The cSINGV modelling system is a high-resolution (1.5 km) regional coupled atmosphere–ocean–wave framework developed for the WMC. It integrates a convection-permitting atmospheric model based on the Met Office Unified Model (UM), an ocean model based on the Nucleus for European Modelling of the Ocean (NEMO), and a spectral wave model based on WAVEWATCH III (WW3), within a unified coupling framework based on the UK Met Office Regional Coupled Suite (RCS) (Castillo et al., 2022).

In this study, the atmospheric component (hereafter referred to as cSINGV_A) is based on the numerical weather prediction system SINGV, used at the Meteorological Service Singapore (MSS). The ocean and wave components (hereafter referred to as cSINGV_O and cSINGV_W, respectively) have been newly developed and configured for the WMC. The model domain spans 94.6–109.36° E and 5.7° S–8.13° N (Fig. 1), covering the complex coastal and shelf regions of the WMC. The system is designed to capture multiscale interactions, including diurnal variability, coastal processes, and air–sea–wave coupling, which are critical for accurately representing weather and climate in this region. The individual component models and the coupling framework are described in detail in the following sections.

The atmosphere and land surface components in cSINGV are based on a fixed-resolution latitude–longitude grid with a horizontal spacing of 1.5 km over the model domain. It is a fully non-hydrostatic model with a terrain-following vertical coordinate system, consisting of 80 vertical levels extending up to a model top of 38.5 km. The vertical resolution is approximately 5 m near the surface and 1.45 km at the model top, with a 60 s time step for model integration. The configuration is consistent with the MSS operational numerical weather prediction system, SINGV (Huang et al., 2019).

Land surface processes are represented using the Joint UK Land Environment Simulator (JULES v7.5; Best et al., 2011), which includes nine surface fraction types. The atmosphere and land surface components are compiled as a single executable and coupled implicitly at each model time step following Best et al. (2004), ensuring consistent land–atmosphere interactions. Physical parameterisations include the Prognostic Cloud fraction and Prognostic Condensate (PC2) scheme (Wilson et al., 2008), as used in previous implementations (Huang et al., 2019). The boundary-layer parameterization follows a blending approach developed by Boutle et al. (2014), which combines the one-dimensional scheme of Lock et al. (2000) with the three-dimensional Smagorinsky–Lilly scheme (Lilly, 1962). The cloud microphysics scheme is based on Wilson and Ballard (1999), while radiation processes are modelled using the Edwards and Slingo (1996) scheme, which includes six shortwave bands and nine longwave bands (Manners et al., 2012).

The configuration used in this study (cSINGV_A) adopts the Regional Atmosphere and Land v2 in the Tropics (RAL2-T) configuration (Bush et al., 2023). This configuration is used as the baseline atmospheric system onto which the ocean and wave components are coupled. SST is prescribed from OSTIA (Donlon et al., 2012) and updated daily in cSINGV_A to provide a consistent reference for comparison with the coupled simulations. Ocean surface currents are set to zero, and the Charnock parameter is prescribed as a constant value of 0.011 over all ocean points (Bush et al., 2023). No additional tuning or modifications are applied beyond this configuration.

The ocean component was newly configured and developed for application over the WMC. The ocean component, cSINGV_O, is based on version 4.0.4 of the NEMO (NEMO development team, 2021), configured at 1.5 km horizontal resolution. NEMO is a primitive-equation, hydrostatic model extensively used in climate and operational forecast applications (Madec et al., 2017). The model employs an orthogonal curvilinear coordinate system with Arakawa C-grid staggering for its horizontal grid, and the bathymetry is derived from the General Bathymetric Chart of the Oceans (GEBCO) dataset (https://www.gebco.net/data-products/historical-data-sets#gebco_2020, last access: 26 February 2026).

Featuring 51 vertical levels within a terrain-following coordinate system, the model incorporates the stretching function by Siddorn and Furner (2013). In our study, the ocean model configurations utilize baroclinic time steps set at 60 s. The Generic Length Scale (GLS) turbulence model (Umlauf and Burchard, 2003) is employed to calculate turbulent viscosities and diffusivities. For the computation of the bottom drag coefficient, an implicit form of non-linear parameterization with a log-layer formulation is applied, setting the minimum and maximum drag coefficients at 0.0001 and 0.15, respectively. The Flather boundary condition (Flather, 1987) is adopted for SSH and barotropic velocities, while a flow relaxation scheme (FRS, Davies, 1976) is implemented for tracers and baroclinic velocities at the lateral open ocean boundaries.

The configuration (cSINGV_O) is largely based on the regional NEMO setup of Thompson et al. (2021), with additional modifications introduced during the upgrade from version 3.6 to 4.0.4 (Table S1 in the Supplement). Prior to coupling, standalone ocean-only sensitivity experiments were conducted to establish a stable and physically credible regional NEMO baseline. These experiments tested configuration choices, including the light-penetration scheme, and the selected configuration improved the representation of SST in both deep-ocean and shallow-shelf regions. The experiments were intended solely to establish suitable behaviour of the standalone ocean component, and the selected configuration was subsequently held fixed in the coupled simulation.

The wave component, cSINGV_W, was newly configured and tuned for the WMC. It is based on version 7.12 of the third-generation spectral wave model WW3 (Tolman, 2014), developed by the US National Centers for Environmental Prediction. To better represent the complex coastlines around Singapore, the model is configured with a two-tier spherical multiple-cell (SMC) grid (Li, 2012), using the same computational domain and bathymetry as the ocean component. The SMC approach enables multi-resolution meshes, allowing fine-scale representation of coastlines and small islands while maintaining coarser resolution in the open ocean. This feature provides more effective refinement of coastal areas compared to a regular WW3 grid (Li, 2019), which is particularly beneficial in the coastal environment of the WMC. The model grid has a resolution of approximately 1.5 km in coastal regions, expanding to 3 km in offshore areas.

The model employs the ST4 source term package (Ardhuin et al., 2010) to represent wave growth and dissipation processes. Nonlinear wave–wave interactions are parameterised using the discrete interaction approximation (NL1; Hasselmann et al., 1985), while shallow-water processes include depth-induced breaking (DB1; Battjes and Janssen, 1978) and bottom friction (BT1; Hasselmann et al., 1973). Wave propagation is computed using a second-order upstream non-oscillatory scheme (UNO; Li, 2008), and dispersion correction is applied through the PR2 option to mitigate the garden sprinkler effect (Booij and Holthuijsen, 1987). Wind input and stress calculations are handled internally within the source term formulation (FLX0), with additional options for wind and current interpolation (WNT1, WNX1, CRT1, CRX1). In the coupled configuration, the spatial remapping of the atmospheric wind and ocean-current fields between the component grids is performed by OASIS3-MCT using precomputed remapping weights, and each component evolves with its own time step between coupling exchanges and therefore compiled with WNT0 and CRT0. A summary of the model configuration and compilation switches is provided in Table S2.

Specific adjustments to key parameters, including the wave growth parameter betamax and shallow-water physics options, were made for application to the WMC. These parameters were selected during the standalone development of the regional WW3 configuration, which was forced by winds from cSINGV_A, before the fully coupled experiments were conducted. They were subsequently held fixed in the coupled simulation and were not adjusted to improve the performance of the coupled model. The tuning strategy and sensitivity experiments are described in Appendix A.

The coupled modelling system is implemented using RCS, which provides a flexible framework for integrating multiple model components using the cylc workflow engine (Oliver et al., 2019). Coupling between the atmosphere, ocean, and wave components is achieved through the OASIS3-MCT coupler (Craig et al., 2017), enabling the exchange of fluxes and state variables between components.

The configuration used in this study is based on the RCS suite (suite ID: u-bf945), with a dedicated cSINGV branch developed for this application and available via the Met Office repository (https://code.metoffice.gov.uk/svn/roses-u/b/f/9/4/5/u-bf945_add_cSINGV_br, last access: 30 March 2026). This branch provides the specific coupling setup and workflow required for the coupled system, ensuring reproducibility of the model configuration.

The atmosphere, ocean, and wave components operate on their native grids and exchange variables at an hourly coupling frequency. The one-hour coupling interval was selected as a practical compromise between computational cost, coupling overhead, and the need to represent evolving regional air–sea interactions. Previous work over the Maritime Continent found that hourly or more frequent coupling better represented SST variability and convection (Zhao and Nasuno, 2020). However, the hourly interval is a design choice rather than a demonstrated optimum. Shorter intervals may better represent rapidly evolving convection and high-frequency air–sea interactions (Makrygianni et al., 2026).

The framework supports two-way interactions between components, allowing for feedbacks such as the influence of SST on atmospheric fluxes, wave-dependent surface roughness effects, and the impact of ocean currents on wave evolution. A schematic overview of the coupling architecture is shown in Fig. 2, and the exchanged variables are listed in Table 1. In the fully coupled atmosphere–ocean–wave configuration (cSINGV), the atmosphere supplies the 10 m zonal and meridional wind components to WW3, and WW3 provides NEMO with the zonal and meridional components of the stress transmitted to the ocean after accounting for the evolving wave field. This coupling strategy is consistent with established atmosphere–wave–ocean coupling frameworks (e.g., Breivik et al., 2015, their Fig. 8).

Simulations are conducted for January–June 2019 using both standalone and coupled configurations:

Atmospheric-only (cSINGV_A)

The atmospheric model lateral boundary conditions are provided by the 3-hourly European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis, ERA5 (Hersbach et al., 2020). Ocean model initial and lateral boundary conditions are obtained from the daily Mercator Ocean reanalysis (Lellouche et al., 2021). Wave boundary conditions are derived from a high-resolution global WW3 simulation using a four-tier spherical multiple-cell (SMC61250) grid, with resolutions of 50, 25, 12, and 6 km from the open ocean to coastal regions (Li, 2019; Kumar et al., 2025), forced by ERA5 winds. Tidal forcing at the ocean lateral boundaries is prescribed using tidal elevations and currents from the Finite Element Solution dataset (FES2014b; Lyard et al., 2021; https://www.aviso.altimetry.fr/en/data/products/auxiliary-products/global-tide-fes.html, last access: 15 May 2024). River runoff is represented using monthly climatological fields from Dai and Trenberth (2002). In addition, monthly chlorophyll concentrations from the Sea-viewing Wide Field-of-view Sensor (SeaWiFS; Hooker and McClain, 2000) are used in the ocean model to parameterise light absorption and its influence on the upper-ocean heat budget.

For model evaluation, multiple observational and reanalysis datasets are used. Precipitation is evaluated against the Integrated Multi-satellite Retrievals for Global Precipitation Measurement (GPM IMERG V06B; Huffman et al., 2019, 10.5067/GPM/IMERG/3B-HH/06), hereafter referred to as GPM, which provides reliable estimates over Southeast Asia (e.g. DaSilva et al., 2021). Three-hourly 10 m wind speed and 2 m air temperature observations are obtained from the World Meteorological Organization (WMO) MetDB database (Met Office, 2008). SST is evaluated using the Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA: 10.48670/moi-00168) dataset and the GLORYS12V1 product from the Copernicus Marine Environment Monitoring Service (CMEMS; Lellouche et al., 2021, 10.48670/moi-00021). Hs is validated using multi-mission satellite altimeter observations (Piollé et al., 2022) obtained from CMEMS.

Model performance is evaluated using a range of statistical metrics. The root mean square error (RMSE) is used to assess 10 m wind speed, 2 m air temperature, SST, and Hs, and is defined as: 1RMSE=1N∑i=1NMi-Oi2 where Mi and Oi denote the modelled and observed values, respectively, and N is the number of samples.

For wave evaluation, additional metrics are applied. The bias is defined as: 2Bias=1N∑i=1NMi-Oi. The Pearson correlation coefficient (r) is given by: 3r=∑i=1NMi-M‾Oi-O‾∑i=1NMi-M‾2∑i=1NOi-O‾2. The scatter index (SI), representing the normalised RMSE, is defined as: 4SI=RMSEO‾. Distributional agreement between modelled and observed Hs is assessed using the Kolmogorov–Smirnov (KS) statistic (Massey, 1951), defined as: 5KS=max∣Fm(x)-Fo(x)∣ where Fm(x) and Fo(x) are the cumulative distribution functions of modelled and observed values, respectively.

Probability distribution functions (PDFs) were estimated using kernel density estimation (KDE; Parzen, 1962), implemented via KDEpy (Odland, 2018). For precipitation, the Fraction Skill Score (FSS; Roberts and Lean, 2008) is used to evaluate the spatial representation of rainfall. It is computed for selected precipitation thresholds and neighbourhood sizes and is defined as: 6FSS=1-∑Pf-Po2∑Pf2+Po2 where Pf and Po represent the fractional coverage of forecast and observed precipitation exceeding a specified threshold within a given neighbourhood. FSS is computed within the evaluation domain indicated by the red box in Fig. 1.

To assess whether differences between cSINGV and cSINGV_A are distinguishable from day-to-day variability, paired analyses were applied to the daily scores of the two configurations. For FSS, the mean paired difference (ΔFSS = cSINGV - cSINGV_A) and its 95 % confidence interval were calculated from the paired daily differences. For the RMSE of 10 m wind speed and 2 m air temperature, the daily paired differences (ΔRMSE = RMSE_{cSINGV_A} - RMSE_cSINGV) were examined using a sign test. Daily precipitation bias relative to GPM was also calculated for both configurations.

The evaluation of atmospheric variables was performed using the MetPlus framework, which provides a standardized environment for model verification and ensures reproducibility of the analysis (https://metplus.readthedocs.io/en/latest/, last access: 26 September 2025).

The ocean component is evaluated through comparisons of SST with OSTIA and CMEMS reanalysis datasets for the period January–June 2019. Given that SST is the key variable linking the ocean to the atmosphere in the coupled system, the evaluation focuses on its representation.

The WMC region is characterised by two distinct ocean regimes: the shallow Sunda Shelf in the South China Sea (SCS) and the deeper Indian Ocean basin west of Sumatra (Fig. 1). To assess model performance across these contrasting environments, the analysis is conducted separately for these regions (Fig. S1 in the Supplement). Figure 3 shows the time series of domain-averaged RMSE of daily SST relative to OSTIA for (i) the full domain, (ii) the Indian Ocean, and (iii) the South China Sea over the six-month simulation period.

The model reproduces SST variability well across both regions, with domain-averaged RMSE values generally below 0.5 °C throughout the simulation period. Some deviations are evident in the Indian Ocean during February and May, but the model captures the overall temporal variability and maintains errors comparable to CMEMS. In the SCS, SST variability is also well represented, with consistent agreement across the simulation period. Additional comparison with the regional NEMO configuration of Thompson et al. (2021) shows that the large-scale SST patterns are well reproduced (Fig. S2), and domain-averaged RMSE values are comparable (Fig. S3), providing further confidence in the ocean configuration.

Following the model tuning described in Appendix A, the cSINGV_W configuration is evaluated against multi-mission satellite altimeter observations for January 2019 over the full domain and two subregions: the Indian Ocean and the South China Sea, representing distinct wave regimes. January 2019 was selected as it includes Tropical Storm Pabuk, providing a strong-wind case suitable for assessing wave growth.

The model reproduces the variability of Hs well across the domain (Fig. 4). Scatter comparisons show strong agreement with observations, with most values distributed close to the 1:1 line. For the full domain, the RMSE is approximately 0.2 m, with a correlation coefficient exceeding 0.926. Performance is slightly reduced in the Indian Ocean (RMSE = 0.21 m), possibly reflecting the greater influence of remotely generated swell, while improved agreement is observed in the South China Sea (RMSE = 0.185 m), where locally generated wind sea is expected to be more important.

To assess the performance of the selected configuration (betamax = 1.75) beyond the January strong-wind case, the cSINGV_W evaluation was extended to January–June 2019 (Fig. S4 and Table S3). Across the six months, RMSE ranges from 0.179 to 0.287 m, the correlation coefficient from 0.910 to 0.966, and SI from 0.150 to 0.270, with a negative monthly bias ranging from -0.018 to -0.195 m. Performance is strongest in January–February and weaker in April–May, but correlations remain above 0.91 in every month. These results support the suitability of betamax = 1.75 for the standalone regional wave configuration over the evaluation period.

Overall, the standalone ocean and wave evaluations demonstrate that the individual components provide a credible representation of key marine processes, forming a strong foundation for the development of the fully coupled cSINGV system.

Building on the evaluation of the standalone ocean and wave components, we now examine the performance of the fully coupled system (cSINGV) by comparing it with the atmosphere-only configuration (cSINGV_A). It should be noted that cSINGV_A uses prescribed OSTIA SST, whereas cSINGV uses dynamically evolving SST from the ocean model. Therefore, the comparison does not represent a fully controlled coupling experiment, but rather provides insight into the combined effects of the interactive coupling processes represented in the coupled configuration.

Two further differences between the configurations should be considered. First, SST in cSINGV_A is updated daily and remains fixed within each day, whereas cSINGV produces appreciable and spatially heterogeneous sub-daily SST variability. The mean daily SST range in cSINGV – calculated as the difference between the maximum and minimum of the 24 hourly values and then averaged over January–June 2019 – is typically 0.5–0.75 °C over the open ocean and exceeds 1 °C in some shallow and coastal waters, including the Malacca Strait and waters surrounding Sumatra (Fig. S5). This variability may modify near-surface air–sea temperature and humidity gradients, thereby influencing the timing and magnitude of turbulent heat fluxes and convection on diurnal timescales. Consequently, differences between the two configurations, particularly at diurnal timescales, cannot be attributed solely to two-way coupling, as they may also reflect the presence of sub-daily SST variability in cSINGV. Second, cSINGV_A uses a fixed Charnock parameter of 0.011, whereas cSINGV uses a WW3-supplied Charnock parameter that varies with the evolving sea state, exhibiting spatial and month-to-month variability (not shown). The configurations therefore differ in their representation of aerodynamic roughness, which may contribute to the simulated differences in surface winds and turbulent fluxes. Accordingly, (cSINGV_A) serves as a reference against which the combined response to interactive SST, wave-dependent roughness, and other coupling pathways in cSINGV is assessed.

To further assess the impact of coupling at the event scale, we examine a set of Sumatra squall events. Sumatra squall lines are among the most frequent and impactful convective systems over the WMC, often producing intense rainfall, strong winds, and localized flooding (Lo and Orton, 2016). These systems typically initiate along the western coast of Sumatra and propagate eastward across the Straits of Malacca toward Peninsular Malaysia and Singapore. Their evolution is influenced by both land-based processes, such as orographic lifting and diurnal heating, and air–sea interactions, making them a useful benchmark for assessing the role of coupling in regional simulations.

In this section, we analyse a set of Sumatra squall events identified from the MSS squall database based on a threshold of maximum 1 h accumulated precipitation exceeding 50 mm. Five observed squall events satisfying this criterion occurred within the study domain during January–June 2019 (4 and 27 April, 8 and 19 May, and 3 June). Of these, only the 4 and 27 April events were reproduced sufficiently well by at least one configuration to permit a process-based comparison; neither simulation reproduced the remaining three events, which were therefore not analysed further. Two additional events (3–4 September 2019 and 2–4 November 2020) were examined using separate simulations outside the continuous January–June 2019 experiment. The analysis is based on rotated Hovmöller diagrams constructed along a transect aligned with the typical propagation pathway from Sumatra to Singapore (Fig. 1), allowing a consistent assessment of the spatiotemporal evolution of convection.

To further investigate the processes underlying the differences between the cSINGV and cSINGV_A simulations, we examine the influence of coupling on surface fluxes, with a focus on latent heat flux (LHF). As discussed in Sect. 4.4, the largest differences between the two configurations occur during the propagation of organised convection, suggesting an important role for air–sea interactions.

Figure 14 shows the mean difference in LHF between the coupled and uncoupled simulations for January–June 2019. The coupled configuration exhibits enhanced upward fluxes across much of the domain, with localised anomalies exceeding 15 W m⁻² along the Sumatra–Java coastal regions. These enhancements indicate increased transfer of moisture and energy from the ocean to the atmosphere, supporting convective development and maintenance.

To better understand the underlying mechanisms, the LHF differences are decomposed into two components: (i) a wind-speed contribution (ΔU), and (ii) a moisture-gradient contribution (Δ(qs-qa)), where qs is the saturation specific humidity at the sea surface and qa is the near-surface atmospheric humidity. The wind-related contribution is relatively modest (approximately ±5 W m⁻²), whereas the moisture-gradient term exhibits larger amplitudes (up to ±12 W m⁻²). To examine how the LHF differences relate to changes in SST, near-surface winds and sea-surface roughness, January–June mean differences in LHF, SST, 10 m wind speed and the Charnock parameter between cSINGV and cSINGV_A were compared (Fig. S17). The widespread increase in LHF broadly coincides with warmer SST in cSINGV, whereas the wind-speed differences are smaller and more spatially heterogeneous. This suggests that SST-related changes in the near-surface thermodynamic gradients contribute more consistently to the mean LHF response than changes in wind speed, although the individual contributions cannot be isolated from these spatial patterns alone.

From a process perspective, the increased LHF in cSINGV supplies additional moisture and energy to the lower atmosphere, which may support the organisation and persistence of convection during its propagation. This mechanism is consistent with the case studies presented in Sect. 4.4. In particular, for the 25–27 April 2019 and 2–4 November 2020 events, cSINGV produced greater LHF over parts of the offshore region and a more sustained propagation of the squall than cSINGV_A. The enhanced LHF was primarily associated with changes in the air–sea humidity gradient, providing a physically plausible mechanism through which the coupled configuration may have contributed to the more sustained convection. In contrast, for land-driven cases such as 2–4 April and 3–4 September 2019, where convection is primarily initiated over the Sumatra landmass, differences in LHF between the two configurations play a smaller role in shaping the system evolution. This is consistent with the relatively similar behaviour of cSINGV and cSINGV_A in these cases, indicating that the influence of coupling on convection is secondary when land-based forcing dominates. However, a larger event sample or ensemble analysis would be needed to establish this systematically.

Overall, the coupled configuration provides a more physically consistent representation of key processes in the WMC, including oceanic precipitation, offshore propagation, and air–sea flux interactions. While near-surface atmospheric variables show no systematic improvement and some biases persist for precipitation, particularly over land, the results indicate that atmosphere–ocean–wave coupling can contribute to improved simulation of precipitation characteristics and associated processes in this region.