Evaluating MPAS-A Performance for Mesoscale Simulation in a Tropical Region: A Case Study of Extreme Heat in Jakarta, Indonesia

Abstract

The Model for Prediction Across Scales–Atmosphere (MPAS-A) has been widely used for larger scale simulations, but its performance in mesoscale, particularly in tropical regions, is less evaluated. This study aimed to assess MPAS-A in simulating extreme surface air temperature in Jakarta during the hot spells of October 2023 with eight different simulation setups. Several validation metrics were applied to near-surface meteorological variables, land surface temperature (LST), and vertical atmospheric profile. From the eight simulations, MPAS-A captured diurnal patterns of the near-surface variables well, except for wind direction. The model also performed well in LST simulations. Moreover, the biases in the vertical profiles varied with height and were sensitive to the initial/boundary conditions used. Simulations with modified terrestrial datasets showed higher LST and air temperatures over the sprawling urban areas. MPAS-A successfully simulated the extreme event, showing higher air temperatures in southern Jakarta (over 36 °C) compared to the northern part. Negative temperature advection by sea breeze helped lower air temperature in the northern area. This study highlights the role of sea breezes as natural cooling mechanisms in coastal cities. Additionally, MPAS-A is feasible for several applications for urban climate studies and climate projection, although further development is needed.

1. Introduction

Extreme hot weather conditions can have several adverse effects on human health [1,2]. Prolonged high temperatures can increase the risk of health issues such as dehydration and heat stroke [3]. High heat stress can also reduce physical work capacity, leading to decreased productivity [4]. Moreover, energy demand for cooling increases during hot days [5]. In Indonesia, studies reported a trend of more frequent warm days and nights [6] and increasing surface air temperatures in Jakarta [7].

Numerical simulation in meteorology is a valuable tool for modeling atmospheric conditions and dynamically explaining specific phenomena [8,9,10]. It is also instrumental in assessing the impact of human activities, such as urbanization, on changes in atmospheric conditions [11,12,13,14]. Numerical simulation is widely used for weather forecasting and simulating extreme conditions [15,16,17,18] to enhance preparedness and prevention actions. One of the most popular models for mesoscale simulation, including urban-scale modeling, is the Weather Research and Forecasting (WRF) model [19]. Several studies cited previously [10,11,12,13,14,16] also utilized the WRF model. This model utilizes rectangular meshes and relies on domain nesting for local refinement. However, this approach can lead to flow distortion at nest boundaries and introduce artifacts into the model results due to abrupt grid transitions [20,21].

In contrast, the Model for Prediction Across Scales (MPAS) employs smooth grid refinement on centroidal Voronoi meshes, offering a much smoother transition between mesh resolutions. This approach effectively removes the distortions commonly found along the lateral boundaries of nested meshes and significantly reduces resolution-related artifacts [20,21]. This model has been recently developed mainly by the National Center for Atmospheric Research (NCAR) and Los Alamos National Laboratory/United States Department of Energy (LANL/DOE) to advance Earth system simulations for climate and weather studies. MPAS includes stand-alone models for the atmosphere, ocean, land, and sea ice. The atmospheric component, MPAS-A [22], solves the fully compressible non-hydrostatic equations of motion.

MPAS-A adopts some the Advanced Research WRF (ARW) model features, including atmospheric physics and time integration using the third-order Runge–Kutta method [23]. As noted above, the primary differences between the two models are in their grid structures and vertical coordinates. WRF utilizes rectangular meshes with a hydrostatic pressure vertical coordinate, whereas MPAS employs unstructured centroidal Voronoi (mostly hexagonal) meshes with C-grid staggering and geometric-height vertical coordinates. MPAS-A can be applied to global and limited-area (regional) simulations, allowing for uniform and variable horizontal resolutions. The variable resolution enables high spatial resolution in any study area, gradually transitioning to coarser resolutions outside the focus areas.

Several studies have been conducted to evaluate the performance of MPAS-A. Li et al. [24] assessed its performance for sub-seasonal winter surface air temperature forecasting over the northern hemisphere using a 120 km uniform mesh for global simulation. Cheng et al. [25] evaluated MPAS-A for dynamically downscaled daily precipitation, near-surface air temperature, and circulation features during summer monsoon over China, comparing global and regional configurations with the highest spatial resolution of 25 km. Maoyi and Abiodun [26] used the model to simulate the Botswana high over sub-tropical southern Africa, analyzing its impact on precipitation with a global model resolution of 240 km. Lui et al. [27] compared the performances of MPAS-A and WRF in simulating tropical cyclone tracks and intensities over the Western North Pacific. MPAS-A has been used extensively for synoptic and global-scale simulations.

On the other hand, the atmospheric motion scales range from microscale phenomena like turbulence and microphysics to planetary-scale processes such as global circulation [28,29]. Each scale possesses distinct characteristics and governing dynamics. However, an atmospheric model must be capable of simulating a broad range of atmospheric phenomena with scales spanning from the model’s resolution to the simulation domain. Currently, MPAS-A offers a maximum spatial resolution of 3 km (this spatial resolution falls within the mesoscale range). Given that MPAS-A is a relatively new model still in development, it is important to evaluate its performance at mesoscales, including urban-scale variations.

On average, the tropical region is characterized by hot and humid conditions due to receiving more direct solar radiation than other parts of the Earth [30], with relatively minor temperature fluctuations [31]. As a result, this region maintains consistently warm temperatures with minimal variations throughout the year and a small horizontal gradient. Due to the persistent high temperatures, heatwaves in the tropics tend to be subtler [15]. Given these unique characteristics, accurately simulating the timing and spatial extent of extreme heat events in the tropics can be challenging. Moreover, heatwaves in the tropical region are less studied compared to those in the mid-latitudes, where the mechanisms of heatwaves are better understood (Figure 2 of Domeisen et al. [15]). By taking a case study of extreme surface temperature in Jakarta (the capital city of Indonesia), this study aimed to assess the performance of MPAS-A in simulating a mesoscale phenomenon, namely extreme heat condition, in a tropical city. Several validations of MPAS-A were quantified for near-surface meteorological variables, land surface temperature, and vertical profiles of atmospheric variables. This study also compared simulation results for different initialization/boundary conditions, urbanization conditions, and simulation domains.

As a coastal city on the northern coast of Java Island, Jakarta is influenced by sea breeze circulation. Previous observational studies [32,33] have characterized this phenomenon in the region, identifying that the sea breeze signal is most pronounced during the dry season months, from July to October. During this period, sea breeze fronts can develop along the northern coastal plain of West Java and propagate inland until encountering complex topography. Due to its dynamic characteristics, it is crucial to analyze the role of the sea breeze in temperature advection in Jakarta. Previous studies have highlighted the potential of sea breezes to mitigate high surface air temperatures during daytime in various coastal cities, including in Greece [34], Australia [35], Japan [36,37], and China [38].

This paper is organized as follows: Section 2 details the methodology by providing explanations about the study area and case study, model configuration and experiment designs, and validation metrics. In addition, a method to analyze surface air temperature variations is also provided in this section. The results, presented in Section 3, emphasize validations in near-surface meteorological variables, land surface temperature, and vertical profile of atmospheric variables. In addition, the effect of terrestrial dataset modifications on simulated land surface and air temperatures is presented in this section. The last two parts of this section are dedicated to spatial and temporal variations in the extreme event and the possible mechanism of the air temperature variations. Section 4 discusses the role of sea breezes as natural cooling mechanisms in Jakarta and the limitations of this study. Lastly, the conclusion is provided in Section 5.

2. Materials and Methods

2.1. Study Area and Case Study

The study area encompasses the west part of Java Island (Indonesia), which includes Jakarta, West Java, and Banten provinces (Figure 1a). As mentioned in [39], those provinces have the densest population in Indonesia. The region has a tropical climate and generally experiences two seasons: the dry season (April to October) and the wet season (November to March). As the largest city in Indonesia, Jakarta is classified as a megapolitan city with over 10 million residents. Given the high population density, an extreme event would have a more significant impact in this region. Geographically, Jakarta is a coastal city facing the Java Sea to the north. As the population has been continuously growing, the urban areas in Jakarta and the surrounding cities (Bogor, Depok, Tangerang, and Bekasi) have been developing and merging into the Greater Jakarta (Jakarta metropolitan area).

Figure 1a also presents the terrain height of the area as one of the static terrestrial data inputs in MPAS-A. The region features diverse terrain characteristics, including relatively flat coastal areas to the north, west, and south sides of the land and complex mountainous terrain in the central region. This variability in terrain naturally leads to a range of dynamic interactions and local wind patterns. For instance, the coastal areas experience land–sea breezes, which are well-documented phenomena [32,33]. In contrast, the mountainous central region is characterized by anabatic and katabatic winds [40]. These diverse terrain features result in unique challenges for numerical simulations.

Figure 1. (a) Terrain height (m) in the MPAS-A with mesh resolution of 3 km over the study area. Black dots with letters (A–G) show the locations of ground-based observation stations for validation (data source: BMKG, Indonesia). Orange polygon shows province boundary, while black indicates regency/municipality boundary (data source: GIA, Indonesia [41]). (b) Monthly climatology of maximum (red), mean (black), and minimum (blue) temperatures. (c) Maximum values by month of daily maximum, mean, and minimum temperatures in September (solid lines) and October (dashed lines) for each year from 1987 to 2023. (d) Daily (maximum, mean, and minimum) temperatures in October 2023. The data in Tangerang Selatan (marked as D in (a)) are used for (b–d).

Figure 1. (a) Terrain height (m) in the MPAS-A with mesh resolution of 3 km over the study area. Black dots with letters (A–G) show the locations of ground-based observation stations for validation (data source: BMKG, Indonesia). Orange polygon shows province boundary, while black indicates regency/municipality boundary (data source: GIA, Indonesia [41]). (b) Monthly climatology of maximum (red), mean (black), and minimum (blue) temperatures. (c) Maximum values by month of daily maximum, mean, and minimum temperatures in September (solid lines) and October (dashed lines) for each year from 1987 to 2023. (d) Daily (maximum, mean, and minimum) temperatures in October 2023. The data in Tangerang Selatan (marked as D in (a)) are used for (b–d).

Figure 1a also shows the distribution of seven observational stations, indicated by black dots with letters, used for model validation of surface meteorological variables (Section 2.3).

Table 1. Information of the stations that are used for model validation (source: BMKG, Indonesia).

Label	Station ID	Station Name	Municipality/Regency	Elevation (m)
A	96749	Soekarno Hatta	Tangerang	11
B	96741	Tanjung Priok	Jakarta Utara	3
C	96745	Kemayoran	Jakarta Pusat	4
D	96733	Banten	Tangerang Selatan	27
E	96753	Jawa Barat	Bogor	207
F	96751	Citeko	Bogor	920
G	96783	Bandung	Bandung	791

summarizes the information on the stations used for the validation. Figure 1b shows that the highest maximum temperatures occur in September and October, based on the monthly climatology of maximum temperature recorded in Tangerang Selatan station (point D in Figure 1a). Long-term records (from 1987 to 2023) of surface air temperature indicate that the highest maximum temperature was recorded in October 2023 (Figure 1c). Daily maximum temperature values in October 2023 (Figure 1d) reveal that the highest temperature occurred on 17 October. Therefore, this study simulated the atmospheric conditions around that date (16–17 October 2023). Global domain simulations using MPAS-A with variable-resolution meshes demand significant computational resources [20]. Due to these computational constraints, the simulation period in this study was limited to 48 h. However, this duration was sufficient to capture the peak of the extreme heat event on October 17, with the first six hours of the preceding day used for spin-up. Previous numerical studies have employed even shorter simulation periods, such as a 30 h simulation for 10 m wind over the Persian Gulf using WRF with a 4 km resolution [42] and a 36 h precipitation simulation around Japan and South Asia using the Japan Meteorological Agency (JMA) non-hydrostatic model (NHM) and WRF with a 20 km resolution [43].

2.2. Model Configuration and Experimental Designs

The MPAS-A model version 8.1.0 was used in this study. The model configurations included a variable mesh that smoothly changed from the spatial resolution of 3 km in the study area to 60 km when moving farther from the focused area. Eight simulations were conducted in this study (

Table 2. Numerical experiments with a combination of simulation domain, initial/boundary condition (IBC), and terrestrial datasets.

Simulation	Combination (Sim. Domain–IBC–Terrestrial Data)	Number of Vertical Levels
01	Global–NCEP–Default	41
02	Global–ERA5–Default	41
03	Global–NCEP–Modified	41
04	Global–ERA5–Modified	41
05	Regional–NCEP–Default	55
06	Regional–ERA5–Default	55
07	Regional–NCEP–Modified	55
08	Regional–ERA5–Modified	55

), combining (1) simulation domain, (2) initial/boundary conditions, and (3) terrestrial datasets (land use and land cover [LULC], albedo, and green vegetation fraction [GVF]). These combinations facilitate ensemble simulations, generating a range of simulated weather conditions. Given that initial and boundary conditions derived from reanalysis or analysis datasets contain inherent uncertainties [44], multiple simulation realizations are preferable to a single deterministic realization. Additionally, these combinations serve to assess the flexibility of MPAS-A in modifying each of the aforementioned three points.

The simulation domains were configured with two options: (a) global and (b) regional (limited area) domains (please see the Supplementary Information [SI] in Figure S1 for both domains). The initial or boundary (only for regional simulations) conditions are from either (a) the National Centers for Environmental Prediction (NCEP) Global Data Assimilation System (GDAS) final analysis [45] or (b) the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5) [46,47,48]. All experiments used sea surface temperature (SST) input from ERA5. Both data sources are widely used and considered reliable choices for the model’s initial/boundary conditions [42,49]. The terrestrial datasets had two alternatives: (a) default and (b) modified versions, as shown in Figure 2a–f. The default terrestrial datasets were sourced from the Moderate Resolution Imaging Spectroradiometer (MODIS) and are freely available on the webpage of the Weather Research and Forecasting Model Preprocessing System (WRF WPS) geographical data [50]. Modifications to these datasets (LULC, GVF, and albedo) were applied only within the area inside the gray polygon in Figure 2a–f.

The modified version of the terrestrial datasets is explained as follows: LULC within the gray polygon was updated by substituting the default data with a classified LULC product from 2019, as provided by previous work [39], to better represent the current LULC conditions (primarily for the urban pixel) in the study area (Figure 2b). In the default LULC (Figure 2a), urban pixels are primarily concentrated in Jakarta, Bogor (near station E), and Bandung (around station G), with most of the study area covered by forest (dark green shading). The modified LULC (Figure 2b) shows an expansion of urban areas, with croplands (green shading) becoming the predominant LULC in the modified dataset.

The GVF values were customized only for the updated urban pixels (Figure 2d). In the modified version, the albedo of urban pixels was set to 0.135, based on the dominant albedo values observed over those pixels as inferred from Figure 2g. Since the extreme hot condition occurred primarily in the urban area around Jakarta, modifications to GVF and albedo were applied only to urban pixels within the gray polygon.

The simulation period is two days, from 16 October 2023 at 00:00 UTC to 18 October 2023 at 00:00 UTC. The model’s top height is set at 30,000 m, with 41 vertical layers for global domain simulations and 55 layers for regional domain simulations. The model integration uses a time step of 20 s, with output data saved every hour. Due to computational constraints, this study directly employed the “convection-permitting” physical suite of parameterization schemes in MPAS-A, as recommended for simulations with a spatial resolution below 10 km according to the MPAS user’s guide (https://www2.mmm.ucar.edu/projects/mpas/mpas_atmosphere_users_guide_8.1.0.pdf, accessed on 22 May 2024). This suite comprises the Grell–Freitas scheme for convection parameterization [51], the Thompson (non-aerosol-aware) microphysics scheme [52], the Noah land surface model, the Mellor–Yamada–Nakanishi–Niino (MYNN) boundary layer scheme [53,54], the MYNN surface layer scheme, the rapid radiative transfer model-GCM applications (RRTMG) for longwave and shortwave radiation parameterizations [55], the Xu–Randall scheme for cloud fraction in radiation parameterization, and the Yonsei University (YSU) scheme for parameterizing gravity wave drag due to orography.

2.3. Validation of Near-Surface Meteorological Variables

The validation was conducted by comparing time series data from the model outputs at grid points overlapping with observational stations. For the seven stations (as shown in Figure 1a), five variables were compared between the model outputs and observations, namely 2 m air temperature (T2m), 2 m relative humidity (RH), surface pressure (SP), and 10 m wind speed (WS) and wind direction (WD). The observational data were sourced from BMKG (Indonesia’s meteorology, climatology, and geophysical agency). The model’s performance was quantified by calculating temporal Pearson correlation and mean absolute error (MAE). The first six hours of the model outputs were excluded from these calculations as part of the model spin-up. The correlation coefficient indicates the correspondence between the model output and observations [24]. Meanwhile, MAE measures the average magnitude of the model’s absolute error.

The formulas for the Pearson correlation (Equation (1)) and MAE (Equation (2)) are as follows:

$$r_j={\displaystyle \frac{\sum _{i=7}^{i=49}\left(o_i^{\prime }{m}_{i,j}^{\prime }\right)}{{\left[\sum _{i=7}^{i=49}{\left(o_i^{\prime }\right)}^2\right]}^{1/2}{\left[\sum _{i=7}^{i=49}{\left(m_{i,j}^{\prime }\right)}^2\right]}^{1/2}}}$$

(1)

$$MA{E}_j={\displaystyle \frac{\sum _{i=7}^{i=49}\left|m_{i,j}-o_i\right|}{N}}$$

(2)

where $r_j$ and ${MAE}_j$ represent the correlation value and mean absolute error, respectively, at a station or grid point for simulation $j$ ($j=1, 2, 3, \dots , 8$). The primes denote anomalies (i.e., subtractions from the mean values). $m_{i,j}$ is the model output at time $i$ for simulation $j$; $o_i$ is the observational value at time $i$, and $N$ is the number of paired model output and observational data points. The correlation value ranges between −1 and +1, with −1 indicating a perfect negative linear correlation, +1 indicating a perfect positive linear correlation, and 0 indicating no correlation. An MAE value of zero signifies a perfect simulation, while higher MAE values indicate greater discrepancies between the simulation and observation [56]. For the calculation of MAE in wind direction, the numerator was adjusted to $360°-\left|m_{i,j}-o_i\right|$ if $\left|m_{i,j}-o_i\right|>180°$, ensuring that the smallest directional difference between the model and observation was used, considering that the deviation cannot be more than 180° in modulus [57]. Validation of wind direction was only conducted for observations where wind speed is greater than zero, as observational errors in wind direction tend to be larger at lower wind speeds [58].

2.4. Validation of Land Surface Temperature

The reference of land surface temperature (LST in °C) was retrieved from the Himawari-8/AHI (advanced Himawari imager) gridded dataset [59], provided by the Center for Environmental Remote Sensing (CEReS), Chiba University, Japan. The LST retrieval utilizes three thermal infrared bands (B13, B14, and B15) from Himawari-8, with inputs consisting of top-of-atmosphere brightness temperature and land surface emissivity corresponding to each band, referred to as the YAM algorithm [59,60]. This dataset has an hourly temporal resolution and a spatial resolution of 0.02° (~2.22 km). Validation was quantified by calculating each grid’s temporal correlation (Equation (1)) and MAE (Equation (2)). Before these calculations, the LST outputs from the model were linearly interpolated from a 3 km resolution to match the spatial resolution of the reference dataset.

2.5. Validation of Vertical Profile of Atmospheric Variables

The reference dataset for this step is the sounding data from a single station—Jakarta observation station at Soekarno Hatta Airport (point A in Figure 1a, station ID 96749)—downloaded from this website: https://weather.uwyo.edu/upperair/sounding.html (accessed on 23 April 2024). Four time points of the sounding dataset were used, namely 16 October 2023 at 12:00 UTC, 17 October 2023 at 00:00 UTC and 12:00 UTC, and 18 October 2023 at 00:00 UTC for three variables: air temperature (°C), relative humidity (%), and mixing ratio (g/kg). The bias, or differences between model and observation ($m-o$), was calculated to assess the systematic discrepancy between the model’s output and observations [24].

2.6. Analysis of Spatio-Temporal Variations in Surface Air Temperature

As mentioned in Oke et al. [61], changes in temperature over time in the air layer or volume near the surface are influenced by four factors, namely the divergence of net radiation, the divergence of turbulent sensible heat (assuming no latent heat release due to fog formation), anthropogenic sources, and the effect of horizontal advection of temperature by wind.

Table 3. List of the output variables from MPAS-A used to calculate the radiation and energy budgets at the surface.

Variables	Description	Unit
gsw	net surface shortwave radiation flux	$\mathrm{W}{\mathrm{m}}^{-2}$
glw	all-sky downward surface longwave radiation	$\mathrm{W}{\mathrm{m}}^{-2}$
hfx	upward heat flux at the surface	$\mathrm{W}{\mathrm{m}}^{-2}$
lh	latent heat flux at the surface	$\mathrm{W}{\mathrm{m}}^{-2}$
skintemp	ground or water surface temperature	K

lists output variables from the MPAS-A related to the radiation and the energy budget at the surface. The net radiation (Q) is calculated as the sum of net shortwave radiation (SW) and net longwave radiation (LW). SW is directly extracted from the variable ‘gsw’, while LW is calculated from ‘glw’ minus the upward longwave radiation, defined as $\epsilon \sigma {\left(\mathrm{s}\mathrm{k}\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{m}\mathrm{p}\right)}^4$. Here, $\epsilon$ (emissivity) is assumed to be 1, and $\sigma$ is the Stefan–Boltzmann constant ($5.67\times {10}^{-8} \mathrm{W}{\mathrm{m}}^{-2}{\mathrm{K}}^{-4}$). Sensible heat was directly extracted from variable ‘hfx’. However, using these variables, quantifying the divergence of net radiation and turbulent sensible heat, especially for the vertical components of the fluxes’ divergence, proved challenging.

Additionally, for specific urban modeling, MPAS lacks the detail found in the WRF-UCM (urban climate model [62]), particularly in parameterizing urban surface processes, high-resolution urban land use, building morphology, and anthropogenic heating. Consequently, this study focuses primarily on analyzing the effects of horizontal advection. Temperature advection [63] is mathematically expressed as:

$$-\mathbf{U}·\mathsf{\nabla }T=-u{\displaystyle \frac{\partial T}{\partial x}}-v{\displaystyle \frac{\partial T}{\partial y}}$$

(3)

where $\mathbf{U}$ is the horizontal wind vector and $\mathsf{\nabla }T$ represents the vectorial form of the air temperature gradient. This study uses 10 m wind data and 2 m air temperature from the simulations. On the right-hand side of Equation 3, the first and second terms represent the zonal and meridional components of the advection effect, respectively. Consistent with Section 2.4, all data were remapped onto a 0.02° grid using a barycentric interpolation through a utility tool provided by MPAS-A.

3. Results

3.1. Validation of Near-Surface Meteorological Variables

Figure 3 presents time series plots of the five variables (T2m, RH, SP, WS, and WD) across the seven stations, comparing observational data with results from the eight simulations (the scatter plots for the data in Figure 3 are presented in Figure S2 of the SI). The model generally performs well in representing T2m, RH, SP, and WS, and it can simulate the temporal variations in those variables. However, it is less effective at accurately capturing the temporal pattern of WD. A previous work [10] using the improved WRF also showed that simulated WS, air temperature, and RH in Jakarta were also in good agreement with the observations. Additionally, studies on Hanoi cases indicated that the performance of WRF in simulating wind direction was less reliable than its performance in simulating air temperature [12,13].

Figure 3 also shows that the diurnal variations in T2m are well captured, with maximum values occurring around noon local time (LT) and minimum values observed before sunrise. At the Soekarno Hatta station (A), the model outputs are overestimated during the nighttime, while at Tanjung Priok (B), all simulations are consistently underestimated. This underestimation is also detected around noon at Citeko (F) and Bandung (G). At Kemayoran (C), the observed T2m shows a dramatic decrease of nearly 10 °C at 12:00 LT. A similar drastic change is seen in RH simultaneously, with a drop of around 40%. These abrupt changes might indicate errors in the observation. Smaller abrupt changes are also noted at Tanjung Priok. At Banten station (D), simulation 6 (represented by a dashed green line) closely matches the observation before noon but fluctuates between underestimation and overestimation from 12:00 LT to 18:00 LT.

From Figure 3, RH has an opposite pattern to T2m, with lower values during daytime and higher values at night. At Soekarno Hatta, the model underestimates nighttime RH. Meanwhile, most simulations overestimate RH during the daytime at Tanjung Priok. The ranges of the simulated T2m and RH are lower than those observed in that station. SP displays a semidiurnal pattern (two peaks in a day), indicating atmospheric tides [64]. In general, all simulations can capture this pattern, with values reasonably close to the observation. However, in Citeko, the model tends to overestimate SP, while in Bandung, the model tends to underestimate it. This discrepancy is possibly due to elevation differences between the actual terrain and the model’s representation, which uses smoother terrain [57] based on the model’s spatial resolution.

Meanwhile, WS generally increases during the daytime and decreases at night. This diurnal pattern is primarily due to the increased downward turbulent mixing of momentum during the day [65]. These observed patterns are also well simulated by the model, though the model output shows a smoother pattern during the daytime. One simulation, namely simulation 6, produced an unreasonably rapid change in WS, reaching 9 m s⁻¹ at station B (Tanjung Priok) after midnight on 17 October 2023. WD is well simulated during the daytime when WS is stronger, with the wind blowing between 180° and 0° (a veering pattern, indicating a clockwise change in WD). However, when WS decreases during the nighttime, many WD observations are missing, making it more difficult to compare the simulation with observation. The general sea breeze pattern is observed in stations A, B, C, and D (those located near the northern coastline, Figure 1a). In the early morning (~7 LT) on 17 October 2023, the wind direction is approximately southerly (~180°) and shifts to approximately northerly (~0°) by noon or afternoon. This sea breeze pattern was also observed by [10] over Jakarta. The MPAS-A performed quite well in representing this pattern at the four stations.

Figure 4 presents each simulation’s quantitative validation measures, namely time correlation values (upper panel) and MAE (lower panel). Both metrices are calculated by combining samples from the seven stations. Consistent with the previous qualitative assessment, strong correlation values are also shown for T2m, RH, SP, and WS, exceeding 0.6. For SP, the correlations nearly reach 1 across all simulations, indicating that the temporal pattern of the simulations closely aligns with the observations. Meanwhile, for WD, the correlation values range from around 0.2 to slightly above 0.4, with a significantly high MAE ranging from just under 55° to approximately 60°. Summarizing the model performance by simulation design, except for SP and WD, is challenging. For SP (Figure 4h), the global simulation domain (simulations 1–4) exhibits a lower absolute error compared to the regional simulation domain (simulations 5–8). Regarding WD, simulations using ERA5 as the initial/lateral boundary source (even-numbered simulation) show a lower MAE than those using NCEP as the source of the initial/boundary condition (odd-numbered simulations).

3.2. Validation of Land Surface Temperature

Figure 5a–h illustrate the spatial pattern of the time correlation value of LST for each grid, with the distribution across all grids presented in Figure 5i for each simulation. Generally, the correlation values are high. Relatively smaller correlation values are found over the northern coast east of the Jakarta province, which corresponds to agricultural areas, as shown in Figure 2a,b. Lower correlation values are also found in the higher elevation areas around the center and southern parts of the study area, as indicated in Figure 1a. Overall, all simulations display similar patterns of spatial correlation values. As shown in Figure 5i, most grids have correlation values greater than 0.8, as inferred from the distribution pattern in the box plots. A limited number of lower correlation values (lower than 0.6), particularly over the higher elevation regions, are indicated by red plus signs. The high correlations suggest that the model outputs align well with the LST observations. In constrast, the generally consistent spatial pattern of high correlation indicates a consistent spatial pattern over time between the observation and the model outputs.

Figure 6a–h display the spatial pattern of gridded MAE of LST for each simulation, while Figure 6i shows the MAE distribution across all grids. Consistent with the lower correlation locations, higher MAE values are generally found on the northern coast east of Jakarta province, in the mountainous areas in the middle of maps, and in smaller regions along the southern coast of the study area. Figure 6i shows that the MAE of LST has average and median values around 3 °C, with most MAEs being below 4 °C. However, some grid points have relatively higher MAE values, around 6 °C or more, as indicated by the red plus signs. The lower correlation and higher MAE over the higher elevation might be due to the simulated LSTs lacking finer spatial details in these regions because of the model’s smoothed terrain, as inferred from previous work [66] over more complex terrain.

3.3. Validation of Vertical Profile of Atmospheric Variables

Figure 7a–c present the observed vertical profile of atmospheric conditions at the four time points, represented by different line colors, for three variables such as temperature (Figure 7a), relative humidity (Figure 7b), and mixing ratio (Figure 7c). Air temperature decreases with height in the lower to middle atmosphere (1000–100 hPa), corresponding to the troposphere. In contrast, between 80 and 10 hPa, the air temperature increases with height, which is typical of the stratosphere. The boundary between these two layers is known as the tropopause, characterized by a relatively constant temperature with height.

In Figure 7a, the tropopause is visually identifiable between 100 and 80 hPa, climatologically close to 100 hPa in the equatorial region [67]. RH varies significantly from the surface to the 50 hPa level. Above the level of 50 hPa, RH values reach 0%. Below 500 hPa, the general pattern of RH has a higher value with a decreasing pattern with height, with lower values observed between 500 and 300 hPa. Between 200 and 100 hPa, RH remains below 50%. The mixing ratio has a decreasing pattern with height, starting near 20 g/kg at the surface and reaching zero above 500 hPa. These patterns reflect the higher availability of water vapor in the lower troposphere, while water vapor is more limited in the upper troposphere and scarcer in the stratosphere.

Figure 7d–f show vertical profiles of biases (differences) in air temperature, relative humidity, and mixing ratio between the simulation outputs and the observed patterns averaged across the four observation times, respectively. The ranges of the biases from eight simulations are detailed in Figure S3 of the SI. In Figure 7d–f, the line colors and types represent different simulation experiments. Figure 7d reveals that the air temperature biases vary with vertical pressure. Larger bias variations are found in the middle to upper atmosphere (above 200 hPa) compared to the lower atmosphere (below 200 hPa), except near the surface. Figure 7d also indicates that all simulations have biases in the ranges of −3 °C to 3 °C, except around 60 hPa (more than 3 °C), 40 hPa (less than −3 °C for simulations 6 and 7), below 30 hPa (for simulations 1–4), and at the top of the observation level. A clearer separation of the biases, grouping simulations with different initial/boundary conditions, is detected in the middle to upper troposphere (between 300 hPa and around 150 hPa). Above 200 hPa to around 150 hPa, simulations with initial/boundary conditions from ERA5 (green and yellow lines for even-numbered simulations) show more overestimated values (more positive biases). Conversely, from 300 hPa to 200 hPa, simulations with initial/boundary conditions from NCEP (red and blue lines for odd-numbered simulations) exhibit more underestimated values (more negative biases).

Figure 7e shows the RH biases across different vertical pressure levels. In most levels, simulations display positive biases of RH, except around 500 hPa, 700 hPa, 900 hPa, and near the surface. The biases are relatively smaller in the troposphere (below 200 hPa) compared to those in the layer above 200 hPa. In the upper troposphere to the lower stratosphere (from 300 hPa to 40 hPa), there is a clearer separation of biases based on the initial/boundary conditions: simulations with NCEP data (red and blue lines) show smaller positive biases compared to those using ERA5 data (green and yellow lines). Regarding the mixing ratio biases, Figure 7f shows that the lower troposphere (a layer below 500 hPa) has larger biases with an alternating pattern than the layer above it. This shows that the accuracy of the mixing ratio is becoming lower in the layer with more water vapor, such as in the lower troposphere, as shown in Figure 7c.

3.4. Effect of Terrestrial Dataset Modifications on Land Surface and Air Temperatures

Figure 8 shows the LST (left panel) and T2m (right panel) differences between paired models that used modified terrestrial inputs (simulations 3, 4, 7, and 8) and those that used default terrestrial inputs (simulations 1, 2, 5, and 6). The differences with the paired output models were conducted to analyze the effect of modification in the terrestrial datasets. The default and modified versions of the terrestrial dataset are shown in Figure 2. By comparing Figure 2a,b and Figure 8a–d, significantly higher positive differences (2 °C–5 °C) in LST are detected over the sprawling (new) urban areas represented in the modified version of the LULC dataset, such as areas west of Jakarta, between Jakarta and Bogor, the southern part of Bandung, and several smaller urban areas (for reference to the region names mentioned, please see the location of the stations used for validation as in Table 1). In the urban area already represented in the default version, smaller increases in LST are detected, such as within Jakarta, Bogor, and Bandung (represented by light green shading). In regions classified as non-urban in both the default and modified LULC datasets, a common pattern of slight LST increases is found in the northern part of the study area, over the agriculture area east of Jakarta (Figure 8a–d). However, inconsistent patterns are found in the southern part of the study area, with negative differences shown in Figure 8a,b and positive differences in Figure 8c,d.

Significant positive differences (approximately 1 °C) in T2m are consistently detected over the urban areas (Figure 8e–h), indicating that the modifications to the terrestrial dataset can simulate higher LST and T2m in the updated urban areas, as shown in Figure 2b. In regions classified as urban in the default LULC, positive differences in T2m are generally detected but with smaller magnitude, such as within Jakarta, Bogor, and Bandung. A notable exception with small negative differences is present in the northern part of Jakarta in Figure 8g. Higher LST and T2m are typically observed in urban built-up regions, suggesting that the modified version of the terrestrial datasets in LULC, GVF, and albedo could simulate more realistic and coherent urban characteristics.

Inconsistent differences are found over non-urban areas, for example, the region east of Jakarta, where small positive differences (yellow shading in Figure 8e,f) contrast with negative differences (green shading in Figure 8h). This suggests that a more detailed classification is necessary for the non-urban regions. The modified LULC dataset provided by previous work [39] distinguishes only between two types of greeneries—low vegetation and dense vegetation/forest—and does not fully represent the non-urban conditions in October 2023. Additionally, more terrestrial parameters, such as soil characteristics, should be updated for model input in non-urban areas, as inferred from [68]. Since this study primarily focused on the extreme conditions that occurred in the urban areas around Jakarta, the inconsistencies in simulated parameters over non-urban regions are unlikely to affect the conclusions significantly. Moreover, validating near-surface meteorological variables and LST around Jakarta remains reliable. However, more detailed LULC classification and terrestrial parameters that better represent surface conditions will be crucial.

3.5. Spatial and Temporal Variations in the Extreme Event

Figure 9 shows the spatial pattern of hourly surface air temperature and 10 m horizontal wind over the study area from 9 LT to 17 LT on 17 October 2023. The figure is based on the output of simulation 8, which is presented due to its smaller MAE in T2m and WD, as shown in Figure 4f,j. Similar plots for the other simulations are available in Figures S4–S10 of the SI. Three boxes are included in each plot to highlight the spatial and temporal variations in the extreme event. Purple and black boxes are located over Jakarta, with the purple box over the northern part and the black box over the southern part of Jakarta. One green box is located over the agricultural field, as indicated in Figure 2a,b.

The simulation indicates that higher temperatures are present over the lower elevation areas in the northern part of the study region. Air temperatures exceeding 36 °C are detected east of Jakarta at 11:00 LT, with these elevated temperatures persisting until the afternoon (15:00 LT). Compared to Figure 2b, the regions with higher temperatures correspond to built-up areas, including industrial towns and estates [69], as well as agricultural fields. In October, the agricultural fields are commonly bare after the harvest season, as indicated in Figure 2d, showing a lower green vegetation fraction (GVF).

The simulation also reveals spatial variations in surface air temperature between Jakarta’s northern and southern areas, as indicated by the purple and black boxes, respectively. The southern region has a higher temperature, greater than 36 °C, with these extreme temperatures persisting until the afternoon (16 LT). Meanwhile, high temperatures over the agricultural area (green box), east of Jakarta, decrease more rapidly than those in the urban areas. This rapid temperature decline in the agricultural region is likely influenced by reduced sensible heat and negative temperature advection, as discussed in detail in the following subsection. During the daytime, the simulation also depicts a sea breeze pattern, with northerly to northeasterly winds blowing from the sea water onto the mainland in the northern part of the study area. The intensification of sea breeze flow occurred between 12:00 and 17:00 LT. This intensification may enhance moisture transport in the region, potentially explaining the increase in relative humidity observed at the coastal stations (Stations A–D), as shown in Figure 3. A previous observational study [32] also noted a significant increase in relative humidity and the strengthening of northerly winds in the lower atmospheric layer during sea breeze intrusions in Jakarta.

3.6. Possible Mechanisms for the Variations in Surface Air Temperature

Figure 10 illustrates the hourly variations in several variables throughout the day of the extreme event (17 October 2023) over the three boxes defined in Figure 9. Specifically, Figure 10a–i display (a) near-surface air temperature, (b) surface radiation budget, (c) sensible and latent heat, (d) time derivative of air temperature, (e) temperature advection, (f) other effects defined by subtracting (d) from (e), (g) zonal temperature advection, (h) meridional temperature advection, and (i) horizontal winds.

Figure 10a shows that during the nighttime to early morning (6:00 LT) on 17 October 2023, air temperatures over the urban areas (black and purple lines) are approximately 4 °C higher than those over the agricultural field (green line). In certain periods (8:00 LT to 14:00 LT), the agricultural land also shows a high temperature with comparable magnitude to that of the southern urban area (black line). Starting at 7:00 LT, air temperature over the agriculture field becomes higher than that over the northern urban area. Both urban areas do not show significant differences in air temperatures before 9:00 LT in the morning. However, from 9:00 LT until nighttime at 20:00 LT, the southern urban area (black line) exhibits a higher temperature than the northern urban area (purple line), as also indicated in Figure 9. Observations of T2m in two stations (A and D) located in the northern and southern urban areas, respectively, also confirm that the southern part shows a higher temperature than the northern part (red and blue dashed lines in Figure 10a).

For the surface radiation budget (Figure 10b), the northern urban area (solid purple line) exhibits significantly higher net radiation than the southern urban area from noon to 15:00 LT. The same pattern was also observed in net SW (dotted lines; positive values mean downward SW). This indicates that higher net SW contributed to net radiation. Meanwhile, there are no significant differences in the LW for the three locations (dashed lines). The temporal patterns of the radiation budget (Net Q, SW, and LW) over the agricultural land are comparable to the northern urban area.

In the sensible and latent heat contexts (Figure 10c), sensible heat in the northern urban area is higher than in the southern part. However, in both urban areas, the latent heat is comparable. It seems that the higher value in the net radiation in the northern part transformed to higher upward sensible heat from the surface over the area. In the agricultural area, the sensible heat is higher than that in the urban areas, from morning at 7:00 LT to approximately 15:00 LT. It might contribute to the fast increase in the morning and fast decrease in the afternoon in surface air temperature. Meanwhile, LH over the agricultural land is higher than the urban areas. This might be due to more moisture availability over the agricultural surface even in dry conditions (after harvesting).

Figure 10d shows the time derivative of surface temperature, discretized using the central difference method, which indicates the rate of change in air temperature over time. It is discernable that there are positive temperature changes before noon and negative temperature changes in the afternoon. In the early morning, urban areas show a slower increase in temperature compared to the agricultural region. From 7:00 LT to noon, the southern urban area shows a greater temperature change than the northern part. In the afternoon, a higher magnitude of negative temperature change is shown over the agricultural area due to the fast decrease in air temperature there.

Despite a lower increase in air temperature (Figure 10a), higher net radiation (Figure 10b), and higher sensible heat (Figure 10c) in the northern urban area compared to the southern urban area, the contribution of temperature advection (Figure 10e) can explain these discrepancies. Significantly higher magnitudes of negative advection are located over the northern area compared to the southern part during the daytime. The minimum cool advection occurred at 12:00 LT, with a magnitude of approximately −8 °C/h in the northern part and approximately −2 °C/h in the southern part. This could result in a lower simulated air temperature over the northern part of Jakarta, even with higher simulated net radiation and surface sensible heat in that area.

Figure 10f illustrates the other factors, outside the effect of temperature advection, that contributed to the temperature changes over time. This pattern was calculated by subtracting Figure 10d from Figure 10e. These other factors might include contributions from radiation and turbulent diffusion. Here, the anthropogenic contribution to urban heat is excluded due to the inability of the current MPAS to explicitly simulate the complex urban metabolism. Figure 10f also infers that the high positive contribution of the other factors to air surface temperature changes in the northern urban area was compensated by large negative temperature advection; otherwise, the area will have higher temperature during the daytime.

Figure 10g,h show the zonal and meridional components of surface temperature advection, respectively. Warm zonal advection with a small magnitude was simulated over the urban areas around 10:00 LT and 14:00–16:00 LT, with the magnitude larger over the northern urban area. This warmer air was advected from the agricultural area (east of Jakarta) as also inferred from Figure 9 and Figure 10a as the temperature over the agricultural land increases faster in the morning and the horizontal wind has a considerable (negative or easterly) zonal component during that time (also shown by the green dashed line in Figure 10i). On the other hand, negative zonal and meridional temperature advections might have also contributed to the rapid decrease in surface air temperature over the agricultural area in the afternoon. Meanwhile, the meridional component of advection over the northern urban area has a large negative magnitude from morning to evening that contributed mostly to the total negative advection in Figure 10e. This indicates that the meridional advection was primarily caused by the sea breeze penetration, as inferred from the meridional wind component in Figure 10i. The sea breeze started at approximately 9:00 LT in the morning, and the meridional wind component strengthened over time until the afternoon (around 15:00 LT). Meanwhile, the negative temperature advection over the southern urban area has a lower magnitude than that over the northern area. The meridional wind component over the southern urban area was also lower compared to that over the northern urban area.

4. Discussion

4.1. The Role of Sea Breezes as Natural Cooling Mechanisms in Jakarta

This numerical simulation study indicates that the sea breeze effectively reduces high surface air temperature in the northern urban area of Jakarta during the extreme event. Several previous observational [34,35,36] and simulation [37,38] studies also reported the potential cooling capacity of sea breeze during extreme temperatures. A decrease in temperature of approximately 4 °C from the maximum temperature was reported in several Greek regions [34]. Another study mentioned that the average cooling capacity is approximately 21.3 °C h per event in an Australian city [35]. Previous studies [35,37,38] also reported that sea breeze cooling penetration varies from approximately 2.5 km to 29 km, with a gradient cooling magnitude based on the distance from the coastline.

While the sea breeze plays an important role in heat transport, studies by [70,71] reported a decreasing trend in the number of sea–land breeze events and their magnitude in China. The magnitude of land–sea breeze was generally influenced by the temperature difference between land and sea. The studies mentioned that both dynamic and thermodynamic factors contribute to the temperature gaps between land and sea. One of the dynamic factors is surface roughness, which is influenced by the average height of city buildings. Meanwhile, the cooling contribution from radiation due to anthropogenic aerosols could decrease the temperature gaps over land and sea.

For the Jakarta case, to mitigate the dynamic factors, it is recommended to minimize the number of high-rise buildings in the city region, as proposed by [38] for a coastal city in China. However, the number of tall buildings in Jakarta is increasing, currently with a total of 165 buildings, as recorded in the “Council on Tall Buildings and Urban Habitat (CTBUH)” database. The high-rise buildings can obstruct and slow down the surface northerly wind of the sea breeze. Meanwhile, for the thermodynamic factors, without policy intervention, Lestari et al. [72] predicted that the total emissions in Jakarta would continue to increase.

4.2. Limitations of This Research

Several validations were conducted in this study, including surface meteorological variables (Section 3.1), land surface temperature (Section 3.2), and vertical profile of the atmosphere (Section 3.3). However, validation of the energy budget was not performed. A previous study by Junnaedhi et al. [10] validated radiation output variables from the improved WRF model, which incorporated new aerodynamic parameterization and anthropogenic heat. Their study reported an overestimation of simulated downward shortwave radiation and sensible heat flux but an underestimation in latent heat flux simulation. Future studies should include validation of these variables if observational data on the energy budgets become available.

Meanwhile, major urban parameters such as anthropogenic heat and building height are not represented in the current MPAS-A model. In contrast, those parameters are already implemented in the WRF urban canopy model (UCM), reflecting advancements in urban modeling with WRF. However, as a relatively recently developed model, MPAS-A presents significant opportunities for the weather and climate research community to refine and advance its capabilities in the future.

5. Conclusions

This study aimed to evaluate the performance of the MPAS-A for mesoscale simulation with a case study of an extreme condition of near-surface temperature in Jakarta, Indonesia. The period of simulation was 16–17 October 2023. Eight numerical experiments were run with a combination of simulation domain, initial/boundary condition, and terrestrial dataset (Table 2). Several validation assessments were applied for near-surface meteorological variables, land surface temperatures, and vertical atmospheric profiles, including time correlations, mean absolute errors, and biases.

The main findings of this study are as follows:

• During the extreme case, from the eight experiments, MPAS-A was able to simulate the diurnal pattern of the near-surface meteorological variables well, namely near-surface temperature, relative humidity, surface pressure, and 10 m wind speed, with correlation values exceeding 0.6. However, the model was less accurate in simulating the wind direction, with correlation values ranging from around 0.2 to slightly above 0.4 and MAE ranging from below 55° to approximately 60°. One of the reasons for the less accurate wind direction is the inaccurate DEM (digital elevation model) and domain resolution (3 km), for example, since the topography and urban obstacles (such as buildings) can affect the wind direction.
• From the eight experiments, validations on land surface temperature also showed good performance, with high time correlation values across most grids in the study area, where the area average and median of the time correlation values exceeded 0.8. Additionally, the area average of mean absolute errors is less than 4 °C.
• The vertical profiles of biases in air temperature, relative humidity, and mixing ratio varied through height. Generally, biases in air temperature and relative humidity were lower in the troposphere than those in the stratosphere, except near the surface, where air temperature biases were higher. In certain vertical layers of the atmosphere, biases were sensitive to the initial and boundary conditions used. For altitudes from 200 hPa to approximately 150 hPa, simulations based on ERA5 initial and boundary conditions showed higher temperatures. In contrast, from 300 hPa to 200 hPa, simulations using NCEP initial and boundary conditions tended to underestimate temperatures. Additionally, simulations with NCEP data exhibited a smaller positive bias in relative humidity compared to those using ERA5 data, particularly from 300 hPa to 40 hPa. Larger biases in mixing ratio with alternating patterns were shown in the lower troposphere (below 500 hPa).
• Modifications to the terrestrial datasets could simulate higher land surface and air surface temperatures with values of 2–5 °C and ~1 °C, respectively, over the updated sprawling urban areas.
• MPAS-A successfully captured the intensity of the extreme temperature event on 17 October 2023. During this event, the southern part of Jakarta had higher surface temperatures (greater than 36 °C) than the northern part, which persisted until the afternoon (16:00 LT).
• The simulations indicated that the lower air temperature over the northern part of Jakarta was due to the higher magnitude of negative advection over that region compared to the southern part during the daytime. The minimum cool advection occurred at 12:00 LT, with a magnitude of approximately −8 °C/h in the northern part and approximately −2 °C/h in the southern part. Separation into zonal and meridional components of the advection indicated that the meridional component contributed more significantly to the cool advection, influenced by the sea breeze.
• From the perspective of air temperature advection, one potential mitigation measure for hot weather conditions in a coastal city like Jakarta is to maximize the benefits of the sea breeze in transporting heat out of the city for city ventilation. This could be achieved by limiting the number of high-rise buildings, particularly in the coastal regions, as they increase surface roughness, reduce wind speed, and hinder sea breeze penetration. Additionally, this numerical study highlights the potential application and development of the MPAS-A model for urban studies. Future advancements could include enhancing urban parameterizations (such as detailing urban surfaces and accounting for building effects and anthropogenic influences) and incorporating urban canopy models, like those implemented in the well-known WRF-ARW model for urban studies. Moreover, MPAS-A is a feasible and promising tool for general applications in weather forecasting and climate prediction/projection, suitable for both research and operational purposes.