Assessment of model projections of climate‐change induced extreme storms on the south‐east coast of Australia

General circulation models (GCMs) and their downscaled regional‐scale equivalents have been important tools for climate‐change studies. However, there has been limited assessment of the performance of GCMs and downscaled models in simulating extreme storms in temperate coastal environments. This study assesses the model characterization of extreme storms on the heavily populated coast of south‐east Australia. Twenty‐year average recurrence interval (ARI) storm intensities derived from generalized extreme value (GEV) distributions based on observed and large‐scale atmospheric model data are compared. Changes in extreme storms from past climate to a high‐emission future scenario are also investigated. Simulations of storm minimum surface pressures compared favourably with measured data. Both the GCMs and downscaled models reproduced the observed decrease with increasing latitude along the coast in the 20 year ARI of minimum surface pressure. Both indicated that the minimum storm surface pressure should change negligibly in a high‐emission future. Although the models underestimated the maximum daily precipitation significantly, models are improving significantly with CMIP epoch and downscaling. In the high‐emission future scenario, the GCMs and NARCliM projected the 20 year ARI maximum daily precipitation would increase in the order of 25%. GCMs and the corresponding downscaled products presently do not represent the extreme value distributions of historical wind speed data well, overestimating at smaller values of ARI and significantly underestimating in larger values of ARI. Significant changes in the magnitude of the 20 year ARI maximum daily‐average onshore wind speed are not projected for the high‐emission future.


| INTRODUCTION
The Australian coastal zone plays an important role in social, economic and ecological environments.Over 80% of the population now live within 50 km of the coastline in Australia, while the south-east coast is occupied by over 55% of the total population (Head et al., 2014;Herold et al., 2021;McInnes et al., 2016;Peirson et al., 2011Peirson et al., , 2014)).In the Australian coastal zone, coastal storms are one of the most common hazards (Walsh et al., 2016), which are extreme weather systems and cause extreme runoff, coastal water level rise and storm surges due to the low atmospheric pressure, strong onshore winds or intense precipitation (Gillanders et al., 2011;Glamore et al., 2016;Peirson et al., 2014;Scavia et al., 2002).Coastal storms cause storm damage and coastal flooding, which threaten public safety, coastal infrastructure, ecological systems and the economy (Bouwer, 2019;Dowdy et al., 2019;Mortlock et al., 2017).Predicting extreme wind, precipitation and water level is of fundamental importance for storm surge and flooding prediction, coastal engineering risk assessment and disaster mitigation and prevention (Peirson et al., 2014).Future coastal planning requires a quantitative understanding of the changes in extreme storms under the envisaged fossil-fuelled socio-economic development.
The Australian continent is vulnerable to climate change (Head et al., 2014).A historical tendency to warming and drying has been observed in southeastern Australia (Head et al., 2014;Murphy & Timbal, 2008).It is highlighted by the Intergovernmental Panel on Climate Change (IPCC) sixth assessment report that an continuing increase of up to 3.7 C in mean temperatures is projected in Australia by the end of 21st century under high emission scenario (IPCC, 2023).
Climate extremes are often crucial in climate change studies (Perkins et al., 2007).The annual extreme temperature is projected to increase over Australia (Alexander & Arblaster, 2017;Head et al., 2014;Perkins & Pitman, 2009).The projection in the extreme precipitation shows great variation.Some studies have found that both the intensity and frequency of extreme precipitation will increase in the future in most regions of Australia (Alexander & Arblaster, 2017;Evans et al., 2016;Herold et al., 2021;Peirson et al., 2011Peirson et al., , 2014;;Pfahl et al., 2017), while others have indicated that the heavy precipitation would decrease in part of the country (e.g., the extratropical region; Dowdy et al., 2013).With these changes in extreme conditions and the important role of the coastal zone in Australia, enhancing our comprehension of regional extreme climate variability across the southeastern coast has become essential.The aims of the present study are firstly to assess the climate models' simulations of the magnitudes of extreme storms by comparing with measurements, then to summarize the projected changes in the characteristics of the changes in extreme storms along the southeastern coast of Australia.
The performance of GCMs in CMIP3 and CMIP5 in simulating extreme daily temperature and precipitation was studied by Sillmann et al. (2013) and they found that there were some improvements in CMIP5 over earlier versions in simulating the magnitude of precipitation.Ding and Zhang (2020) compared extreme precipitation derived from observations and the averaged ensembles outputs from five GCMs in CMIP5 to study the changes in extreme precipitation under the influence of climate change.The averaged GCM results projected well the extreme precipitation, after bias-correction, in the Lanchang River Basin in China.
In Australia, Perkins et al. (2007) assessed the performance of various CMIP3 climate models in simulating precipitation and extreme temperatures by calculating probability density functions (PDFs) in 12 regions of Australia.They found that the simulations of precipitation, minimum temperature and maximum temperature performed reasonably well; several models captured over 80% of the observed PDFs.The projections of extreme precipitation was not undertaken in this study.Peirson et al. (2011Peirson et al. ( , 2014) ) assessed the models' accuracy by comparing measurements with data from the CMIP3 GCMs along the south-east Australian coast.These studies calculated the generalized extreme-value (GEV) distributions and recurrence intervals for precipitation, surface pressure and wind speed, and found that GCMs provided suitable results for coastal engineering design, except for precipitation for which bias needed to be reduced by applying downscaling.Alexander and Arblaster (2017) examined CMIP5 GCMs simulations of temperature and precipitation extremes in Australia by comparing the modelling outputs and multiple observations in various indicators.They found that the models generally provided accurate simulations of temperature extremes, including climatology, variability and trend patterns, but had great variability in extreme precipitation.However, an evaluation of the consequence for regional areas and of the uncertainty in the climate projections was not undertaken.They also pointed out that there is a need to improve the precipitation simulations.Grose et al. (2020) compared the ensemble differences between CMIP5 and CMIP6 in climate-change projections in Australia.They found that the projections of the changes in Australian mean and extreme climate variables derived from the CMIP6 ensemble generally showed an agreement with those from CMIP5.Both ensembles simulated comparable extreme precipitation, but in terms of extreme temperature, CMIP6 projected a warmer future than CMIP5.
The global domain of GCMs requires them to have a horizontal grid spacing of 100-500 km, with a consequent poor spatial resolution for extreme storms and coastal topography (Manage et al., 2016;Nishant et al., 2021;Olson et al., 2016;Peirson et al., 2011Peirson et al., , 2014)).A process of nesting regional-scale models, termed downscaling, can be used to overcome the coarse spatial resolution of the GCMs (Evans et al., 2014;Nishant et al., 2021;Olson et al., 2016).
NARCliM (the New South Wales (NSW) and Australian Capital Territory (ACT) Regional Climate Model) is a regional dynamically downscaled modelling project applicable to south-east Australia (Evans et al., 2014;Olson et al., 2016).This project was developed in collaboration with governments, the environmental agencies and research organizations to provide information on the regional climate to support policy and decision making at local scales (Evans et al., 2014;Nishant et al., 2021).Selected GCMs were dynamically downscaled by Regional Climate Models (RCMs) to yield projections at 10 km resolution (Fita et al., 2017;Nishant et al., 2021;Olson et al., 2016).
The spatial patterns and validation with observations of the temperature and average precipitation data from NARCliM have been extensively studied.According to Olson et al. (2016), NARCliM better captured spatial distributions of the temperature and precipitation and reduced the temperature root-mean-square error, compared to GCMs.Similarly, Herold et al. (2021) found that the spatial characteristics of 20 year ARI for the NAR-CliM simulated extreme temperature and precipitation were consistent with observation.However, Herold et al. (2021) indicated that the model ensemble underestimated the magnitude of extreme heat and overestimated the extreme precipitation.Manage et al. (2016) investigated the NARCliM performance in a precipitation simulation in the Goulburn River catchment in the Upper Hunter region of New South Wales in Australia.They found poor agreement between the NARCliM data and ground measurements.In addition, the projections of the frequency of extreme precipitation over south-east Australia had been studied.Herold et al. (2021), in studying projections of the frequency of extreme precipitation over south-east Australia, found that the 20 year ARI extreme precipitation would occur as twice as frequently in the future, using NARCliM 1.0 datasets downscaled from CMIP3.
Few studies evaluated the projections of regional extreme storm climate and its changes by both the GCMs and the NARCliM downscaled models along the coastal zone.This article focuses on projecting the extreme storms in the south-east Australian coast and estuarine systems by examining the GCMs and the latest NARCliM downscaled model outputs.Extreme precipitations along the east Australian coast are usually associated with the occurrence of East Coast Lows (ECL) (di Luca et al., 2016;Dowdy et al., 2013;Pepler & Dowdy, 2022).Dowdy et al. (2013) found that the changes in the upper-tropospheric vorticity were closely related to the changes in the frequency of extreme precipitation by analysing CMIP3 outputs.In the south-east Australian coast, Pepler et al. (2016) found that the ECL frequency was projected to increase which resulted in the amplification of associated extreme precipitation intensity using NARCliM downscaled output.Pepler and Dowdy (2022) found that the ECL intensity was projected to decline, but the ECL associated average maximum precipitation intensity is projected to increase by 7% K −1 in the south-east Australian.Very high precipitation causes strong inundation.Low surface pressure induces barometric storm surge.Strong onshore winds generate wind set up as well as large waves which can cause surge due to wave set up (Antunes et al., 2019;Fang et al., 2020;Peirson et al., 2014).
This article specifically evaluates the performance of the CMIP3, CMIP5, CMIP6 GCMs and NARCliM 1.5 downscaled models in characterizing extreme coastal storm events along the south-east coast of Australia in terms of minimum surface pressure, maximum precipitation and maximum onshore wind speed.Climate-induced changes for an envisaged high-emission future are quantified.The structure of the paper is as follows.Date sources and methods are described in Section 2; the results of evaluating the GCMs projections for extreme climate in the historical scenario are presented in Section 3; and the changes in extreme climate in a highemission future are presented in Section 4. Section 5 discusses the results of this study, and the final section presents conclusions.

| Observation data
The observation datasets were obtained from available long-term meteorological observation stations on the coast, known to have captured high-quality data over a substantial sufficient observation period (Peirson et al., 2011(Peirson et al., , 2014)), covering the period from 1960 to 2000, minimizing the possibility of contaminations by climate change signal.Figure 1a shows the location and spatial distribution of the stations.The meridional spatial distance between each station is approximately 2 , ranging from latitudes 27 S to 43 S.For each station, surface pressure, precipitation and wind data were extracted.All datasets were averaged to a daily resolution for comparison between observations and model values (Peirson et al., 2011(Peirson et al., , 2014)).
The individual wind data points were decomposed into their zonal (u) and meridional (v) components.The onshore wind speed (westward wind [−u] along east coast of Australia) was analysed in this study.Any averaging was applied to the decomposed components.Typically, wind speed was recorded at 9 a.m. and 3 p.m. at each station, and averaged to a daily interval.This averaging might cause uncertainty on the validity of comparing the model winds with the daily average.Topography can affect the measurements recorded by coastal anemometers (Peirson et al., 2014).Thus, the westward design wind from AS1170:2011 was also included to compare with the model data at 20 year ARI.For simplicity and data consistency, the GEV distribution results for the observation datasets were obtained from figures 3.5B, 3.6B, 3.7B and C.4d in Peirson et al. (2011) and figures 4-6 in Peirson et al. (2014).These observations were compared statistically with data obtained from GCMs and downscaled model results.The compared model datasets are extracted from the grid point adjacent to each observation station.

| Global climate model (GCM)
Daily datasets of surface pressure, precipitation and westward wind speed were extracted from the multiple GCM model outputs listed in Table 1 in three CMIP phases: the third phase (CMIP3); fifth phase (CMIP5); and sixth phase (CMIP6).
For comparison with meteorological records, the model simulations in historical experiment were used.We examined all available ensembles of the selected models in the CMIP3 historical experiment cover the period of 1960 to 2000.The simulation forced by first realization, first initialization, first physics and first forcing "r1i1p1f1" of the selected models in the CMIP5 and CMIP6 historical experiments were examined, covering the period 1850 to 2005 and the period of 1850 to 2014, respectively.
The impacts of climate change in the worst-case scenarios need to be paid more attention and systematically studied (Head et al., 2014).Given that changes in extreme storms have great importance in coastal management and decision-making, the high-emission future scenarios projected were analysed to highlight the potential changes with climate.For the CMIP6, the projections of future scenario are shown in Shared Socioeconomic Pathways (SSP) which cover the period of 2015 to 2100 (Eyring et al., 2016;Meinshausen et al., 2020;O'Neill et al., 2016).The SSP5-85 experiments, used in present study, simulate climate for fossil-fuelled socio-economic development with high greenhouse gas emissions that are high enough to produce a radiative forcing of 8.5 W/ m 2 in 2100 (Eyring et al., 2016;Meinshausen et al., 2020;O'Neill et al., 2016;Taylor et al., 2012;van Vuuren et al., 2011).

| NARCliM 1.5 downscaled models
The NARCliM project is a climate downscaling with RCMs based on the Weather Forecasting and Research (WRF) modelling system (Olson et al., 2016).NARCliM provides downscaled climate projections for south-east Australia covering New South Wales and Australian Capital Territory at a 10 km resolution (Evans et al., 2014;Fita et al., 2017;Manage et al., 2016;Nishant et al., 2021;Olson et al., 2016).The domain of 10-km resolution is from 27 S to 39 S and 148 E to 154 E, shown as the orange box in Figure 1a.This downscaled dataset has been published in two generations.The first generation NARCliM 1.0, downscaled from CMIP3, has not been considered here due to the short simulation periods.The second generation NARCliM 1.5, downscaled from CMIP5, is used.The downscaling models cover both land and sea.
To evaluate the NARCliM 1.5 model performances, the simulations of historical scenario were used to compare with the observation data.The changes in the extreme surface pressure, precipitation and westward wind speed in the high emission (business-as-usual) futures simulations (RCP8.5)were investigated by using the NARCliM 1.5 downscaled RCP8.5 simulation.Details of these fine-scale downscaled simulations are listed in Table 1.

| GEV distributions
The GEV distributions of Jenkinson (1955) are widely used for modelling extremes (Hosking et al., 1985).GEV distributions have been used to estimate the ARIs for extreme events, with cumulative distribution function described by Kotz and Nadarajah (2000).In this paper, annual extreme values, more specifically minimum values of surface pressure and maximum values of precipitation and wind, within the recorded length are fitted T A B L E 1 Summary of the GCMs and their downscaled models.

Model name GCM generation Resolution
Total simulation period (years) Historical CSIRO-MK3.0CMIP3 1.9 × 1.9 Pressure and wind: 80 Precipitation: 20 to the GEV distribution using the method of probabilityweighted moments (PWM) to estimate the location, scale and shape parameters (Hosking et al., 1985).Return values between 1 and 100 years are then estimated for a given ARI. Figure 2 illustrates the approach using the GEV distribution.It shows the minimum surface pressure projected by CanESM5 model in CMIP6 historical experiment at the grid point adjacent to the Sydney Airport Station (Station No. 66037,151.15 E,33.95S; see Figure 1).Using the constructed annual series equivalent to a 165-year record, the plotted data show a well-defined GEV distribution, the associated confidence limits and no evidence of any numerical damping of the most extreme events.
For assessment, a 20 year ARI was selected to provide a reasonable level of confidence in the context of the observational and CMIP3 record lengths.Confidence intervals applicable to an estimate of the extreme value can be determined by a negative log-likelihood function and the delta method (Coles, 2001).While other confidence intervals are possible, a 90% confidence interval was selected for use during this study.The 20 year ARI was selected as this represents a substantial level of extreme storm activity.For much larger ARIs, the confidence limits become very large.

| Skill score
A skill score was calculated to evaluate model performance relative to the recorded observational data (Allen et al., 2007;Yang et al., 2020).
where X m is the model value and X o is the corresponding observed value, both at 20 year ARI.The overbar represents the measurement mean.Performance levels were categorized as: >0.65 excellent; 0.65-0.5 very good; 0.5-0.2good; < 0.2 poor (Allen et al., 2007;Maréchal, 2004;Yang et al., 2020).

| EVALUATION OF MODEL PROJECTIONS IN EXTREME STORMS
The skill scores, in Table 2, shows the performance of the selected GCMs and NARCliM 1.5 downscaled models in simulating the 20 year ARI extreme storms along the southeast Australian coast.Compared with extreme precipitation and westward wind speed, the climate models show better simulation for minimum surface pressure.
All GCMs and downscaled-model simulations for minimum surface pressure are excellent, except MIROC6 in CMIP6.All estimates of precipitation and westward wind speed were poor.

| Surface pressure
The model simulations of the minimum surface pressure were the best of all the variables evaluated.The GEV distributions of the minimum surface pressure simulated by the GCMs and NARCliM at the Sydney Airport Observation Station (66037) (151.15E, 33.95 S) are presented in Figure 3.In the CMIP3, GFDL2.1 showed good agreement with the observations for small-ARI events and 2 hPa overestimated in extreme events larger than 10 year ARI.In contrast, the minimum surface pressure estimated by GFDL2.0 was 2 hPa lower in the small-ARI events but showed a good agreement in extreme events >20-year ARI.CSIRO-MK3.0projections were 3 hPa overestimated in the extreme events smaller than 6-ARI but underestimated 2-5 hPa for the larger-ARI events.CSIRO-MK3.5,CGCM3.1,MIROC and ECHAM were inaccurate at all recurrence interval levels.
The GFDL, CanESM2 and CSIRO-MK3.6showed improvement in the CMIP5 when compared to CMIP3.In addition, ACCESS1.0,ACCESS1.3 and CanESM2 replicated the GEV distribution better than the other three models, estimating the pressure 1 hPa higher than observation in all extreme events.GFDL overestimated the minimum pressure by 2-3 hPa, whereas CSIRO-MK3.6 and IPSL underestimated it.
Compared with the CMIP3 and CMIP5 GCMs, the estimations of minimum surface pressure were much improved in CMIP6.The GCM results were similar and consistently showed good agreement with observation in larger-ARI events, except for MIROC6 which estimated the surface pressure 5 hPa lower than observation and the other models.
After being downscaled by R1 and R2 in NARCliM 1.5, ACCESS1.3 showed improvement in simulating minimum surface pressure.The GEV distribution derived from ACCESS1.3 downscaled by both R1 and R2 showed a better agreement with the observations for all ARI events.Compared with the parent GCMs in CMIP5, the downscaled ACCESS1.0 and CanESM2 were less consistent with observation in both R1 and R2.
There was a clear and systematic decrease in the 20 year ARI of the minimum surface pressure with increasing latitude of approximately 1.1 hPa/ in the observations at the selected stations along the east coast of Australia (Figure 4).All GCMs closely replicated the observed systematic decrease.The GCM model ensembles in CMIP5, CMIP6 and NARCliM 1.5-R2 well captured the magnitude of the decrease: 1.2 hPa/ in the CMIP5 models; 1.1 hPa/ in the CMIP6 models; and 1.3 hPa/ in the NARCliM 1.5-R2 models.The magnitudes of the decrease were underestimated in CMIP3 and NARCliM 1.5-R1, by 0.9 hPa/ .
All the GCMs model simulations of the changes in the 20 year ARI minimum surface pressure showed good agreement in CMIP3, CMIP5 and CMIP6 at relatively lower latitudes (<34 S).At higher latitudes, the GCMs simulations showed greater bias.Most cases in CMIP3 showed underestimation, whereas most GCMs in CMIP5 and CMIP6 tended to overestimate.The higher the latitude, the less accurate the surface pressure projected by the GCMs in CMIP3 and CMIP5, with error at 2-5 hPa.In CMIP6, ACCESS, GFDL and EC-Earth3 achieved improvement for 20 year ARI minimum surface pressure projections at most stations; MIROC6 underestimated it by 5 hPa from 27 S to 34 S.
The downscaled NARCliM 1.5 models, in both R1 and R2, also demonstrated improvement in simulations of 20 year ARI minimum surface pressure, though the models downscaled by R1 overestimated the 20 year ARI minimum surface pressure by 1-2 hPa ar the stations at latitudes lower than 28 S and the models downscaled by R2 overestimated by 2-5 hPa.ACCESS1.0 and CanESM2 downscaled by both R1 and R2 were 3 hPa lower than the observations in the region between 29 S and 39 S, where the ACCESS1.3 provided better estimates of the extreme value of 20 year ARI minimum surface pressure.Compared with the GCMs of CMIP5, ACCESS1.0 and CanESM2 estimated lower 20 year ARI minimum surface pressure at most stations and were less accurate than the parent GCMs.the GEV distributions and ARIs of the maximum daily precipitation at the Sydney Airport Observation Station (No. 66037) (151.15E, 33.95 S).All the GCMs and their downscaled models underestimated the magnitude of the maximum daily precipitation at all levels of extreme events by a factor of about 2. Importantly, the shapes of the extreme value distributions derived from the models were similar to the shapes of the observations.

| Precipitation
In the CMIP3, the maximum precipitations estimated by models were 80 and 100 mm lower than the observations.The GFDL2.0, GFDL2.1,CSIRO-MK3.0,CSIRO-MK3.5 and CGCM3.1 projected maximum precipitation was 100 mm/day at 100 year ARI level, the MIROC and ECHAM were 70-80 mm/day, whereas the observed extreme precipitation was 250 mm/day.The GEV distribution of maximum precipitation simulated by the GCMs in the CMIP5 showed large differences between the different models, with ACCESS1.0 and ACCESS1.3 values were 30 mm/day higher than other models.However, compared with observation, ACCESS1.0 and ACCESS1.3 still underestimated the extreme precipitation at all ARI events by 70-130 mm/day.The GCMs performed better in CMIP6 than in CMIP3 and CMIP5.Nonetheless, the maximum precipitation was still greatly underestimated: 50-100 mm/day lower than observed at ARIs of 10 to 100 years.
Compared with the precipitation simulations by ACCESS1.0,ACCESS1.3 and CanESM2 models in CMIP5, the extreme daily precipitation simulation was much improved after model downscaling by both RCMs in the NARCliM 1.5.ACCESS1.0 performed better than the others for ARIs lower than 10 years, with underestimates of only 10-30 mm/day.The underestimates of ACCESS1.0 and CanESM2 were 50 mm/day, those of ACCESS1.3 were up to 100 mm/day for the events larger than 20 year ARI.
The latitudinal distributions of the 20 year ARI maximum daily precipitation are shown in Figure 4.The observed maximum daily precipitation decreased from low latitudes to higher latitudes, with the change approximately 9 mm/ for the 20 year ARI events.Although both the GCMs and downscaled models indicated a systematic decrease with latitude, the rate of decrease was only 3-6 mm/ , principally because of the systematic underestimation of extreme precipitation as shown in Figure 3 (middle column).
Consequently, underestimates were especially large in the region from 27 S to 30 S and 34 S to 38 S; the greatest difference between model results and observation was 140 mm/day in the CMIP3 models, 120 mm/ day in the CMIP5 models, 90 mm/day in the CMIP6 models, 90 mm/day in the NARCliM 1.5-R1 and NAR-CliM 1.5-R2.

| Near-surface wind
The GEV distributions for maximum wind speed in the vicinity of the Sydney Airport Observation Station (No. 66037) (151.15E, 33.95 S) are shown in the right column of Figure 3.Of greatest concern is the poor form of the extreme value distributions in comparison with the observations.In all cases, both the GCMs and the downscaled models tended to overestimate the maximum onshore wind speeds at lower ARIs.Although the maximum wind speed must, necessarily, increase with the ARI, the model results showed very weak increases in comparison with the observations.The GCMs in CMIP3 overestimated the maximum westward wind speeds at this location by 2-3 m/s for lower ARI events, except GFDL2.0,GFDL2.1 and CSIRO-MK3.0,which slightly underestimated extreme westward wind speeds by 1-2 m/s.ECHAM tended to overestimate by 2 m/s for all ARIs, while the other GCMs underestimated by up to 8 m/s in the larger-ARI events.CanESM2 in the CMIP5 significantly overestimated the maximum onshore wind speed for all ARIs events, specifically in lower ARIs which errors are up to 5 m/s.Others GCMs underestimated the extreme onshore wind speed.In the CMIP6, CanESM5, MIROC6 and EC-Earth3 tended to overestimate the maximum westwards wind speed in lower ARIs, with the errors of 2-3 m/s compared with observation and underestimate for the ARIs larger than 50 year by 2-3 m/s.The maximum onshore wind speeds projected by other GCMs were much weaker than observation for all ARIs, with the errors up to 10 m/s.
Compared with CMIP5, the differences between the estimations of maximum onshore wind speed and observations were much reduced after downscaling in NAR-CliM 1.5.In both R1 and R2, ACCESS1.0 slightly overestimated the maximum onshore wind speed by 1-2 m/s for lower ARIs.ACCSS1.3 and CanESM2 underestimated all ARIs, with the errors approximately 2 m/s for the lower ARIs and 3-5 m/s for the larger ARIs.
The distribution of 20 year ARI of the maximum onshore wind speed are shown in the right column of Figure 4.According to the recorded data, 20 year ARI maximum onshore wind speed was stronger at latitudes between 27 S and 33 S and 36 S and 40 S, up to 24 and 18 m/s, respectively, than other locations where the wind speed was 10 m/s lower.Neither GCMs in CMIP3 nor CMIP5 capture this spatial distribution, whereas the GCMs in CMIP6 and NARCliM 1.5 downscaling clearly replicated the spatial distribution: 20 year ARI of onshore wind was stronger in both lower and higher latitude and was 50% weaker in mid-latitude.Although all the GCMs and downscaling significantly underestimated the maximum onshore wind speed, the projections of 20 year maximum onshore wind speed were improved after downscaling in NARCliM 1.5, with the errors reduced to 7 m/s in lower latitudes (<33 S) and <5 m/s in higher latitudes.

| CHANGES IN EXTREME VALUES WITH CHANGING CLIMATE
According to the previous section, the GCMs in CMIP6 and NARCliM1.5 are much improved compared with CMIP3 and CMIP5.Minimum surface pressures, maximum precipitations and maximum onshore wind speeds projected by CMIP6 and NARCliM 1.5 are now comparable with those observed historically.In this section, future changes are quantified as the future projected value minus historical value at the same ARI.

| Surface pressure
The left column of Figure 5 shows the changes in the 20 year ARI of the minimum surface pressure projected by the selected GCMs and the downscaled models for the high-emission future.According to model projections, the 20 year ARI minimum surface pressure will have negligible changes in the future, with the variation <2 hPa projected by model ensemble average along the south-east coast, with a reduction in pressure of 2 hPa being approximately a 20 mm increase in water under inverse barometer conditions.This is different with the changes in mean sea level pressure presented in IPCC ( 2023), which indicates that the mean sea level pressure would increase at high latitudes and decrease at midlatitudes.
Both model ensemble average for GCMs in CMIP6 and NARCliM downscaled models projected the 20 year ARI minimum surface pressure slightly decrease in lower latitudes (<32 S), whereas no change or slightly increase in the higher latitudes (Figure 5).The standard deviations of the projected changes in minimum surface pressure were summarized in Table 3.
At lower latitudes (27-33 S), the changes in the 20 year ARI minimum surface pressure projection by the GCMs of CMIP6 showed great variation with larger standard deviation (Table 3).ACCESS, CanESM5, GFDL and MIROC6 in CMIP6 projected the minimum surface pressure would be 2-7 hPa lower, while EC-Earth and IPSL projected the 20 year ARI minimum pressure would be 2-4 hPa higher.At higher latitudes (>33 S), ACCESS projected no change; CanESM5 and GFDL projected the minimum surface pressure would be 1-3 hPa higher, whereas IPSL, MIROC6 and EC_Earth3 projected the minimum surface pressure would be 1-2 hPa lower.At the Sydney Airport Observation Station, the 20 year surface pressure will vary <2 hPa.
Although the spatial patterns of the changes projected by model ensemble average in both R1 and R1 in NAR-CliM1.5 are similar, the changes signs were different between individual models.ACCESS1.3 downscaled by the NARCliM 1.5-R1 and NARCliM 1.5-R2 presented similar projections which indicated the minimum sea lever pressure would be 1-2 hPa lower in the high emission future along the south-eastern coast.Both ACCESS1.0 downscaled by R1 and R2 projected the 20 year ARI surface pressure would be 1-3 hPa higher in all locations.CanESM2 downscaled by R1 projected 20 year ARI surface pressure would change slightly with no more than 1 hPa lower, while the model downscaled by R2 projected the 20 year ARI minimum surface pressure would be 1 hPa lower in the lower latitudes (<32 S) but 1 hPa higher in the higher latitudes.All models in NARCliM1.5-R1projected an increase of 0.5-3 hPa, whereas ACCESS1.3 and CanESM2 in R2 projected a decrease at the Sydney Airport Observation station.

| Precipitation
Systematic increases in the 20 year ARI maximum daily precipitation were projected by most GCMs and NAR-CliM downscaled models (Figure 5, middle column).These increases were larger at lower latitudes (<40 S), with the model ensemble average projecting an increase of 35 mm/day, and smaller in the higher latitudes which were no more than 20 mm/day in most cases in CMIP6.The model ensemble average of NARCliM1.5-R1and R2 also projected an increase of 20 mm/day at low latitude (28 S), but the changing patterns were different in the higher latitude.NARCliM1.5-R1projected a greater increase along the south-east coast.The increase of 20 year ARI maximum precipitation was notably larger in the region between 29 S and 32 S, with the average increase up to 50 mm/day.Whereas NARCliM1.5-R2projected a decline at mid-latitude (31 S) and no more than 20 mm/day increase in latitudes higher than 32 S.
The precipitation projections were of great uncertain, with the standard deviation higher than 15 mm (Table 3).ACCESS, GFDL, IPSL and MIROC6 projected the 20 year ARI maximum daily precipitation would be increased by 15%-20% at lower latitudes.The CanESM5 in CMIP6 projected a significant increase in the region between 27 S and 34 S and 65% in the region between 34 S and 43 S. The changes projected by EC_Earth3 showed the same magnitude with other models, except at 33 S and 36 S, where the 20 year ARI maximum precipitation would decrease by 15%.At the Sydney Airport Observation station, the GCMs in CMIP6 projected the 20 year ARI maximum precipitation would increase more than15%.
After downscaled in both R1 and R2, the uncertainty of precipitation projection remained significant, especially in the lower latitude region (<32 S) where the standard deviation was over 30 mm in most cases (Table 3).In NARCliM-R1, all models projected that the 20-year ARI maximum precipitation would be 18%-37% greater in the future at the latitudes higher than 32 S. At lower latitudes, different results were found.The ACCESS1.0 and CanESM2 projections showed a 4% and 11% decrease in lower latitude (<29 S), respectively.ACCESS1.3 projection showed a 30% increase in this region.The NARCliM1.5-R2downscaled model from ACCESS1.3 and CanESM22 project the increase by 10%-35% over the whole south-east coast.Both NARCliM1.5 models in R1 and R2 projected the 20 year ARI for maximum precipitation would be greatly at Sydney Airport Observation station, except ACCESS1.3 downscaled by R2 which projected that the precipitation would be slightly decrease.
These are significant increases in precipitation, with projected increases of 25% in some cases.It is also important to note that the increase in runoff may be proportionally greater.Although the initial and continuing losses of the catchment were projected to increase with the changing climate (Ho et al., 2023), the projected increases in precipitation still can be over and above existing precipitation depth with these losses deducted.

| Near-surface wind
Changes in the maximum westward wind speed projected by the GCMs were negligible in the high-emission T A B L E 3 Summary of the standard deviation of the changes in 20 year ARI minimum surface pressure, maximum precipitation and maximum westward wind speed projected by CMIP6 and NARCliM1.5 models.future.The model ensemble averages of CMIP6, NARCliM1.5-R1and R2 (Figure 5, right column) showed agreement in projecting the 20 year ARI westward wind speed.There is no change in the lower latitude region (<36 S).In higher latitude, the model ensemble averages of CMIP6 projected a decrease of approximately 1 m/s.The projections of downscaled models both in R1 and R2 in NARCliM1.5 showed uncertainty of the change sign in 20 year ARI maximum westward wind speed over the south-east coast.However, these model projections were great consistent in the region between 32 S and 34 S with the standard deviation <0.4 m/s (Table 3).Models projected that the 20 year ARI for maximum onshore wind would not change at Sydney Airport Observation station, with the changes <0.3 m/s.

| DISCUSSION
The projections of the GCM and downscaled models of latitudinal variations in the 20 year ARI minimum surface pressure agreed closely with the observations, consistent with the findings in Peirson et al. (2011Peirson et al. ( , 2014)).
Although the systematic decrease trend in the 20 year ARI maximum precipitation with increasing latitude is captured well by both GCMs and downscaled projections, these modelling outcomes still significantly underestimate the observations, as found in previous studies (Peirson et al., 2011(Peirson et al., , 2014;;Perkins & Pitman, 2009).Perhaps of greatest concern is the poor representation of the extreme value distributions for onshore wind speed in comparison with the observations.Although the latitudinal distribution projected by GCMs in CMIP6 and NAR-CliM1.5 downscaled models showed general agreement with the observations, some models tended to overestimate onshore wind speeds at low ARIs, and all models underestimate at higher ARIs.Compared with the design onshore wind value from Australian Standards AS1170:2011 (Figure 5 in Peirson et al., 2014), one of the important criteria of coastal structures, the results are encouraging.Although the average of all the GCM models projected 20 year ARI onshore wind speeds was lower than the design wind value, the NARCliM 1.5 projections of 20 year ARI onshore wind intensities showed a good approximation of the design wind values.
Wind speeds are known to be higher over sea than land (Peirson et al., 2014).And the precipitation is strongly affected by the local topographic features (Olson et al., 2016).Consequently, the sensitivity to the location of grid cells in the GCMs and downscaling models were examined.Table 4 shows the changes in the 20 year ARI of all three variables between adjacent land and ocean grid cells (Figure A1) at mid-latitude region of the study domain.Both GCMs and NARCliM 1.5 downscaled simulations showed the onshore wind speeds were systematically higher over sea than land, whereas the simulations of maximum precipitation indicate the 20 year ARI precipitation were lower over sea than land.Compared GCMs and the high-resolution downscaled models, the spatial resolution has great impact on the magnitude of the maximum precipitation and onshore wind speed at both land and sea cells.In most land-cell cases, the simulations of the 20 year ARI the maximum precipitation and westward wind speed show stronger intensities in the NARCliM 1.5 downscaled models than GCMs in historical scenario.The GCMs estimated the 20 year ARI precipitations are approximately 10-50 mm/day higher over the sea cells than over land, whereas the maximum precipitation was less sensitive to the change in terrain, with the differences between sea and land cells were only 10-20 mm/day projected by NARCliM 1.5.According to the GCMs and downscaled models, the changes pattern of minimum surface pressure, maximum precipitation and maximum onshore wind speed are similar over sea and land.The minimum surface pressure would change negligibly, while the maximum daily precipitation increase significantly over both sea and land.Changes in maximum westward wind speed showed great uncertainty on both sea and land.
The changes in the 50 year ARI values in the future climate at the Sydney Airport Observation station, located at the mid-latitude of this study domain, were also investigated (Table 5).The characteristics of the projected changes in three variables in the high-emission future were similar to those at 20 year ARI.Both GCMs and NARCliM 1.5 downscaled models showed that the minimum surface pressure change negligibly at 50 year ARI, with the changes <4 hPa.The changes in onshore wind speed were uncertain, ranging from −9% to 10%.However, the 50 year ARI maximum precipitation intensity had a larger increase than that of the 20 year ARI.The model ensemble average of CMIP6 and NARCliM1.5-R1and R2 indicated the changes in the 20 year ARI of around 23.1%-33.6%and changes in the 50 year ARI of around 27.8%-34.5%,excluding ACCESS1.0-R2 which projected a decrease extreme precipitation.This finding is consistent with CSIRO and Bureau of Meteorology (2015) and Grose et al. (2020).
Although different indicators or model outputs were used for evaluating extreme precipitation projection, the results of changes in extreme precipitation were consistent with previous studies (Dowdy et al., 2013;Herold et al., 2021;Peirson et al., 2011Peirson et al., , 2014;;Pfahl et al., 2017).There is great uncertainty of the changes in the magnitude of the increasing maximum precipitation by the end of 21st (Alexander & Arblaster, 2017;Evans et al., 2016;Herold et al., 2021;Peirson et al., 2011Peirson et al., , 2014;;Pfahl et al., 2017).Note: The locations of grid cells are shown in Figure A1.
Precipitation variability in southeast of Australia is closely related the sea level pressure anomalies and dynamical process such as fronts interactions and extratropical circulation (Dowdy et al., 2013;Murphy & Timbal, 2008;Pfahl et al., 2017).Some investigations have attributed the variations of extreme precipitation projections to the uncertainty in the changes in atmospheric dynamic and regional dynamical circulation (Pfahl et al., 2017).For example, the duration and intensification of ECLs strongly depends on the physical representation of climate processes (di Luca et al., 2016).
The results according to standard deviation of changes in precipitation show that the uncertainty was not addressed by downscaling in the lower latitude region along southeast Australian coast.The uncertainty of extreme precipitation seems not only affected by local conditions and processes, but also possible affected by the larger scale dynamical process.Identifying the key processes that result to the different patterns between individual models would contribute to the improvement of extreme precipitation projection.One of the major challenges in improving maximum precipitation projection is to understand the changes in climate drivers (e.g., ENSO) and the response dynamical process of regional circulation (Grose et al., 2020;Head et al., 2014;Pfahl et al., 2017).
In addition, the factors that influence precipitation in southeast Australia are not restricted to topography and atmospheric disturbances (Olson et al., 2016).The temperature and local soil moisture are important drivers.The higher temperature can contribute to an increasing extreme precipitation by amplifying the atmospheric moisture level in accordance with the Clausius-Clapeyron relationship (Panthou et al., 2014;Roderick et al., 2019;Trenberth et al., 2003;Vazquez et al., 2020).Given the wind speeds would remain unchanged in most part of the study region, the increasing extreme heatwave frequency in south-east Australia might be responsible for the increasing extreme precipitation (Herold et al., 2021).
The skill scores are calculated by using the 20 year ARI at all stations along the south-east coast.These results should be taken with caution.For example, MIROC6 in CMIP6 received a skill score of 0.7 for surface pressure which classified as excellent, but this model showed an obvious underestimation at Station No. 66036 (Figure 3).The reasons leading to this inconsistency were that the bias of MIROC6's projections was small (<5 hPa) and MIROC6 estimated 20 year ARI minimum surface pressure well in the station at higher latitude.Therefore, MIROC6 in CMIP6 had a high skill score overall, but it did not replicated the GEV distribution at all ARI with The simulations by the GCMs and its downscaled counterparts of minimum surface pressure, maximum precipitation and maximum westward wind along south-east Australian coast were fully investigated by derived GEV distribution for the variables.This study did not attempt to identify the ensemble differences to specific processes or mechanisms.Instead, we focused on presenting the differences across the models in extreme simulations of all three variables.The GCMs in all CMIP phases and their downscaled counterparts successfully capture the changing with latitude along the south-east coast of Australia in daily minimum surface pressure, maximum precipitation and maximum onshore wind speed.Overall, both the selected GCMs in every CMIP phases and the NARCliM 1.5 downscaled show agreement in the regional climate change.Compared with CMIP3, the GCMs in CMIP5 showed an improvement as supported by previous studies (Sillmann et al., 2013).Furthermore, the GCMs in CMIP6 and downscaled in NARCliM 1.5 also significantly improved.
The extreme surface pressure outputs provided by GCMs, especially in CMIP6 phase, seemed to be more reliable and can be applied reliably in studying the extreme events in this coastal region.The maximum precipitation projections improved significantly in the later GCMs generations and their downscaled models.Nonetheless, the GCM-based projections underestimate the historical maximum precipitation, and bias corrections are still required to incorporate these in climate change studies.Although, the GCMs in CMIP6 and their downscaled models projected extreme onshore wind speeds did improve, the model-derived extreme value distributions do not replicate the form of the extreme value distributions and the increase of onshore wind speed in rare events is significantly underestimated.The greatest concern is the overestimation of the wind at ARIs <10 which significantly changes the form of the extreme value distributions.
The model projections of extreme climate change in the high-emission future were also investigated.The CMIP6 and NARCliM1.5 downscaled models indicated that the projected future changes in minimum surface pressure and onshore wind speed would be small.Precipitation was projected to increase significantly in this region, which would contribute to an increased runoff.Despite the uncertainties in the magnitude of changes in extreme precipitation, the increase of precipitation is relatively robust across the model.The analysis of changes in extreme storm events is thus of great importance in coastal planning and management.The significance of all these findings is presently being incorporated into an estuarine inundation assessment on the south-east coast of Australia.
Some models with high equilibrium climate sensitivity (ECS) do not perform well in climate projections (Scafetta, 2022).Current study includes models with all ECS level, including low, middle and high ECS.Taking account of the variations of ECS and its impact can help to improve the robustness and reliability of model assessment.The normalization of changes by the degree of warming should be conducted in further research.

DATA AVAILABILITY STATEMENT
The GCMs model data used in this paper are available at the official website of the Earth System Grid Federation (https://esgf.nci.org.au/projects/esgf-nci/).The NARCliM downscaled datasets can be downloaded from the NSW Climate Data Portal (https://climatedatabeta.environment.nsw.gov.au/datasets/gov.au).The extracted observed data can be accessed at https://doi.org/10.5281/zenodo.7508723.The full codes for reproducing the analysis have been uploaded to Zenodo and can be accessed at https://doi.org/10.5281/zenodo.7508699.

ORCID
Wenjun Zhu https://orcid.org/0009-0007-6817-7602 Map showing the study area as denoted by the red box in the inset with the spatial distribution of the observation stations (red stars) with their respective station numbers.The orange box indicates the NARCliM domain of 10 km resolution; (b) Grids of ACCESS1.0 in CMIP5, blue dots indicate the data selected grid cell; (c) Grids of ACCESS1.0 in NARCliM 1.5, blue dots indicate the data selected grid cells.[Colour figure can be viewed at wileyonlinelibrary.com]

F
I G U R E 2 GEV distribution of minimum surface pressure at Sydney Airport Observation Station (No. 66037) derived by historical experiment simulation from CanESM5 in CMIP6.Red circles represent the model outputs, the blue line indicates the fitted GEV distribution.90% confidence limits are shown as black dashed lines.[Colour figure can be viewed at wileyonlinelibrary.com] Precipitation biases in GCMs are well known(Alexander & Arblaster, 2017;Peirson et al., 2014;Perkins & Pitman, 2009).Figure3(middle column) show F I G U R E 3 GEV distributions from GCM historical experiments proximate to Sydney Airport Observation Station (No. 66037).GEV distribution for minimum surface pressure (left column), maximum precipitation (middle column) and maximum westward wind speed (right column) extracted from CMIP3, CMIP5, CMIP6, NARCliM 1.5-R1 and NARCliM 1.5-R2.Black solid lines represent the GEV distributions derived from observation.Dashed lines indicate the observation 90% confidence intervals.[Colour figure can be viewed at wileyonlinelibrary.com]

F
I G U R E 4 The 20 year ARI values with 90% confidence intervals as shown.Minimum surface pressure (left column), maximum precipitation (middle column) and maximum westward wind speed (right column) extracted from historical experiment in CMIP3, CMIP5, CMIP6, NARCliM 1.5-R1 and NARCliM 1.5-R2.Black dots indicate the corresponding observations.Black lines indicate the design wind value from AS1170:2011 using CEM (2006) figure II-2-1.[Colour figure can be viewed at wileyonlinelibrary.com]

F
I G U R E 5 Changes in 20 year ARI minimum surface pressure (left column), maximum precipitation (middle column) and maximum westward wind speed (right column) projected by CMIP6, NARCliM 1.5-R1 and NARCliM 1.5-R2.The red dash lines value is zero and indicate no changes.The black dash lines indicated the changes projected by model ensemble average.[Colour figure can be viewed at wileyonlinelibrary.com]

T
A B L E 4 Comparison of the 20 year ARI values in historical scenario derived from adjacent land cells and ocean cells and the changes in the high-emission future at each grids at the mid-latitude region near Sydney.
T A B L E 2 Summary of the skill scores of selected GCMs and NARCliM 1.5 models in 20 year ARI precipitation, surface pressure, westward wind speed along south-east coast.[Colour table can be viewed at wileyonlinelibrary.com] Green shading indicates model performance level is excellent, light green shading indicates model performance level is very good.Orange shading indicates the model performance level is good.Blue shading indicates the model performance level is poor. Note: T A B L E 5 Changes in 20 year ARI and 50 year ARI at Sydney airport observation station (No. 66037) for: minimum surface pressure, maximum precipitation and maximum westward wind speed projected by GMCs in CMIP6 and NARCliM1.5R1andR2.[Colour table can be viewed at wileyonlinelibrary.com]Orange shading indicates increase, green shading indicates decrease and blue shading indicates no change.2154observationsat Station No. 66036.Thus, both skill scores and GEV distributions at the stations should be taken into consideration when the model skills were assessed for low pressure, precipitation and westward wind. Note: