To address the first objective of this study, this section documents the experiment design of the CMIP6 dynamically downscaled climate projection over SEA. Table 1 lists the CMIP6-downscaled simulations generated within the CORDEX-SEA framework. At the time of this analysis (September 2025), 62 simulations (denoted by cross) were available and are included in this study. As of July 2026, the CORDEX-SEA archive has expanded with additional simulations (e.g., denoted by ticks) completed and others still being produced (e.g., CMIP6 CORDEX-CORE) as part of this ongoing modelling effort.

These comprise three regional climate models (RCMs), eight model configurations, and eleven driving CMIP6 GCMs running across historical and future scenarios. A detailed description of the model configurations is provided in Table 2. The subset of GCMs was selected either based on their performance in reproducing key climatological features (Desmet and Ngo-Duc, 2022) or their ability to simulate dominant precipitation drivers (e.g., the monsoon) and teleconnections with major climate modes (e.g., IOD, ENSO), while considering model independence and the spread of future projections (Nguyen et al., 2024).

Most simulations cover the extended SEA domain (17.3° S–31.9° N, 86.5–149.3° E), which fully covers Myanmar, except for the CCAM-2021 simulations that span a slightly smaller region (14.8° S–27° N, 89.5–146.5° E). Simulation periods vary slightly across models but generally extend from 1960 to 2099, following the CMIP6 CORDEX Experiment Guidance (CORDEX, 2021), and include both SSP1-2.6 and SSP3-7.0 scenarios.

Eight RCM configurations are utilised in this study (Table 2). Non-hydrostatic RegCM (RegCM4-NH; Coppola et al., 2021) simulations were conducted by several modelling institutions in SEA. Among the 44 physical parameterisation combinations conducted using RegCM4-NH in the sensitivity analysis phase with ERA5 forcing, two configurations – EXP16 and EXP28 – were selected based on their performance in reproducing near-surface temperature and precipitation (Ngo-Duc et al., 2024). Both use the University of Washington (UW) PBL scheme (Bretherton et al., 2004), but differ in their convective parameterisations: EXP16 employs the Tiedtke scheme (Tiedtke, 1989), while EXP28 uses the Kain–Fritsch scheme (Kain, 2004). Resolved-scale precipitation is represented using the SUBEX scheme (Pal et al., 2000). Other parameterisations follow the default RegCM4-NH setup, including the Community Climate Model version 3 (CCM3) radiative transfer scheme (Kiehl et al., 1998), the Community Land Surface Model version 4.5 (CLM4.5; Lawrence et al., 2019) and the Zeng ocean flux scheme with roughness option 1 (Zeng et al., 1998). Recently, RegCM version 5 (RegCM5) has been released, featuring a new non-hydrostatic dynamical core derived from the MOLOCH weather prediction model (Giorgi et al., 2023). Accordingly, two additional experiments were performed using this configuration.

The Conformal Cubic Atmospheric Model (CCAM) is a global stretched-grid atmospheric model developed by the Commonwealth Scientific and Industrial Research Organisation (CSIRO), in which the target region is simulated at higher resolution (McGregor, 2005, 2015; McGregor and Dix, 2008; Nguyen et al., 2012; Katzfey et al., 2016; Truong et al., 2025). CCAM employs a non-hydrostatic semi-Lagrangian dynamical core with multiple physical parameterisation options (Truong and Thatcher, 2025). Three CCAM configurations were applied by two institutions: the Climate Change Research Centre at the University of New South Wales (UNSW-CCRC) and CSIRO Australia.

The UNSW-CCRC CCAM configuration uses a C192 stretched grid, providing 25 km resolution over Southeast Asia. It employs 54 vertical atmospheric levels (from 20 m to 40 km) and 40 oceanic levels (to 5 km depth). Spectral nudging (Thatcher and McGregor, 2009; Truong et al., 2026) is applied to large-scale wind, temperature, and surface pressure fields from the host GCMs (e.g., ACCESS-CM2 and EC-Earth3-Veg) at a 6-hourly frequency and scales of ∼ 3000 km, beginning around 850 hPa (∼ 1.5 km above the surface). The UNSW-CCRC CCAM 2017 and 2021 versions differ in their convective and planetary boundary layer (PBL) schemes: the 2017 version uses a prognostic Turbulent Kinetic Energy (TKE)-based PBL scheme (Hurley, 2007), whereas the 2021 version applies a diagnostic Richardson-number-based scheme and an updated mass-flux convective closure including downdraft, entrainment, and detrainment processes (McGregor, 2003).

The CSIRO-CCAM configuration follows an AMIP-style setup driven by bias- and variance-corrected sea-surface temperature (SST) and sea-ice fields from CMIP6 GCMs (Hoffmann et al., 2016). It employs a C96 stretched grid and a modified version of the 2021 convective scheme (Mod2021), with 27 vertical levels extending from 20 m to 40 km.

Both UNSW-CCRC and CSIRO-CCAM share similar subsystems for the ocean and land surface (e.g., the Community Atmosphere Biosphere Land Exchange System – CABLE; Kowalczyk et al., 2006) and a prognostic aerosol scheme implemented within the Simplified Edwards–Slingo four-band (SE4) (Freidenreich and Ramaswamy, 1999; Schwarzkopf and Ramaswamy, 1999) radiation scheme based on CSIRO-Mk3.6 aerosols.

The UK Met Office RCM HadREM3-GA7.05 (Tucker et al., 2022) is a limited-area configuration of the global GA7.05 model (Walters et al., 2019) with some modifications described in Buonomo et al. (2024). The RCM has a grid with 0.11° × 0.11° horizontal resolution (∼ 12 km), a 4 min timestep, and 63 levels (with an upper lid at ∼ 39 km). It is driven at its lateral boundaries using a one-way nesting approach every 6 h using variables from the driving GCMs (UKESM1-0-LL, EC-Earth3-Veg and NorESM2-MM). The RCM adjusts to the surface and lateral boundary forcing across a 13-point external rim. No nudging or parameter perturbations were used in either model configuration. In HadREM3-GA7.05, convection is parameterised using a mass-flux approach based on Gregory and Rowntree (1990). Large scale microphysics is a single-moment scheme derived from that of Wilson and Ballard (1999), and large-scale cloud is based on the prognostic cloud fraction and prognostic condensate PC2 scheme. The land surface model used is the Joint UK Land Environment Simulator (JULES) (Best et al., 2011; Clark et al., 2011). Aerosol radiation and cloud effects are derived from the MACv2-SP data set for the historical (Stevens et al., 2017) and SSP3-7.0 (O'Neill et al., 2016) scenarios. Monthly 3D fields of shortwave and longwave optical properties (absorption, extinction, scattering, and asymmetry) and cloud droplet number concentration on the 12 km model grid were calculated, and then these were prescribed as a time series in the RCM simulation for use in its calculation of time-varying radiative forcing.

To address second and third objectives of this study on the evaluation and benchmarking of the performance of the CMIP6 CORDEX-SEA ensemble, we use multiple global and regional gridded observational datasets to account for observational uncertainties in model evaluation (Nguyen et al., 2020). It is worth noting that we focus on precipitation and temperature over land, as the temporal coverage of oceanic observational products is limited. The selected datasets are summarised in Table 3. Precipitation datasets include Asian Precipitation-Highly-Resolved Observational Data Integration Towards Evaluation of water resources (APHRODITE) (V1101 and V1101EX) at 0.25° resolution (Yatagai et al., 2012); Rainfall Estimates on a Gridded Network with 1° × 1° resolution (REGEN version Allstns V1 2019; Contractor et al., 2020) and Climate Hazards Group InfraRed Precipitation with Station at 0.25° × 0.25° resolution (CHIRPS version 2.0; Funk et al., 2015) They were chosen based on their demonstrated consistency in capturing daily precipitation and extremes over SEA (Nguyen et al., 2020, 2023). For daily near-surface temperature, we also use APHRODITE_MA V1808 at 0.25° resolution (Yasutomi et al., 2011); Berkeley Earth Surface Temperatures (BEST; Rohde and Hausfather, 2020); the Climate Research Unit (CRU) TS v4.08 dataset at 0.5° resolution (Harris et al., 2020); and the ECMWF Reanalysis v5 (ERA5; Hersbach et al., 2020) at 0.25° resolution. Recognizing the sparse observational network over the Maritime Continent in both global datasets and the regional high-resolution dataset of APHRODITE (Yatagai et al., 2012), we additionally include the Southeast Asia Climate Assessment and Dataset (SACAD) version 2.0 at 0.25° resolution. Providing precipitation, mean, maximum, and minimum temperature at a daily scale, this dataset enhances the station network density and coverage, particularly over Indonesia, through contributions from the Indonesian meteorological institutes (Van Den Besselaar et al., 2017).

Hereafter, APHRODITE is selected as the primary baseline for all main precipitation figures, as it utilises the largest number of rain gauges among global datasets and compensates for the limited spatial coverage of SACAD. Note that for temperature, APHRODITE and CRUTS do not provide maximum, minimum, and mean temperature data at a daily time scale, while SACAD has limited spatial coverage (Fig. S1 in the Supplement). Therefore, BEST is chosen as the reference dataset because it is an observational dataset that provides daily mean, maximum, and minimum temperature data across our Southeast Asia domain (15.14° S–27.26° N, 89.26–127.28° E). It should also be noted that inhomogeneities in the raw observations are not fully corrected in the BEST algorithm, which may introduce systematic temperature biases (e.g., the underestimation of temperature trend and magnitude), as previously reported in Canada (Way et al., 2017). Results from all other observational datasets are included in the Supplement (Figs. S2–S11 in the Supplement), with detailed explanations provided within the main text for intercomparison purposes.

All simulations are analysed over the climatological period 1982–2014, with overlapping coverage of multiple observations and simulations. For trend analysis, the period is extended to 1960–2014 to maximise the common period across all RCM simulations. Prior to assessment, all observational datasets and model outputs are regridded to the common 0.22° models' grid using conservative remapping for precipitation and bilinear interpolation for temperature for a fair comparison.

We evaluated the performance of the CMIP6-driven CORDEX-SEA simulations by comparing simulated and observed annual mean daily temperature (tas) from BEST over 1982–2014 (Fig. 2). Overall, the simulations reproduce the spatial patterns of observed temperature reasonably well, with Pearson's correlations (Scor) exceeding 0.8 in all cases. However, despite this consistency, all simulations systematically underestimate near-surface temperature. The area-averaged bias across terrestrial SEA ranges from -0.58 to -3.04 °C, which is substantially larger than the spread among reference datasets (-0.45 to 0.45 °C; Fig. S1). Seasonally, the cold bias is more pronounced during boreal winter (DJF) than boreal summer (JJA) (Fig. S2a). Most RCMs show statistically significant cold biases between approximately -5 and -1.5 °C across most grid points of the domain. The spatial distribution of these biases is strongly dependent on the RCM rather than the driving GCM. For example, RegCM simulations show cold biases over the mainland and western Indonesia but warm biases over parts of northwestern Mainland SEA (e.g., Myanmar), whereas CCAM and HadGEM3-RA7 simulations exhibit widespread cold biases across nearly the entire region. CCAM simulations, in particular, stand out with substantial cold biases exceeding 3 °C, likely linked to excessive simulated precipitation (see Sect. 3.4). It is worth noting that some driving GCMs (e.g., EC-Earth3-Veg, NorESM2-MM, MPI-ESM1-2-HR) also exhibit cold biases, although these are less pronounced (Nguyen et al., 2024), whereas others (e.g., CMCC-ESM2, ACCESS-CM2, UKESM1-0-LL) exhibit warm biases. These results suggest that systematic errors in daily mean temperature are strongly influenced by the RCM rather than by the driving GCMs. In terms of mean absolute errors (MAE) to observation, CCAM simulations tend to show the biggest errors compared with other simulations, with the biggest difference over the Mainland and during the boreal winter (DJF) (Fig. S3 in the Supplement).

Figure S4 in the Supplement shows the seasonal cycle of near-surface temperature over two subregions: Mainland and MC, highlighting the systematic underestimation by CMIP6 CORDEX-SEA simulations. Overall, the RCMs capture the phase of the seasonal cycle well but tend to underestimate temperature magnitude throughout most months of the year, particularly during the boreal winter (DJF) over the Mainland. While most RCMs realistically simulate both the lower (DJF) and higher (MJJ) peaks over the Mainland (Fig. S4a), several simulations (e.g., RegCM5–MIROC6-exp16 and RegCM4–MIROC–ESM2L-exp28) fail to reproduce these features over MC (Fig. S4b). The mainland exhibits a clearly unimodal seasonal cycle, with a wide amplitude (e.g., 12–26 °C) between lower and higher peaks, whereas the MC displays a bimodal cycle with a narrower range (e.g., 23–26.5 °C).

To better understand the biases in mean temperature, we further examined model performance in daily maximum temperature (tasmax) and minimum temperature (tasmin) (Figs. 3 and 4, respectively). In general, the spatial coherence of RCMs (e.g., Scor) with observations in tasmax is weaker than that for tas, although Scor values exceed 0.7 across models. In addition, many RCMs substantially underestimate tasmax both annually (Fig. 3) and seasonally (Fig. S5 in the Supplement), where regionally averaged tasmax biases vary widely across simulations (e.g., ranging from -4.92 to -0.38 °C at the annual scale; Fig. 3). These are the major contributions to the cold biases in tas found in Fig. 2. Notably, there are significant differences between the Mainland and MC. In particular, while all RCMs show a significant cold bias over MC, most RCMs show a mix of warm biases over the northwest and cold biases over the remainder of the Mainland. Interestingly, simulations from the same RCM setup show very similar spatial biases but differ in magnitude. For example, CCAM-Mod2021 simulations exhibit considerable warm biases north of 15° N, while RegCM simulations show strong warm biases exceeding 4 °C over Myanmar and pronounced cold biases over the MC. Further seasonal and sub-regional analyses reveal that tasmax cold biases intensify (Fig. S5) and expand spatially during boreal summer (figures not shown), whereas during winter these cold biases weaken, often disappearing or even reversing to warm biases over parts of the Mainland. Consistent with these patterns, MAE values for tasmax peak during boreal summer in most simulations (Fig. S6 in the Supplement).

In contrast, biases in daily minimum temperature (tasmin) are generally much smaller in magnitude but exhibit greater spatial heterogeneity across models and grid points than those in tas and tasmax (Fig. 4). CMIP6-downscaled simulations typically underestimate tasmin annually, with regionally averaged biases ranging from -0.04 to -2.52 °C. Spatially, cold biases in tasmin are more pronounced over the northern and eastern Mainland. Meanwhile, over the MC, the signal is more heterogeneous, with weaker cold biases or even warm biases in several simulations (e.g., HadGEM3-RA7, RegCM5-CMCC-ESM-exp28, RegCM4-CanESM5-exp28, and RegCM4-NorESM2-MM-exp28). Seasonally, tasmin cold biases are generally stronger during DJF (Fig. S7 in the Supplement), with MAE values also peaking in the boreal winter (DJF) for most simulations (Fig. S8 in the Supplement). In contrast, during boreal summer (JJA), nearly half of the simulations (10 out of 22) exhibit warm biases in tasmin (Fig. S7).

The ability of the CMIP6-driven CORDEX-SEA simulations to reproduce the whole observed temperature distribution at the daily time scale was further evaluated using quantile–quantile (Q–Q) analysis and ASM (Sect. 2.3.1) (Fig. 5). The higher the ASM, the further away it is from the observations. ASM was calculated for the entire SEA domain in boreal summer and winter (Figs. S9 and S10 in the Supplement, respectively) and for tas, tasmax and tasmin. To provide a detailed assessment across the full distribution, temperature was evaluated across three intervals: the lower half (0–50th percentile), the central-to-upper range (50–99th percentile), and the extreme upper tail (> 99th percentile).

Consistent with the mean bias results, the Q–Q plots for daily mean temperature (tas, Fig. 5a–c) show that most RCM simulations lie below the observational reference line, confirming a systematic cold bias across much of the distribution. These discrepancies are most pronounced at the lower end of the distribution (below the 50th percentile), indicating that models particularly underestimate cooler-than-average conditions. This deficiency is strongest over Mainland SEA and during boreal winter (Fig. S10d–f), consistent with the seasonal and regional bias patterns identified earlier. In contrast, model performance improves at higher percentiles, where simulated temperatures are closer to observations, and in some cases even exceed observed values at the extreme upper tail (> 99th percentile). Model skill varies substantially among simulations. For example, RegCM5-CMCC-CESM2-exp28, HadGEM3-RA7-NorESM2-MM, and RegCM4-MPI-ESM1-2-HR-exp28 consistently rank among the best-performing simulations, with the lowest ASM values across regions and seasons. In contrast, RegCM4-CNRM-ESM2-1-exp16 and CCAM-Mod2021 simulations driven by ACCESS-CM2 and GFDL-CM4 exhibit the largest ASM values, reflecting pronounced cold biases and broader distributional errors (Fig. S3).

These distributional biases in tas are closely linked to deficiencies in simulating tasmax, which shows larger and more systematic departures from observations (Fig. 5d–f). Most RCM simulations fall well below the observational reference line across much of the distribution over SEA and its two subregions, confirming widespread cold biases. These biases are particularly pronounced in the mid-range of the distribution (50th–90th percentiles, corresponding to approximately 32–34 °C), indicating that models struggle to reproduce typical daytime temperature conditions, particularly over the mainland (Fig. 5e). Interestingly, CCAM-Mod2021 simulations exhibit distinct behaviour in simulating the distribution of tasmax, with warm biases emerging at the extreme upper tail (> 99th percentile), particularly over the Mainland. Among CMIP6-downscaled RCMs, CCAM-Mod2021-ACCESS-CM2 generally shows better agreement with observations than most other simulations. Meanwhile, RegCM4-CNRM-ESM2-1-exp16 consistently ranks among the five poorest-performing simulations based on ASM across both annual and seasonal timescales (Figs. S6, S9, and S10 in the Supplement). More broadly, RegCM simulations tend to reproduce the tasmax distribution less accurately than other RCM families. These results highlight the earlier finding that cold biases in tasmax are a primary driver of errors in tas particularly due to the models' inability to accurately represent the central and upper portions of the temperature distribution.

In contrast, the RCMs generally demonstrate better skill in reproducing the distribution of tasmin, although systematic biases remain (Fig. 5g–i). Unlike tas and tasmax, tasmin biases exhibit a distinct pattern, with models tending to underestimate cooler-than-average conditions (below the 50th percentile) and overestimate warmer-than-average conditions (above the 50th percentile). This indicates an overall compression of the temperature distribution, with insufficient representation of the full observed variability. Despite this, several simulations (e.g., CCAM-Mod2021-ACCESS-CM2, CCAM-2017-ACCESS-CM2, CCAM-2017-EC-Earth3-Veg, and HadGEM3-RA7-EC-Earth3-Veg) show relatively good agreement with observations across the full distribution and across multiple reference datasets (Figs. S8–S10). In contrast, some RegCM simulations, particularly RegCM4-CNRM-ESM2-1-exp28 and RegCM5-MIROC6-exp16, exhibit substantially poorer performance, with large ASM values reflecting systematic underestimation across much of the distribution, especially on annual and winter timescales.

Regional differences are also evident in tasmin performance. Most simulations show better agreement with observations over the MC. In contrast, larger errors over the Mainland suggest greater challenges in representing land-surface processes, boundary-layer dynamics, and regional circulation influences that strongly affect nighttime temperature.

Since there are large observational uncertainties in temperature variables over Southeast Asia (Fig. S1), the additional results for multiple reference products are presented in the supplementary material (Figs. S2–S10). It is noted that the temperature from BEST is higher than any other product. In general,, the main conclusion remains the same, regardless of the choice of reference. In summary, all simulations actually capture well the spatial distribution of temperature across SEA. However, cold biases of CMIP6-downscaled RCMs in tas primarily originate from deficiencies in simulating tasmax, which exhibits larger magnitude errors and stronger regional and seasonal variability than those in tas and tasmin. While tasmin biases are generally smaller and more spatially heterogeneous, tasmax biases show more systematic and pronounced errors, particularly over the MC and during JJA. Importantly, across all temperature variables (tas, tasmax, and tasmin), the spatial structure of biases is strongly determined by the RCM, highlighting the dominant role of regional model physics and parameterisations in shaping near-surface temperature performance in CORDEX-SEA simulations.

We then compared the seasonal mean total precipitation on wet days from RCM simulations (1982–2014) against APHRODITE (Figs. 6 and 7). Precipitation biases exhibit larger spatial and inter-RCM variability than temperature biases. Most models capture the spatial distribution of precipitation reasonably well, with 16 out of 21 simulations showing spatial correlations exceeding 0.4. In terms of regional averages, almost all RCMs exhibit wet biases for MJJASO total precipitation, ranging from 3.18 to 158.95 mmyr-1. The only exception is RegCM4-CESM2-exp28, which shows a dry bias of -16.84 mmyr-1 and a low spatial correlation score of 0.08.

Despite generally reasonable spatial correlations, precipitation bias patterns are highly spatially heterogeneous, characterised by a combination of wet and dry regions. Distinct groups of RCM-type-related biases can be identified. For instance, CCAM-2021 and CCAM-2017 simulations exhibit wet biases over topographically complex regions, particularly the MC, where numerous islands of varying sizes are present. Similarly, biases in HadGEM3-RA7 and CCAM-Mod2021 simulations remain consistent regardless of the driving GCM. For CCAM-Mod2021, this similar behaviour is likely due to the use of bias- and variance-corrected SST forcing, as previously noted by Evans et al. (2021). The underlying cause of the comparable behaviour in HadGEM3-RA7, however, remains unclear. In contrast, RegCM's simulations show pronounced wet biases over high wind-speed regions such as Thailand, along with dry biases over Borneo. Comparison of RegCM5 and RegCM4 simulations forced by EC-Earth3-Veg (Fig. 6d and g, respectively) shows an improvement in presentation of precipitation with substantial reductions of wet biases over Thailand. Interestingly, although RegCM-MIROC6 (Fig. 6m) and RegCM4-CanESM5-exp28 (Fig. 6n) exhibit similar domain-mean biases (approximately 83 mmyr-1, Fig. 6), their spatial bias structures differ markedly. RegCM4-CanESM5-exp28 shows a predominantly wet bias across most of the domain, except over the north-western part of SEA, whereas RegCM5-MIROC6-exp16 exhibits dry biases over much of the domain, with wet biases confined primarily to north-western Myanmar and southern Thailand. Using the same GCM nudging technique with different convective parameterisations, CCAM-2017 simulations (Fig. 6c and k) show significant improvement compared with CCAM-2021 (Fig. 6p and u). These findings highlight the critical role of RCM setup, particularly the convective parameterisation in shaping precipitation biases over the SEA domain.

The rainfall pattern of the NDJFMA reveals some interesting features (Fig. 7). RCMs show better skill (e.g., higher spatial correlation with APHRODITE) in capturing the spatial distribution of precipitation than that during the MJJASO season, with most models exhibit dry biases over the mainland and wet biases over MC. Interestingly, the higher biases compared with that during the MJJASO season are more prevalent in MC in all simulations. Specifically, the RegCM simulations cannot capture the spatial distribution of precipitation well over MC, with all models showing a low correlation coefficient of less than 0.2 (figure not shown).

Due to large observational uncertainties in precipitation during both seasons (Fig. S1), the similar analyses of mean bias are applied for other different observational datasets (Fig. S11 in the Supplement). Interestingly, there are at least 14 out of 21 RCM simulations exhibiting the wet biases in terms of regionally averaged total wet-day precipitation, regardless of the choice of reference data.

The seasonal cycle of monthly mean total rainfall over two sub-regions of SEA is presented in Fig. S12 in the Supplement, confirming the substantial overestimation of precipitation intensity in both regions as discussed previously. In term of timing, observed rainfall peaks during May–October over the Mainland and minimal rainfall are observed during the boreal winter (December–February). RCMs generally capture well the peak time of rainfall over the Mainland. However, most RegCM simulations fail to adequately capture the precipitation peak over the MC. It is important to note that different sub-regions of the Maritime Continent (MC) exhibit multiple distinct climate regimes, including monsoonal regions (e.g., Java) with a dominant rainfall peak during the austral summer (December–February), semi-monsoonal regions (e.g., parts of Borneo and Sumatra) with equatorial bimodal peaks during the transitional seasons (March–May and September–November), and anti-monsoonal regions (e.g., eastern part of the Maritime Continent) characterised by an opposite seasonal phase that peaks during the boreal summer (June–August) due to its local topography, specific mountain ranges, and ocean-atmosphere interactions (Tangang et al., 2020; Aldrian et al., 2004) These regimes are further modulated by synoptic processes such as cold surges and the Borneo vortex over Peninsular Malaysia and surrounding areas, leading to substantial spatial variability in rainfall seasonality. As a result, rainfall peaks occur at different times across sub-regions, and no single unimodal seasonal-cycle representation is appropriate for the entire Maritime Continent. This deficiency highlights the models' inability to correctly simulate the complex circulations and precipitation dynamics over MC, thereby warranting further sub-regional analyses of atmospheric circulation patterns across localised zone within MC.

Figure 8 compares regionally-averaged daily precipitation quantiles from CORDEX-SEA RCMs with APHRODITE observations over the whole of SEA, the Mainland, and the MC for the MJJASO and NDJFMA seasons during 1982–2014. Precipitation is also considered in three percentile ranges: 0–50th percentile (left of the blue dashed line), 50–99th percentile (between the dashed and dot-dashed lines), and 99th percentile (right of the dot-dashed line), allowing a detailed assessment of model performance across the distribution.

Overall, almost all RCM (coloured) lines are above the observational line (black), indicating more intense precipitation in simulations except over the Mainland during the NDJFMA season. This systematic shift toward higher precipitation values in RCMs compared to APHRODITE is more evident after the 50th percentile. At the highest percentiles, RCMs diverge markedly from the reference dataset, highlighting the ongoing challenges for RCMs to capture the intensity of extreme precipitation. Note that some RCMs (CCAM-2017-ACCESS-CM2, CCAM-2021-EC-Earth3-Veg) more closely match observed extremes during dry seasons, while others remain biased, either overestimating (e.g., RegCM4-EC-Earth3-Veg -exp28 and RegCM4-MIROC-ES2L-exp28) or underestimating (RegCM4-CESM2-exp28) the heaviest rainfall during the dry seasons. An exception is RegCM-CESM2, which has persistent dry biases in the RCM in both seasons. Interestingly, all simulations are below the observational line for the Mainland during DNJFMA, indicating the dry biases as mentioned above (Fig. 8e). Quantitative evaluation using the area score shows that CORDEX-SEA simulation aligns more closely with observations during the dry season compared to the wet season (17 out of 21 simulations exhibit smaller area scores, regardless of the reference dataset, Fig. S13 in the Supplement). Among all simulations, CCAM-2021-ACCESS-CM2 displays the largest deviations from observations across both seasons, as reflected in its area scores, which are higher than those of other RCMs.

There is an interesting difference between the two sub-regions. While the RCMs generally capture precipitation below the 50th percentile well over the Mainland, the distribution over the MC is more divergent, even at percentiles below the 50th. This highlights the challenges in simulating precipitation over the MC, which consists of numerous islands of varying sizes and complex atmospheric-oceanic interactions. Moreover, due to the dominant role of convective rainfall over the MC, this region is highly sensitive to the cumulus parameterisations schemes used in RCMs (Ngo-Duc et al., 2017)

In this section, we move to the next step of sub-setting the 21 RCM simulations from the CMIP6 CORDEX-SEA ensemble for further dynamical downscaling at the kilometre-scale over SEA megacities, under the CARE for SEA megacities project using the minimum standard metrics (MSMs) from the benchmarking framework (BMF) (Isphording et al., 2024) as mentioned in Sect. 2.2.

In this section, we examine the projected future climate change spread from the CMIP6 CORDEX-SEA ensemble, which provides simulations under the SSP3-7.0 scenario (Table 1). Figure 13 shows the distribution of projected changes in mean precipitation (y-axis) versus mean temperature (x-axis) during the wet and dry seasons. In this study, we classify regional temperature change into low, medium, and high warming categories. Specifically, low, medium, and high regional warming are defined as regional mean temperature increases of approximately +1.5, +2, and > +3 °C relative to the baseline period (1981–2010), respectively. Meanwhile, precipitation thresholds are set at dry (below 5 %), neutral (5 %–15 %), and wet (above 15 %), respectively.

In general, both variables exhibit an increasing tendency, with temperature changes sitting in the mid to high ranges (e.g., changes from 2 to 5 °C) and precipitation changes from 1.5 % to 35 %. However, several RegCM experiments (RegCM4-MPI-ESM1-2-HR-exp28, RegCM4-NorESM2-MM-exp28, and RegCM4-CanESM5-exp28) indicate decreasing trends in daily precipitation during NDJFMA, despite lying within the mid- to low-range precipitation projections during MJJASO.

Among the 15 models passing the BMF criteria, 11 experiments provide SSP3-7.0 simulations. These experiments cluster around the mid-range of temperature projections but span a wide range of precipitation responses. Wet models projecting precipitation increases include RegCM4-EC-Earth3-Veg-exp28, RegCM4-NorESM2-MM-exp16 (e.g. above 15 %), CCAM-2017-ACCESS-CM2, CCAM-2021-ACCESS-CM2, HadGEM3-RA7-EC-Earth3-Veg, HadGEM3-RA7-NorESM2-MM, HadGEM3-RA7-UKESM1-0-LL, and RegCM4-NorESM2-MM-exp28 (e.g., 5 %–15 %); while no or dry changes (< 5 %) and are represented by CCAM-2017-EC-Earth3-Veg, RegCM4-MPI-ESM1-2-HR-exp28, and RegCM4-CanESM5-exp28.

In the next step, the seasonal spatial bias patterns (e.g., Figs. 2, 6 and 7) were subjected to hierarchical clustering using Pearson's correlation matrix as the distance measure between models. This aims to group similar spatial patterns of bias. Clusters shown in the dendrogram (Fig. 14) are designated by different colours. The horizontal axis on the dendrogram is a measure of similarity between individual models and clusters, where models positioned on the same branches share spatial patterns similarities. Hierarchical clustering of historical NDJFMA total wet-day precipitation and annual mean temperature reveals three dominant spatial clusters (Fig. 14b and c). These cluster, distinguished by three distinct colours, broadly corresponding to the RCM families: CCAM, HadGEM3-RA7, and RegCM (Fig. 14b and c). In contrast, the MJJASO precipitation pattern (Fig. 14a) exhibits a strong seasonal dependence, forming three distinct clusters that separate CCAM, specific RegCM simulations (e.g., RegCM4 forced by NorESM2-MM using experiment No.16, see Ngo-Duc et al., 2024; MICROC-ES2L, CESM2, and RegCM5 forced by MIROC6), and HadGEM3-RA7 simulations grouped alongside the majority of RegCM. This seasonal variation in clustering behaviour aligns with the findings of Nguyen et al. (2024) and Gibson et al. (2024).

Overall, based on model performance, future spread under the SSP3-7.0 scenario, model dependency, and data availability, we select at least two experiments from three independent model groups to be considered for further km-scale dynamical downscaling over SEA megacities. These include:

RegCM4: EC-Earth3-Veg-exp28 – high range; NorESM2-MM-exp28 – mid range; MPI-ESM1-2-HR-exp28 and CanESM5-exp28 – low range