Tropical cyclone simulations over Bangladesh at convection permitting 4.4 km & 1.5 km resolution

High resolution simulations at 4.4 km and 1.5 km resolution have been performed for 12 historical tropical cyclones impacting Bangladesh. We use the European Centre for Medium-Range Weather Forecasting 5th generation Re-Analysis (ERA5) to provide a 9-member ensemble of initial and boundary conditions for the regional configuration of the Met Office Unified Model. The simulations are compared to the original ERA5 data and the International Best Track Archive for Climate Stewardship (IBTrACS) tropical cyclone database for wind speed, gust speed and mean sea-level pressure. The 4.4 km simulations show a typical increase in peak gust speed of 41 to 118 knots relative to ERA5, and a deepening of minimum mean sea-level pressure of up to −27 hPa, relative to ERA5 and IBTrACS data. The downscaled simulations compare more favourably with IBTrACS data than the ERA5 data suggesting tropical cyclone hazards in the ERA5 deterministic output may be underestimated. The dataset is freely available from 10.5281/zenodo.3600201.


Background & Summary
To construct this dynamically simulated tropical cyclone dataset we use the latest generation Met Office regional model to simulate tropical cyclones (TCs) over the Bay of Bengal (BoB) at grid-box resolutions of 4.4 km and 1.5 km. Using the ERA5 reanalysis data 1,2 to initialise and provide boundary conditions for our regional models, we dynamically downscale 12 historical TCs that made land-fall over Bangladesh between 1991 and 2019, using an ensemble approach.
Downscaling of ERA5 is reported in a few other studies: Bonanno et al. 3 downscale ERA5 using the Weather Research and Forecasting (WRF) model to produce a new 7 km reanalysis over Italy; preliminary work by Taddei et al. 4 use ERA5 to force the BOlogna Limited Area Model-MOdello LOCale (BOLAM-MOLOCH) regional model for the purposes of coastal risk assessment in the North Western Mediterranean sea, and Wang et al. 5 use ERA5 to run a 10 km WRF domain over high mountain Asia. Specifically examining tropical cyclones, many studies use variations of the Weather Research and Forecasting (WRF) Model 6 , such as Kaur et al. 7 who use WRF to downscale the National Center for Environment Prediction (NCEP) Climate Forecast System (CFSv2) and its atmospheric component Global Forecast System (GFS) to 9 km over the north Indian Ocean for two historical cases (Mora and Ockhi), with analysis focusing on the spatial accuracy of rainfall and 850 hPa vorticity, and the vertical profiles of wind and temperature. They conclude that the downscaled model significantly improves the spatial distribution of rainfall, maximum vorticity evolution, wind, and temperature profiles for mature phase cyclones. Studies specifically examining the BoB simulations [8][9][10] typically make empirical comparisons of TC simulations at ~10 km resolution against observationally based data, but often with an India-centric domain that contains a larger number of landfalling events. By contrast, in this study we specifically focus on Bangladesh, with simulations at higher resolution.
We make 12 variables available, including: air temperature, maximum wind gust speed, minimum air pressure at sea level and precipitation amounts (see Table 1 Table 2 lists approximate landfall times and their International Best Track Archive for Climate Stewardship (IBTrACS) 11,12 ID number. Note that for this paper, the name refers to a shorthand identifier, used for file naming purposes, but it does not necessarily reflect the official storm identifier. At the time of writing, ERA5 data was only available from 1979 onwards, so our new catalogue excludes cyclones prior to 1979, most notably Cyclone Bhola of November 1970.
Whilst this data descriptor paper focuses on validating this new model dataset, a companion paper 13 uses the data presented here to produce spatially-consistent exceedance probability estimates for extreme gust speeds over Bangladesh, and demonstrates a method for producing a transparent decision-making framework for tropical cyclone warnings

Methods
Numerical modelling. Our high-resolution convection-permitting modelling utilises the latest generation Met Office Unified Model 14 v11.1, regional atmosphere configuration RAL2-T, a further development of RAL1-T 15 -hereafter referred to as RAL2. For each historical tropical cyclone case listed in Table 2, we run the RAL2 model in a 'downscaling' configuration, using ERA5 data to initialise and provide boundary conditions for a series of 9 time-lagged ensembles (see Fig. 1 for a visual representation of this configuration).
As there is no data assimilation process or nudging, the initial conditions imposed by ERA5 are found to have significant influence on the resulting tropical cyclone development. The time-lagged configuration is designed to limit the free-running model time to 72 hours, whilst ensuring that the central 24-hour period of interest (centred on the tropical cyclone landfall time) is sufficiently sampled from a range of ERA5 initial conditions. This initial condition ensemble approach produces a set of 9 plausible tropical cyclone development scenarios associated with each named event. After initialisation, each ensemble member is free running for 72 hours, with hourly boundary conditions provided by ERA5. Each run requires a 24-hour spin-up period as the regional model adjusts from the weak initial state inherited from the ERA5 driving global model. This initial 24 hours of model  www.nature.com/scientificdata www.nature.com/scientificdata/ data are discarded in subsequent analysis and data files. Together, the amassed ensemble provides 9 simulations of the central 24 hours, but covers a total period of 72 hours.
The RAL2 4.4 km domain avoids placing model boundaries over the Himalayas and covers Nepal, Bhutan, Myanmar, most of India, and parts of the Tibetan plateau; the RAL2 1.5 km domain is limited to Bangladesh only (Fig. 2). To ensure model stability over this mountainous terrain, the RAL2 model was run with a 30 second time-step for both 4.4 km and 1.5 km simulations with additional orographic smoothing applied (using a 1-2-1 filter) to model cells 1500 m above mean sea level.

Storm tracking.
Storm tracking is performed on 3-hourly fields of RAL2 mean sea-level pressure (MSLP), 400 hPa temperature and 10 m wind speed, using the Tempest extremes software 16 . Unfortunately, 3-hourly fields are not frequent enough to estimate landfall time using the Tempest tracking algorithm for RAL2 data. The tracking algorithm has two parts -the initial feature detection and the stitching of these features to calculate tracks.
Feature detection is based on finding minima in air pressure at sea level, with features within a radius of 6° of each other being merged. The features are then further refined with a two 'closed contour criteria' . First an increase in sea level pressure of at least 200 Pa (2 hPa) within 5.5° of the candidate node, and second a decrease in 400 hPa air temperature of 0.4 K within 8° of the node within 1.1° of the candidate with maximum air temperature.
Stitching, to combine the individual features into tracks, uses a maximum distance between features of 3°, a minimum track length of 2 points (equivalent to 6 hours) and a minimum path distance of 0.1°. We also apply a topographic filter and a filter on maximum wind speed: tracks were rejected if they did not have at least one time-step and last at least 24 hours at an altitude less than 10 m; and if they did not have maximum wind speed of at least 17 m s −1 at one time-step.
Datasets. ERA5 1,2 is the fifth and latest generation reanalysis dataset issued by the European Centre for Medium-Range Weather Forecasts (ECMWF). It combines both model data and observations on a real-time basis in a data assimilation process. Like a forecast, newly available observations are combined with model data to produce the best estimate of the state of the atmosphere. ERA5 data offers many improvements on the previous reanalysis, ERA-Interim, including more developed model physics and dynamics and an increased horizontal resolution of 30 km. In term of vertical resolution and extent, it has 137 model levels up to 80 km. For ERA5, we compare our simulated storm data with '10 metre wind gust since previous post-processing' defined as the maximum 3-second wind for each hour (parameter ID 49) and MSLP (parameter ID 151). Prior to 30th Sep 2008, ERA5 gust estimates only include turbulent contributions; the convective contribution was added to the wind gusts in post-processing for events after this date 17 .
International Best Track Archive for Climate Stewardship 11,12 forecasts are made by numerous forecasting centres around the world, and consists of the positions and intensities of tropical cyclones 18 . For our validation purposes, two Regional Specialized Meteorological Center (RSMC) datasets are used: the India Meteorological Department, New Delhi (IMD), and the Central Pacific Hurricane Center, Honolulu (CPHC). IBTrACS best track initialise the simulation start point (grey dots). Each ensemble member then has a 24 hour spin-up period (grey dashed lines) which is discarded from all analysis. The 48-hour simulation that is kept is represented by the solid blue line. ERA5 lateral boundary conditions (LBCs, black dots) feed into the 4.4 km domain every hour. The lagged ensemble is designed to simulate a central 24-hour period (shaded grey), common to all ensemble members and centred on the tropical cyclone land-fall time (orange star), but also sample a range of ERA5 initial conditions. data are typically calculated using a post-season reanalysis of storm positions and intensities from all available data, including ship, surface and satellite observations 18 . Typically, best track data consist of a time series of the storm's position, maximum sustained wind speed (in knots) and minimum central pressure. Estimated uncertainty of the IBTrACS forecast wind speed are ± 10 to ± 20 knots, with positional uncertainty radiuses of 10 km to 40 km, dependent on wind speed intensity 19 . No uncertainty information is provided for pressure, but we note that the World Meteorological Organisation typically assume reporting precision of ± 3 hPa. We also note that IBTrACS data is subject to forecaster best judgement and best track data typically lags the provisional operational data cyclone estimates by some months, subject to the availability of reanalysis data.
For the IBTrACS dataset we compare with 'maximum sustained wind speed' and MSLP. Although the World Meteorological Organisation 20 defines sustained wind speed as a 10-minute average windspeed at 10 m height above ground, it is reported as 1-minute averages by US forecast centres, and 3-minute averages by IMD. Some agencies, including CPHC, estimate gust speeds; however this data is not available for the BoB basin. Methods for obtaining maximum wind speed in IBTrACS vary by agency, as do their availability of TC observation data. IBTrACS minimum central pressure is generally estimated with both subjective and objective satellite analysis as well as automated buoys that may be present 19 . Note that IBTrACS estimates usually end once the cyclone makes landfall.
Comparing datasets. For the purposes of comparing RAL2 simulated winds and gusts with IBTrACS and ERA5, the RAL2 maximum sustained wind speed is taken as the maximum of a single RAL2 model timestep windspeed over the accumulation period (1 hour). This is broadly comparable to a sustained maximum windspeed calculated with 30-second averaging period. In contrast, the parameterised RAL2 gust diagnostic represents a prediction of the 3-second average windspeed at every timestep. The maximum of this 3-second average speed over an hour is then taken to give the hourly maximum 3-second gust speed.
Considering the ERA5 and RAL2 model physics, ERA5 uses a mass flux scheme for cumulus parameterisation 21 whereas RAL2, while not truly resolving deep convection, is able to explicitly represent deep convective processes within the resolved dynamics. At these kilometre-scale resolutions the lower horizontal size limit of convective cells is still set by the effective resolution (5 to 10 times the grid length 22,23 ). More generally, only grid spacings on the order of 1 km are comparable to the size of particularly energetic eddies in the planetary boundary layer 24 , so the turbulent processes as well as the dominant turbulent length scale will be under resolved in both our downscaled model and ERA5. ERA5 gusts are parametrised based on the 10 m wind speed, friction velocity, atmospheric stability, roughness length and a convective contribution based on wind shear between the model levels at 850 hPa and 925 hPa 17 . It is known that extreme gusts associated with vigorous convection in ERA5 are generally under-estimated, sometimes by a factor of two 25 . The RAL2 model uses a gust parametrisation based on 10 m wind speed with scaling proportional to the standard deviation of the horizontal wind that also accounts for friction velocity, atmospheric stability and roughness length 26 .
Comparisons of minimum MSLP are more straightforward. We compare the RAL2 hourly minimum MSLP estimated every 30-seconds, with the hourly minimum MSLP from ERA5, and the 3-hourly minimum MSLP from IBTrACS.

Data Records
Our RAL2 model output is available from Zenodo in NetCDF format 27   www.nature.com/scientificdata www.nature.com/scientificdata/ The ERA5 data used to drive our RAL2 model and used for validation is available from the Copernicus Climate Change Service portal https://climate.copernicus.eu/climate-reanalysis.
To facilitate integration with loss modelling processing necessary for risk management and risk transfer, we also make our RAL2 data available in a format compatible with the open source Oasis loss model 28 . This data format is designed to be used as one component of a loss model and is formed of CSV and binary files. This data is available under CC-BY 4.0 licence from https://oasishub.co/dataset/bangladesh-tropical-cyclone-historical-catalogue.

technical Validation
A lack of reliable, high-frequency and consistent meteorological observation data available for Bangladesh mean that verification of modelling results against in-situ observational data is not possible. Instead we establish the validity of the RAL2 4.4 km data relative to ERA5 and the IBTrACS catalogue. It is important to recognise the differences in how the data are collected, their processing and resolution (see Table 3). Comparison of storm tracks is performed against the IBTrACS best track data only.
For the purposes of validation, we focus on three key variables: maximum wind speed, maximum gust speed and minimum pressure at mean sea level (MSLP). We have a particular interest in gust speed, as these are strongly related to storm damage 29,30 and are commonly used within the catastrophe modelling industry. All comparisons against IBTrACS compare hourly maximum wind from our RAL2 4.4 km model versus 3-hourly maximum wind speed estimates from IBTrACS. For maximum gust speed, we compare the RAL2 hourly maximum 3-second gust diagnostic with ERA5 hourly maximum 3-second gust speed diagnostic. MSLP estimates are comparable across all three datasets. In each case, the comparison is performed over a land-masked longitude-latitude domain that extends [79, 100]°E and [10,25]°N -see Fig. 2. This domain explicitly seeks to focus on the Bay of Bengal so as to compare model fields without land effects. In all cases, excluding land areas has very minor impact on the validation comparison (not shown) as peak wind, gust and minimum MSLP all occur over the ocean. Although our storm tracking output does not allow us to explicitly compare the time of landfall between datasets, we expect that differences in the time of peak wind speeds would be mirrored in the differences in the time of landfall across datasets as RAL2 peak wind speeds tend to occur just prior to landfall.
Each validation plot (Fig. 3, and Supplementary Information Figures A1-A11) displays the gust speed, wind speed and MSLP from the ERA5, IBTrACS and RAL2 4.4 km. We resample the IBTrACS 3-hourly data by forward filling to 1-hourly intervals to aid the comparison of max/min timing with ERA5 and RAL2 datasets. Where IBTrACS maxima (minima) persist over several hours, the time differences reported in Sections 4.1 and 4.2 are then the minimum time difference between padded IBTrACS data and RAL2. The actual difference of RAL2 with respect to ERA5 (RAL2 -ERA5) is denoted ΔERA5 for brevity. For IBTrACS, actual differences with respect to IMD and CPHC are denoted ΔND and ΔUS respectively.
The statistical robustness of differences between datasets are assessed using the percentile bootstrap hierarchical shift function [31][32][33][34] . Given the potential skewness of the data, rather than looking at the differences of a single estimate of central tendency across all events (e.g. the median), differences are assessed for deciles (or percentiles) across the full distribution of the data, calculated using the distribution-free Harrell-Davis estimator 35 . This method explicitly deals with the hierarchical setting of data representing the same event, sharing common synoptic atmospheric conditions, but where different events are independent in time. The robustness of differences is assessed using bootstrapped (n = 1000) uncertainty intervals for each decile difference. Where the 95% highest density interval (HDI) of uncertainty does not intersect zero, decile differences are considered statistically robust.
Intensity and timing of maximum sustained wind speed. For all events, RAL2 maximum sustained wind speeds are faster than ERA5 wind speeds (Fig. 4), with median (across all events) ΔERA5 = 35 kn (18 m s −1 ), with the 5 th to 95 th percentiles of the data spanning [10,70] kn ([5, 36] m s −1 ). Comparing IBTrACS, median ΔUS = −6 kn (−3 m s −1 ) and ΔND = 10 kn, (5 m s −1 ). Assessing the robustness of differences, the distribution of ΔERA5 is robustly slower than RAL2 across all deciles (based on 95% HDI for each decile difference). ΔND is also robustly slower for differences greater than the 40 th percentile; however, note that at the time of writing, IBTrACS IMD maximum sustained wind speed data for Fani and Bulbul were unavailable. Although IBTrACS US data has a tendency toward faster sustained wind than RAL2 (i.e. negative ΔUS) these differences are not robustly different to zero at the 95% HDI.
The timing of maximum wind speed shows significant variation between events, with no clear correlation to peak wind intensity differences; however, generally RAL2 peaks are delayed relative to ERA5 and IBTrACS data.

Dataset
Data  www.nature.com/scientificdata www.nature.com/scientificdata/ Across all events, median ΔERA5 = 5.5 hours delay, with ΔUS = 2.5 hours and ΔND = 0.5 hours. Only ΔERA5 and ΔUS times are robustly different to RAL2 (evaluated at the 95% HDI). The largest time differences occur against ERA5 data: e.g. for Fani, some RAL2 ensemble members show maximum wind intensities delayed by over 20 hours relative to ERA5, but it is noted that for these cases the ERA5 tropical cyclone simulation seems especially weak (for maximum wind, gust and minimum MSLP) compared to IBTrACS data. Some of the variance in peak times will also derive from the differences in data frequency (1-hourly for RAL2 versus 3-hourly for IBTrACS) but this requires further investigation to quantify.
Intensity and timing of mean sea-level pressure. For most events, the RAL2 ensemble produces deeper MSLP minima than the ERA5 and IBTrACS data (Fig. 5), but whilst ΔERA5 (median = −18 hPa) and ΔND (median = −10 hPa) differences with RAL2 are robustly different to zero, ΔUS (median = −2 hPa) is not (all evaluated at the 95% HDI). At the time of writing, IBTrACS MSLP data for Fani and Bulbul are unavailable from IMD, and BOB01, BOB07 and TC01B are unavailable from CPHC.
As for wind speeds, the timing of RAL2 MSLP minimum is typically delayed relative to IBTrACS or ERA5 data. Median time difference of MSLP minima are similar to wind speed maxima differences: ΔERA5 = 7.5 hours delay, ΔUS = 3.5 hours and ΔND = 0.5 hours. Again, only ΔERA5 and ΔUS times are robustly different to RAL2 (evaluated at the 95% HDI). As for the timing of gust peaks (Section 3.1), the RAL2 simulation of Fani shows median delays in MSLP minima of 14 hours (ΔERA5) and 11 hours (ΔUS). BOB01 also has an equivalent delay of 13 to 14 hours (ΔND and ΔERA5 respectively).
Intensity and timing of maximum 3-second gust speed. The distribution of RAL2 gust speeds across events, are uniformly higher than ERA5 (Fig. 6). The median difference across all events is 63 kn (32 m s −1 , Fig. 6), with some particularly strong individual events showing median differences up to 93 kn (48 m s −1 , BOB01) and 118 kn (61 m s −1 , Sidr). Comparing differences in the RAL2 and ERA5 gust speed distributions using bootstrapped median difference by percentile across all events, shows that these differences are robustly different to zero at the 95% HDI.
As with wind and MSLP, differences in the timing of maximum 3-second gust speed vary considerably between events with no clear correlation between the magnitude of the gust difference and the absolute time differences. The median time difference across all events is 2.5 hours (Fig. 6), but this is not robustly different to zero at the 95% HDI.

Storm tracks.
We compare the track density of our nine downscaled ensemble members to IBTrACS in 30 × 30km spatial bins. Typically, the area influenced by the tropical cyclone wind hazard is in excess of 200 × 200 km, so this assessment of storm tracks plays a more important role in evaluating storm surge, primarily influenced by the area of low pressure at the centre of the cyclone. www.nature.com/scientificdata www.nature.com/scientificdata/ Comparing storm tracks (Fig. 7) shows that for 8 of 12 cyclones, the RAL2 storm tracks have at least one ensemble member that makes landfall with the bounds of an IBTrACS track. Notable exceptions to this are: BOB07, which shows high consistency in storm track amongst the RAL2 ensemble, but makes landfall to the north of the IBTrACS estimates; TC01B and Viyaru, which show greater spread amongst the RAL2 ensemble members, but consistently make landfall to the south of the IBTrACS estimate. Note that no IBTrACS track data are available for cyclone Fani at the time of writing.

Fig. 4
Differences in maximum wind speed intensity (left) and timing of maximum (right) for IBTrACS US (blue) ND (orange) and ERA5 (green) relative to RAL2 ensemble members, ordered by magnitude of the intensity difference. Comparisons are made only within the period of RAL2 data, up to 36 hours pre and post landfall. Differences are calculated relative to RAL2 maximum, such that a positive intensity (time) difference indicates that the RAL2 model is faster (ahead) of the respective ERA5 or IBTrACS data. IBTrACS data is resampled by forward padding data to hourly intervals to aid comparison with RAL2. Where there are joint maxima in the IBTrACS data over multiple timesteps, we plot the smallest differences. Individual model differences are shown by coloured circles, with median difference per storm are show by coloured bars. Lower boxplots aggregate differences across all storms, with the 50 th percentile marked by the black bar and whiskers extending to the 5 th and 9 th percentiles of the data. and ERA5 (green) relative to RAL2 ensemble members, ordered by magnitude of the MSLP intensity difference. Details as for Fig. 4. A negative (positive) difference in MSLP indicates that the RAL2 MSLP minima are deeper (shallower) than the respective ERA5 or IBTrACS data. Note that IBTrACS ND MSLP data was not available for Fani or BulBul, and US MSLP not available for BOB01, BOB07 and TC01B, at the time of writing.
www.nature.com/scientificdata www.nature.com/scientificdata/ Differences between 1.5 km and 4.4 km model output. We don't explicitly validate the 1.5 km data but summarise differences between the distributions of maximum gust speed and minimum MSLP on a quantile basis, in relation to the 4.4 km data (Fig. 8). In order to facilitate a fair comparison, we compare identical spatial domains roughly equivalent to the 1.5 km model domain (see Fig. 2), but with a reduced northern extent to exclude as much mountainous terrain as possible, whilst encompassing the full geographic extent of Bangladesh.
Differences in maximum gust speed footprints, for the 1 st to 80 th percentiles, of the 1.5 km data are order 1 kn faster than the 4.4 km data. In all cases these differences are sufficiently robust that the 90% HDI of the differences amongst storms does not overlap zero ([0.3, 1.7] kn; [0.14, 0.86] m/s). For the very highest gust speeds (90 th , 95 th and 99 th percentiles of the 1.5 km data) the differences with the 4.4 km data shows much greater variability. The 90% HDI does overlap zero, with extremes of the quantile differences ranging from −2.4 kn to 1 kn ([−1.22, 0.50] m/s). Compared to lower percentiles, there are comparatively less data in the extreme upper percentiles, so the large range in this case is expected. Given the relatively robust speed increase seen in the 1.5 km data, compared to the 4.4 km data, for lower percentiles, we suspect that the minimal difference seen in the upper extreme percentiles results from under sampling rather than a systematic difference. Although we might expect the speed increase in the 1.5 km data to be consistent across all percentiles given better sampling, we cannot draw this conclusion based on these 12 storms alone.
For minimum MSLP footprints, the 1 st and 5 th percentiles of the 1.5 km data are [50, 87] hPa and [10,37] hPa shallower respectively (90% HDI), but note that the equivalent under sampling observed for high percentiles of gust speeds is likely to be prevalent in the low percentiles of MSLP. All other percentiles do not show any robust differences -the 90% HDI ranges [−11, 12] hPa. We do not feel these results show robust evidence for a systematic difference in MSLP between the 1.5 km and 4.4 km data.
The percentile differences suggest that the environmental MSLP (i.e. high percentiles) on the edge of the cyclone are similar in both the 1.5 km and 4.4 km simulations. Given the relationship between central pressure deficit (i.e. the difference between the tropical cyclone central pressure and the environmental pressure outside the tropical cyclone), peak wind speed and tropical cyclone size 36 , this comparisons suggests that 1.5 km storms may also be smaller in size than the 4.4 km storms. This result is commonly cited in analyses of general circulation models [37][38][39] and reanalysis data 40,41 .
In general, the substantial increase in computing effort required to generate 1.5 km gust speed simulations, over and above the 4.4 km simulations, is probably not merited for most catastrophe risk applications given the nature of the parametrisation. Comparing other parameters, such as vorticity or maximum wind speed, may yield different conclusions, and this would be an interesting direction for further research.
Other notable results. There is a semi-diurnal sea level pressure oscillation which occurs in the days preceding the minimum in MSLP. This oscillation is particularly noticeable in the ERA5 dataset for storms Aila, Bulbul, Rashmi, Roanu, Sidr and Viyaru, and to a lesser extent in RAL2 cyclones Akash, Mora, Rashmi, Roanu and TC01B (see Appendix A). The IBTrACS data does not capture this oscillation, probably due to the limited time sampling. This may be a manifestation of the diurnal radiation cycle as noted by Tang & Zhang 42 , Dunion et al. 43,44 and Knaff et al. 45 , amongst others. From simulation studies, Tang & Zhang 42 in particular note that the absence of a diurnal cycle (principally night time cooling) fails to trigger convection outside the cyclone inner core. Night-time cooling and associated destabilization typically enhance the primary storm vortex, eventually promoting the development of outer rain bands and increasing the size of the storm. Where this process is not evident in model simulations, it could diagnose simulations that have not correctly simulated the cyclogenesis Fig. 6 Differences in maximum 3-second gust speed (left) and timing of maximum gust speed (right) for ERA5 relative to RAL2 ensemble members, ordered by the magnitude of the gust speed difference. Details as for Fig. 4. A positive (negative) difference in gust speed indicates that the RAL2 gust speed maximum is faster (slower) than ERA5 data. Note that at the time of writing gust speed data was not available from IBTrACS for any of these events.
www.nature.com/scientificdata www.nature.com/scientificdata/ stage and are therefore likely to underestimate cyclone intensity. In our case, for most RAL2 simulations, as shown in the Supplementary Information Figures A1-A11, we do not allocate enough simulation time to the tropical cyclone pre-landfall to make this assessment (for computational efficiency reasons) and we have trimmed the spin-up period from the plots. Assessment of future tropical cyclone simulations could benefit from earlier initialisation times to investigate this further.
It is worth emphasising that the RAL2 model wind speed typically compare more favourably with IBTrACS wind speed data than to ERA5 wind speed. Based on the evaluation of these 12 events, tropical cyclone hazards in the ERA5 deterministic output may underestimate wind and gust intensity, and MSLP depth for tropical cyclones. For some specific cases, despite the ERA5 representation of Fani and Bulbul being less intense compared to the IBTrACS estimates, our RAL2 ensemble has sufficient model freedom (over a 24 hour spin-up period) to develop the ERA5 initial conditions into peak gust and minimum MSLP intensities that have much greater agreement with the IBTrACS data than the ERA5 data. This adds credibility to the spread of the RAL2 model ensembles: where there is substantial RAL2 ensemble spread (e.g. Viyaru or Mora) we suggest this reflects greater atmospheric variability associated with these events, such that the RAL2 ensemble might producing a wider range of counterfactual storm outcomes than would otherwise be seen in the driving reanalysis. Comparing these event ensembles with the ERA5 ensemble spread would be an interesting avenue of future work. www.nature.com/scientificdata www.nature.com/scientificdata/ Future work. Further work comparing the spread of RAL2 ensembles with the ERA5 uncertainty information would contextualise the range of variability that is introduced by the RAL2 model ensemble configuration. Additionally, work is needed to identify landfall times based on the RAL2 tracks. Future downscaling simulations would benefit from outputting variables required for tracking at hourly intervals, to facilitate hourly storm tracking.

Usage Notes
RAL2 time methods. Time methods are defined by the sampling period of the data and the sampling type applied to this period. The sampling period (or sampling interval) is one of: hourly (T1H), 3-hourly (T3H) or 24-hourly (T24H). The sampling type is one of max (maximum), min (minimum), mean or point. Point sampling is an instantaneous sample taken from the model time-step (which is typically much less than the sample period). Together then, T1Hmax is interpreted as hourly maximum data; T3Hmean is interpreted a 3-hourly mean data, and T1Hpoint are model instantaneous time-step output taken every hour.
In addition to timeseries data, we produce time-aggregated data for each ensemble member. Referred to as event 'footprints' , variables are aggregated by minima or maxima over the entire time period. These are commonly used within the catastrophe modelling industry.

RAL2 File naming.
Model time-series files are named according to the following convention: .
. . . . . VAR TIMEMETHOD UMRA2T TIMEPERIOD NAME RES nc where: VAR is a short variable identifier of the variable contained within the netCDF file; TIMEMETHOD is the time method, specifying if the var is a mean, min, max or point and the period of time over which the mean, min, max or point measure is found (as described above); UMRA2T is an identifier for the Met Office regional model type; TIMEPERIOD is the time period that the data spans, in the form START_END formatted as YYYYMMDD; NAME is the common name of the storm for the given time period; RES is the resolution of the dataset, either 4p4km = 4.4 km or 1p5km = 1.5 km grid size.
Files relating to ensemble footprints have a simpler file naming structure: . . . . . fpens VAR TIMEMETHOD NAME RES nc

Code availability
The Met Office Unified Model is available for use under licence. Several research organisations and national meteorological services use the UM in collaboration with the Met Office to undertake basic atmospheric process research, produce forecasts, develop the UM code, and build and evaluate Earth system models. For further information on how to apply for a licence, see https://www.metoffice.gov.uk/research/approach/collaboration/ unified-model/partnership Python and R code used to process the RAL2 data is available from Zenodo 46 .

Fig. 8
Percentile differences between 1.5 km and 4.4 km tropical cyclone data for (a) maximum gust speed and (b) minimum mean sea-level pressure (MSLP) footprints. Differences between resolutions are assessed on a quantile basis, in a hierarchical manner to account for dependence between storm ensemble members sampled from multiple storms. Quantile median estimates are shown by black circles, with 95% highest density intervals (HDI) shown by black bars. Where the 95% HDI overlaps 0, the median circles are filled white. The bootstrapped difference distribution (n = 1000) at each quantile is shaded turquoise (gust speeds) and orange (MSLP).