Impact of Domain Size on Tropical Precipitation Within Explicit Convection Simulations

We investigate the sensitivity of modeled tropical precipitation accumulation, intensity and structures to the extent of convection‐permitting limited area model (LAM) domain size. Our comparison focusses on two LAM domains, with identical physical parameterization schemes and using 2.2 km grid spacing. One LAM domain spans almost the full tropical belt while the other focusses on southeast Asia. We show that the LAMs both capture the complex diurnal cycle of precipitation and that the timing and intensity of precipitation are comparable with satellite observations. Systematic differences between the LAMs are largest within ∼1,000 km of the western and eastern boundaries of the southeast Asia LAM. This is due to convective spin‐up at the western boundary of the southeast Asia LAM and a lack of propagating deep convection. We highlight that showing the added value of global storm‐resolving models by comparing with LAMs will help to accelerate their operational implementation.

• We compare precipitation in a limited area model (LAM) to a novel 2.2 km resolution Tropical Channel domain, focusing on southeast Asia • A LAM can adequately reproduce the precipitation characteristics of a Tropical wide simulation at 15% of the cost • Further work within our model framework will investigate the upscale impacts of organized convection within the Tropical Channel represented in the coarser resolution parent model. Therefore, operational LAMs have often increased domain sizes or deployed variable resolution meshes to move the lateral boundaries further from the area of interest (Rennie et al., 2022;Tang et al., 2013).
Here we present results from a large 2.2 km resolution Tropical Channel model and compare with a continental-scale LAM focussed on southeast Asia (SE Asia). Both LAMs use the same physics (and dynamics) configuration and horizontal mesh but the Tropical Channel LAM has boundaries that are far distant from the focus region for our analysis. We look to assess the added value of a large tropical LAM-a domain in which we would expect tropical waves to propagate coherently into the SE Asia region as they are internally generated rather than prescribed in the boundary conditions. The SE Asia focussed LAM is large compared with most operational regional NWP models, and its size is similar to cutting-edge regional climate modeling domains for example, CORDEX simulations (Diez-sierra et al., 2022). Any benefits over the traditional nested LAM approach arising from explicitly simulating convection across the Tropical Channel would also be expected to be realized in GSRMs.

Model Setup
The Tropical Channel domain is 17,000 grid points east-west by 3,300 north-south, spanning 170°W-170°E and 40°S-26°N. The SE Asia LAM, spans 90°E-154°E and 18°S-26°N (see Figure 1a for domain extent and Table 1 for further model details). The 2.2 km Tropical Channel simulation is computationally expensive; 1 simulated day uses as much computational resource as running the SE Asia LAM (at the same horizontal resolution) for 6.5 days (or the global N768 driving model configuration for ∼5 months).
Met Office Unified Model (MetUM) version 12.0 (Brown et al., 2012) is used, coupled to the Joint UK Land Environment Simulator to represent the land surface (Best et al., 2011). Both LAMs use a 2.2 km resolution horizontal mesh with 90 vertical levels up to a model top at 40 km, and use the Regional Atmosphere and Land version 3 (RAL3) physics configuration. RAL3 builds upon the RAL2 configuration (Bush et al., 2023), with addition of a double-moment cloud microphysics parameterization (Field et al., 2023) and a new bimodal large-scale cloud scheme (Van Weverberg et al., 2021). In both simulations the lower boundary sea-surface-temperature is updated daily using 1/20th degree Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA) data (Donlon et al., 2012), making the simulations more similar to hindcasts than traditional forecasts.
The Tropical Channel and SE Asia LAM are both forced with hourly lateral boundary conditions. These are provided by a global MetUM simulation with a grid spacing of approximately 17 km in the mid-latitudes (N768), which drives a 12 km resolution intermediate nest. The 12 km nest bridges the gap in horizontal resolution between global and LAM simulations and critically allows prescribed dust and aerosol ancillary data to be downscaled to 2.2 km resolution for the LAMs. Both the global model and 12 km intermediate nest parameterize convection and use global atmosphere land science configurations with 70 vertical levels and a model top at 80 km (Table 1). The global model uses the GAL6.1 configuration that was used in operational forecasts in 2017 (Walters et al., 2017), and the intermediate nest uses the GAL7.2 science configuration (Walters et al., 2019).
All simulations are initialized at 0600 UTC on 30 December 2017 from the Met Office global analysis and run for a 12-day integration period. We chose this period as it coincides with impactful flooding events in SE Asia. We exclude the first 48 hr of the simulations from our analysis to allow for the Tropical Channel and SE Asia LAM to spin-up realistic convective structures.

Observations
Given improvements to the water cycle are one of the key drivers of convection-permitting simulations, this study focusses on precipitation. Satellite-derived precipitation measurements from the Global Precipitation Measurement (GPM) are used for comparison with the MetUM simulations. The integrated multi-satellite retrievals (IMERG) data set is used, calculated using multiple satellite passes and later calibrated using rain gauge data to produce a half-hourly 0.1° resolution gridded precipitation field (Hou et al., 2014). MetUM precipitation fields are regridded on to the GPM grid prior to analysis.
Recent work has shown GPM to be reliable for assessing NWP model performance in SE Asia, although there are biases when comparing with observed daily rainfall thresholds above the 95th percentile (Da Silva et al., 2021).
Previous work has also highlighted that we should be cautious when comparing the size of precipitating cells in GPM to those in MetUM simulations (Hanley et al., 2021). GPM uses cloud cover to estimate rainfall which can lead to an overestimate of the size of features (Feng et al., 2023).

Observed and Modeled Spatial Distribution of Precipitation
Early January 2018 was dominated by a very active MJO propagating from Phase 2 to Phase 3 across the Indian Ocean toward SE Asia (Lafleur et al., 2015). Precipitation totals across the west of SE Asia are typically wetter than average in MJO phases 2 and 3 (Peatman et al., 2014). Observed precipitation totals during the case study are >8 mm day −1 across the eastern Indian Ocean and much of Sumatra, Java and Borneo (Figure 1a  To the southwest of Sumatra a depression forms in the Tropical Channel simulation on 6 January 2018-and persists for 72 hr-leading to larger precipitation totals compared with the SE Asia LAM. A tropical depression is also observed in a similar region but moves southwest more rapidly. The SE Asia LAM dry bias extends more than 10° (around 500 grid points) into the western side of the domain. This is much further than the typical value to allow for convective spin-up at the lateral boundaries in idealized model simulations (Ahrens & Leps, 2021).
Zonal variability in model-observation differences (Figure 1e), show that between 92°E and 104°E the SE Asia LAM (5.27 mm day −1 ) has less precipitation than both GPM (8.35 mm day −1 ) and the Tropical Channel (8.45 mm day −1 ). In contrast, between 135°E and the boundary of the SE Asia LAM domain (at 154°E) both the Tropical Channel (5.89 mm day −1 ) and SE Asia LAM (6.84 mm day −1 ) overestimate precipitation compared with GPM observations (4.11 mm day −1 ). The largest wet biases are close to the eastern boundary, and the magnitude of the bias is larger in the SE Asia LAM. Spurious convection propagates into the eastern portion of the LAM from the driving model with parameterized convection. MetUM simulations (along with many climate models) are known to overestimate precipitation and some produce a double ITCZ within the west Pacific region (Stanfield et al., 2016).
As highlighted in Figures 1d and 1e, toward the center of the domain (105°E to 135°E) differences between the two model configurations are small and both simulations are comparable with GPM observations. Domain mean precipitation amounts show relatively small model biases with the Tropical Channel (6.94 mm day −1 ) and SE Asia LAM (6.83 mm day −1 ) producing slightly more rainfall than GPM (6.59 mm day −1 ), but this masks significant regional biases discussed above.

Domain Mean Precipitation Characteristics, Cell Sizes and Timeseries
Timeseries of the domain mean precipitation through the final 10 days of the simulations (Figure 2a)  large-scale precipitation variability (R = 0.75). Both MetUM simulations fail to fully capture the reduction in GPM observed precipitation between 6 and 8 January 2018. This period sees both larger model-observation differences and larger model-model differences.
The histogram of hourly precipitation rate in Figure 2b shows the contribution of different precipitation rates to the total precipitation. The Tropical Channel and SE Asia LAM simulations have higher peak intensity compared with GPM observations, with modeled rainfall rates between 4 and 30 mm/hr contributing more to the mean than in GPM observations. Both produce less light precipitation (<2 mm hr −1 ) than GPM observations. The precipitation distribution of the two MetUM simulations are very similar to each other-the Tropical Channel has a marginally reduced peak intensity compared with the SE Asia LAM and slightly larger contributions to the mean for rainfall rates less than 10 mm hr −1 .
The aggregation of precipitation is assessed by computing the number of adjacent cells in the MetUM simulations and GPM observations using a representative rainfall threshold of 4 mm hr −1 and after regridding the model data on to the GPM 0.1° grid (Figure 2c; note analysis of other rainfall thresholds produce qualitatively similar results). A limitation of this methodology is that the same cell will be counted multiple times as cells are identified at an hourly frequency through the simulations. The Tropical Channel and SE Asia LAM are in very close agreement-both overestimate the frequency of small cell sizes (<30 km radius) and underestimate the number of large cells (>70 km radius) compared with GPM observations. The tendency for MetUM simulations to overly fragment precipitation into smaller cells than GPM observations has previously been noted in convection-permitting MetUM simulations over east Africa (Hanley et al., 2021). The tendency for GPM to produce too much light precipitation , and hence overestimate cell sizes (Feng et al., 2023), may also be contributing to the model-observation differences described here.

Diurnal Cycle of Precipitation
The geography of SE Asia with a network of islands that have steep and high orography surrounded by warm oceans create a complex diurnal cycle of precipitation (Love et al., 2011;Peatman et al., 2023). Figure 3 shows the timing and amplitude of the first harmonic of the observed and simulated diurnal cycle. Areas where the amplitude of the diurnal cycle is small such as parts of the West Pacific along with Vietnam and Thailand tend to be the areas where there are larger errors in the timing of the diurnal maximum. Figures 3a-3c show both MetUM simulations accurately capture the contrasting timing of the diurnal maximum of precipitation over islands such as Java, Borneo and Sumatra (∼0800UTC or 1500 local) compared with the later peak (∼1800 UTC or 0100 local) over surrounding ocean regions. Differences in the diurnal cycle between the Tropical Channel and SE Asia LAM simulations are negligible, both when looking at the first harmonic ( Figure 3) and broader domain mean statistics (not shown).
The amplitude of the diurnal cycle is overestimated compared with GPM observations over the complex terrain of New Guinea, Sulawesi and over the mountainous interior of Java (Figures 3d-3f). This tendency to produce more precipitation than GPM observations over steep orography in SE Asia has previously been highlighted in MetUM convection permitting ensemble simulations (Ferrett et al., 2021). In general, the MetUM simulations have less spatial coherence in the timing of the diurnal cycle (Figures 3a-3c) than the GPM observations-this may be linked to the overly fragmented precipitation cells (see Section 3.2) producing more small-scale variability than is seen in GPM observations.

Propagation of Precipitation
One known deficiency of parameterized convection configurations of the MetUM is that the MJO amplitude and propagation are much weaker than observed (Holloway et al., 2013). The boundary conditions of the Tropical Channel domain are distant from the region of interest, and we hypothesize that organized convection associated with tropical waves will propagate more coherently in the Tropical Channel as waves are internally generated rather than prescribed through coarser scale lateral boundaries.
A precipitation Hovmoller plot averaged between 15°S and 15°N using GPM observations (Figure 4a) shows active deep convection at 90°E at 0000 UTC on 2nd January, propagating eastwards with embedded diurnal variability. It can be traced as far as 120°E at 0000 UTC 7 January 2018. Both Tropical Channel ( Figure 4b) and SE Asia LAM (Figure 4c) simulations capture some of the initial convective activity at 90°E, though its intensity and spatial scale are underestimated. There is little evidence of subsequent propagation. The Tropical Channel (Figure 4b) shows enhanced precipitation propagating eastwards later in the simulation (from 98°E at 0000 UTC 7th January). This highlights that the larger Tropical Channel model is able to organize and propagate deep convection at long lead-times-although the feature is displaced compared with GPM observations.
In agreement with the results seen in Figures 1d and 1e, the SE Asia LAM simulation fails to produce or propagate organized deep convection eastwards between 90°E and 104°E after 0000 UTC 5 January 2018. Differences between the SE Asia LAM and Tropical Channel simulation are also due to a lack of eastwards propagating 10.1029/2023GL104672 7 of 9 convection into the SE Asia LAM from the driving model. MetUM configurations similar to the driving model have previously been shown to underestimate convective activity and rainfall associated with the MJO (Holloway et al., 2013). The impact of the driving model can also be seen at the eastern boundary of the SE Asia LAM with erroneous convection (compared with the Tropical Channel and GPM) propagating westwards into the domain between 2nd and 6th January.

Conclusions
For the case study analyzed here, with a focus on precipitation over SE Asia, we find that the behavior and performance of a convection-permitting Tropical Channel model is similar to that of a SE Asia focussed LAM over a 10-day simulation period. Both models produce a realistic diurnal cycle, but both have more frequent small cells and less frequent large cells than GPM observations with precipitation overly fragmented in the MetUM. Both the Tropical Channel and SE Asia LAM produce more frequent moderate and intense precipitation than GPM observations and less frequent light rain.
Convective spin-up and a lack of organized propagating convection in the 90°E to 104°E region contribute to a dry bias in the SE Asia LAM that extends >500 grid points (>1,100 km) into the western portion of the domain. Between 145°E and the SE Asia LAM domain eastern boundary at 154°E (450 grid points) both MetUM simulations are wetter than GPM observations, largely due to excess precipitation in the west Pacific ITCZ. The magnitude of this wet bias is enhanced in the SE Asia LAM due to erroneous features propagating in from the eastern lateral boundary conditions (inherited from the parameterized convection driving model). This study provides an initial deterministic assessment of the sensitivity of tropical precipitation to domain extent. Use of ensembles would enable us to strengthen our conclusions on whether precipitation propagates more realistically in the Tropical Channel, however deploying ensembles for the Tropical Channel is prohibitively expensive and unfeasible on current compute resources. Both the Tropical Channel and SE Asia LAM show skill in capturing the regions impacted by tropical deep convection in this case study. Our results highlight, that away from the domain boundaries, the SE Asia LAM simulation is able to reproduce the same precipitation characteristics as seen in our Tropical Channel domain. Similar LAM experiments focusing on Africa and South America were carried out as part of this study (not shown), and findings are consistent with those shown here-suggesting the results are applicable across the tropics. For comparison, the Tropical Channel uses >6 times more computational resource than the SE Asia LAM. Persistent model biases, such as overly intense precipitation within the ITCZ, are seen in both simulations. Improvements to underpinning physical parameterization schemes therefore remain a priority to enable the benefits of convection-permitting global models (Tomassini et al., 2023), and will accelerate progress toward their practical operational implementation.