The group I experiments cover the years 1601–1849 with pre-industrial CRFs (“picontrol”) and fixed 1850 direct human forcings (“1850soc”) described in the grey column 3 of the ISIMIP3b scenario Table 2 as well as the subsequent years 1850–2014 considering pre-industrial and historical CRF (“picontrol” or “historical”) and different assumptions about DHF (“histsoc”, “2015soc”, “1850soc”, and “nat”) as described in the grey column 4 of Table 2. The reasoning behind the individual experiments are introduced below.

To estimate the impacts of historical and future changes in the CRFs, the protocol includes reference simulations based on pre-industrial CRFs and DHF identical to those considered in the climate change scenario runs. These reference simulations represent large samples (at least 250 years) of impact distributions based on stable pre-industrial CRFs (picontrol) and constant DHFs (see “picontrol + 1850soc”, “picontrol + 2015soc”, and “picontrol + nat” experiments in Table 2). Compared to the often used much shorter historical reference periods, this allows for a more robust fitting of extreme value distributions such as Gumbel distributions to simulation of e.g. annual maximum discharge to estimate reference 100 year return levels.

In addition, the protocol includes a reference experiment for the historical period (1850–2014) with DHF changing over time (histsoc) and 1850–2014 pre-industrial CRF (picontrol), while fixed 2015 DHF is considered afterwards (2015–2100) (“picontrol + 2015soc-from-histsoc”). This run may be different from the “picontrol + 2015soc” simulation for this time window because of the lagged effects of increasing DHF from 1850 to 2014. The “histsoc” DHF is identical to ISIMIP3a (Frieler et al., 2024).

The complete pre-industrial reference runs are divided in two parts. Only the first parts from the start until 2014 belong to group I (grey fields in Table 2), while the second parts covering the period 2015–2100 belong to group II (red parts of Table 2).

Comparing these reference simulations to the scenario experiments using historical CRFs (historical + histsoc, historical + 2015soc, historical + 1850soc, historical + nat; default (see below)) allows for the estimation of the effects of simulated historical climate change conditional on the assumed DHF. The historical CRF (“historical”) starts from the pre-industrial climate simulation in 1850, i.e. the “picontrol” and “historical” versions of the experiments have a common starting point. As some impact indicators may have “internal” trends not necessarily forced by external drivers (e.g. re-growth of forests), the comparison of the 1850–2014 impact simulations forced by the “historical” CRF to parallel simulations using the “picontrol” CRF is more appropriate to estimate the effects of historical climate change than comparing an early period of the historical impact simulation to the end of the historical period. The comparison of the simulations assuming “historical” CRF with the reference simulations based on “picontrol” CRF does not allow for a separation of the impacts of the natural and anthropogenic historical climate forcing. To allow for a quantification of the effects of the anthropogenic CRFs, we also support historical reference simulation accounting for the natural CRF only (“hist-nat” simulations generated within the Detection and Attribution Model Intercomparison Project (DAMIP) as sub-MIP of CMIP6, (Gillett et al., 2016) While this set-up is not an official part of the ISIMIP3b protocol, we provide the associated bias-adjusted CRF as secondary climate input data (Lange et al., 2023).

For models requiring a spin-up, the “picontrol” CRFs should be used in combinations with DHF (i) at 1850 levels to spin-up for the “1850soc” and “histsoc” experiments, (ii) at 2015 levels to initialise the “2015soc” experiment, and (iii) set to zero to start the “nat” experiments. For the spin-up all years of the “picontrol” CRF should be copied as often as needed. The “picontrol + 1850soc” run from 1601–1849 is part of the regular experiments that should be reported and hence the spin-up has to be finished before this pre-industrial period.

This experiment covering the historical period (1850–2014) is determined by historical (“historical”) CRFs and DHFs evolving according to observations (ISIMIP3a “histsoc” DHF). The ISIMIP3b “historical + histsoc” experiment is comparable to the default “obsclim + histsoc” experiment considered within ISIMIP3a but based on simulated CRFs. The simulated climate is different from the observed realisation due to differences in the internal variability of the observed and simulated historical climate and potential deficits in the climate model simulations or the observational data. A comparison between the default ISIMIP3b “historical + histsoc” impact model simulations to the associated ISIMIP3a results allows for a quantification of the effects of the discrepancies between the observed and simulated CRFs on the considered impact indicators. These simulations can be initialised from the spin-up of the associated pre-industrial reference simulations in case a spin-up is needed.

This historical experiment is similar to the standard historical experiment except that it is forced by fixed 2015 DHF. It is introduced into the “first priority” scenario-set-up to generate an ensemble of historical cross-sectorally consistent impact simulations that is as large as possible by not excluding impact models that are not able to handle varying DHF. If a spin-up is needed the associated simulations can be initialised from the spin-up of the associated pre-industrial reference simulation (picontrol + 2015soc, default) described at the beginning of this section.

This historical experiment is also similar to the standard historical experiment but it is forced by the fixed 1850 DHFs. It corresponds to the “obsclim + 1901soc” experiment of ISIMIP3a. Here in ISIMIP3b we consider the year 1850 instead of 1901 used in ISIMIP3a as this is the year where the “historical” climate simulations with observed natural and human forcings start, i.e. a branch from the pre-industrial climate simulations assuming constant pre-industrial forcings (“picontrol”). The “historical + 1850soc, default” impact model simulations allow for the quantification of the “pure effect of climate change” over the historical period. In contrast, the comparison of the “historical + 1850soc, default” impacts simulations to the “historical + histsoc” simulations allows for the quantification of the “pure effect of historical changes in DHF”. If a spin-up is required it does not have to be newly generated as it is identical to the spin-up for the default “picontrol + 1850soc”, “picontrol + histsoc”, and “historical + histsoc” simulations and described in the beginning of this section.

Considering no DHF (nature run) allows quantifying the effect of the simulated historical climate change conditional on otherwise natural conditions, i.e. no direct human influences on land use, water management etc.. This experiment is introduced as a companion experiment to the “obsclim + nat” experiment of ISIMIP3a. The comparison with the three historical simulations with constant DHF allows for testing to what degree the impact of climate change on the simulated natural or human systems is conditional on the underlying DHF. This experiment is only included for the biomes and fisheries and marine ecosystems fisheries sectors as models from other sectors usually need some basic information such as vegetation patterns that are not available for natural-only conditions. The biomes models generate their own natural-only vegetation patterns based on their dynamic representation of vegetation. A spin-up does not have to be newly generated but is identical to the spin-up for the “picontrol + nat” experiment described above.

So far, the default oceanic forcing is not bias-adjusted as globally the observational data are too sparse to be used in a similar empirical way as for the bias-adjustment of the atmospheric forcing. However, the biases in the forcing are expected to also induce biases in the historical and future impact simulations. To quantify these effects and to test a suggested bias-adjustment method based on comprehensive ocean-biogeochemistry model simulations forced by bias-adjusted atmospheric forcings we include a sensitivity experiment where the default CRF is replaced by input data generated by a dynamical de-biasing approach (Lengaigne et al., 2025) using the NEMO-PISCES physical-biogeochemical ocean model (Madec, 2015), which is the oceanic component of the IPSL-CM6A-LR climate model. Thus, the forcing data will first be generated for IPSL-CM6A-LR, but later extended to other ISIMIP-GCMs as described in Sect. 3.4.2.

In contrast, the oceanic forcing for the regional component of the marine ecosystems and fisheries sector have been bias-adjusted by regional observational oceanic data as described in Sect. 3.4.3. In this case most models only use the bias-adjusted inputs and not the raw ones. Nevertheless the experiments are labeled as “de-biased” sensitivity experiments to ensure a consistent naming across scales.

The ISIMIP3b, group II simulations comprise a set of future impact projections (2015–2100) using fixed levels of DHF as considered in the historical simulations (“2015soc”, “1850soc”, and “nat”) or reached at the end (2014) of the historical period in the “historical + histsoc” runs (“2015soc-from-histsoc”). These runs are described in the red cells of Table 2.

These simulations are included into the ISIMIP3, group II part of the protocol to allow for the estimation of the effect of climate change by comparing the future impact projections to simulations assuming the same background DHF but pre-industrial levels of CRF (see description of baseline simulations in Sect. 2.1).

This experiment represents an expansion of the group I “historical + histsoc” experiment based on fixed 2015 DHF for the future. Note that this experiment is different from the experiment with fixed 2015 DHF for the future where the associated impact simulations start from the “historical + 2015soc” group I simulations (see description below).

These experiments have been introduced to describe the impacts of different scenarios of changes in the climate-related systems on today's natural systems and societies, i.e. assuming present day population levels and distributions, land use patterns, water, and agricultural management measures etc. In many cases, the projected changes in natural and human systems can be interpreted as the pure effect of the prescribed changes in the climate-related systems. However, they could also partly result from lagged effects of the historical changes in DHFs (“histsoc”), CRF (“historical”), or natural temporal trends induced e.g. by re-growth of forests. To be able to separate natural trends from the effects of changing CRFs, the associated simulations can be compared to reference impact simulations with pre-industrial CRF forced with the same DHF (“picontrol + 2015soc-from-histsoc”, see description in group I section).

These experiments expand the “historical + 2015soc” experiments from ISIMIP3b, group I using DHFs that are held constant at 2015 levels for the historical period into the future. Although the DHF in the future period is identical to the future simulations described above, the difference in historical forcing may affect the impact simulations in the future period. These simulations are also considered first priority as some of the impact models may not be able to handle varying DHF and therefore can only perform these experiments. Models participating in the “2015soc-from-histsoc” experiments described above are also asked to complete the “2015soc” runs to generate an ensemble of consistent impact model simulations with as many members as possible.

These experiments continue the default “historical + 1850soc” experiments considered in ISIMIP3b, group I. They are included to estimate the impact of changes in the climate-related systems conditional on 1850 levels of DHF that can be compared to the impact conditional on today's levels of DHF (“2015soc”).

These experiments continue the default “historical + nat” experiments in ISIMIP3b, group I. They are included to estimate the effect of changes in the climate-related systems (here climate change itself and increasing CO2 concentrations) assuming no DHF.

To separate the effects of increasing atmospheric CO2 concentrations (in particular the CO2 fertilisation or water use efficiency effects on vegetation) from the effects of other changes in the climate-related systems, the ISIMIP3b protocol includes sensitivity experiments where atmospheric CO2 concentrations are held constant at 2015 levels. For SSP1-2.6 and SSP3-7.0, they are only introduced as deviations from the default “2015soc-from-histsoc” experiments while for SSP5-8.5 the effect can also be quantified conditional on all levels of direct human influences considered in the previous experiments.

To study the effects of future changes in lightning flash rates as opposed to using a stationary lightning climatology, the ISIMIP3b protocol includes sensitivity experiments where future lightning flash rates change along the RCPs. The future lightning data sensitivity experiment is introduced as a deviation from the default “2015soc-from-histsoc” experiment and only for one climate model (UK-ESM). This sensitivity experiment has been introduced for the fire sector.

To study the effects of high atmospheric CO2 concentration without accompanying changes in climate, the ISIMIP3b protocol includes a sensitivity experiment where the atmospheric CO2 concentration are prescribed according to RCP8.5, while the other CRF, in particular the atmospheric forcings are from SSP1-2.6. The future climate sensitivity experiment is introduced as a deviation from the default “ssp126 + 2015soc-from-histsoc” experiment. This sensitivity experiment has been introduced for the peat sector.

These simulations represent the future extensions of the “de-biased” group I simulations described above. They are designed to test the dynamical bias-adjustment suggested for the global oceanic forcings under different levels of climate change (ssp126, ssp370, ssp585). The regional impact projections within the sector are also based on de-biased oceanic forcings and are therefore also labeled as “de-biased” sensitivity experiments to ensure a consistent labeling across scales.

For ISIMIP3b we provide the daily atmospheric forcings for the same variables as in ISIMIP3a on the default 0.5° grid (see Table 3). These variables are from the output of CMIP6 climate model simulations, selected and processed as described below. We use the climate simulations from the picontrol (for pre-industrial conditions), historical (for historical conditions), ssp126, ssp370, and ssp585 (for future conditions under the scenarios SSP1-2.6, SSP3-7.0, and SSP5-8.5, respectively) CMIP6 experiments.

For the pre-industrial conditions, 500 years of picontrol output data are used and harmonised across GCMs with respect to the time range they cover. This is possible because picontrol data only carry nominal year labels without historical reference points (see original nominal start and end years in Table 4). We changed the GCM-specific picontrol time ranges listed to 1601–2100.For the historical and future climate conditions, we provide input data for 1850–2014 and 2015–2100, respectively, in line with the time ranges covered by the corresponding CMIP6 experiments. The common time axis is important as the use of the input data should be harmonised across all sectors. In particular, the year-by-year combination of the pre-industrial CRF with the historical DHF should be done in the same way across all sectors and models.

To limit the number of mandatory impact simulations and hence lower the barrier to participation in ISIMIP3b, we provide climate input data for only five selected CMIP6 climate models. The basic characteristics of the five GCMs are listed in Table 4. The models were selected based on data availability at the selection time (late 2019 to early 2020), performance in the historical period, structural independence, process representation and equilibrium climate sensitivity (ECS).

To be included in ISIMIP3b, a GCM had to provide daily data for all variables listed in Table 3 except for near-surface specific humidity (huss) (which was derived from near-surface relative humidity (hurs), surface air pressure (ps) and near-surface air temperature (tas), see below), ps if sea level pressure (psl) was available, so a proxy for ps could be computed based on psl and tas, and near-surface wind speed (sfcwind) if zonal and meridional near-surface wind components (uas, vas) were available, so a proxy for sfcwind could be computed based on uas and vas. Those daily data had to cover 500 picontrol years and all years of the historical, SSP1-2.6, SSP3-7.0, and SSP5-8.5. In addition, we favoured GCMs that provided the additional input data needed for the tropical cyclone modelling (Table 5) and the fisheries and marine ecosystems sector (FishMIP; Table 10).

According to a skill analysis (see Fig. 2), the GCMs ACCESS-CM2, AWI-CM-1-1-MR, CESM2, CESM2-WACCM, CMCC-ESM2, EC-Earth3-AerChem, GFDL-CM4, GFDL-ESM4, HadGEM3-GC31-LL, HadGEM3-GC31-MM, MPI-ESM1-2-HR, MPI-ESM1-2-LR, MRI-ESM2-0, NorESM2-MM, SAM0-UNICON, TaiESM1, and UKESM1-0-LL perform relatively well in reproducing the main historically observed characteristics of the atmosphere. From that list, only GFDL-ESM4, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL provided all required daily data at the time of model selection. Another model that fulfilled all those data requirements and shows an average performance in the historical period is IPSL-CM6A-LR. These five GCMs were selected to be used in ISIMIP3b. With the exception of GFDL-ESM4, these models also provide the data needed for tropical cyclone modelling. GFDL-ESM4 is the model providing the most comprehensive oceanic bio-geochemical forcings for FishMIP while other models cover less and partly other oceanic variables (see Table 8). Three of the climate models (GFDL-ESM4, IPSL-CM6A-LR, UKESM1-0-LL) are successors of models already used in ISIMP2b and in the ISIMIP Fast Track.

The five GCMs are structurally independent in terms of their ocean and atmosphere model components. Furthermore, all of them have a coupled climate and carbon cycle and in some cases, fully interactive chemistry and aerosol components. We favoured models that applied prognostic couplings between processes and model domains wherever possible to maximise the coverage of simulated feedbacks.

The five GCMs provide a good representation of both the mean and the range of the full CMIP6 multi-model ensemble ECS. According to Meehl et al. (2020), the CMIP6 multi-model mean ECS is 3.7 °C, which is precisely met by the mean ECS of the five ISIMIP3b GCMs. The transient climate response (TCR) of 2.0 °C is also precisely met. This provides an improvement over ISIMIP2b, in the sense of the selected GCM subset reflecting the statistics of the larger CMIP ensemble. In ISIMIP2b the mean ECS for the full CMIP5 was 3.2 °C compared with a mean ECS of 3.72 °C for the four ISMIP2b GCMs (see Tables S1 and S2 in Jägermeyr et al. (2021). The ISIMIP3b ensemble includes three models with below-average ECS (GFDL-ESM4, MPI-ESM1-2-HR, MRI-ESM2-0) and two models with above-average ECS (IPSL-CM6A-LR, UKESM1-0-LL) (see Table 4). In line with their ECS values, we find GFDL-ESM4 and UKESM1-0-LL to project the weakest and strongest global warming, respectively, under any future scenario considered (see Fig. 3). Under SSP5-8.5, the global mean near-surface temperature in 2100 is about 3 °C larger in UKESM1-0-LL than in GFDL-ESM4. Under SSP1-2.6, the projections are about 1.5 °C apart. The ensemble mean warming of the ISIMIP3b CMIP6 models is significantly higher than the warming of the ISIMIP2b CMIP5 models, across global land area by an average of 0.3 °C, but over the main breadbasket cropland regions by more than 0.5 °C between 1983–2013 and 2069–2099, under both SSP1-2.6 and SSP5-8.5 (Table S1 in Jägermeyr et al., 2021). This is in line with the higher median ECS in CMIP6 compared to CMIP5; indeed, some CMIP6 models have an ECS above the assessed likely (2.5 to 4 °C) and very likely (2 to 5 °C) ranges in the IPCC's sixth assessment report (AR6) (Forster et al., 2021). The reasons for these higher estimates of ECS are complex, with cloud feedback processes playing an important role (Zelinka et al., 2020). While the plausibility of the very high ECS estimates has been questioned, recent studies indicate CMIP6 models with high ECS tend to simulate cloud properties better than low ECS models (Bock and Lauer, 2024); also, unaccounted natural variability may have biased the IPCC's assessed ranges somewhat low (Liang et al., 2024; Watanabe et al., 2024).

The ISIMIP3b ensemble reflects the spread in ECS of the overall CMIP6 ensemble, with two models above the AR6 likely range and one of these (UKESM1-0-LL) above the very likely range. The strong warming response of these models should be kept in mind when conducting ISIMIP3b-based impacts studies. However, depending on the region and variable of interest, the high ECS does not necessarily have any bearing on the magnitude or realism of projected regional impacts, and any further selection of models should not be based solely on ECS but on the models' suitability for the impacts variables in question (Swaminathan et al., 2024). In many applications, results can be harmonized by describing the simulated impacts in terms of global mean temperature changes instead of time for the different emission scenarios.

To make the GCM-based climate forcing usable for the impact modellers we apply a bias adjustment ensuring that the GCM simulations match the observed distribution of climate data over the historical reference period (1979–2014). In addition to the bias adjustment a statistical downscaling to our standard 0.5° grid is included in the pre-processing of the surface and near-surface atmospheric variables (see Table 3). The method used for the bias adjustment and statistical downscaling (BASD) in ISIMIP3b is version 2.5 of ISIMIP3BASD (Lange, 2019b, 2021a).

ISIMP3BASD has several advantages compared to the method used for bias adjustment and statistical downscaling in ISIMIP2b (Frieler et al., 2017; Lange, 2017, 2018). First, it clearly separates the adjustment of biases in climate model output at 1 or 2° resolution, whatever is closest to the original output data, from the statistical downscaling to the target resolution of 0.5°. Compared to ISIMIP2b, where climate model output was first spatially interpolated to the target resolution and then bias-adjusted, the new approach avoids the associated underestimation of the spatial variability at the target resolution (Lange, 2019b). Second, the new quantile mapping method preserves trends in each quantile of the distribution of the daily data and adjusts biases in distribution quantiles of the daily data more accurately than the ISIMIP2b bias adjustment methods (Lange, 2019b).

For trend preservation, we first produce future pseudo-observations by shifting the historically observed daily data by the simulated future climate change. Here, the signal of climate change is the difference or the ratio between the inverse empirical cumulative distribution function of the historical period and the respective distribution functions of each 36-year period of the future. Using the difference ensures additive trend preservation and using the ratio ensures multiplicative trend preservation under bias adjustment. We apply additive trend preservation for near-surface air temperature (tas), sea level pressure (psl, see Table 3), and surface downwelling longwave radiation (rlds). We apply primarily multiplicative trend preservation for precipitation including snowfall (pr), near-surface wind speed (sfcWind), and the range (tasrange = tasmax - tasmin) between the daily maximum and minimum near-surface air temperatures (tasmax and tasmin, respectively) that can transition smoothly to additive trend preservation for data with large negative biases in the historical period (Lange, 2019b). In a second step, the future simulations are mapped onto the future pseudo-observations by quantile mapping. Both steps, the generation of the future pseudo-observations and the quantile mapping of the future simulations onto the pseudo-observations, are applied for each day of the year separately. The distributions include data from the 31 d around the considered day and all years of the reference or future period, respectively. This means a sample size of 31×36 values for each day of the year. Through this approach the bias adjustment implicitly also adjusts the multi-year mean annual cycle and a mix of year-to-year and day-to-day variability (Haerter et al., 2011).

In addition, the method adjusts the frequency of daily data falling outside of the inner bounds specified in Table 3 (e.g. the dry day frequency, i.e. the number of days with precipitation below 0.1 mm d⁻¹).

Four variables were adjusted and downscaled indirectly: near-surface specific humidity (huss) was derived from adjusted and downscaled near-surface relative humidity (hurs), surface air pressure (ps), and near-surface air temperature (tas) using the equations of Buck (1981) as described in Weedon et al. (2010), snowfall (prsn) was derived from adjusted and downscaled precipitation including snow (pr) and the snowfall ratio (prsnratio = prsn / pr), and daily maximum and daily minimum near-surface air temperatures (tasmax and tasmin, respectively) were derived from adjusted and downscaled tas, and the tasrange = tasmax - tasmin and skewness of the daily temperature cycle tasskew = (tas - tasmin) / (tasmax - tasmin).

The basic characteristics of ISIMIP3BASD (version 1.0) are described in Lange (2019b). However, the method finally used to generate the forcing data now provided within ISIMIP3b (ISIMIP3BASD version 2.5) deviates from the original version in some aspects. In the following we describe the most important updates of the procedure relative to the one described in Lange (2019b). For a complete list of differences between the two versions of the BASD method and the full history of which feature was added in which update, see the CHANGELOG included in the archive of code version 2.5 (Lange, 2021a).

In Lange (2019b) the bias-adjustment was applied on a monthly basis, i.e. the future pseudo-observations and the quantile mapping described above was applied to all daily January data, February data and so forth. This approach can introduce discontinuities at the transition from one month to another (see Fig. 4). That is why for ISIMIP3b the adjustment is done in the running window mode with steps of 1 d and a window width of 31 d as described above. This approach resolves the discontinuity issue (see Fig. 4), as suggested by Themeßl et al. (2012) Thrasher et al. (2012) Gennaretti et al. (2015) and Grenier (2018).

Since ps, rlds and tas can show significant trends within the 36-year training and application periods ISIMIP3BASD v1.0 includes a detrending of these variables within these intervals before the pseudo-future observations and the transfer functions are estimated and applied. Afterwards the trend is added back again. This is done to prevent the confusion of trends with interannual variability during quantile mapping (Lange, 2019b; Maraun, 2013). In contrast to v1.0, in v2.5, applied to generate the ISIMIP3b forcing data, the detrending is only applied if the trend is significantly different from zero at the 5 % level.

We also changed the method used to generate future pseudo-observations of bounded variables (Eqs. 8 and 9 of Lange, 2019b), in order to stabilise results in some edge cases. If, e.g., the historically observed relative dry-day frequency was 0.0 while the simulated frequency was 0.8 for the historical period and 0.9 for some future period, then, according to Eq. (9) of Lange (2019b), the future pseudo-observed frequency would be equal to 1-(1-0.0)(1-0.9)/(1-0.8)=0.5. As this is considered unrealistic we apply a revised version of Eq. (9) of (Lange, 2019b) that reads 1Pfutobs=Pfutsim if Phistsim=Phistobs,0+(Phistobs-0)(Pfutsim-0)/(Phistsim-0)if Pfutsim≤Phistsim>Phistobs,1-(1-Phistobs)(1-Pfutsim)/(1-Phistsim)if Pfutsim≥Phistsim<Phistobs,Phistobs+Pfutsim-Phistsim otherwise. In this revised relation, the otherwise case applies if Pfutsim<Phistsim<Phistobs or Pfutsim>Phistsim>Phistobs. Hence it applies to the aforementioned edge case, where it produces a less extreme future pseudo-observed relative frequency of 0.0+0.9-0.8=0.1. Equation (8) of Lange (2019b) was revised analogously to Eq. (9).

Furthermore, we refined the method used to generate future pseudo-observations (step 5 of the bias adjustment algorithm of Lange, 2019b) for all variables with at least one bound: In v1.0, the future pseudo-observations were generated by transferring simulated trends in all distribution quantiles to the observational reference data. That included trends in, e.g., precipitation quantiles below the wet-day threshold. However, in some cases, the trend transfer turned many dry days into wet days, with a profound impact on the shape of the distribution of future pseudo-observed wet-day precipitation. As a result, simulated trends in wet-day precipitation intensity were not well preserved. In v2.5, trend transfers are restricted to values within threshold. This particularly improves the preservation of trends in wet-day precipitation intensities.

We also modified the bias adjustment method for Near-Surface Relative Humidity (hurs) because ISIMIP3BASD v1.0 turned out to produce unrealistic distributions of hurs under climate change if there are too many cases of supersaturation (hurs ≥100 %) in the simulated data. This is the case for several of the CMIP6 GCMs selected for ISIMIP3b, particularly in high-latitude winter: While no supersaturations are found in the observational reference data, the GCM simulates many supersaturations in the historical reference period and even more so in a future period, under SSP5-8.5 (see Fig. 5). ISIMIP3BASD v1.0 preserves this projected trend and hence produces future bias-adjusted hurs data with many supersaturations. In v2.5, this trend is no longer preserved. Instead, the supersaturation probability is fixed at the observed level, which is zero or very close to zero in all seasons and grid cells for W5E5. Future pseudo-observations of hurs are generated by applying the revised (see above) Eq. (8) of Lange (2019b) to all hurs values after capping them at 100 %. The new approach was motivated by findings from Ruosteenoja et al. (2017, 2018). They analysed hurs data from CMIP5 and showed that (i) supersaturations in those data are mostly spurious, resulting from, e.g., inconsistencies in the interpolation of temperature and specific humidity to the near-surface level, and (ii) climatological mean value trends of hurs become more consistent with trends in relative humidity from the lowest model level if hurs is capped at 100 % before trends are calculated.

In addition, while ISIMIP3BASD v1.0 applies parametric quantile mapping to all climate variables, we used a nonparametric approach for the bias adjustment of near-surface relative humidity (hurs), the snowfall ratio (prsnratio), surface downwelling shortwave radiation (rsds), and the skewness of the daily temperature (tasskew) since the parametric quantile mapping method previously used for those variables suffered from occasionally unstable beta distribution fits.

Moreover, the parametric quantile mapping described in Lange (2019b) does not only adjust biases in quantiles of the simulated daily data but also adjusts biases in the likelihood of individual events, as in Switanek et al. (2017). To avoid overfitting artifacts we did not adjust event likelihoods for ISIMIP3b.

Finally, the diurnal temperature range (tasrange) was ultimately bias-adjusted using a Weibull distribution, not a Rice distribution as described in Lange (2019b) because the Weibull distribution fits the data better in most cases, in particular in the upper tail.

For further details of the application of ISIMIP3BASD v2.5 for ISIMIP3b, including the exact Python commands and application periods used per CMIP6 experiment, see the ISIMIP3b bias adjustment fact sheet (Lange, 2021b).

In addition, we use a new observational target dataset. Instead of using the EWEMBI dataset (E2OBS, WFDEI and ERAI data merged and bias-corrected for ISIMIP; (Lange, 2019a)) in ISIMIP3b we adjust the climate forcing data to version 2.0 of the W5E5 dataset (WFDE5 over land merged with ERA5 over the ocean; Lange et al., 2021). The data cover the entire globe at 0.5° horizontal and daily temporal resolution from 1979 to 2019. W5E5 v2.0 is derived by applying version 2.0 of the WATCH Forcing Data methodology (WFDE5; Cucchi et al., 2020) to ERA5 reanalysis data (Hersbach et al., 2020) and precipitation data from version 2.3 of the Global Precipitation Climatology Project (GPCP; Adler et al., 2003).

The statistical downscaling method did not change between v1.0 and v2.5 of ISIMIP3BASD, i.e. for ISIMIP3b we use the approach described by Lange (2019b). This method adds the spatiotemporal variability that is missing at the low spatial resolution at which the bias adjustment is done (1 or 2°, depending on the GCM), compared to the target resolution of the downscaling (0.5°). The method is a modified version of the MBCn algorithm from Cannon (2018), which in turn is a stochastic, multivariate, non-parametric quantile mapping method. We use it to transfer the statistical relationship between low-resolution and high-resolution W5E5 data to the GCM output that was previously bias-adjusted using low-resolution W5E5 data. In comparison to the approach used in ISIMIP2b (a spatial interpolation to the target resolution followed by a bias adjustment at that resolution), the approach used in ISIMIP3b is less prone to inflate temporal variability and deflate spatial variability, i.e. the ISIMIP3b approach produces more realistic spatiotemporal variability patterns at the target resolution (Lange, 2019b).

We provide large ensembles of potential realisations of TC tracks and intensities that are consistent with the large-scale atmospheric and oceanic conditions simulated by four of the five ISIMIP3b GCMs (see Table 6) and for a selection of scenarios considered in ISIMIP3b (see Table 1). The tracks are generated by two different statistical-dynamical approaches, the MIT approach and the CHAZ approach detailed below, forced by data from the ISIMIP3b GCMs listed in Table 6. For the MIT approach, we provide gridded wind (maximum 1 min sustained wind speeds during the whole duration of the TC) and rainfall (maximum 24-hourly amounts of rain during the whole duration of the TC) fields at a spatial resolution of 300 arcsec (approximately 10 km) using the same approaches also applied to the historically observed tracks (Frieler et al., 2024), Sect. 4.2).

Both methods to generate the TC tracks are assessed in Meiler et al. (2025). The modeling approaches consist of a genesis, a track, and an intensity module:

MIT approach. Within MIT (Emanuel et al., 2008), the time-evolving state of the atmosphere and ocean surface given by the GCMs is randomly (uniformly distributed in time and space) seeded by weak proto-cyclones (genesis module). The seed disturbances are assumed to move with the GCM-provided large-scale flow in which they are embedded, plus a westward and poleward component owing to planetary curvature and rotation (track module). Their intensity is calculated using the Coupled Hurricane Intensity Prediction System (CHIPS; Emanuel et al., 2004), a simple axisymmetric hurricane model coupled to a reduced upper ocean model to account for the effects of upper ocean mixing of cold water to the surface (intensity module). Applied to the synthetically generated tracks, this model simulates which of the seeded proto-cyclones develop into TCs, reaching maximum 1 min sustained wind speeds of at least 35 knots, or dissipate due to unfavourable environments. The probabilistic seeding of proto-cyclones is repeated until the desired number of storms per year is reached (in our case, 1500). For each year, the share of proto-cyclones that dissipated in the process is used to derive an estimate of annual TC occurrences (freqyear). Extensive comparisons to historical events (Emanuel et al., 2008) have revealed that the statistical properties of the simulated events are consistent with historical TC genesis.

For each year of the ISIMIP3b period 1850—2100 (except for GFDL-ESM4, where tracks were only generated for 1850–2014 and 2061–2100, and MRI-ESM2-0 for 1950–2100, see Table 1), 1500 tracks were generated, globally. Depending on the application, a simple subsampling (Meiler et al., 2022) or a more advanced bias-correction and emulation procedure (Geiger et al., 2021) might be necessary to extract properly-sized sets of potential realisations from the MIT ensembles.

The “ISIMIP3b tropical cyclone tracks MIT” dataset shall only be used for noncommercial purposes, including teaching and research at universities, colleges and other educational institutions, research at non-profit research institutions, and personal non-profit purposes. It is accessible through the ISIMIP data portal (https://data.isimip.org/access/ISIMIP3b/InputData/climate/tropical_cyclones/MIT/tracks/, last access: 20 December 2025) after agreeing to the corresponding license.

For using the tracks for commercial purposes, including but not restricted to consulting activities, software or data products, and a commercial entity participating in research projects, please contact Kerry Emanuel (MIT, email: emanuel@mit.edu) for an appropriate license.

CHAZ approach. CHAZ (Lee et al., 2018) seed disturbances are also initialised randomly, but, in contrast to the MIT model, the global seeding rate and the local probabilities are derived from two versions of a TC genesis index (TCGI, Tippett et al., 2011 (genesis module)) and intended to represent the environmental conditions instead of being adjusted to produce a prescribed number of TCs. It is noted that CHAZ's projection of global and basin-wide TC annual frequency is sensitive to the choice of the particular variable used to represent moisture in its genesis module. Simulations using column relative humidity (CRH) as the moisture variable tend to project an overall increase in global TC frequency, while those using saturation deficit (SD) show a decrease (Camargo et al., 2014; Lee et al., 2020). Both parameters describe how far the atmosphere is from saturation, and they have very similar spatial patterns in the present climate, so historical data cannot be used to determine which variable is the best choice to represent the climate. These two configurations reflect the uncertainty of TC frequency projections (Sobel et al., 2021). Here we provide CHAZ downscaling using both choices of moisture variable to account for this uncertainty.

Similar to MIT, CHAZ then moves the synthetic storms by advection of the environmental steering flow plus a beta drift (track module). The evolution of synthetic storm intensity is calculated using the surrounding atmospheric conditions through an empirical multiple linear regression model plus a stochastic component (intensity module, Lee et al., 2015, 2016). The stochastic component accounts for internal storm dynamics that do not depend explicitly on the environment. While, in MIT, TC occurrence frequency is provided as an additional variable, in CHAZ, this information is implicitly contained in the number of TCs that were seeded by the genesis module and that reached TC strength according to the intensity module.

For ISIMIP3b, 20 different CHAZ realisations of the genesis and subsequent tracks are generated with 40 ensemble members each from the intensity module for the historical period and for all RCP-SSP combinations considered within ISIMIP3b. For each of the 20 realisations, we compute wind and rain fields for the first ensemble member from the intensity ensemble. The design of 20 realisations allows CHAZ to generate similar numbers (∼1800) of synthetic storms per year per GCM as the MIT models over the historical period. The exact number of storms per year in CHAZ varies by GCM, by scenario, by the choice of humidity variables in CHAZ's genesis component (Lee et al., 2020). On average, CHAZ generates 1817, 1802, 1820, 1810, 1842 storms per year for GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL, respectively. The CHAZ model has been shown to capture the statistical properties of the observed storms when forced by global reanalysis data (Lee et al., 2018). Its CMIP6 downscaling results are reported in Fosu et al. (2024). Sobel et al. (2019) used both models to study cyclone risk for Mumbai, India and showed that MIT and CHAZ generate comparable return periods (frequency of exceedance) of maximum wind speeds at landfall. However, a frequency bias-correction might still be necessary, depending on the application (Meiler et al., 2022).

The “ISIMIP3b tropical cyclone tracks (CHAZ)” dataset shall only be used for noncommercial purposes, including teaching and research at universities, colleges and other educational institutions, research at non-profit research institutions, and personal non-profit purposes. It is accessible through the ISIMIP data portal (https://data.isimip.org/access/ISIMIP3b/InputData/climate/tropical_cyclones/CHAZ/tracks/, last access: 6 March 2026) after agreeing to the corresponding license.

For using the tracks for commercial purposes, including but not restricted to consulting activities, software or data products, and a commercial entity participating in research projects, please contact Chia-Ying Lee (Columbia University, email: cl3225@columbia.edu) for an appropriate license.

We do not yet provide coastal water levels as forcing data for ISIMIP3b. However, we plan to generate time series of coastal water levels from 1900 to 2100 at hourly resolution or for daily maxima. The data set and method will be described in a separate manuscript. Similar to the hourly water level dataset of ISIMIP3a (see Sect. 3.3 of Frieler et al., 2024, and Treu et al., 2024), we will combine long-term annual sea level change with estimates of short-term coastal water level variation. Concerning the long-term sea level change component, we go beyond the ISIMIP2b approach (Frieler et al., 2017) and use tide gauge, satellite, vertical land motion and global climate model data to constrain a model with observations and IPCC AR6 projections in a Bayesian setting building on Perrette and Mengel (2025). The Perrette and Mengel (2025) model allows for smooth projections of relative sea level from observational time series collected at tide gauge stations with an explicit representation of the different components of sea level rise. To become usable for ISIMIP3 we will (a) extend the approach to all coastlines and (b) use GCM output directly for the global thermosteric and local sterodynamic components (Gregory et al., 2019) of sea level rise, which are modeled by the GCMs. Extension to all coastlines is possible for the ice sheet, glacier and sterodynamic components as they rely on spatial fingerprints or GCM output. Processes driving vertical land motion that are not related to large scale climate processes are however more difficult to model. They are estimated from data as residual vertical land motion in Perrette and Mengel (2025). As we do not have data at all coastlines we will extrapolate in time and space the historical rates from tide gauge sites or apply zero rates for this component. Using the explicit component structure of the model, the method will provide relative sea level projections (including vertical land motion), which can be directly used in coastal impact studies. We plan to estimate the short-term coastal water level variation by a machine-learning approach that is trained to reproduce simulations of the Global Tide and Surge Model (GTSM) driven by ERA5 reanalysis data (Muis et al., 2020) or simulations from HighResMIP (Muis et al., 2023). We are currently testing the dependency of the short-term water level variation on available atmospheric information at GCM resolution. If the predictive power is high enough we will use the findings to provide computationally efficient water level projections specific for the ISIMIP GCMs.

In the default experiments, the ocean variables provided by the GCMs are not subject to bias-adjustment, unlike the atmospheric forcing data (Sect. 3.4.1). This is due to the absence of a comprehensive global observational oceanic dataset to serve as a reference for the adjustment.

However, in order to mitigate potential biases in global impact model simulations stemming from biases in raw oceanic forcing data, we plan to provide a de-biased version to be used in a sensitivity experiment (see Table 2). They will be derived from an ocean-biogeochemistry model forced by bias-adjusted monthly atmospheric surface flux data from four of the five ISIMIP3b GCMs. The approach preserves the monthly variability of the underlying GCM while the daily variability is added from an independent simulation (see Sect. 3.4.2).

For the regional impact model simulations, observational data for individual variables have either been applied directly (if the required forcing was observed) to rectify biases in regional oceanic forcings by the delta method or have first been translated into the required forcing variable by model simulations (see Sect. 3.4.3). The regional bias-adjustment is independent from the generation of the global de-biased forcing data.

In order to gauge the effects of these adjustments on the corresponding impact simulations, the protocol includes sensitivity experiments (“de-biased”) grounded on these adjusted CRF (see Table 2). The comparison of associated impact simulation to the default ones is expected to provide valuable insights into the effects of potential biases in the CRF. The “de-biased” experiments are considered a starting point to develop methods to bias-adjust the oceanic forcings in further ISIMIP simulation rounds and make these simulations the default ones. Following the ISIMIP “consistency framing” the bias-adjustment should also preserve the daily variability of the original GCM simulations to allow for a cross-sectoral integration on daily time scale.

For the “varlightning” sensitivity experiment we provide temporally varying lightning density (strokes km⁻² d⁻¹) for the period 2015–2100 on monthly resolution (monthly mean of daily lightning stroke density) and the standard 0.5° global grid. This dataset may be used in a range of applications, for example, to understand the influence of lightning on wildfire ignition or atmospheric composition.

The lightning density is derived from future climate simulations of UKESM1-0-LL and an empirical relationship between Convective Available Potential Energy (CAPE) and lightning strokes based on the WWLLN Global Lightning Climatology and time-series (WGLC) (Kaplan and Lau, 2021, 2022). Daily mean CAPE is calculated from non bias-adjusted air temperature, air pressure, and specific humidity on pressure levels from the surface to the top of the troposphere.

The relationship between daily CAPE and daily lightning is estimated by linear regression of log-transformed CAPE derived from the GCM-calculated CAPE during the period of overlapping model output and observed daily lightning from WGLC (2015–2020) for each gridcell and month of the year. Where <10 observations of daily lightning are available over the calibration period, we use global mean regression parameters.

The empirical relationships are applied to the daily CAPE data from the UKESM1-0-LL simulations for all three climate scenarios SSP1-2.6, SSP3-7.0, and SSP5-8.5. The associated lightning densities were monthly averaged. To maintain the spatial structure of lightning observed at present, lightning anomalies compared to the simulated 2015–2020 climatological reference were added to the observed present-day lightning climatology from WGLC for 2015–2020. The “varlightning” sensitivity experiment is assumed to start from the default historical group I simulation, assuming the Flash Rate Monthly Climatology (Cecil, 2006), not changing with climate change.