Global climate models (GCMs) are essential for investigating climate change, but their coarse scale limits their efficacy for climate adaptation planning at the regional scales where climate impacts manifest. Dynamical downscaling of GCM outputs better resolves regional climate and thus provides improved guidance for climate policy at regional scales. Being expensive to run, downscaling uses a subset of GCMs, necessitating careful GCM selection. This evaluation identifies a suitable subset of CMIP6 GCMs for downscaling over Australia by assessing individual GCMs against three criteria: (a) performance simulating daily climate variable distributions, climate means, extremes, and modes; (b) model independence; and (c) climate change signal diversity. Over Australia, GCMs are generally biased cold (warm) for maximum (minimum) temperature, with larger biases for minimum temperature. GCMs are generally wet biased, especially over the monsoonal north, but dry biased over eastern regions. Most GCMs show larger biases for temperature and precipitation over geographically complex, heavily populated eastern regions, relative to other regions. Evaluations identify a distinct group of 11 GCMs that perform consistently poorly across climate variables, statistics, and timescales with widespread, statistically significant biases, versus 13 GCMs that show consistent adequate‐to‐good performance with substantially reduced errors. Assessment of model independence highlights the lack of independence between several high‐performing GCMs, particularly from allied modeling groups, demonstrating the importance of careful ensemble selection when making selective samples of climate space. Once GCM climate signal diversity is considered, 6–8 mid‐to‐high‐performing, independent GCMs occupy the full range of the future climate space and, thus, are suitable for dynamical downscaling over CORDEX‐Australasia.

Global climate models (GCMs) are commonly downscaled to understand future local climate change. The high computational cost of regional climate models (RCMs) limits how many GCMs can be dynamically downscaled, restricting uncertainty assessment. While statistical downscaling is cheaper, its validity in a changing climate is unclear. We combine these approaches to build an emulator leveraging the merits of dynamical and statistical downscaling. A machine learning model is developed for each coarse grid cell to predict fine grid variables, using coarse‐scale climate predictors with fine grid land characteristics. Two RCM emulators, one Multilayer Perceptron (MLP) and one Multiple Linear Regression error‐reduced with Random Forest (MLR‐RF), are developed to downscale daily evapotranspiration from 12.5 km (coarse‐scale) to 1.5 km (fine‐scale). Out‐of‐sample tests for the MLP and MLR‐RF achieve Kling‐Gupta‐Efficiency of 0.86 and 0.83, correlation of 0.89 and 0.86, and coefficient of determination (R²) of 0.78 and 0.75, with a relative bias of −6% to 5% and −5% to 4%, respectively. Using histogram match for spatial efficiency, both emulators achieve a median score of ∼0.77. This is generally better than a common statistical downscaling method in a range of metrics. Additionally, through “spatial transitivity,” we can downscale GCMs for new regions at negligible cost and only minor performance loss. The framework offers a cheap and quick way to downscale large ensembles of GCMs. This could enable high‐resolution climate projections from a larger number of global models, enabling uncertainty quantification, and so better support for resilience and adaptation planning.

The NARCliM project contributes to the CORDEX initiative for Australasia. The first generation of NARCliM (N1.0) used CMIP3 global climate models (GCMs) and provided near and far future estimates of climate change across Australasia at 50-km and southeast Australia at 10-km resolution under a business-as-usual climate scenario. However, multiple sets of 20-year periods in N1.0 did not permit analysis of long-term, inter-annual to decadal trends across the 21^st century. Feedback on user needs for regional climate information revealed the desire for multiple emission scenarios and use of newer CMIP5 GCMs for dynamical downscaling. These limitations led to development of the second iteration of NARCliM, namely NARCliM1.5 (N1.5). The N1.5 downscaling exercise uses CMIP5 GCMs and is temporally expanded to cover 150 years (1950 to 2100) for two future Representative Concentration Pathways (RCP4.5 and RCP8.5). N1.5 simulations remain at the 50-km and 10 km resolutions over the same domains as N1.0, thus producing an expanded and complementary dataset for regional climate change. N1.5 simulations substantially improve over N1.0 in capturing the seasonal patterns and magnitudes of precipitation, including improvements in overall bias. Conversely, N1.5 shows similar results to N1.0 for maximum and minimum temperature, with no substantial improvement in overall bias. N1.5 projections project a hotter and drier future relative to N1.0. The combined N1.0 and N1.5 ensemble provides a wider spread of future climates more representative of that found in the full CMIP5 ensemble. Together, N1.0 and N1.5 ensembles provide an improved, more comprehensive data set for studying climate change.

Atmospheric reanalysis products offer high-resolution and long-term gridded datasets that can often be used as an alternative or a supplement to observational data. Although more accessible than typical observational data and deemed fit for climate change studies, reanalysis data can show biases resulting from data assimilation approaches. Thus, a thorough evaluation of the reanalysis product over the region and metric of study is critical. Here, we evaluate the performance of the latest generation of ECMWF reanalysis, ERA5, in simulating mean and extreme temperatures over Australia for 1979–2020 versus high-quality gridded observations. We find ERA5 generally simulates maximum and minimum temperatures reasonably well (mean bias ~1.5 °C), even though it underestimates/overestimates the daily maximum/minimum temperatures, leading to a cold bias for Tmax and a warm bias for Tmin. ERA5 also underestimates the decadal warming trend in both Tmax and Tmin compared to the observations. Furthermore, ERA5 struggles to simulate the temporal variability of Tmin, leading to a markedly worse skill in Tmin than Tmax. In terms of extreme indices, ERA5 is skilled at capturing the spatial and temporal patterns and trends of extremes, albeit with the presence of biases in each index. This can partially be attributed to the warm bias in the minimum temperature. Overall, ERA5 captures the mean and extreme temperature indices over the Australian continent reasonably well, warranting its potential to supplement observations in aiding climate change-related studies, downscaling for boundary conditions, and climate model evaluation.

Australia experiences a variety of climate extremes that result in loss of life and economic and environmental damage. This paper provides a first evaluation of the performance of state-of-the-art Coupled Model Intercomparison Project Phase 6 (CMIP6) global climate models (GCMs) in simulating climate extremes over Australia. Here, we evaluate how well 37 individual CMIP6 GCMs simulate the spatiotemporal patterns of 12 climate extremes over Australia by comparing the GCMs against gridded observations (Australian Gridded Climate Dataset). This evaluation is crucial for informing, interpreting, and constructing multimodel ensemble future projections of climate extremes over Australia, climate-resilience planning, and GCM selection while conducting exercises like dynamical downscaling via GCMs. We find that temperature extremes (maximum-maximum temperature -TXx, number of summer days -SU, and number of days when maximum temperature is greater than 35 °C -Txge35) are reasonably well-simulated in comparison to precipitation extremes. However, GCMs tend to overestimate (underestimate) minimum (maximum) temperature extremes. GCMs also typically struggle to capture both extremely dry (consecutive dry days -CDD) and wet (99th percentile of precipitation -R99p) precipitation extremes, thus highlighting the underlying uncertainty of GCMs in capturing regional drought and flood conditions. Typically for both precipitation and temperature extremes, UKESM1-0-LL, FGOALS-g3, and GCMs from Met office Hadley Centre (HadGEM3-GC31-MM and HadGEM3-GC31-LL) and NOAA (GFDL-ESM4 and GFDL-CM4) consistently tend to show good performance. Our results also show that GCMs from the same modelling group and GCMs sharing key modelling components tend to have similar biases and thus are not highly independent.

Dynamical downscaling (DD), and machine learning (ML) based techniques have been widely applied to downscale global climate models and reanalyses to a finer spatiotemporal scale, but the relative performance of these two methods remains unclear. We implement an ML regression approach using a multi-layer perceptron (MLP) with a novel loss function to downscale coarse-resolution precipitation from the Bureau of Meteorology Atmospheric high-resolution Regional Reanalysis for Australia from grids of 12–48 km to 5 km, using the Australia Gridded Climate Data observations as the target. A separate MLP is developed for each coarse grid to predict the fine grid values within it, by combining coarse-scale time-varying meteorological variables with fine-scale static surface properties as predictors. The resulting predictions (on out-of-sample test periods) are more accurate than DD in capturing the rainfall climatology, as well as the frequency distribution and spatiotemporal variability of daily precipitation, reducing biases in daily extremes by 15%–85% with 12 km prediction fields. When prediction fields are coarsened, the skill of the MLP decreases—at 24 km relative bias increases by ∼10%, and at 48 km it increases by another ∼4%—but skill remains comparable to or, for some metrics, much better than DD. These results show that ML-based downscaling benefits from higher-resolution driving data but can still improve on DD (and at far less computational cost) when downscaling from a global climate model grid of ∼50 km.