1. Introduction

Regional climate modelling plays a critical role in understanding and projecting the complex climate processes that influence Australasia, which is a region encompassing Australia, New Zealand, and neighboring Pacific and Southeast Asian areas. The diverse geography of this region including arid interiors, mountain ranges, maritime tropics, and temperate coasts, generates highly variable climatic conditions that global climate models (GCMs) cannot fully resolve. This limits the effectiveness of GCMs for regional climate change resilience planning, which are the scales at which climate change impacts are experienced, and mitigation and adaptation actions are implemented [1]. Downscaling techniques, such as regional climate models (RCMs) and variable resolution GCMs (VRGCMs), bridge this gap by providing fine-resolution climate information that is essential for assessing regional climate change impacts, informing adaptation strategies, and guiding policy decisions [2]. Over the past fifteen years, regional climate modelling in Australasia has evolved from a series of isolated efforts into a coordinated scientific enterprise producing multi-model ensembles and long-term datasets at higher resolution for research and applications.

At the centre of this coordinated effort is the Coordinated Regional Climate Downscaling Experiment (CORDEX), an international initiative launched under the World Climate Research Programme (WCRP). CORDEX provides a common framework for regional climate downscaling around the world, facilitating the comparison, evaluation, and application of regional climate projections [3]. By defining standard domains, experiments [4], and evaluation and data protocols, CORDEX has enabled a step change in the consistency and reproducibility of regional projections. The CORDEX Australasia domain (Fig 1) is the regional expression of this global initiative, bringing together multiple institutions across Australia, New Zealand, and internationally [5]. Its outputs have supported a wide range of studies, from hydrological and agricultural impact assessments [6, 7] to the development of national climate services [8].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. CORDEX-Australasia domain showing the topography and bathymetry.

The base-map coastlines were obtained from Natural Earth 1:10m public-domain vector data and rendered using Cartopy. www.naturalearthdata.com.

https://doi.org/10.1371/journal.pclm.0001066.g001

While CORDEX provides a valuable international framework, significant advances in regional climate science in Australasia have also emerged from non-CORDEX efforts. Various studies have contributed to methodological innovation [9], the development of new datasets [10], and the exploration of previously underrepresented phenomena such as tropical convection [11], extreme heat [12], and compound events [13]. Including these efforts in this synthesis ensures a comprehensive account of the region’s scientific progress and highlights the complementary roles of CORDEX and related modelling programs.

This paper reviews the progress made since the initiation of CORDEX in 2009. It synthesises contributions to CORDEX-Australasia and related modelling efforts, identifying how regional collaboration has improved the robustness and utility of climate information. While the CORDEX framework includes Empirical-Statistical downscaling (ESD), there is no application of ESD that covers the CORDEX-Australasia domain and the many applications of ESD within subregions of the CORDEX region are considered outside the scope of this study. Broader discussions on downscaling and regionalising projections are found in various reviews [3,14,15]. We first describe significant projects in regional climate downscaling in Australasia. We then summarise the achievements within CORDEX Australasia and the scientific insights gained. This is followed by an outline of current progress, challenges and gaps, and future directions that can address these gaps. By consolidating fifteen years of regional climate downscaling research, this paper provides a foundation for the next phase of high-resolution climate projections in Australasia, aimed at delivering more consistent, actionable information for adaptation and risk management.

2. Evolution of regional climate modelling in Australasia

From the moment Limited Area Models were first proposed as tools for downscaling global climate models for climate applications in the late 1980s [16–18], researchers in Australia were among the first to apply regional climate models [19–23], positioning the region at the forefront of dynamical downscaling efforts. One of the earliest large-scale dynamical efforts was the CSIRO (Commonwealth Scientific and Industrial Research Organisation) regional model, which provided climate projections for Australian states at ~60 km resolution [22,24], with a CSIRO-NIWA (National Institute of Water and Atmospheric research) collaboration later extending this to New Zealand [25,26]. In parallel, the Australian Bureau of Meteorology (BoM) developed analogue-based statistical downscaling techniques [27], complemented by state-based initiatives, illustrating the diversity of methods being explored prior to coordinated frameworks.

The late 2000s and early 2010s saw a marked increase in the scale and ambition of dynamical downscaling activities. Programs such as Climate Futures for Tasmania represented a step change by combining multiple emissions scenarios, dynamical downscaling, and bias correction to quantify uncertainty in future climate projections [28]. Here, uncertainty refers to the degree to which the future state of the climate cannot be determined uniquely, reflecting the combined effects of scenario uncertainty, model uncertainty, internal climate variability, and observational limitations [29]. Targeted tools such as weather generators and sector-specific downscaling systems were developed to support agricultural and resource management applications, particularly in New South Wales [30] and Queensland. The NARCliM (New South Wales and Australian Regional Climate Modelling project) project delivered a large multi-model dynamical ensemble explicitly designed to comply with the emerging CORDEX protocol, including experimental design, ensemble construction, and data standards, first at ~50 km for NARCliM 1.0 and 1.5 [31,32] and then 20 km for NARCliM2.0 [33]. The most recent iteration, NARCliM 2.0, is a collaborative initiative contributing to CORDEX-CMIP6 (Coupled Model Intercomparison Project phase 6) led by the New South Wales (NSW) state government in partnership with universities (University of NSW and Murdoch University), state and territory governments (Australian Capital Territory, South Australia, Western Australia, Victoria), and the National Computing Infrastructure (NCI). Notably, it produces projections at convection-permitting resolution (4 km) over southeastern Australia, capturing approximately 70% of the Australian population. The Queensland Future Climate Science Program (QldFCP) is a collaboration between the Queensland Government and the University of Queensland. Using CCAM (Conformal Cubic Atmospheric Model), in 2015–2016, QldFCP-1 downscaled CMIP5 at ~50km resolution, while in 2025 QldFCP-2 contributed to CORDEX-CMIP6 at ~10km resolution.

Regional climate modelling studies also addressed other parts of the CORDEX-Australasia domain, including the Maritime Continent and Pacific archipelagos. In New Zealand, Earth Sciences New Zealand (formerly ‘NIWA’) undertook systematic dynamical downscaling experiments using the Unified Model, initially from CMIP3 and later extended to CMIP5 [34–37]. More recently, the Conformal Cubic Atmospheric Model (CCAM) was applied for CMIP6 downscaling at 12 km resolution over New Zealand and at ~12–30 km resolution over the wider South Pacific [38–41]. High-resolution downscaling has also been performed for some Pacific Island regions within the CORDEX Australasia domain [42,43] but no studies to date have used CORDEX-Australasia to investigate climate change over Pacific Islands. The downscaled CMIP6 dataset enabled the development of a generative-AI-based regional climate model emulator that operates orders of magnitude faster than RCMs and VRGCMs while demonstrating considerable skill in simulating climate change signals for precipitation means and extremes [44–46]. Other targeted efforts examined tropical cyclones [47,48] and East Coast Lows [49,50], atmospheric rivers [51], orographic rainfall [52], heat extremes [53] and drought [54].

By the early 2010s, the growing number of independent modelling efforts had generated a wealth of regional climate information but also highlighted challenges related to consistency, comparability, and data accessibility. It was within this context that CORDEX-Australasia was established, building on more than two decades of regional modelling experience and providing a common framework alongside focused initiatives targeting higher resolution, specific applications, and under-represented regions. These include the BoM Atmospheric Regional Projections for Australia (BARPA) system (based on an adaptation of the Met Office’s Unified Model for the Australasia region) developed in the Energy Sector for Climate Information (ESCI) project from 2019 to 2021 [55], the Western Australian Climate Science Initiative (CSI) - which extends NARCliM 2.0 to convection-permitting resolution over Western Australia - and the Australian Climate Service (ACS), established in 2021, which integrates expertise from national agencies to link regional CMIP6-based projections from CCAM and BARPA with hazard intelligence and socio-economic information. The recently established National Partnership for Climate Projections strengthens communication, coordination and collaboration amongst the regional modelling groups in Australia. Together these initiatives reflect a dynamic and mature regional climate modelling landscape in Australasia (Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Timeline of significant regional climate projections efforts in the CORDEX-Australasia region.

Details of individual efforts can be found in the supplementary material.

https://doi.org/10.1371/journal.pclm.0001066.g002

3. CORDEX Australasia: Achievements and outcomes

CORDEX–Australasia advances the overarching CORDEX objectives by: (i) improving understanding of regional climate variability and change; (ii) evaluating and refining regional climate models (RCMs); (iii) producing coordinated Australasian climate projection datasets; and (iv) facilitating knowledge exchange between modelling groups and end users. In practice, this has supported the development of multiple RCM systems, shared analysis frameworks, and collaborative investigations of alternative downscaling strategies, model parameterisations, and their implications for regional projections.

The CORDEX-CMIP5 Australasia ensemble comprised 20 simulations generated by downscaling 11 CMIP5 global climate models (GCMs) with six RCMs [5]. While overall this was an ensemble of convenience, parts of it were designed to account for model performance, independence and the spread of future change in temperature and precipitation across the CMIP5 ensemble [31]. Evaluation was performed against a gridded observation dataset over Australia with varying station density supporting it [56]. Station density is sparse over north-western Australia and this area is masked when the density is considered too low. Over the historical reference period (1976–2005), the ensemble reproduced the principal features of observed climate but exhibited systematic biases, including underestimation of daily maximum temperature, overestimation of minimum temperature (and thus a reduced diurnal temperature range), and excessive precipitation over southern Australia. Despite these biases, the RCMs generally added value relative to their driving GCMs [57]. Here “added value” refers to improvements in the representation of climate statistics by a regional climate model relative to the driving data, which may be a global reanalysis or a coarser-resolution climate model simulation. Under the high-emissions scenario RCP8.5, late 21st-century projections indicate substantial warming across Australasia, reaching ~5 K over inland Australia with smaller coastal increases. Precipitation projections suggest predominantly drier conditions, particularly in southwestern and southeastern Australia during the cool season. Summer changes are weaker and regionally variable, with some areas indicating potential increases.

The current CORDEX-CMIP6 ensemble includes contributions from five modelling groups, producing 40 simulations by downscaling 19 CMIP6 GCMs with six RCMs over the Australasian domain [58]. For CORDEX-CMIP6, ensemble design preferentially sampled different GCMs rather than multiple ensemble members of the same model, with certain key models sampled by all RCMs, in a ‘sparse matrix’ approach [8,59], as supported by previous studies [60,61]. Horizontal resolution ranges from 12.5 to 20 km; for intercomparison, outputs are remapped to a common AUS-20i grid at approximately 20 km resolution. All models met established performance benchmarks for the historical period (1985–2014), although systematic biases remain [62] when measured against gridded observations [56], acknowledging that setting benchmarks is challenging and currently subjective. Notably, all RCMs demonstrated added value in simulating daily minimum temperature [63]. Direct comparison between NARCliM CMIP6 to CMIP5 generation downscaling has shown improvements in the simulation of maximum temperature and precipitation [64,65]. Analysis of CMIP6-based projections is ongoing. Under SSP3–7.0, Australian temperatures are projected to increase by approximately 2–5 K by the late 21st century, broadly consistent with the driving GCMs (Fig 3). This temperature increase is slightly less than that projected by the CORDEX-CMIP5 ensemble under the RCP8.5 emission scenario, primarily due to the lower forcing. Analysis of heat extremes under high forcing by the end of the century indicates results similar to those from CMIP5 projections, where extremes scale roughly similarly or in some cases more than the mean [66]. Assessments of sub-ensembles indicate minimum temperature extremes increasing most in northern Australia and maximum temperature extremes increasing most in central and southern Australia, which is similar to the CORDEX-CMIP5 ensemble [67]. Projected changes in the full CORDEX-CMIP6 ensemble indicate widespread winter (JJA) drying, while summer (DJF) precipitation responses show substantial inter-model disagreement similar to the precipitation projections in CORDEX-CMIP5. This inter-model disagreement results in the projections of annual precipitation extremes having less confidence than temperature changes. The CORDEX-CMIP6 ensemble indicates larger increases in precipitation extremes in northern Australia [67] in contrast to the CORDEX-CMIP5 ensemble that indicated larger (non-significant) increases in southern Australia [5]. RCMs driven by the same GCM tend to exhibit similar, though not identical, responses [58]. Projections of rare extreme rainfall show greater consistency, with increases to 1-in-100 year events expected across Australia and its capital cities [68].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Projection of area-averaged temperature change for SSP3-7.0 and SSP1-2.6 from CMIP6 and CORDEX for the Eastern Australia (EA), showing time series for an example region (a) and bar plot for the 2080-2099 period, each showing the 10-90% model range (b); (c,d) as for (a,b) but rainfall change in %.

Here 35 members of the CORDEX-CMIP6 ensemble are used for each SSP. This is created from 15 GCMs being downscaled by 5 RCMs in a sparse matrix. Observations for temperature are from ACORN-SATv2.6 and observations for rainfall are from AGCD. Change is from the 1995-2014 baseline. All data are available through the Earth System Grid Federation (https://metagrid.esgf-west.org/search/cmip6/) or Australia’s National Computational Infrastructure (https://opus.nci.org.au/spaces/CMIP/pages/213713626/CORDEX-CMIP6+and+Regional+Downscaling+Datasets).

https://doi.org/10.1371/journal.pclm.0001066.g003

Comparison of the added value between the CORDEX-CMIP6 and CORDEX-CMIP5 ensembles show that CORDEX-CMIP6 has higher added value for maximum and minimum temperature extremes but less spatial consistency of added value in precipitation extremes [57,63]. This shows that, despite increased resolution, precipitation processes, including convection and circulation related phenomena, remain challenging to simulate.

CORDEX-Australasia employs a “sparse matrix” design in which multiple RCMs downscale multiple GCMs, though several higher-performing GCMs were downscaled by all modelling groups. GCM selection is based on regional performance metrics, including temperature, precipitation, representation of key modes of variability (e.g., El Nino-Southern Oscillation, Indian Ocean Dipole, Southern Annular Mode), storm tracks, and drought indices, as well as the availability of sub-daily output [8,33]. Collaboration among regional modelling teams has supported coordinated evaluation studies [58,68,69] and, for CMIP6, the development of fit-for-purpose performance benchmarks [62].

4. Scientific Insights and Applications

Numerous studies have quantified the added value (AV) of RCMs over Australasia and advanced methodological frameworks for its assessment [41,57,63,70,71]. While AV is typically evaluated for historical simulations relative to observed reference datasets, the potential added value (PAV) has been estimated from future simulations, extending the AV framework to the climate change signal [72]. The concept of realised added value (RAV) combines historical AV and future PAV into a unified metric [57], later applied to the CORDEX-CMIP6 ensemble [63]. The Integrated Added Value (IAV) framework further incorporates changes at daily and seasonal scales in addition to climatological means [71]. Collectively, these methodological developments support the capacity of RCMs to provide enhanced regional information.

Performance improvements are most pronounced in coastal and mountainous regions, where finer resolution better resolves orography and land-sea contrasts [70]. For example, in Southeast Queensland, which is home to the majority of this state’s population, RCMs improved performance by approximately 150% relative to their driving GCMs [71]. Enhanced representation of seasonal rainfall contrasts in Tasmania [73], and improved simulation of rainfall extremes [28] have similarly been attributed to better-resolved topography. Convection-permitting models further refine land-surface processes and explicitly simulate convective moisture fluxes, reducing reliance on parameterisation schemes [67,74], and providing a better representation of extreme wind gusts [75]. In New Zealand, kilometre-scale models improve the simulation of orographic enhancement of both mean and extreme rainfall [52]. Higher temporal resolution also enables a better detection of synoptic scale processes like east coast lows [69].

RCMs have been shown to provide a better representation of teleconnections with climate modes compared to their driving global models [39,76,77]. However, because RCMs inherit large-scale boundary conditions from GCMs, deficiencies in external climate drivers are typically retained. For example, in the CORDEX–CMIP6 ensembles (BARPA, CSIRO-CCAM, and NARCliM2.0), the relationship between Australian spring precipitation and the Niño3.4 index largely reflects that of the host GCMs. Bias correction of boundary conditions, as applied in QldFCP-2 [78], can reduce these inherited errors.

The Australian monsoon, characterised by strong internally generated variability [79], remains challenging for RCMs to capture. Many models, including BARPA-C, QldFCP-2, and NARCliM2.0, systematically underestimate northern Australian summer rainfall [71,74,80]. In CCAM, this bias has been linked to overestimation of wind and moisture fluxes north of Australia [77].

Multi-model benchmarking clarifies the implications of differing downscaling strategies and helps identify priorities for model development. For example, comparisons between CSIRO’s CCAM configuration employing spectral nudging and the QldFCP-2 CCAM configuration using bias-corrected SSTs [78] show that the latter is less influenced by biases in the driving CMIP6 models and exhibits greater independence in projected trends. However, it also displays persistent systematic biases not evident in the nudged configuration [62]. Similarly, the use of alternative physical parameterisations in NARCliM2.0 provides insight into boundary-layer processes and their influence on projected regional climate change.

5. Ongoing progress and integration

The transition from CMIP5 to CMIP6, and from Representative Concentration Pathways (RCPs) to Shared Socioeconomic Pathways (SSPs), presents both opportunities and challenges for the production of application-ready climate projections. CMIP6 introduces updated model physics, higher resolution, and expanded scenario design, but differences in scenario structure complicate continuity with earlier impact assessments. Under CMIP5, emphasis was placed primarily on RCP4.5 and RCP8.5, whereas CMIP6 centres on SSP1-2.6, SSP2-4.5, and SSP3-7.0, alongside the introduction of the Global Warming Levels framework to complement time-slice analysis. The limited number of SSP5-8.5 experiments conducted for CMIP6 has been important for maintaining partial continuity with RCP8.5-based applications across modelling generations.

The National Computational Infrastructure (NCI) plays a central enabling role in Australian regional climate modelling. NCI hosts an Earth System Grid Federation node providing access to CMIP datasets and Australian-produced CORDEX simulations, ensuring both domestic accessibility and international discoverability. Co-located high-performance computing and curated shared storage allow efficient analysis, post-processing, and visualisation of large multi-model ensembles. This integrated environment supports coordinated experimental design, benchmarking, and collaborative evaluation across universities and government agencies, and underpins much of Australia’s regional downscaling activity.

The National Partnership for Climate Projections (NPCP) has been established to strengthen coordination across Australia’s climate projection community. The NPCP provides a national framework to align modelling strategies, scenario selection, benchmarking standards, and data delivery protocols across institutions. Its objectives include improving transparency in experimental design, enhancing consistency between modelling efforts, reducing duplication, and facilitating clearer communication of projection uncertainty to end users. The NPCP also promotes collaboration between modelling groups, data providers, and sectoral stakeholders, helping to translate advances in climate science into decision-relevant information. By fostering communication and shared technical standards, the NPCP supports a more coherent national projection capability including global model evaluation and regional downscaling.

CORDEX Flagship Pilot Studies (FPS) address targeted scientific challenges of regional and international relevance. CORDEX-Australasia has a leading role in two such initiatives. The Sub-Hourly Extreme Precipitation (SHEP) project evaluates the feasibility and robustness of projecting sub-hourly rainfall extremes using convection-permitting models over major Australian and New Zealand cities. Here robustness is the extent to which a projected climate change signal is consistently reproduced across an ensemble of climate models and is statistically distinguishable from internal climate variability and model uncertainty. This multi-institutional effort responds directly to stakeholder demand for improved information on short-duration extreme rainfall and associated hazards such as flash flooding. The Island Climate - Pacific (IC-Pac) project focuses on developing and assessing downscaling methodologies tailored to the needs of Pacific Island countries, where data scarcity, complex topography, and high exposure to climate risks require carefully designed regional approaches. Both initiatives involve close collaboration among modelling centres and engagement with governments, reflecting the increasing integration of scientific development with user needs across the Australasian and Pacific regions.

6. Challenges and gaps

Despite substantial advances over the past fifteen years, regional climate downscaling in Australasia continues to face several persistent challenges and gaps. Foremost among these are technical and computational constraints. High-resolution, multi-decadal regional climate simulations require sustained access to national-scale high-performance computing infrastructure. They also generate very large data volumes, often reaching petabytes per project (e.g., NARCliM2.0 produced ~15PB). Long-term storage, curation, analysis, and accessibility of these datasets remain limiting factors, particularly for smaller institutions and downstream users. In addition, limited resources and focus are directed towards development and testing of regional climate models.

Observational datasets that are used in model evaluation have their own uncertainties due to sparse station coverage, sensor limitations, and interpolation techniques used to grid the data [115,116]. This observational uncertainty has rarely been considered explicitly in these studies. Not all climate variables are observed well enough, i.e., with sufficient spatial and temporal coverage, to create gridded products for model evaluation and hence reanalysis may be used as a point of comparison in these cases. However, this should be accompanied by acknowledgment of the limitations of the reanalyses [117,118].

Model biases remain a central scientific challenge. Although RCMs often improve the representation of topography, coastlines, and extremes, they can inherit systematic biases from driving GCMs and introduce their own additional biases [5,64,118]. This highlights the importance of treating downscaling as part of a modelling chain that may include bias correction. In several studies, downscaling has not consistently reduced uncertainty in precipitation projections and, in some cases, has increased ensemble spread, particularly for convective rainfall and regions of complex terrain [40,57,58]. In addition, limited availability of suitable (e.g., subdaily) GCM data constrains regional modelling, underscoring the need for better coordination of future CMIP datasets.

Identifying robust regional climate change signals remains challenging. While temperature projections are generally robust and spatially coherent, precipitation responses are often heterogeneous and sensitive to model configuration and driving GCM choice, complicating interpretation and communication [46,101,102]. Credible regional projections require AV, PAV, and RAV signals that are consistent across RCMs and clearly linked to physical drivers such as orography. Australian CMIP5 regional modelling suggested enhanced summer rainfall over the Alps and cool-season drying on windward slopes, supported by observations [106], but these signals are not clearly reproduced in some CORDEX-CMIP6 projections [33]. This illustrates the challenge in identifying clear consistent insights, and the need for a multi-model ensemble. Despite improved coordination through CORDEX and national initiatives, differences in experimental design and data delivery still limit synthesis. CORDEX modelling must of course remain aware of the tradeoffs among spatial and temporal resolution, spatial coverage, model complexity, and ensemble size to achieve its goals, and the interactions with types of models in the ‘multiverse of modelling approaches’ [119].

Finally, a persistent science-policy gap remains. Although regional projections increasingly underpin impact assessments and climate services [120], translating complex ensemble information into clear, decision-ready guidance remains challenging [8,121,122]. Conflicting outcomes across RCMs and the availability of different model subsets across platforms create confusion and hinder the use of consistent, credible data. Efforts to improve uncertainty communication and ensemble interpretation in terms that are understandable and actionable by decision-makers are needed. Bridging this gap will require earlier and deeper engagement between modellers, impact scientists, and decision-makers, alongside stronger climate literacy and better coordination in the delivery, interpretation, and guidance for the use of downscaled climate information.

7. Future directions

Further development and use of regional climate models remains a focus. These models simulate the processes that underpin the climate system and allow interrogation of the processes causing change. Understanding these processes provides new avenues to evaluate the models and can increase our confidence in the simulated future change. This allows us to test whether the models are getting the right answer for the right reasons.

A major research priority is the progression toward kilometre-scale resolution. Convection-permitting models provide demonstrable improvements in the simulation of convective processes, extreme precipitation, and mesoscale circulation features, and are particularly valuable for urban climate applications, land-use change studies, and complex terrain [10,52,123,124]. High spatial and temporal resolution are essential for representing sub-hourly extreme precipitation which requires explicit treatment of convective dynamics and improved representation of microphysical processes.

Initiatives such as the CORDEX Flagship Pilot Study (FPS) on SHEP exemplify this next phase of development. Convection-permitting, high-frequency simulations over major Australian and New Zealand cities enable explicit representation of short-duration extremes and investigation of small-scale processes, including urban influences. Through co-design with key stakeholders, the SHEP FPS will strengthen links between climate projections, flood modelling, and adaptation planning.

Enhancing model complexity through improved coupling with additional Earth system components is another key direction. The inclusion of interactive three-dimensional ocean models has improved the representation of coastal climates, air-sea interactions, and marine heatwaves - critical for Australia [125,126]. Further coupling with hydrological, vegetation, and aerosol processes is expected to increase physical consistency and realism in regional projections [127].

Given that even the best performing GCMs have biases that are passed to the RCM via the boundary conditions, bias-correcting GCM boundary conditions before they are used in dynamical downscaling is an avenue for future research. Recent work has demonstrated that bias correction of the GCM data before use results in improved RCM simulations [9,78,128–135] and can reduce the range, standard deviation and mean absolute deviation of the change in climate, especially for temperature variables over land [136]. However, the choice of physical parameterisations within RCMs has a larger influence on model performance under historical climate, as well as future projections, than bias correction, highlighting that model physics remains key [137].

Although higher resolution improves the realism of extreme events, their rarity complicates robust estimation of future changes. Addressing this challenge will require large ensembles, targeted ensemble design, and efficient sampling strategies that maximise information content within computational constraints. These approaches are essential for quantifying uncertainty and delivering actionable climate information.

Expanding the range of driving GCMs and ensemble members would better sample plausible futures, but computational costs limit comprehensive downscaling. Coordinated efforts among modelling groups, with integration of outputs into a shared “super-ensemble,” offer a pathway to improved robustness. Complementary machine-learning-based downscaling approaches provide computationally efficient means to extend ensemble size, explore internal variability, and emulate RCM behaviour [45,46,138]. While artificial intelligence/machine learning (AI/ML) methods show promise, particularly for extremes, their physical consistency and reliability under extrapolation beyond training data require careful evaluation and benchmarking [139,140]. Questions concerning the interpretability and uncertainty quantification of these methods also need to be addressed [141]. Nevertheless, the use of AI/ML within model parametrisations and hybrid dynamical-statistical approaches to ensemble generation represent promising medium-term research directions.

Advances in regional projection fidelity depend equally on improved observational datasets. Sustained investment in high-quality observational analyses and high-resolution reanalyses is critical for resolving regional processes and extremes, supporting rigorous model evaluation, and enhancing bias-correction frameworks. Improved observational constraints with uncertainty quantification will increase confidence in both dynamical and data-driven approaches.

CORDEX-Australasia includes several remote Australian islands and Pacific island nations that are vulnerable to the impacts of climate change [142]. Extending high-quality projections and tailored analyses to these under-represented regions is essential for strengthening regional collaboration, supporting adaptation, and building technical capacity. This could be achieved by supporting and contributing to the CORDEX FPS Island Climate - Pacific (IC-Pac).

Looking ahead to CMIP7, designing an optimal regional ensemble will require coordinated selection of GCMs, ensemble members, RCM configurations, emulation techniques, and bias-correction strategies. Although increasingly resource-intensive, such an integrated and strategically designed framework offers the potential to provide more balanced sampling of uncertainty and more robust, decision-relevant climate information. This work will be done in coordination with CORDEX and other World Climate Research Program activities.

Ongoing strategic thinking and planning are needed, including on the relative strengths, weaknesses and best use of dynamical, statistical and AI/ML models, and how the ‘multiverse’ of modeling is evolving.

8. Conclusions

Over the past fifteen years, CORDEX-Australasia has established a more coordinated and scientifically robust regional climate downscaling capability for Australasia. Through successive CMIP5 and CMIP6 ensembles, the initiative has delivered high-resolution, multi-model projection datasets, advanced evaluation and benchmarking frameworks, and demonstrated clear added value over driving global models.

A defining achievement has been the synergy between CORDEX and non-CORDEX efforts. National and state-based initiatives - including BARPA, CCAM, NARCliM2.0, QldFCP-2, the Australian Climate Service, and the Western Australian Climate Science Initiative - have operated both within and alongside CORDEX, strengthening experimental design, expanding ensemble coverage, and accelerating methodological innovation. Shared infrastructure provided by the National Computational Infrastructure (NCI) and coordination mechanisms such as the National Partnership for Climate Projections (NPCP) have further integrated modelling, evaluation, and service delivery across institutions. Together, these collaborations have fostered a mature and dynamic regional modelling ecosystem that links research, operations, and applications.

Key findings from the current generation of CMIP6-downscaled ensembles indicate robust multi-model agreement on substantial warming through the 21st century, with uncertainty dominated by scenario choice, while precipitation changes remain more uncertain except in specific regions such as southwestern Australia. Regional models continue to add value, particularly for extremes and finer-scale climate features, relative to their GCM drivers.

Looking ahead, strategic priorities for the next 15 years are clear. First, continued progression toward kilometre-scale, convection-permitting modelling will be essential for improving the representation of extreme events and supporting impact-based applications. Second, enhanced Earth system coupling and expanded ensemble strategies, including coordinated “super-ensembles” and AI-enabled emulation, will be needed to better characterise uncertainty. Third, sustained development of benchmarking frameworks, bias-correction methods, and observational datasets will remain critical for maintaining projection credibility. Fourth, extending high-quality regional projections to vulnerable Pacific Island nations should remain a priority, strengthening regional scientific capacity and climate resilience.

Achieving these goals will require deeper cross-jurisdictional collaboration. Building further alignment across Commonwealth agencies, states and territories, universities, and service providers will improve efficiency, reduce duplication, and ensure nationally consistent climate information. Coordinated ensemble design, shared infrastructure, and common communication strategies will be particularly important as the community prepares for CMIP7 and beyond.

Continued collaboration across modelling groups, operational agencies, and end users will ensure that advances in regional climate science translate into actionable, decision-relevant information. With coordinated effort and long-term commitment, partners contributing to CORDEX-Australasia are well positioned to deliver the next generation of robust, high-resolution climate projections needed to support adaptation, risk management, and resilience across Australasia and the Pacific.

Supporting information

Acknowledgments

Part of this research was undertaken with the assistance of resources from the National Computational Infrastructure (NCI Australia), an NCRIS enabled capability supported by the Australian Government. We would also like to thank the CSIRO HPC (Petrichor) and the Pawsey Supercomputing Research Centre for the computational resources for CSIRO CCAM simulations and WRF simulations over Western Australia.