The IRIS model has been used to quantify the additional impact on maximum wind speed by warming and potential intensity increase on major damaging hurricanes over the last four decades.

Summary
The wind speed increased by up to +13% and the return period reduced by up to -60% across the gates and cases examined. The annual probability of a major hurricane making US landfall is now about 30% larger compared to the 1980 baseline. The probability of a Category 5 landfall has nearly tripled from a 30-year to 10-year event,

1. IRIS (v1.0)
IRIS is a new model driven with several key innovations. It recognises that the key step for estimating landfall wind speed is the location and value of the life-time maximum intensity (LMI). It redefines the problem as one of decay only (1). The initial value is physically constrained by the thermodynamic state as defined by the potential intensity (PI). Tracks are based on observations. The model (v0.1) description is undergoing minor revisions under peer review (2).

In July 2023 Maximum Information Ltd. performed independent validation of IRIS (v0.1). They identified a low bias in landfall counts (all storms) particularly near Florida. We confirmed their findings. To improve the model, we added several features to a new version, v1.0, used for this report. There are physically informed changes in v1.0 compared to v0.1:

1. The track leading up to LMI is now included. V0.1 only has tracks post-LMI.  This increased landfall counts. When a TC crosses, for example Florida, then the LMI  occurs after re-entering the ocean before making a 2nd landfall. This now counts as two landfalls.
2. We allow energy production by the ocean heat flux which reduces the rate of decay. Providing the local potential intensity along the track exceeds tropical storm strength, cyclones do not decay below storm level (17.5 m/s). In v0.1 storms less than 17.5 m/s over the ocean are terminated and some thus did not make landfall.
3. When the TC re-enter the ocean (“sea fall”) from land they are re-intensified and continue their decay from the previous landfall intensity. V0.1 has continuous decay so that after sea fall TC are unrealistically weak, and some were terminated.
4. The child tracks are only shifted relative to the parent track as in v0.1. V0.1 also creates additional dispersion by using the “forecast” cone. This was found to be unnecessary.

We compare the mean climatology of IRIS(v1.0) and IBTRACS for 1980-2021. Figure 1 shows the eleven US gates which are used for the validation. The gate size was chosen to capture an adequate sample size while being approximately similar in scale.

Figure 1. Landfall locations and gate number. Coloured dots are IRIS simulations and  circles are observations.

Figures 2-4. show the frequency of storms making landfall by gate and minimum intensity.
The storm rates are captured well at the continental scale 2.52 yr-1 modeled compared to 2.40 yr-1 observed (Fig. 2). The rates at the different gates are also well captured. Noted differences are near the Carolinas (Gate 7) there are limited observations in Gate 7 and the model count is higher but within 2 standard deviations (Fig. 3). Figure 4 shows the analysis for major hurricanes storms above Category 3. The continental rate of 0.54 yr-1 is also in good agreement with the observations of 0.57 yr-1. As expected, the model against observed variability asserts itself at the local scale. There is some difference between the gates. This is largely due to hurricanes entering a gate slightly more/less than the adjacent one. Gates 7,9 and 10 have no observations. Overall, there is no clear regional bias. The observations (one realisation) of 40 years are within the model variability.

Figure 2. Total storm (vmax>17.5 m/s) count rate (yr^-1) for each gate.  Intervals and bars are the 95%,5% tile range based on 42-year samples of the 10,000-year simulation.

Figure 3. As Figure 2 but for hurricanes (vmax> 33 m/s).

Figure 4. As Figure 2 but for major hurricanes (vmax>50 m/s).

2. Attribution Methodology
The IRIS model can be used to infer the additional strengthening of hurricanes that can be attributed to recent warming or more specifically to changes in only the potential intensity. 17 Hurricanes are considered making 18 landfalls which had major impacts (Figure 5).

Figure 5. The tracks of the 17 most damaging hurricanes since 1980. Red dots show landfall locations which are allocated to the appropriate gate (see Figure1).

We first need to consider the change in the thermodynamic environment. ERA – reanalysis is used to calculate monthly mean PI fields during the satellite era since 1980. We consider global warming to manifest itself differently with latitude. We have low confidence in attributing regional or longitude trends to global or anthropogenic warming. The regional changes are more likely to be caused by decadal variations and less likely to be sustained or representative of global warming. It is for example very unlikely that the North Atlantic will continue to warm relative to the globe at the current rate. We are assuming that the underlying anthropogenic trend in PI is best represented by the global zonal mean. We know of no evidence, e.g. from climate models, that the anthropogenic portion of the increase in the PI in the North Atlantic is enhanced compared to the zonal mean. If it is, then our attribution estimate would be conservative. To calculate the PI field in any given year we apply the corresponding monthly global zonal mean trend to the 1980-2020 monthly mean PI field. In this way we can estimate the regressed anomalous PI field in any month and year. This regressed value is not the actual PI but that portion due to a linear change since “1980”. We use quotation marks around the year to signify that the year is not the actual PI for that year but the regressed value. Figure 6 shows the difference in potential intensity between “2021” and “1980” in the peak hurricane month of September based on the trend. There are large changes in the tropics which reduce in magnitude towards the subtropics and then increase again at larger latitudes. This meridional structure is interesting and is different to the SST trends which tend to gradually increase from the tropics to higher latitude.

Figure 6. Change in global zonal mean ERA-5 September regressed potential intensity between “2021” and “1980”.

The frequency of landfall is the next consideration. This version of the IRIS model does not change the number of events in the Atlantic, only the initial life-time maximum intensity is modified by the PI. However, the landfall rate will change because on average storms last longer if they have a larger initial value. Since we do not account for the relative warming of the North Atlantic nor the increase in the number of hurricanes (which are likely related) our attribution is therefore not complete but may be conservative. It is important to note that we are not examining the counterfactual of an individual hurricane, but the counterfactual (“1980”) of that gate climatology in the year of the landfall. For reference we take the observed intensity at landfall.

3. Results
3.1 US landfalls

Table 1 shows the results for the 18 landfall cases. When interpreting the results it is important that we are comparing events of the type of the named Hurricane. We are not re-running the individual case with higher potential intensity (see Sandy case study below as an example of this). Rather the cases are representative of events of that intensity in the respective gate for different potential intensity states.

Table 1. Hurricane wind speed and return period for the gate fr the most damaging hurricanes since 1980. Change is the difference relative to the “1980” baseline. Georges made two damaging landfalls in Gate 11 (“Georges 1”) and 3 (“Georges 2”).

Relative to a baseline PI climate of “1980” the relative wind speed increase ranges up to 13% (Sandy 2012). The biggest absolute change of 6 m/s is found for Ida (2021). The largest absolute decreases of return period are found for the largest return periods such as for Andrew (1992), whereas the largest relative change of -60% is found for Michael (2018). The absolute change in wind speed with years follows the PI trend, whereas the reduction in return period is much more variable over time. The return periods are shown up to 200 years (except Andrew which is up to 1000 years) for the appropriate gates for all 18 landfall cases are attached in the supplementary figures.

As well as considering the individual cases it is also informative to understand how the overall US landfall risk has changed. The general increase in wind speed for major hurricanes is about 4-6 m/s. Figure 7 shows the return period of the intensity for all US landfall for now (“2022”) and past (“1980”) climate. The annual probability of a major hurricane (vmax> 50 m/s) at landfall has changed by 31% from a return period of 2.1-year to 1.6-year. The change in Category 5 is even more dramatic: the probability (return period) has approximately tripled from 30-year event to a 10-year event. The observed climatology lies largely in between the “years” supporting the IRIS model assumptions.

Figure 7. Mean maximum wind speed (m/s) of US landfall storms vs return period (years) for observations (black)  “2022”  (orange) and “1980”  (blue). 200 year samples (grey) out of 10,000 year “2022” simulation. Observations

3.2 Case studies: Sandy and Andrew
Sandy (2012) caused an estimated economic damage of $60B. It is assumed that wind speed attribution is not possible. Instead a study attributed only the anthropogenic sea-level rise contribution to the surge damage to be approximately 13% (3). Our study here reveals that at the time of Sandy a type of cyclone with Sandy’s wind speed has become much more likely by about 37%. Conversely, the expected wind speed at a 30 year return period was about 13% higher than in 1980 baseline. We now have an attribution. This is the largest percentage increase in wind speed of all the hurricanes studied. If we remove the Sandy (track) from IRIS, then the landfall rate for Gate 10 decreases by only about 12%. The model climatology at that gate is not very sensitive to one event.

Does IRIS suggest climate change has made Sandy 13% more intense? We could use IRIS differently (not as originally intended) and simulate 1000 possible intensities at landfall of Sandy’s track post LMI as an ensemble of counterfactuals. Figure 12 shows the distribution of counterfactual landfall wind speeds for a “1980” and “2012” climate. The probability of reaching the observed Category 1 landfall wind speed, 38.5 m/s is about 20%. (Figure 8). For that probability there is a shift to the larger intensity in “2012” compared to the “1980” baseline of +1.7 m/s or only 4% compared to 13% for the gate. The probability of reaching the same intensity in 2012 increases by +0.03 (+18%) compared to “1980”. This change is only driven by the potential intensity change at a fixed Sandy’s LMI location and the small magnitude of wind speed would not be detectable. It is important to state the landfall climatology at gate 10 is caused by many LMI locations and different PI trends. The counterfactual Sandy case (barely detectable) and the IRIS climatology are not the same. Sandy counterfactuals contribute to the gate climatology, but in essence IRIS is a climate model. Removing Sandy from the observation or model has little effect (Figure 9).

Figure 8. CDF against maximum wind speed (m/s) for 5000 member ensemble of Hurricane Sandy (2012) simulations in 1980 and 2012 climate.

Figure 9. The effect of removing Sandy on the gate return period from the model (blue) and observations data sets (black-dashed) compared the base model (orange) and all observations (black).

Andrew (1992) was an exceptional Category 5 Hurricane at landfall in Florida. The IRIS model places this type of event at a very low probability with a return period in excess of about 500 years. There is no meaningful attributable change in the wind speed. The LMI of Andrew was near the Bahamas (25oN) which lies in the belt of minimal PI trend (Figure 6), so that any change of PI and LMIs by 1992 was modest. The observed rates in Andrew’s gate 6 are very sensitive to individual events because the sample size is small. Andrew itself had a substantial effect on the observed major hurricane rate. The model rate is about 50% lower than that observed (Figure 4). However, if we remove just Andrew from observational analysis, then the observed rate of major hurricanes more than halves and then agrees much better with the model (Figure 10.) The overall model landfall rate at this gate is only reduced a little when Andrew is removed from the model.

Figure 10. The effect of removing Andrew on the gate return period from the model (blue) and observations data sets (black-dashed) compared the base model (orange) and all observations (black). Major hurricane lines are shown to illustrate impact of Andrew removal on gate rate shown in Figure 4.

4.0 Conclusions

We find a wide range of increases in wind speeds (up to +13%) and reduction in return periods (up to -60%) for the 18 landfalls. Overall, the probability of major hurricane (vmax>50 m/s) landfalls in the US has increased by about 31% since 1980. The probability of a Category 5 landfall has nearly tripled. This model result is at least consistent with the observed doubling of major tropical cyclones making landfall globally since 1980 (4). For the whole North Atlantic basin, a doubling of the major hurricane fraction since 1980 has been reported (5). The model major hurricane landfall rate trend is somewhat smaller than the trend of the basin fraction, but the near tripling of model landfall Category 5s is plausible given these observations. It is noteworthy that the model absolute wind speed increase is only about +5 m/s . This is within the estimated IBTRACS uncertainty of the maximum wind speed reported in the 1980s. The economic damage per year is extremely variable. It is thus to be expected that the detection of both historic landfall wind speed trends (4) and damage trends has been both challenging and controversial. We note from the IRIS model simulations, that the intensity changes are indeed found to be small, but that there has likely been a substantial change in the probability of damaging hurricanes.

Here we assumed a constant total hurricane rate and a global zonal mean increase in PI. It is worth noting that the actual annual probability can be expected to be larger because the Atlantic potential intensity has increased more than the global mean and the number of North Atlantic hurricanes has also increased. However, it is not clear that these two changes are due to anthropogenic forcing or variability (more likely). It therefore appears that the expected annual loss, driven by major hurricanes, has increased by at least 30% over the last four decades due to global warming.

References

1. Shuai Wang,Ralf Toumi, Recent migration of tropical cyclones toward coasts. Science 371,514-517(2021).DOI:10.1126/science.abb9038
2. Sparks, N., Toumi, R. IRIS: Imperial College Storm Model, Scientific Data, under review.
3. Strauss, B.H., Orton, P.M., Bittermann, K. et al. Economic damages from Hurricane Sandy attributable to sea level rise caused by anthropogenic climate change. Nat Commun 12, 2720 (2021). https://doi.org/10.1038/s41467-021-22838-1
4. Wang, S., Toumi, R. More tropical cyclones are striking coasts with major intensities at landfall. Sci Rep 12, 5236 (2022). https://doi.org/10.1038/s41598-022-09287-6
5. Kossin JP, Knapp KR, Olander TL, Velden CS. Global increase in major tropical cyclone exceedance probability over the past four decades. Proc Natl Acad Sci U S A. 2020 Jun 2;117(22):11975-11980. doi: 10.1073/pnas.1920849117. Epub 2020 May 18. Erratum in: Proc Natl Acad Sci U S A. 2020 Nov 24;117(47):29990. PMID: 32424081; PMCID: PMC727571

Supplementary Figures

Mean maximum wind speed (m/s) at the gate of the named hurricane vs return period (in years) for “year” of the hurricane (orange line) and “1980” (blue line). 200-year samples (grey lines) out of 10,000 years of the “year” of the hurricane simulation.

The uncertainties of this study depend on some of the key assumptions. The potential intensity calculation and changes are sensitive to ERA5 reanalysis errors.The assumption of constant uniform probability distribution of the relative intensity (LMI/PI) and constant tracks are important sources of uncertainty which could either decrease or increase the scale of the attribution presented here.

The final methodology and results are provided for general information purposes only, and do not represent the opinion of any individual within the Lighthill Risk Network or its members. The information is not intended to be taken as advice with respect to any individual situation and cannot be relied upon as such.

Notes to editors

About Lighthill Risk Network

The Lighthill Risk Network is an all-encompassing and inclusive organisation with the specific aim of facilitating and enhancing knowledge transfer into business from academic, government and commercial experts at the forefront of risk-related research.

For more information, please visit: https://lighthillrisknetwork.org/

PR Contacts

Helen Wright, Lysander PR
helen@lysanderpr.com
07842 729 579

Roddy Langley, Lysander PR
roddy@lysanderpr.com
07547 901 618

Steve Colton, Lysander PR
steve@lysanderpr.com
07554 889 056