Disentangling the uncertainties in regional projections for Australia

2.1. Global CMIP6 models

CMIP6 provides global climate change projections of the Earth system from state-of-the-art GCMs forced with a range of natural and anthropogenic-influenced scenarios. In the present study, we analyse the first ensemble member of 35 GCMs from the CMIP6 archive with monthly data available for the required Shared Socioeconomic Pathways (SSPs; Meinshausen et al. 2020). The scenarios considered in the present study are the historical scenario, with realistic greenhouse gas and aerosol forcings from 1960 to 2014, as well as a low future emissions (SSP1-2.6) and high future emissions scenario, SSP3-7.0, spanning 2015–2100. Wherever a different ensemble member for a GCM was selected for downscaling, we also used that same ensemble member in our GCM sample for consistency. A full list of GCMs used in the present study is provided in Table 1.

In addition to the GCM simulations listed in Table 1, we also analysed two GCM large ensembles. The 40-member ACCESS-ESM1.5 large ensemble was used to provide context for the special case of a simulated much drier projection for Australia (ACCESS-ESM1.5 r6i1p1f1). The 58-member EC-Earth3 large ensemble was used to provide context for the special case of a simulated much wetter projection for Australia (EC-Earth3 ri1p1f1). The selection of these two GCM simulations as representative wetter and drier futures for Australia is described in Grose et al. (2023).

2.2. Scenarios and change periods

In the present study, we analyse a low emissions (SSP1-2.6) and a high emissions (SSP3-7.0) scenario from the ScenarioMIP experiments, comparing the projected changes from the historical CMIP scenario experiments for the same models. Two other scenarios, a medium emissions scenario (SSP2-4.5) and a very high emissions scenario (SSP5-8.5) are shown for GCMs below (see section 4.2.3 for context), but are not used elsewhere in this study for consistency with the CORDEX-Australasia data set. The future period analysed here is a 20-year time slice from 2080 to 2099, representing late 21st Century conditions. The historical period used here is a 20-year time slice from 1995 to 2014. These periods are consistent with sampling choices applied in the Intergovernmental Panel on Climate Change (IPCC) AR6 report and Atlas (Gutiérrez et al. 2021; Masson-Delmotte et al. 2021), and allow a difference in climate forcing that is large enough to provide a clear picture of anthropogenic forcing. Extra analysis of longer 50-year periods is provided for some indices in the body and others in the Supplementary material to assess the effect of period length (Fig. S5–S12, S15).

2.3. Regional climate model ensembles

The Australian climate modelling community has active collaboration and intercomparison through the National Partnership for Climate Projections (NPCP) supporting the Climate Projections Roadmap for Australia (Department of Climate Change, Energy, the Environment and Water 2023). There are currently four modelling initiatives active as part of NPCP undertaking dynamical downscaling, with each running a single RCM:

The focus of each of these initiatives is the delivery of climate projections that can be used to inform future planning and decision making as part of a public climate service. They provide nationally consistent regional projections at spatial scales of 10–20 km for a coordinated set of scenarios over the Australasia region following CORDEX guidelines (Giorgi and Gutowski 2015), hereafter referred to as the CORDEX-Australasia ensemble. We note that ACS-CCAM and QldFCP-2 make use of the same model (CCAM), but in rather different configurations, meaning they are related but not identical. CCAM is a global model, but here it is used for regional downscaling. Hereafter, the ACS-CCAM and QldFCP-2 models are referred to as regional models for simplicity, reflecting their use in the CORDEX-Australasia ensemble. The ensemble of RCMs provides a rich source of future climate information, while also being applied individually to specific regions and for various customers. In the present study, we analyse three different RCMs in a total of five model configurations (Table 1).

CMIP6 GCMs were selected for downscaling for each of the four RCM ensembles following objective criteria for each domain of interest, including an evaluation of performance, consideration of model independence and representativeness of the range of projections (Di Virgilio et al. 2022; Chapman et al. 2023, 2024; Grose et al. 2023). We note that each of the RCM ensembles was designed to stand alone as a representative set of projections, although the model selection for ACS-BARPA and ACS-CCAM noted the opportunity to form a complimentary ‘sparse matrix’ of projections to optimise the sampling of projection uncertainty with limited resources available (Grose et al. 2023).

The ACS uses BARPA-R to dynamically downscale seven CMIP6 GCMs over a limited domain. BARPA-R couples the UK Met Office’s Unified Model (MetUM) atmosphere with the Joint UK Land Environment Simulator (JULES) land surface model. BARPA-R operates on a limited-area horizontal grid with a spacing of 0.1545° (~17 km) and employs the HadREM3-GA7-05 global atmosphere/land (GAL) physics configuration, with additional enhancements for land surface characterisation and convection (Su et al. 2022; Howard et al. 2024). The model’s vertical grid varies slightly, with either 64 levels up to 41 km or 61 levels up to 32 km, depending on the availability of global model forcing data. BARPA-R simulations are constrained by unadjusted lateral boundary and sea surface temperature (SST) data derived from global models and are dynamically nudged towards their temperature and wind fields between 11 and 37 km above the surface to reduce inconsistency between the regional and global models at the lateral boundary (Stassen et al. 2023). Physical parameterisations in BARPA-R include the single-moment microphysics scheme of Wilson and Ballard (1999), radiative transfer scheme of Edwards and Slingo (1996), prognostic cloud fraction and condensate (PC2) scheme of Wilson et al. (2008), gravity-wave drags scheme of Scaife et al. (2002), boundary layer turbulent mixing scheme of Lock et al. (2000) and a mass-flux convection scheme of Gregory and Rowntree (1990), with updates by Walters et al. (2019). As a prognostic aerosol scheme is not used, the influence of aerosol radiation and cloud effects are prescribed following Tucker et al. (2022). This approach prescribes scenario-dependent pathways of 4-D aerosol properties and cloud droplet concentration numbers, derived from offline simulations following the approach of the EasyAerosol scheme of Stevens et al. (2017).

The ACS provides a second set of dynamically downscaled CMIP6 projections using CCAM, which uses spectral nudging (Thatcher and McGregor 2009). These downscaled simulations differ from the QldFCP-2 simulations (described below) by spectrally nudging the air temperature, winds and surface pressure back towards the host GCM state at a length scale of 3000 km. Hence, these CCAM simulations are constrained to have closer agreement with the host GCM at large spatial scales, but without bias adjustment. The ACS-CCAM simulations used a C384 grid with regional stretching using a Schmidt transformation (Schmidt 1977), reaching 12.5-km resolution over Australasia. An inline ocean model is also used with the SSTs spectrally nudged to the driving GCM. CCAM includes parameterisations for radiation (Freidenreich and Ramaswamy 1999; Schwarzkopf and Ramaswamy 1999), single-moment cloud microphysics (Lin et al. 1983; Rotstayn 1997), gravity wave drag (Chouinard et al. 1986) and boundary layer turbulent mixing (Hurley 2007). CCAM also includes the Australian Community Atmosphere Biosphere Land Exchange (CABLE) land-surface scheme (Kowalczyk et al. 2013) and an urban parameterisation (Thatcher and Hurley 2011). Prognostic aerosols were included to represent direct and indirect effects (Rotstayn and Lohmann 2002; Rotstayn et al. 2011). The atmosphere includes 54 vertical levels, reaching an altitude of ~35 km, whereas the ocean model consists of 40 vertical levels extending to a depth of 5 km. All CCAM-ACS simulations use coupled atmosphere–ocean CCAM model configuration.

The design and specifications of the NARCliM2.0 programme are described in detail in Di Virgilio et al. (2025). In summary here, NARCliM2.0 uses the Weather Research and Forecasting (WRF) version 4.1.2 model (Skamarock et al. 2008) to simulate the climate over the CORDEX-Australasia domain at a 20- and a 4-km inner domain over south-east Australia by one-way nesting. This inner domain uses a 4-km resolution with the aim of rendering these simulations convection permitting (Kendon 2021; Lucas‐Picher 2021). Hence, whereas the 20-km-resolution outer domain uses convective parameterisation, simulations over the 4-km domain do not. In the present study, we only analyse the 20-km resolution NARCliM2.0 data. NARCliM2.0 includes several other features that differ from previous generations of NARCliM: it uses new RCM physical parameterisations and increases vertical levels from 30 to 45, and urban physics is activated to represent surface energy balance in urban areas by a single-layer urban canopy model (Kusaka and Kimura 2004). NARCliM2.0 downscaling is applied to five CMIP6 GCMs selected by a comprehensive process that considers model evaluation, the independence of model errors and spanning the future change space (Di Virgilio et al. 2022).

The QldFCP-2 has produced a set of 15 dynamically downscaled simulations using the CCAM C288 grid stretched to 10 km over the Australian region model (McGregor 2015; Chapman et al. 2023). The model has 35 vertical levels in the atmosphere and 30 levels in the ocean (Thatcher et al. 2015). QldFCP-2 uses bias and variance corrected SSTs and sea ice, as documented by Hoffmann et al. (2016) and Chapman et al. (2023) and Atmospheric/High Resolution Model Intercomparison Project (AMIP/HighResMIP) style integrations (Gates 1992; Haarsma et al. 2016). In addition, the CMIP6 radiative forcings, which consist of time-varying solar forcing, greenhouse gases (CO₂, N₂O, CH₄, CFCs), ozone change, aerosols (sulfate, organic, black carbon, dust, volcanic, dimethylsulfide) and transient land cover change were used. CCAM was run in atmosphere-only and ocean-coupled modes. Five out of 15 simulations were run in ocean-coupled mode, with bias-corrected SSTs with spectral nudging to ensure SSTs agreed with the host GCM at a length scale of 1000 km (Thatcher and McGregor 2009; Thatcher et al. 2015). The data for the historical period (1960–2014) and for the 2015–2100 period for three emission scenarios (SSP1-2.6, SSP2-4.5 and SSP3-7.0) were produced. The AMIP-style downscaling where host model SSTs and sea ice are used for downscaling instead of high frequency (6 h) 3-day data for winds and temperature allows sampling of additional CMIP6 models that did not save high-frequency data required in traditional RCM downscaling. In the present study, we do not analyse the QldFCP-2 ocean-coupled experiments.

The diversity in approaches to regional dynamical downscaling documented above is seen as a positive here, given that the science is not clear on what approaches might provide the most realistic projections and use of multiple RCMs is required for robust regional projections.

3.1. Regional averages

Although it is common to interpolate GCMs to the RCM resolution for the purposes of added-value analysis (e.g. di Luca 2015), we adopt a different approach as the aim is to better quantify uncertainties at the larger scales. We follow an approach similar to that of Bador et al. (2020), whereby we interpolate all the GCMs and RCMs to a common 1.5 by 1.5° resolution, which provides a fair basis for intercomparison. The data are then averaged to the Australian continental area, as well as four large-scale natural resource management (NRM) regions of Australia (CSIRO and Bureau of Meteorology 2015, Fig. 1), and averaged both annually and seasonally following the calendar definition. For brevity, we primarily present the Austral summer (December to February, DJF) and Austral winter (June to August, JJA) averaged quantities. Results for other seasons can be found in the Supplementary material (Fig. S1–S15). The variables analysed in this study are near-surface air temperature (Tas) and precipitation (Pr). The NRM regions are Eastern Australia (EA), Northern Australia (NA), Rangelands (R) and Southern Australia (SA).

Fig. 1.

Natural Resource Management (NRM) super-cluster regions of Australia. The NRM regions are Eastern Australia (EA), Northern Australia (NA), Rangelands (R) and Southern Australia (SA). (Source: https://www.climatechangeinaustralia.gov.au/en/overview/methodology/nrm-regions/).

3.2. Partitioning uncertainty

We follow the approach of Hawkins and Sutton (2009), who quantify and attribute projection uncertainty into three categories: variance in climate due to internal variability (V), forcing scenarios (S) and GCM response to forcing (G). Following Evin et al. (2021), we introduce a fourth category to account for RCM model-to-model uncertainty (R). Uncertainty for each category varies in time (t), except for internal variability (V), which we estimate as a single number for each simulation. Though this is an oversimplification, we found that calculating V as a function of time did not change the substance of our results (not shown). We note that this does not imply that variability does not change, but rather that it does not change much as a fraction of variance in future projection uncertainty. We investigate changes in variability in later sections of this study. Following Hawkins and Sutton (2009):

3.3. Equilibrium Climate Sensitivity

Equilibrium Climate Sensitivity (ECS) is a measure of the sensitivity of global temperature change to a doubling of globally averaged atmospheric carbon dioxide (CO₂) concentration at equilibrium. The CMIP6 ensemble is known to have an unbalanced sampling of ECS compared with previous generations of models (e.g. CMIP5) and other lines of evidence (Sherwood et al. 2020), with a larger range in ECS values as well as very ‘hot models’ projecting more than 5°C global warming per doubling of CO₂ concentration (i.e. ‘hot model’ problem; see Hausfather et al. 2022). In the present study, ECS values for each model are taken from Masson-Delmotte et al. (2021) and Meehl et al.(2020) and classified into the likely ECS range (with robustness tests) of 2.3–4.5°C, or considered as high ECS (>4.5°C), in line with Sherwood et al. (2020).

4.1. Region-averaged projections of temperature and precipitation

We begin this study with a presentation of Australian continental and annual averaged changes in surface temperature and precipitation relative to a 1995–2014 historical base period, smoothed with a centred 20-year running mean (Fig. 2). Although unconventional, in the first column, we show all GCMs, RCMs and SSPs together without any distinguishing colours or markers in order to emphasise the range and variability of plausible futures found in future climate simulations for Australia. The bar plots following in Fig. 2 are all presentations of the same late 21st Century projections from GCMs and RCMs, but grouped in three different ways: first by SSP, next by GCM or RCM ensemble and finally by individual GCM or driving GCM (i.e. in the last column of Fig. 2, a bar has all SSPs for a given GCM, and also includes any CORDEX simulations driven with that same GCM).

Fig. 2.

Projected changes in average annual surface temperature (a–d), and precipitation (e–h) relative to 1995–2014, for all models, SSPs and ensembles together (a, e). Box plots of changes by 2090 categorised by SSP (b, f), by 2090 separated by RCM and GCM ensemble (c, g), and by 2090 separated by GCM or driving GCM (d, h). All quantities are smoothed with a centred 20-year running mean.

Continental annual averaged temperature changes by 2090 range between 0.2 and 5.9°C in the models analysed (Fig. 2a). The 20-year smoothed time series of continental annual averaged temperature change are reasonably smooth, indicating little decadal variability. The range of temperature change is strongly related to the future scenario considered, with large changes seen for higher emissions. Each of the RCM ensembles provides a similar range in late 21st Century projections of temperature to that seen in the CMIP6 GCM ensemble. The range of changes in temperature varies by GCM (or driving GCM); however, there is strong overlap between the individual GCM ranges.

Continental annual averaged precipitation changes by 2090 range from ~−0.6 to +0.6 mm day^–1 in the models analysed. The 20-year smoothed time series of precipitation change are highly variable, indicating a high level of decadal variability. In contrast to temperature changes, for continental annual averaged precipitation, we find that changes by 2090 are not strongly related to SSPs, although the range in projected changes does appear to increase slightly with forcing scenarios. The range in precipitation changes varies by GCM or RCM ensemble; however, these differences are small in comparison with the variation in precipitation changes between GCM or driving GCM groupings.

4.2. Relative sources of uncertainty in regional projections

As shown in the previous section, the projected climate in any future 20-year period can be significantly different from the historical base period owing to a range of factors. These factors lead to considerable uncertainty in future projections and are of differing importance depending on the variable of interest.

In Fig. 3 and 4, we present the fraction of uncertainty in climate projections attributed to each of the four sources described in the Methods section. It is important to note that the presentation of projection uncertainty relative to a modern base period by definition ignores the forced response up until the base period that is used. This is appropriate common practice for projection analysis but should not be interpreted as a lack of detected forced change in the present climate.

Fig. 3.

Fraction of temperature projection range for each NRM region due to internal variability (orange), forcing scenarios (S, green), RCM differences (R, light blue), and GCM differences (G, dark blue) for DJF (top) and JJA (bottom).

Fig. 4.

As in Fig. 3, but for precipitation.

In Fig. 3, we show the fraction of temperature projection range for each of the four NRM regions associated with different sources of uncertainty for DJF and JJA. By definition, G and R are small in the near future because all changes are relative to a common base period. Initially, the uncertainty in projected temperature is largely due to V; however, over time, the uncertainty due to S and G grows. By the middle of the 21st Century, G is the largest factor explaining uncertainty in projections, likely owing to differences in the pattern of circulation change with global warming (e.g. Chadwick et al. 2012) as well as differences in climate sensitivity. R is also largest in the middle of the 21st Century in general, although still small in comparison with G. By the end of the 21st Century, the main factor explaining uncertainty in projections of temperature is S. It is notable that R is small, suggesting that the GCM-derived changes are quite robust in this framing for the remainder of the century and downscaling adds little new information at the broad NRM scales considered here, as expected.

The Hawkins and Sutton diagrams for precipitation changes (Fig. 4) show a rather different story compared with temperature. G is the main source of uncertainty for all time horizons except in the near term (for which V is the most important factor). R is larger during summer, consistent with warm season precipitation processes, which tend to be on smaller scales and hence likely to be more sensitive to model resolution and the way in which convection is represented, even if it is parameterised, though this does not explain the same result for temperature (see Fig. 3). We do note that the coarse spatial scales considered here may mean that the uncertainty due to R is underestimated for local or fine-scale applications. The large contribution of G to precipitation uncertainty is likely a reflection of an inability to constrain regional circulation change because at the regional scale, dynamic effects are far less certain than thermodynamic effects (Shaw et al. 2024a, 2024b). Even by the end of the 21st Century, there is still a significant amount of uncertainty attributable to both G and R.

We now turn our attention to each of these sources of projection uncertainty individually.

4.2.3. Projection uncertainty due to model selection

Owing to the high cost of running dynamical downscaling simulations, each ensemble typically only considers a subset of CMIP global models to keep computation and data storage requirements manageable (see Data section – Regional climate model ensembles). Each regional modelling group conducted their own GCM selection for downscaling, focusing on their own criteria of model evaluation and model independence, while attempting to span the range of plausible projection uncertainty for their region of interest. As shown in Table 1, each RCM ensemble consists of a different set of model simulations, and the RCM ensembles are not of equal size. How does the selection of GCM for downscaling across the four regional modelling efforts affect the spread of projections at the national scale for each ensemble?

In Fig. 5, we turn our attention to the implications of model selection, focusing on the Australian continental temperature changes by the late 21st Century. We focus our discussion on the high emissions scenario (SSP3-7.0) for each regional modelling ensemble (Fig. 5c); however, we note that the results are similar for a low emissions scenario (Fig. 5b). The central estimate and 5–95% range of mean temperature change are higher when weighting CMIP6 models equally, sometimes referred to as ‘one model–one vote’ 4.6°C (3.3–5.4°C) than when restricting by likely ECS 4.1°C (3.2–4.9°C), as expected owing to the over-representation of the ‘hot models’ in the CMIP6 model range.

Fig. 5.

Change in Australian mean annual temperature relative to 1850–1900 in CMIP6 and CORDEX; (a) the average in 2080–2099 for four SSPs using ‘one model–one vote’ of r1 simulations from 35 CMIP6 models; (b) the equivalent bars for change in 2080–2099 for SSP1-2.6 – the first bar also using ‘one model–one vote’ but also showing the multi-model mean (white circle) and all 35 models (black dots); the following bar (CMIP6 constr.) shows the same but with the bar constrained to only models with Equilibrium Climate Sensitivity (ECS) in the likely range (2.3–4.5°C) and high ECS shown as red dots (>4.5°C), followed by the selected CMIP6 subsets (including members that are not r1) as brown bars, and RCM simulations themselves as green bars. The NARCLIM2.0 ensemble has two green bars because it is run in two different configurations (R3 and R5). (c) as in (b), but for SSP3-7.0. GCM ensembles use the 10–90% range, RCM model selection and RCM results show 0–100% range owing to smaller sample size.

Each regional modelling group made different choices around model selection, resulting in differences in the projected temperature change; however, a focus on ensuring ‘representativeness’ in the spread of climate sensitivity and mean warming taken by each group means that each model selection gives a similar result. Broadly, each ensemble samples a similar range of projected change when we separate the sampled high-ECS models from each subset. All subsets show 3.3–5.1°C range (mean 4.2°C) to within 0.1°C. Therefore, reporting the range of change in models within likely ECS range (from Sherwood et al. 2020), then treating high cases as ‘low likelihood high warming’ outcomes is justified. One model–one vote was not used for global temperature by the IPCC, but the IPCC weighting scheme was not used regionally. Alternatively, a simpler scheme of simply categorising models by climate sensitivity was proposed by Hausfather et al. (2022). They suggest that by separating models with climate sensitivity outside the ‘likely range’, the remaining ensemble of models is broadly representative of the IPCC AR6 projections.

Projections for mean warming of Australia are similar in the RCM ensembles compared with their driving GCMs, with generally small modification of the change signal through downscaling at this national scale (Fig. 2, see additional bars). At a large scale, this suggests that the RCMs do not significantly modify the projected temperature change. Importantly, the level of climate sensitivity in the host GCM largely sets the climate sensitivity in the RCM results.

Each model has unique factors that relate to projection uncertainty in each ensemble. The ACCESS-CM2 model r2i1p1f1 ensemble member downscaled by QldFCP-2 has lower warming (5.3°C) than the r4i1p1f1 ensemble member downscaled using ACS-BARPA-R and CCAM-ACS (6.0°C) – r4i1p1f1 is the outlier and values of change in the various five ensemble members downscaled are 5.2, 5.3, 5.1, 6.0 and 5.6°C. The extremely high points in the CMIP6 bar are the CanESM5 and CanESM5-Canoe models (however, these models were not selected for downscaling by any group here). The UK-ESM model has the highest ECS of any model downscaled in the CORDEX-Australasia ensemble, and does indeed project relatively high warming for Australia (5.8°C), but this is in fact lower than the projected change from ACCESS-CM2 r4. ECS correlates with temperature change at the regional scale, but local factors also influence regional warming, meaning that the highest ECS global model does not necessarily result in the highest regional temperature change. Two clear examples here are the ACCESS-ESM1.5 r6i1p1f1 ensemble member (ACCESS-ESM1.5 is within the likely range of ECS but at the very top of the bar), which has enhanced drying leading to higher local warming, and the EC-Earth3 and EC-Earth3-Veg r1i1p1f1 simulations with ECS in the high range but Australian warming quite near the multi-model mean (with increased precipitation suppressing local warming). NorESM2-MM has the lowest ECS and shows the lowest warming in all subsets.

Overall, we find strong agreement in the ranges of Australian continental temperature projections found in each RCM ensemble, consistent with the larger CMIP6 GCM ensemble, once the ECSs of the GCMs (or driving GCMs) are accounted for.

4.3. Range of projections in the RCM and GCM ensembles

The subjective choices in downscaling approaches associated with each RCM also further influence the uncertainty in projected changes at the regional scale. Examples of important choices include paramaterisations, resolution, pre-processing steps for boundary conditions, spin-up length, model domain and nudging options. These factors may lead to increases, decreases, or shifts in the range of projected changes owing to local process representation in each downscaling model, and differences between the downscaling models themselves. Fig. 5 shows the projections for Australian averaged temperature change relative to 1850–1900 (using GCM warming until 1995–2014 prior to RCM simulations starting). This shows a very similar projection in all model ensembles, with a likely range of ~3.3 to ~5°C (mean of ~4°C) warming, and projections from ‘hot models’ representing a low likelihood high warming future of between ~5 and ~6°C.

Next, we investigate the ranges of temperature and precipitation change in each ensemble at a seasonal and NRM region scale. In Fig. 6, we show the changes (relative to the 1995–2014 climatology) in surface air temperature by the late 21st Century under a high emissions scenario for the GCM and RCM ensembles, averaged over each of the four NRM regions of Australia for DJF and JJA. Overall, both GCMs and RCMs project temperature increases of ~2–6°C. The high-ECS GCM models (indicated with red symbols in Fig. 6) project higher temperature increases, in excess of 5°C for EA and R. The range of projected temperature changes from individual RCM ensembles (coloured bars) is marginally smaller than the CMIP6 GCMs (grey bars); however, collectively, the projected temperature range from all the RCMs is similar to the CMIP6 GCMs.

Fig. 6.

Changes in December to February (DJF) (top) and June to August (JJA) (bottom) area-averaged surface air temperature (°C) by the late 21st Century under a high emissions scenario (SSP3-7.0) for each model ensemble. Each column represents an NRM region. Red symbols indicate high ECS models and blue symbols indicate low ECS models. Changes are calculated as the difference between 2080–2099 and 1995–2014.

The projected changes for DJF and JJA are largely similar in the EA and R regions. In SA, there is a larger projected increase of ~1°C during DJF compared with JJA, whereas in NA, larger increases (~1°C) are projected during JJA. The range of projected temperature changes at the regional scale reflects the uncertainty from the driving global model projections. In comparison, the differences in projected range between RCM ensembles are relatively small (of the order of 1°C), and even then are partly due to differences in driving GCM choices.

Next, we consider the changes in precipitation by the late 21st Century under a high emissions scenario for the GCM and RCM ensembles, averaged over each of the NRM regions of Australia for DJF and JJA (Fig. 7). Although the RCM ensembles generally project a range of precipitation change that is similar to the CMIP6 GCM range, there are some notable differences.

Fig. 7.

As in Fig. 6, but for precipitation change (%).

For the DJF season (Fig. 7 top row), there is little agreement on the direction of precipitation change in any of the NRM regions in the CMIP6 GCM ensemble. In most regions, the GCM and RCM show both substantial increases and also substantial decreases.

For SA during DJF, the QldFCP2 ensemble mostly projects wetting, whereas the NARCLIM R5 ensemble is on average drying, and other ensembles do not show strong agreement on the direction of change. For both the R and NA regions, the majority of RCM simulations project wetting during DJF, although the range in most ensembles also includes drying (with the exception of QldFCP-2 for R in DJF, which are all wetter than their historical climatology).

The projected precipitation changes in JJA Pr (Fig. 7 bottom row) also show a wide spread, but with a tendency towards drying in most regions and model ensembles. The projected drying in SA for JJA is similar in all ensembles, suggesting that this result is robust (consistent with the conclusions of Rauniyar et al. (2023) for south-west Australia). Such cases of inter-ensemble agreement provide some level of confidence in the projected changes, and may suggest that in those cases the sampling across ensembles is less critical.

Similarly, the projections of Pr for EA, R and NA for JJA are all on average lower in each RCM ensemble, although higher Pr is found for all regions in at least some RCM simulations.

Unlike the case of surface air temperature projections, we found that ECS has no clear relationship to the region-averaged Pr change in any RCM or GCM ensemble for any season (Fig. 7). This result is not entirely surprising, because circulation change is known to be a more influential factor for regional Pr projections than the thermodynamic-related changes (e.g. Chadwick et al. 2012).

In summary, the choice of RCM ensemble may have strong implications for the range of projections at the NRM scale for some variables, such as Pr, but is less of a concern for other variables, such as Tas, which is found to be consistent at the NRM scale in all ensembles. Individual GCMs are found to drive a similar response across RCMs, both for Pr and for Tas at the regional scale. For example, the high-ECS GCMs (e.g. ACCESS-CM2, UK-ESM-LL, CESM2) are generally the ‘hottest projections’ in all RCM ensembles (result in the highest Tas increases). The EC-Earth3 r1i1p1f1 and EC-Earth3-Veg r1i1p1f1 GCMs drive the ‘wettest projections’ at regional scales across different RCM ensembles, whereas the ACCESS-ESM1.5 r6i1p1f1 GCM drives one of the ‘driest projections’ at regional scales in all RCM ensembles. Although the projected changes are similar for a common driving GCM, they are not identical in all RCMs.

4.4. How do RCMs modify the signal from their driving GCMs?

We investigated how the Tas and Pr change signals are modified by RCMs for two diverging forced responses from GCMs. To do this, we examined the driest and wettest projected changes under the SSP3-7.0 scenario downscaled for the Australian region. These extreme projections are found in individual members of the ACCESS-ESM1.5 (drying) and EC-Earth3 (wetting) models respectively.

4.4.1. The ACCESS-ESM1.5 r6i1p1f1 drier future scenario

The r6i1p1f1 ensemble member of the ACCESS-ESM1.5 large ensemble provides one of the driest regional projections for Australia by the late 21st Century under a high emissions scenario (SSP3-7.0). The output from this particular simulation was selected for downscaling by multiple regional modelling efforts for Australia to ensure we captured plausible high-impact futures.

In Fig. 8, we show the changes in seasonally averaged Tas by the late 21st Century under SSP3-7.0 for the ACCESS ESM1.5 GCM r6i1p1f1 simulation (Fig. 8a–d), the ACCESS ESM1.5 GCM large-ensemble mean (Fig. 8e–h) and each RCM downscaling the r6i1p1f1 simulation (Fig. 8i–bb). This is a particularly valuable case study because this GCM simulation is dynamically downscaled by all groups. Comparing Fig. 8a–h, the r6i1p1f1 ensemble member is broadly representative of the ACCESS-ESM1.5 GCM large-ensemble mean projection, though it is on average drier than the large-ensemble mean from December to May. Model-to-model RCM differences are apparent at the regional scale. The RCM downscaled simulations of ACCESS ESM1.5 GCM r6i1p1f1 vary from the GCM in different ways:

• The BARPA-R, NARCliM2.0 models reduce the warming over the west coast of Australia in DJF and March, April, May (MAM).

• The ACS-CCAM model generally increases the warming over northern Australia in JJA and SON.

• The QldFCP-2 and NARCliM2.0 models projects enhanced warming over south-eastern Australia in JJA and SON.

• The NARCliM2.0 simulations tend to have weaker warming over central and western Australia, and stronger warming over eastern Australia, compared with the driving GCM simulation and other RCMs.

Fig. 8.

Change in seasonal Tas. by the late 21st Century (2080–2099 minus 1995–2014) under SSP3-7.0 using ACCESS-ESM1.5 GCM r6i1p1f1 (a–d), the relative change for the GCM large ensemble (e–h), and the relative change for RCM simulations downscaling the ACCESS-ESM1.5 r6i1p1f1 projections (i–bb). Relative changes are calculated by subtracting the GCM changes.

In Fig. 9, we show the same as Fig. 8, but for Pr projections. Overall, the global driving model projects continental drying in all seasons, with some regional wetting near Tasmania. The drying is strongest in DJF and MAM. The r6i1p1f1 member is mostly consistent with (albeit stronger than) the large-ensemble mean for the same model, suggesting that for this model, the change pattern is a forced response due to global warming, and not caused by multidecadal variability. In general, the RCMs show continental scale changes similar to the global driving model.

Fig. 9.

Change in seasonal average Pr (mm day^–1) by the late 21st Century (2080–2099 minus 1995–2014) under SSP3-7.0 using ACCESS-ESM1.5 GCM r6i1p1f1 (a–d), the relative change for the GCM large ensemble (e–h), and the relative change for RCM simulations downscaling the ACCESS-ESM1.5 r6i1p1f1 projections (i–bb). Stippling in (e–h) indicates locations where less than two-thirds of ensemble members agree on the direction of change. Relative changes are calculated by subtracting the GCM changes. Magenta contours indicate locations where the relative change is associated with a change in sign. Absolute changes can be found in the Supplementary material (Fig. S13).

The difference between the GCM and RCM projections is larger in DJF and MAM, and in some cases, large regions experience a switch in sign from negative to positive (e.g. parts of EA in ACS-CCAM and QldFCP-2, and parts of NA in ACS-BARPA-R in DJF). Changes in warm season Pr are likely to be more closely associated with convective processes in the models, suggesting that convection parameterisations may have an important role in modulating the change signal in RCMs. QldFCP-2 projections show consistent tendency towards a less dry future compared with the GCM across all seasons, which may be related to application of bias correction to host models SSTs.

So although the ACCESS-ESM1.5 r6i1p1f1 GCM simulation was selected for downscaling owing to its strong drying and warming projection, the process of dynamical downscaling has modified the change signal in each of the RCMs. As a result, regional variations in the change signal range from stronger drying to localised wetting.

4.4.2. The EC-Earth3 r1i1p1f1 wetter future scenario

Next, we turn our attention to another key GCM projection, produced by the EC-Earth3 r1i1p1f1 ensemble member under SSP3-7.0, which was downscaled by multiple regional modelling groups. This simulation represents a wetter future for most of Australia under strong global warming, in contrast to the earlier scenario. NARCliM models are not included in Fig. 10 and 11 because they did not downscale the EC-Earth3 GCM simulations.

Fig. 10.

Change in seasonal Tas. by the late 21st Century (2080–2099 minus 1995–2014) under SSP3-7.0 using EC-Earth3 GCM r1i1p1f1 (a–d), the relative change for the GCM large ensemble (e–h), and the relative change for RCM simulations downscaling the EC-Earth3 r1i1p1f1 projections (i–t). Relative changes are calculated by subtracting the GCM changes.

Fig. 11.

Change in seasonal average Pr (mm day^–1) by the late 21st Century (2080–2099 minus 1995–2014) under SSP3-7.0 using EC-Earth3 GCM r1i1p1f1 (a–d), the relative change for the GCM large ensemble (e–h), and the relative change for RCM simulations downscaling the EC-Earth3 GCM r1i1p1f1 projections (i–t). Stippling in (e–h) indicates locations where less than two-thirds of ensemble members agree on the direction of change. Relative changes are calculated by subtracting the GCM changes. Magenta contours indicate locations where the relative change is associated with a change in sign. Absolute changes can be found in the Supplementary material (Fig. S14).

In Fig. 10, we show the changes in seasonally averaged temperature by late 21st Century from the EC-Earth3 GCM r1i1p1f1 member (Fig. 10a–d), the EC-Earth3 GCM large ensemble change relative to the r1i1p1f1 member (Fig. 10e–h) and the RCM changes (relative to the GCM) using the r1i1p1f1 simulation (Fig. 10i–t). The warming in r1i1p1f1 is broadly representative of the large-ensemble mean, with relatively weak warming over northern Australia in DJF and MAM as precipitation increases, and moderate warming (compared with other GCMs, as shown in Fig. 6) with a pattern centred over east-central Australia in JJA and September, October, November (SON). Overall r1i1p1f1 warms less than other simulations in the GCM large ensemble, especially over north-western Australia.

RCM differences are apparent at the regional scale in all seasons:

• The ACS-BARPA-R and QldFCP-2 models generally warm more than the GCM over the northern parts of the continent in DJF, MAM and JJA.

• ACS-CCAM shows a similar pattern to the GCM but with weaker warming than the GCM over the south-east of Australia.

In Fig. 11, we show the same as Fig. 10, but for seasonal Pr change. The pattern of seasonal Pr change in the selected GCM ensemble member is similar to the large-ensemble mean in JJA and SON, but the wetting signal over western Australia in DJF is a notable difference. Even in JJA and SON, the wetting pattern extends further south and to the east in the selected GCM ensemble member compared with the large ensemble mean. For most seasons in most locations, there is a lack of agreement on the sign of precipitation change in the EC-Earth3 large ensemble. These results suggest that, with the exception of northern and central Australia, the changes seen in the selected ensemble member may be partly due to multidecadal internal variability, rather than simply a forced mean-state response. This is in contrast to the previous ‘dry future’ example, in which the drying trend appeared consistently in the selected ensemble member, multi-ensemble mean and all RCMs for most seasons and locations.

The RCM downscaled simulations of EC-Earth3 r1i1p1f1 vary from the GCM in different ways:

• The ACS-BARPA-R model projects wetting in most places consistent with the GCM, although the wetting during DJF in particular is larger and more extensive than the GCM. Also evident in the ACS-BARPA-R model is a drying signal over the south-western tip of Western Australia in DJF, not found in the other RCMs.

• ACS-CCAM projects weaker wetting overall, with a pattern of change similar to the GCM except for some localised differences. These include a drying tendency over the northern coast in DJF and MAM (reversing the sign of change from the GCM). Similarly, the QldFCP-2 model has a pattern of precipitation change resembling the GCM albeit with some localised differences. These include a drying tendency over Cape York in DJF and parts of Queensland in MAM.

In this section, we examined two important and diverging projected changes, a wetter and a drier scenario, based on ensemble members from two different GCMs. The analysis of the change signal in RCMs compared with the GCM simulation downscaled illustrates that variations in the projected changes can depend on choice of RCM, although it is unclear if any model is systematically different to or better than others. The comparison of the selected GCM run with the same GCM large ensemble for these two key GCM storylines suggests that another source of uncertainty in the projected changes is internal variability within the GCM itself. Therefore, the downscaling approach of using a small subset of GCM simulations, and only one run of each GCM, may itself lead to a biased projection with internal variability a factor.

4.5. Projected change contrasts across topography

One key feature of the increased spatial resolution offered by RCMs is a more accurate representation of topography. We investigated the RCM and GCM differences over the complex topography found on the eastern seaboard of Australia.

Previous work has identified two regions in which resolution significantly modifies projections of seasonal precipitation change (Grose et al. 2010, 2015; Dowdy et al. 2015).

The first is between the eastern seaboard and inland New South Wales (Dowdy et al. 2015; Grose et al. 2015), represented here by the difference between the Central Slopes cluster and the East Coast South sub-cluster. As the Eastern Seaboard is a distinct climatic entity compared with inland, there are reasons to expect a different projection (Dowdy et al. 2015). In previous generations of dynamical and statistical downscaling, there was a suggestion of a difference in some model ensembles for some seasons, but with no strong model agreement (Grose et al. 2015). The new modelling from CMIP6 and CORDEX-Australasia CMIP6 also shows a range of responses with contrasts of either sign (Fig. 12 and 13). Some models show distinctly enhanced drying on the Eastern Seaboard, notably NARCliM2.0–EC-Earth3-Veg simulations, but also the BARPA and CCAM–EC-Earth3 simulations (Fig. 13b). In contrast, other RCMs have little change or an increase in precipitation on the east coast in contrast to drying elsewhere. Interestingly, although NorESM2-MM has no clear regional variations in trend, the NARCliM2 simulations produce enhanced drying on the east coast, whereas QldFCP-2 produces coastal wetting.

Fig. 12.

Mean winter precipitation percentage change (SSP3-7.0 1995–2014 to 2080–2099) over south-eastern Australia for RCM downscaled GCM projections, along with the mean change for each ensemble of RCM output. Blank maps indicate that output for that combination of GCM and RCM is not available.

Fig. 13.

Mean seasonal precipitation projection (SSP3-7.0 1995–2014 to 2080–2099) in Central Slopes and East Coast South (a,b), and also Southern Slopes Tasmania East and West NRM sub-clusters (c,d), and the difference between them. GCMs are shown on the left, and the corresponding RCM projections in the right column.

The second example is between western and eastern Tasmania, which exhibit two very different climate zones under the influence of different synoptic processes, not well resolved by GCMs. A large contrast was found in DJF (Grose et al. 2010), which showed a contrast between high model agreement for a drier west coast land region and no model agreement on the east coast land region under high emissions (Special Report on Emissions Scenarios A2) over the century. Southern Slopes Tasmania West v. Tasmania East in CMIP6 and CORDEX largely show the opposite contrast, with greater drying in the east than west (Fig. 13), partly reflecting model selection but seemingly also owing to differences in regional climate processes.

Overall, we do not see a large or systematic departure of the RCMs from the GCMs in these regions, even though the RCMs have a more realistic representation of topography. These results suggest that improved resolution of the order of 10 km does not have a systematic effect on region-averaged climate change simulations of precipitation over south-east Australia.

4.6. The relationship between historical climate and projected changes

As both RCMs and GCMs show considerable spread in projected changes at the regional scale, it would be ideal to be able to constrain projections using observational or theoretical evidence – sometimes referred to as emergent constraints (Hall et al. 2019; Brient 2020). A first step towards achieving this might be to analyse the relationship between historical climate (or bias) and the projected changes in model simulations. A finding of either systematic biases, or systematic relationships between bias and projected changes, would suggest that further analysis of observational or emergent constraints would help constrain the distribution of projection uncertainty.

A related issue is the question of interdependence of some GCMs and RCM simulations (e.g. Abramowitz et al. 2019). This arises as multiple GCMs in the CMIP6 ensemble share code, parameterisations, or key model components. As RCM simulations downscale the same runs from the same models, they are also affected by this issue of model interdependence (as shown in Fig. 8–11). If we cannot constrain projections completely, then it would be beneficial to weight model simulations to appropriately reflect the distribution of projection uncertainty due to this issue. Similarly to emergent constraints, a first step towards assessing model interdependence is to analyse the similarity in their historical biases and projected changes.

Fig. 14 shows a scatter-plot of historical surface air temperature and the projected future change by the late 21st Century for GCMs and RCMs. The general lack of significant correlation between historical and projected temperatures indicates that there is little, if any, relationship between the two. This is true for most seasons, for individual model GCM and RCM ensembles, as well as varied RCM simulations using the same individual driving GCMs. A weak correlation (r ≅ 0.3) is seen for NA in both DJF and JJA. RCM simulations downscaling the same GCM do tend to cluster together, implying that RCM projections at the regional scale inherit much of their change signal from the driving global model.

Fig. 14.

Historical DJF (top) and JJA (bottom) surface air temperature v. projected change in surface air temperature (K) by late 21st Century under the high emissions scenario for each NRM region of Australia. Colours indicate model ensemble and symbols are used to indicate specific driving GCMs. The Pearson correlation r and P-value are shown in the top right of each panel.

For precipitation (Fig. 15), projected changes do appear to correlate with historical climatology in several regions and seasons, where higher historical precipitation tends to be associated with more negative projected changes to precipitation. The strongest relationship is seen for NA in JJA (r ~= 0.6), though JJA is the dry season in NA so perhaps less important to consider. Caution should be taken here to not immediately interpret this connection as an emergent constraint, because the causal process connection between historical bias and projected precipitation change is not clear. For example, it is possible that both quantities are connected here to a third factor such as multidecadal internal variability, which is known to moderate Australian springtime climate conditions (e.g. Power et al. 1999). Nevertheless, the correlation between precipitation change and historical precipitation is also seen when considering 50-year periods, especially for the SON season (see Supplementary material, Fig. S8, S12). Unlike for temperature, the QldFCP-2 ensemble has a reduced range in precipitation projections at the regional scale compared with the other RCM and GCM ensembles analysed here. However, the reduced range of precipitation change is at least in part due to the different GCMs selected for downscaling for QldFCP-2.

Fig. 15.

As in Fig. 14, but for Pr (mm day^–1).

There is no clear separation between RCM ensembles for most regions in most seasons for either surface air temperature or precipitation. However, one noticeable pattern in Fig. 14 and 15 is the systematic vertical alignment of the QldFCP-2 simulations. This implies a low range in historical climate (and therefore a low range in bias) in that ensemble. The low range in QldFCP-2 historical climate is expected, owing to its downscaling methodology of bias correcting the driving GCM inputs to the RCM. The QldFCP-2 approach does not appear to constrain the projected change in surface air temperature.

The results presented here indicate that the relationship between historical values and projected changes is not informative for temperature; however, they are a potentially useful avenue for investigating constraints on precipitation for some regions and seasons.

4.7. Variability uncertainty

In this section, we focus on precipitation variability, which we earlier showed to be an important source of uncertainty even towards the end of the 21st Century. Changes to temperature variability are not explored here for brevity.

4.7.2. Change in interannual precipitation variability

Thus far, we have focused on mean state change of precipitation; however, it is also important to quantify the change in year-to-year variability around the mean for precipitation, as these changes can often be much larger than the mean state change. This is particularly important in the near term, where interannual variability may be large compared with mean state changes. On the global scale, precipitation variability is projected to increase in various ways, through an increase in daily and sub-daily extremes, daily variability and seasonal variability (Pendergrass et al. 2017). Variability is projected to increase even in some regions where the mean precipitation is projected to decline. However, projected changes vary by region, with a possible decrease in variability projected in some regions. Pendergrass et al. (2017) found that for Australia, there is low CMIP5 model agreement in the direction of change in seasonal variability (inter-annual variability in seasonal mean precipitation), except for a decrease in south-west Western Australia where precipitation is also robustly projected to decrease.

Here, we focus on the projected change in interannual variability in 35 CMIP6 GCMs and the 39-member CORDEX-Australasia CMIP6 ensemble, measured as the s.d. of seasonally averaged precipitation. We focus particularly on two of the models selected as a ‘representative climate future’ of a ‘hotter and much wetter future’ for most regions of Australia: EC-Earth3 and EC-Earth3-Veg (Grose et al. 2023), as well as the large ensemble of 58 members of EC-Earth3 and CORDEX downscaling from the two models. We use the longer 50-year periods to better quantify variability.

The projected change in interannual variability in CMIP6 and the CORDEX-Australasia CMIP6 ensemble (Fig. 16a–d) broadly supports the CMIP5 results reported in Pendergrass et al. (2017). There is no systematic model agreement on the direction of change in interannual variability in Australia. In fact, there is a large spread across CMIP6 models in all NRM super-clusters in both winter and summer, from large increase to large decrease in variability (Fig. 16a, b). This is also largely reflected in the CORDEX simulations (Fig. 16c, d), as much of the interannual variability will be inherited from the host model. In the EC-Earth3 and EC-Earth3-Veg r1i1p1f1 simulations, chosen as a wet representative climate future, there is an increase in mean precipitation but also a notable increase in variability in the four NRM super-clusters (Fig. 16a, b). For example, for EA in JJA, the change in the mean is 12% and the change in the s.d. is 27% in the EC-Earth3 simulation. There is also an increase in the frequency of wet years, and a larger increase in the median compared with the mean in this model. This means that this simulation should more aptly be described as a more variable or less stable climate, with more very wet years, rather than as a consistently wetter mean state. The downscaled projection is similar to the driving GCMs (Fig. 16c, d).

Fig. 16.

Change in precipitation s.d. 1965–2014 and 2050–2099 under SSP3-7.0 in CMIP6 (a, b), CORDEX (c, d) for JJA (left) and DJF (right) for NRM cluster averages. The EC-Earth3 GCM large ensemble changes in s.d. are shown in panels (e) and (f).

There is a broad spread in the projected change in variability in the 58 members of the EC-Earth3 large ensemble (Fig. 16e, f), including members showing a strong decrease for some regions. There is also a large spread in the projection of mean precipitation between the two 20-year periods (e.g. −40 to +38% for JJA in EA; see Fig. 17). There is a spread in wet years across the members; for example, in EA in JJA, we find an increase over the historical 95th percentile change from 5% of years to 7%, with a range of 0–22% (not shown). The results from the EC-Earth3 large ensemble provide a projection of higher future precipitation variability, with the potential for more frequent extreme and unprecedented wet years, indicating more frequent wetter and drier periods than in our current climate.

Fig. 17.

Change in mean precipitation v. change in precipitation variability (s.d.) between 1965–2014 and 2050–2099 under SSP3-7.0, for JJA (a–d), and DJF (e–h), with regions as marked.

In Fig. 17, we examine the relationship between the projected change in mean and s.d. across the model members. In most regions, the explained variance, R2, ranges from 0.4 to 0.9, which is statistically significant at the 95% confidence level except during DJF in NA and SA. This connection is generally reflected in the CMIP6 ensemble, the EC-Earth3 large ensemble, and CORDEX members as well. This supports the connection between a change in mean and variability at this time and spatial scale. This relationship has been noted in climate projections since at least the IPCC Third Assessment Report (Doblas-Reyes et al. 2021). However, effects are known to be scale-dependent and vary by region (Schwarzwald et al. 2021). For the south-west of Australia cool season specifically, the results support the finding from Pendergrass et al. (2017) that a significant decrease in variability is plausible, linked to the decrease in the mean (e.g. the ACCESS-ESM1.5 large ensemble shows a notable decrease in JJA precipitation s.d. in virtually all members). Note that NA and R are seasonally dry in the JJA season, so proportional (%) changes are exaggerated.