significant at the 5 % level. All of the SAI scenarios significantly reduce storm activity relative to RCP4.5, while G4 and G4NH exhibit insignificant major hurricane changes with respect to HIST at the 1 % level (Supplementary Table 5). Table 3     The manuscript's Methods section contains detail of the feature-tracking methodology. As with BE07, we test different permutations of (ξ 1 , ξ V , n = 4) for ERA-I [12] and HadGEM2-ES against HURDAT2 [30] annual TC frequency. For ERA-I, we find that a TC selection criterion of (6, 5.5, 4) provides a good fit to HURDAT2 data. This criterion closely resembles the BE07 criterion of (6, 6, 4). Previous studies have found that HadGEM2-ES has a low bias in terms of TC intensity and frequency in the North Atlantic basin, which is possibly due to the coarse spatial resolution of the model [5,31,32]. In this study, we also observe this bias when using the ERA-I criterion (6, 5.5, 4) to identify TCs in the HadGEM2-ES simulations ( Supplementary Fig. 11 for HadGEM2-ES simulations, which produces a similar number of TCs in HIST to the HURDAT2 observations ( Fig. 4 and Supplementary Tables 6 and 7). It is important to note that the characteristics (e.g. the intensity) of the model storms will differ from those of the observed storms due to the different criteria applied. Nevertheless, the model capably produces historical TC trends (Fig. 4), which gives us confidence that this approach allows us insight into future TC frequency trends under the RCP4.5 and SAI scenarios. The correlation coefficients between the modelled TC frequencies using the two different TC selection criteria (i.e. column r in Tables S6 and S7) Table 7. Annual TC fre quency in the HadGEM RCP4.5 and solar geoengineering simulations for two configurations of TRACK. TC frequency in the 1950-2000 (HIST) and 2020-2070 periods. *Note that the specified configuration of TRACK (i.e. (6,5.5,4) or (4.5,3.5,4)) is not relevant to the HURDAT2 observations Suppleme ntary Figure 11. Time-series of raw annual TC frequency determined using two TC selection criteria. a-c, TRACK @ (4.5,3.5,4) and d-f, TRACK @ (6, 5.5, 4). Purple and light-blue shading indicate the range of RCP4.5 and G4 values respectively Supplementary Figure 11 shows the raw TC counts for the RCP4.5, G4, G4NH and G4SH simulations using the two TC criteria. It is clear that G4SH produces significantly more TCs per year than G4NH (from 2020-2070) for both TC selection criteria.