Dynamical processes and flux exchanges between Earth system components are better represented in coupled modelling systems rather than the single-component models (e.g. Meehl, 1990). Hence, coupled models, particularly with dynamically interactive atmosphere, ocean, land surface, and sea ice models, are increasingly employed for climate research as well as operational forecast applications (e.g. Miller et al., 2017; Lewis et al., 2018, 2019a). The atmosphere and ocean are two major components of the Earth's climate system, and interactions between these two systems are key drivers of climate and weather. In the past, efforts toward the development of atmosphere–ocean coupled models were largely constrained by their high computational requirements, limited understanding of air-sea coupled processes, and lower computational efficiency (Meehl, 1990). During the last 3 decades, there have been significant advancements in the computational power of supercomputers and the computational efficiency of atmosphere–ocean circulation models. Presently, global atmosphere–ocean–wave–land surface–sea ice coupled operational forecasts are available at spatial resolutions of 0.1∘ in the Integrated Forecast Systems (IFS) developed by the European Centre for Medium Range Weather Forecasting (ECMWF) to 0.25∘ in the Global Forecast System (GFS) developed by the National Center for Environmental Prediction (NCEP). Moreover, the accessibility of high-performance computers (HPCs) to researchers has considerably increased in the last decade. Several regional and global atmosphere–ocean coupled modelling systems have been developed worldwide during this period (see reviews by Giorgi and Gutowsky, 2015, and Xue et al., 2020).

The tropical region lying between the eastern Indian Ocean and western Pacific Ocean, encompassing the Malay Peninsula, Philippine Archipelago, Indonesian Archipelago, and surrounding oceanic and island region is generally referred to as the Maritime Continent (MC). This region is characterised by complex orography and shallow seas interconnected by numerous straits (Fig. 1). The MC region is characterised by strong atmosphere–ocean coupled processes across multiple timescales. The El Niño–Southern Oscillation (Bjerknes, 1969) and the Indian Ocean Dipole–Zonal Mode (Saji et al., 1999; Webster et al., 1999) represent two dominant climate modes of variability that influence the MC on inter-annual timescales. Meanwhile, the monsoons and Madden Julian Oscillation (MJO, Madden and Julian, 1994) manifest the coupled processes over the MC in seasonal and intra-seasonal scales, respectively. Because of its geographical location in the middle of the Indo-Pacific warm pool and in the ascending branch of global atmospheric Walker circulation, the MC has been identified as an area of climatic importance both in regional and global environments (Neale and Slingo, 2003; Qu et al., 2005).

The development of regional coupled models is mainly driven by the idea that by resolving fine-scale orographic, ocean circulation, and coastal ocean features, a more accurate representation of atmosphere–ocean dynamics and coupled processes can be achieved. The prediction of atmospheric and oceanic variables over the MC is challenging because of its complex geography, strong air–sea coupling, and remote ocean influences. Earlier studies suggested that the accuracy of atmospheric and ocean hindcast and forecast significantly improves when the simulations are performed using coupled models (Xue et al., 2014; Thompson et al., 2018; Lewis et al., 2019a). There have been a few coupled modelling studies over the MC focusing on the climate and weather research or short-range atmosphere–ocean forecasting (e.g. Aldrian et al., 2005; Wei et al., 2014; Li et al., 2017; Thompson et al., 2018). Recently, Xue et al. (2020) presented a review of atmosphere–ocean coupled modelling studies over the MC region.

Besides coupling, the model skill in simulating atmosphere and ocean state shows a strong relation to the grid resolution also (e.g. Li et al., 2017). Local mesoscale processes (e.g. land and sea breezes) also play an important role in the adequate simulation of upscale processes such as the MJO (Birch et al., 2016). Convection plays a fundamental role, either locally or embedded in bigger envelopes such as the MJO, influencing the diurnal cycle of precipitation and moving squall lines (Love et al., 2011). Therefore, the simulation of weather and climate processes over the MC requires sufficient resolution to resolve these scales and their interactions. Generally, a horizontal resolution of approx. 4 km, so-called convection permitting, has been effective in representing fine-scale processes over the MC (Love et al., 2011; Birch et al., 2014, 2016; Vincent and Lane, 2017). The first attempt towards the development of a convection-permitting atmosphere–ocean coupled model over the MC was undertaken by Thompson et al. (2018, hereafter T18). T18 used a regional version of the UK Met Office Unified Model (MetUM) atmospheric model and Nucleus for European Modelling of the Ocean (NEMO) ocean model configured for the western MC (WMC). For simplicity, the WMC coupled model configuration used in T18 is referred to as WMCao hereafter. The atmosphere and ocean components of WMCao were configured for the same domain and similar horizontal resolution of 4.5 km × 4.5 km (Fig. 1). The model resolution is fine enough to represent the complex coastal geography, ocean bathymetry at shallow oceans and straits, and orographic features, such as mountain ranges, with enormous influences in local weather (e.g. Bukit Barisan in Sumatra or the Sierra Madre in Luzon).

Work to develop and perform a first-hand evaluation of the WMCao, described in T18, was a preliminary step aimed to establish a high-resolution atmosphere–ocean coupled model focusing on the Southeast Asian region for both operational forecasts and climate research applications. Since the overall objective of the development was to simulate both atmospheric and oceanic variables, the coupling has provided a better consistency between the atmospheric conditions and that of the ocean underneath rather than employing stand-alone models. The case studies conducted as part of the WMCao evaluation suggested that the zonal extent of the domain might not be sufficient for an accurate prediction of weather events such as cold surges or typhoons. For instance, cyclogenesis inside the South China Sea (SCS) is relatively low and most of the cyclones and typhoons that appear over the SCS originated in the northwestern Pacific Ocean (e.g. Ling et al., 2011). The northern Pacific Ocean region between 100∘ E and 180∘E/W is the most active tropical cyclone basin on Earth and it accounts for about one-third of the world's tropical cyclones annually (e.g. Lee et al., 2020). Hence, rather than internally coupled dynamics, the predicted track and typhoon characteristics are dominantly driven by the lateral boundary conditions (LBCs) in WMCao. Similarly, the simulation of MJO- and cold-surge-related weather parameters may also be heavily influenced by LBCs. Hence, to address the issues encountered in T18 and incorporate the latest model scientific developments, the present study aims to bring several key upgrades to the WMCao configuration, and test its feasibility in operational forecast application. The main updates to the coupled modelling system include extending the eastern boundary of the model domain to the western Pacific Ocean, upgrading MetUM to the latest science configuration, and incorporating tide boundary forcing into the NEMO.

This study presents details of the atmosphere–ocean coupled prediction system developed for the MC and an evaluation of the ocean forecast from the system using a 6 d pre-operational forecast for October 2019. Following the method employed in many earlier coupled modelling studies (e.g. Li et al., 2014; Lewis et al., 2018), the evaluation of WMCao in T18 has been performed by using short case study simulations of selected weather events spanning over 5 d . In the present study, instead of case studies, we assess surface and subsurface oceanic variables predicted by the coupled system across different forecast lead times.

The next section of this paper presents an overview of the model setup, including a brief description of the model domain, atmospheric, ocean, and coupled model configurations. A brief discussion of the pre-operational forecast system setup is also presented. Section 3 provides the details of datasets used for the atmosphere and ocean model forcing and evaluation. Section 4 presents an assessment of the sea surface variables simulated by the stand-alone ocean model and both surface and subsurface ocean forecasts delivered by the MC coupled model. Finally, Sect. 5 summarises the results obtained from the study and suggests future developments.

A brief description of the reanalysis, forecast, and observational datasets used for the model initialisation, forcing, and evaluation is presented in this section.

An evaluation of the MCO hindcast and MCOao forecast simulations are presented in this section of the paper. Direct comparison of model simulations with observation or analysis data has been performed. Based on the availability of in situ or satellite observation at the time of data analysis, only a few variables are selected for assessing the model performance. In addition, to maintain consistency between the evaluation of hindcast and forecast simulations, analyses of the same set of variables and observation data have been performed where possible. Oceanic variables employed for the evaluation are SST, SSH, and the subsurface temperature and salinity. Model hindcasts and forecasts over the region 16∘ S to 23∘ N and 92 to 138∘ E and defined as the analysis domain (Figs. 1 and 3), is further used for the analysis. The MC model domain includes oceanic basins with different geographical and climatological characteristics. For the evaluation purpose, we have divided the analysis-domain into 10 sub-regions based on their geographical distribution (Fig. 3). These sub-regions are the (1) Andaman Sea–Malacca Strait (ASMS), (2) southern South China Sea (SSCS), (3) Gulf of Thailand (GoT), (4) the rest of the South China Sea (RSCS), (5) the tropical western Pacific Ocean (TWPO), (6) Sulu–Celebes seas (SuCeS), (7) Banda Sea (BS), (8) Java Sea (JS), (9) Timor–Arafura seas (TAS), and (10) the tropical eastern Indian Ocean (TEIO).

The Maritime Continent has a profound influence on the global climate system because of its complex topography and unique geographic location within the tropical Indo-Pacific warm pool. The MC region is characterised by strong atmosphere–ocean coupled processes across multiple timescales. A coupled convective-scale, eddy-resolving atmosphere–ocean modelling system for the western MC, described in T18, was able to improve the simulation of a cold-surge event, the intensity of Typhoon Sarika, and its atmosphere–ocean interactions. Several upgrades have been added to the T18 model for a future operational implementation, such as extending the eastern boundary of the model domain to the western Pacific Ocean, improved science configuration of the atmospheric model (MetUM), and incorporation of tidal boundary forcing to the ocean model (NEMO). Furthermore, the coupled model's feasibility for use as an operational forecast system is also being tested. Typical runtimes of the daily forecast simulations in our experiments are found to be suitable for the operational forecast applications.

The MCao coupled prediction system was run as a pre-operational forecast system from 1 to 31 October 2019. Hindcast simulations performed for the period 1 January 2014 to 30 September 2019, using the uncoupled ocean model MCO, provided the initial condition to the MCOao. This paper presents details of an atmosphere–ocean coupled prediction system developed for the MC and evaluations of ocean-only model hindcasts and 6 d ocean forecast simulations performed using the coupled system.

The evaluation of MCO hindcast is intended to understand the model's performance in reproducing the past ocean variability, particularly at the ocean surface where the exchange of fluxes between the atmosphere and ocean takes place. The ERA5-driven simulations during the period from 1 January 2018 to 30 June 2019 are utilised for the evaluation of MCO. The SST RMSD between the model hindcast and OSTIA is less than 0.5 ∘C for about 97 % of the analysis domain. Their correlation is above the 99.9 % confidence level, with about 88 % of the analysis domain displaying correlation higher than 0.8. An SST cold bias is seen in the Andaman Sea, while most of the South China Sea, equatorial western Pacific Ocean, and Australian coast of the Timor Sea show a warm bias. Overall, the mean SST bias, RMSD, and mean correlation over the domain are 0.07 ∘C, 0.34 ∘C, and 0.90, respectively.

Comparison of model SST with the RAMA moored buoy observations located at the southeastern tropical Indian Ocean shows good agreement. The SST bias, RMSD, and correlation coefficient between the model and observation are 0.17 ∘C, 0.29 ∘C, and 0.94, respectively, for M1 and 0.12 ∘C, 0.41 ∘C, and 0.92, respectively, for M2 locations. The SSH RMSD is less than 0.10 m for 18 of total 20 tide gauge stations analysed, and 14 stations show RMSD less than 0.05 m. Comparison of model-simulated SST and SSH fields show good agreement with observation and analysis data. Statistically significant correlation with observation suggests that both the spatial and temporal patterns of variability are reasonably well reproduced by the model.

For the evaluation of MCOao, comparisons of ocean forecast for different forecast lead times with OSTIA SST and in situ observations have been performed. The forecasted SST over most of the sub-regions is within the error standard deviation of the OSTIA. Though the model forecast exhibits a warm SST bias, the RMSD is less than 0.45 ∘C over most of the sub-regions during the analysis period. Generally, the forecasted SST tends to be cooler with an increase in forecast lead time, denoting a lower warm bias and RMSD relative to fcst_day1. Overall, the SST correlation over the analysis domain is above the 99 % confidence level across all forecast lead times, while the bias and RMSD are less than 0.19 and 0.35 ∘C, respectively.

The diurnal variability of SST at the RAMA moored buoy locations M1 and M2 are reasonably well reproduced by the model across all forecast lead times. SST bias and RMSD at M1 are less than 0.0 and 0.20 ∘C, respectively, and remains fairly constant across all forecast lead times. Meanwhile, the RMSD at M2 is relatively low, with a maximum of 0.18 ∘C across all forecast lead times. The depth of the upper ocean isothermal and mixed layer at the M1 is well forecasted by the model. To understand the model skill in predicting subsurface temperature and salinity, in situ profile observations are compared with the model forecast for different forecast lead times. For both temperature and salinity, the RMSD remains fairly constant during the entire analysis period.

Comparison of model-forecasted SSH shows good agreement with the tide gauge observations. About 75 % of the stations show bias within ±0.05 m at all forecast lead times. No significant variation in model forecast accuracy or RMSD is seen with the increase in forecast lead time. About 73 % of total stations show RMSD less than 0.10 m for all forecast lead times. Overall, the model forecast deviation of SST, SSH, and subsurface temperature and salinity fields relative to observation is within acceptable error limits of operational forecast models (e.g. Zhang et al., 2010; Yang et al., 2016).

Analysis of subsurface fields revealed that significant model temperature biases exist in the region below the mixed layer. The representation of stratification in the thermocline region is relatively weak in the model forecast than the observations. This subsurface temperature bias remains across the model domain with varying magnitudes. Similar temperature biases were reported earlier in simulations with identical model configurations (e.g. Graham et al., 2018). Further modelling work is needed to improve the thermal stratification through fine-tuning the mixing coefficients or modifying the vertical mixing parameterisation.

Further analysis of model forecast fields using longer forecast simulations with increased observations will be undertaken to assess the model predictability across different seasons and during typical weather events such as a cold surge, a typhoon, or the MJO. In addition, the impact of coupling on the forecast will be investigated by performing simultaneous stand-alone ocean model forecast simulations. In addition, work is ongoing to improve our understanding of the forecast skill of MCAao, and results from that analysis will be presented as research publications (Kumar et al., 2021). The analysis of model-forecasted precipitation, surface wind, pressure, relative humidity, etc., will be undertaken. The comparisons of coupled and atmosphere-only forecasts will also be performed in that study.

The evaluation of sea surface salinity and ocean current fields is not included in the present study mainly due to the lack of in situ and satellite observations. The South China Sea remains as an especially data-sparse region in terms of the subsurface observations, which highlights the necessity of coordinated efforts from the scientific community to fill these spatial data gaps.

In our analyses, the RMSD and positive or negative biases of the ocean forecasts are generally comparable to those observed from the hindcast statistics. This suggests that, up to a certain extent, the model forecast deviation is inherited from the MCO hindcast or the MCOao initial condition. The dependency of model forecast quality on the initial state is well established in the numerical weather prediction studies. Over the tropical oceans specifically, the initialisation of the ocean state is an important element of the forecast systems. The data assimilation techniques help to acquire an improved estimate of the ocean state by combining the model-simulated fields and observations (King et al., 2018). Both the uncoupled and coupled ocean configurations used in our study are free-running models with no restoring or relaxation to the real world. Hence, to provide a better initial condition to MCOao, implementation of data assimilation capability for MCO will be a key priority in our future developments. In addition, earlier studies have shown that including wave-induced mixing in ocean circulation model yields a better representation of the upper-ocean temperature (Lewis et al., 2019b). Thus, work towards the development of a three-way atmosphere–ocean–wave coupled system will be undertaken in the future.