Recent advances in HPC systems have substantially improved the computational throughput of weather and climate models. We use the refactored ICON code described in that uses a GT4Py-based dynamical core developed within the EXCLAIM project. We performed global ICON experiments on the new ALPS HPC infrastructure of the Swiss National Computing Center at grid spacings of 80, 40, and 10 km from 2006–2016 including 1 year of spin-up. The 10 km simulation uses 32 Grace Hopper-200 GPU nodes, achieving a computational throughput of approximately 0.7 simulated years per wall clock day. The 40 km and 80 km run used 24 and 1 node, respectively, with throughputs of 5 and 4.9 simulated years per wall clock day. Initial conditions for all simulations are derived from European Center for Medium-Range Weather Forecasts (ECMWF) operational analysis on 1 November 2005, 00:00 UTC, with sea surface temperatures and sea ice data taken from Program for Climate Model Diagnosis and Intercomparison (PCMDI) AMIP II . This boundary condition data is available at approximately 1° horizontal grid spacing and is updated monthly, with the ICON model performing a linear temporal interpolation to daily resolution.

The simulations use 120 vertical levels with a terrain-following hybrid setup based on smooth level vertical (SLEVE) coordinate . The convection is parametrized using the Tiedtke–Bechtold bulk mass flux convection scheme, which treats shallow, mid, and deep convection with large-scale omega, boundary layer equilibrium, and convective available potential energy closure, respectively . Cloud microphysics are calculated using a single-moment scheme for cloud water, cloud ice, snow, and rain . The turbulent kinetic energy-based surface transfer and planetary boundary layer parameterization are used for turbulence representation with a second-order closure . Land surface processes are simulated by the soil–vegetation–atmosphere-transfer submodel TERRA , which utilizes eight model levels and is responsible for exchanges of heat, moisture, and momentum. The TERRA uses a tile approach to account for the effects of subgrid surface heterogeneities .

A numerical time step of 60, 180, and 360 s is used for the 10, 40, and 80 km runs, respectively. The radiation scheme is called every 6 min. In contrast, the convection, cloud cover, sub-grid-scale orography, and non-orographic gravity wave drag schemes were called every time step in all simulations. The fast physics such as saturation adjustment, surface transfer scheme, land-surface scheme, boundary-layer/turbulent vertical diffusion scheme, and microphysics scheme are called every model timestep. The Tegen aerosols climatology is used, while Global and regional Earth-system Monitoring using Satellite and in-situ data (GEMS) and Monitoring Atmospheric Composition and Climate (MACC) ozone climatologies are used . The external parameters such as topography, land-use land cover, etc., required by ICON were prepared by the External Parameters for Numerical Weather Prediction and Climate Application software . In Fig. S1 in the Supplement, we show the topography of SAsiaM, WAfriM, and NAmerM domains in various ICON simulations. The finer grid spacing resolves complex topographic structures such as the Himalayas, Western and Eastern Ghats of SAsiaM, Ethiopian Highlands of WAfriM, and the slopes of the Sierra Madre Occidental over NAmerM very well compared to its coarser grid counterparts (Fig. S2).

We validate the ICON experiments across the global monsoon domain and the representation of its regional characteristics with an ensemble of gridded precipitation datasets shown in Table  for a period of 10 years (2007–2016). The precipitation observations include products derived from satellites, satellite observations corrected with in situ measurements, station-based, and reanalysis products. In addition to the precipitation, we use the Fifth-generation ECMWF Reanalysis (ERA5) data to verify the moisture transport . Furthermore, we utilize the Indian Meteorological database to validate monsoon low-pressure systems over the SAsiaM (Indian sub-continent) region. A detailed description of these datasets, including their advantages and limitations are provided in Appendix A.

To identify the global monsoon domain in observations and our simulations, we follow a standard definition from . First, we calculate the difference between local summer and local winter precipitation and identify regions that exceed 2 mm d⁻¹, and thereafter filter the regions where the contribution of local summer rainfall exceeds 55 % of total annual rainfall. Here, local summer is defined as June–July–August–September (JJAS) for the Northern Hemisphere and December–January–February–March (DJFM) for the Southern Hemisphere. The first criterion, i.e., the annual range (local summer minus winter) distinguishes the monsoon climatic regions from the equatorial perennial rainfall, while the second criterion (percentage of summer contribution) distinguishes monsoon from subtropical arid and semi-arid regimes. We use standard metrics such as spatial correlation and root mean square error to validate the annual range of precipitation over the global monsoon domain with observations.

We use standard categorical verification metrics, such as the Probability of Detection (POD), False Alarm Ratio (FAR), bias score, and accuracy, to quantitatively assess ICON's skill in simulating regional monsoon domains (see Table ).

Monsoon onset is a key metric impacting sectoral planning across all monsoon regions. We identify monsoon onset as the first date, after 1 March in the northern hemisphere and 1 November in the southern hemisphere, on which the pentad-mean (5 d running mean) precipitation exceeds 5 mm d⁻¹ for at least three consecutive pentads. This criterion accounts for transient events and follows the definition of . Furthermore, tested various monsoon onset criteria across regional monsoon domains, and concluded that the methodology by agrees well with the observations.

The precipitation over various monsoon domains is largely fed by moist low-level jets advected from the oceans. Hence, we quantify the vertically integrated horizontal moisture transport in the lower troposphere from the surface to 850 hPa, 1Q=1g∫psfcp850hPaVqdp, where Q is the moisture transport vector (kg ms⁻¹), g is the gravitational acceleration (m s⁻²), V is the wind vector (m s⁻¹), and q is the water vapor mixing ratio (kg kg⁻¹). The integration is performed from the model surface pressure (psfc) to 850 hPa. Over elevated terrain where psfc < 850 hPa, the layer has zero thickness and Q is not computed. We therefore restrict our analysis of low-level jets to oceanic and low-terrain regions, following .

We delineated the global monsoon domain using the criteria of in our ICON grid spacing runs and in the IMERG reference dataset (Fig. a–d). The contours mark the monsoon boundaries, while the shading depicts the annual range (local summer minus winter), a key metric of monsoon intensity. All ICON simulations broadly capture the spatial pattern of the monsoon domains representing the well-known monsoon regions such as the NAmerM, SAmerM, WAfriM, SAsiaM, EAsiaM, AusMCM, SAfriM, including their adjacent oceans .

The IMERG data shows the most intense annual range in monsoon precipitation over the SAsiaM domain, with values exceeding 14 mm d⁻¹ over the head Bay of Bengal, the eastern Arabian Sea, the western Pacific, and the orographically dominated regions of the Himalayas and Western Ghats. All the ICON simulations broadly capture these features. A notable resolution dependence is evident over the Himalayan foothills, where only the highest-resolution (10 km) simulation reproduces the magnitude observed in IMERG, while the 40 and 80 km runs are unable to replicate the orographically induced precipitation pattern. Furthermore, while all simulations capture the annual range maximum over the continental core of the Indian monsoon region, they consistently show a slight overestimation in this area. All ICON simulations reproduce the core EAsiaM domain and its general annual precipitation range. However, they exhibit a systematic bias in capturing the full spatial extent of the monsoon domain along its northwestern margin, specifically the transitional zone bordering the Tibetan Plateau and continental China.

Over the WAfriM domain, IMERG shows a large annual range (> 12 mm d⁻¹) over the eastern tropical Atlantic, off the coast of Guinea. Across the continental interior, a north-south gradient characterizes the monsoon, with the maximum annual range is visible in the southern part of the domain. All the simulations overestimate both the spatial extent and the annual range over the tropical Atlantic. However, the continental spatial patterns, i.e., the north-south gradient, are broadly captured well in all simulations. For the NAmerM domain, IMERG shows a peak annual range (> 12 mm d⁻¹) over the equatorial eastern Pacific Ocean, followed by a secondary continental peak over northwestern Mexico and the southwestern United States, marking the core monsoon region. All ICON simulations consistently overestimate the annual range over the equatorial eastern Pacific Ocean. While the general features over the continent are captured, the spatial extent of the domain is underestimated, particularly near the interior western United States, and southeastern Mexico.

In the Southern Hemisphere region, IMERG shows peak annual range (> 10 mm d⁻¹) over the southern Pacific ocean, off the coast of northern Australia, followed by the southern Indian ocean, off the coast of South Africa. All simulations show a consistent overestimation of the annual range over these oceanic regions, with a larger magnitude in the finest-grid (10 km) simulation. Over the continental regions of SAmerM, SAfriM and AusMCM, all simulations align well with the IMERG reference data. All the qualitative results presented above remain robust when evaluated against an ensemble of observational datasets (MSWEP, CMORPH, ERA5, GPCC, and CPC; see Table  and Fig. S3).

We quantified model performance using the spatial correlation coefficient and root-mean-square error of annual precipitation range, comparing the ICON simulations with an ensemble of reference datasets to account for observational uncertainty. For the global monsoon domain, all ICON simulations show significant spatial correlations (> 0.7) with all observational datasets (except CPC, a land-only in situ dataset), indicating a robust representation of the spatial precipitation pattern. The 10 and 40 km runs consistently showed greater correlations than the 80 km simulation (except for the 10 km compared to IMERG).

Regarding the RMSE, the 10 km simulation produces the highest values (> 2.5 mm d⁻¹) across all observational datasets, exceeding those of the 40 and 80 km runs. Analysis of the error distribution (Fig. S4) reveals that the elevated RMSE largely stems from the greater annual range over oceanic regions, suggesting greater sensitivity of simulated oceanic precipitation to model grid spacing. Furthermore, to assess the uncertainty associated with the grid remapping methodology, we also performed the analysis on a common coarse grid of 80 km spacing. As shown in Fig. S5, the qualitative conclusions remain unchanged, confirming the robustness of our results.

We next assess the representation of regional monsoon characteristics by examining the monsoon domains and local summer precipitation for each major regional system (Fig. a–f). For the purposes of this article, we focus on tropical monsoon systems in the Northern Hemisphere, but we show a preliminary investigation of monsoon domain spatial extent from various ICON simulations for Southern Hemisphere monsoons. A detailed investigation of the characteristics of Southern Hemisphere monsoons is, however, beyond the scope of this article. The black contours represent the IMERG reference, and the red, green, and blue contours represent the 10, 40, and 80 km grid spacing runs regional monsoon boundaries, respectively, while the shading represents the observed IMERG local summer precipitation. Over the combined SAsiaM/EAsiaM region, all simulations broadly capture the IMERG monsoon domain. However, all simulations exhibit a systematic bias in capturing the full spatial extent of monsoon domain along key margins. Specifically, the transitional zone connecting the western Pacific coastal basins to the deep ocean near 120° E, and the inland boundary bordering the Tibetan Plateau and continental China. Quantitative verification scores for the monsoon domain, including the hit rate, false alarm ratio, bias, and accuracy, are shown in Figs. S6–S8 for various observational datasets. All runs show high-quality, consistent performance (e.g., accuracy ≈ 0.85), indicating reliable monsoon domain detection in all simulations. The spatial pattern of IMERG local summer monsoon precipitation largely follows that of the annual range shown in Fig. , confirming the dominance of summer precipitation in delineating the monsoon seasonal signal.

For the WAfriM domain, all simulations show an extended monsoon domain over the Atlantic Ocean, off the coast of Guinea, and a slight northward shift toward the Sahara. This misalignment results in a slightly higher FAR of > 20 % compared to other monsoon domains. However, all ICON runs show high scores (e.g., accuracy > 0.82), with a slight added value in 10 km simulation compared to 40 and 80 km. Over the NAmerM domain, almost all the ICON simulations are unable to capture the full inland of continental southern United States, specifically towards the northern and the eastern boundaries compared to IMERG. Furthermore, an overestimation of the equatorial Pacific Ocean is detected. Nevertheless, again the verification scores for the NAmerM remain high, with an accuracy of ≈ 0.85. The summer monsoon precipitation patterns for both NAmerM and WAfriM domains again broadly follow the spatial patterns of the annual range precipitation.

Overall, for detecting regional monsoon domain metrics, we conclude that all ICON runs show robust, high quality results independent of grid spacing. Furthermore, these findings are consistent across observational datasets, confirming their robustness.

Onset timing, progression and direction are some primary characteristics of monsoon systems, hence their reliable representation in the model is essential. Figure  shows the monsoon onset in IMERG compared to our ICON runs for the SAsiaM (a–d), WAfriM (e–h), and NAmerM (i–l) domains. IMERG shows a characteristic southeast-to-northwest propagation of the SAsiaM, with the progression beginning over the Bay of Bengal, the southeast Arabian Sea off the coast of Eastern Ghats in late April, and then progressing towards the continental northwest Indian continent by the end of June. An early onset signal in March over the foothills of the Himalayas and near Bangladesh is likely due to persistent orographic rainfall interacting with the low-level monsoon flow. Within this domain, the EAsiaM also follows a southeast to northwest progression from continental China towards the Tibetan Plateau. The IMERG results agree well with the general propagation features of the monsoon , confirming the reliability of our monsoon onset definition. All the ICON simulations capture the large-scale propagation features, but with notable discrepancies. For example, they exhibit a delayed and less homogeneous progression over the Bay of Bengal, and the 80 km run show a delayed onset over central India compared to fine-grid spacing runs. Furthermore, all runs are unable to capture the inland propagation over continental China, consistent with their unrealistic representation of the EAsiaM rainfall domain (Fig. a).

The WAfriM domain features a northward propagation primarily linked to the movement of the Intertropical Convergence Zone (ITCZ), and local land–atmosphere feedbacks. The progression begins in late March to early April near the Guinea region, advances into the Sudanian region by late April to May, and finally reaches the Sahel region by June–July. All simulations capture this large-scale feature. The IMERG onset progression within the NAmerM domain shows two distinct regimes: (1) an early April–May onset over southeastern Mexico and the adjacent western Caribbean, associated with the northward migration of the ITCZ and the bimodal rainfall pattern of Central America, and (2) the canonical late June–July onset of the core North American Monsoon. The core monsoon onset is characterized by a northwestward progression of organized convection from the eastern Pacific Ocean into the Sierra Madre Occidental of northwestern Mexico and the southwestern United States (e.g., ). All simulations capture the progression of the NAmerM from the eastern Pacific to the southwestern United States. However, the inland penetration of the NAmerM into the southern United States is absent (Fig. d).

While the large-scale patterns of the monsoon onset are well captured in ICON runs, they exhibit some systematic biases in timing. In Fig. , we show the onset biases (ICON minus IMERG), along with a statistical investigation of errors over the core monsoon domains using violin plots. Over the SAsiaM domain (Fig. a–d), two major large scale biases emerge in all runs: (1) a systematic late onset (≈ 10–20 d) over continental southeast Asia, Bay of Bengal, and continental central India suggesting a delay in the model's representation of the monsoon trough's northward progression and associated moisture convergence (2) an early onset (≈ 30–40 d) over the Himalayas, Tibetan Plateau and continental China, a common bias often linked to unrealistic snow representation, excessive springtime heating and spurious convection over complex topography . The improvement with finer grid spacing is not systematic across all regions. While the early onset over the Tibetan Plateau is partially reduced in the 10 km run, the delayed onset over southern India is smallest in the 40 km run. Over the core monsoon region, the median (mean) delayed onset for ICON 10, 40, and 80 km are 7 (7), 3 (5), and 12 (12) d, respectively. This indicates that the 10 and 40 km run marginally improve the onset bias compared to 80 km. The difference in the onset bias is very small between 10 and 40 km grid spacings highlighting a non-monotonic relationship with grid spacing.

Over the WAfriM (Fig. e–h) domain, all runs show a systematic early-onset bias (≈ 20–30 d) over the eastern Sahel and Sudanian region, that might be related to land-atmosphere feedback errors in these interior zones. This bias is systematically reduced with finer grid spacing. In contrast, over the western African coast and the adjacent eastern Atlantic, all simulations show substantially smaller biases compared to the eastern continental regions, suggesting model performance is more constrained by observed oceanic boundary conditions. Over the core monsoon region, the 10, 40, and 80 km runs show a median (mean) early onset bias of 6 (4), 8 (7), and 10 (11) d respectively, representing a systematic improvement of ≈ 60 % for the 10 km run compared to the 80 km simulation. Furthermore, the bimodal (double-peak) bias distribution evident in the 80 km simulation (Fig. h), indicating two distinct bias patterns, is consolidated into a single, sharper peak in the 10 km run. This showcases a marginal improvement in the representation of monsoon onset characteristics with fine grid spacing over WAfriM.

The bias patterns over the NAmerM domain (Fig. i–l) are less coherent, underscoring the challenge of representing the intricate dynamic and thermodynamic processes that govern it. Despite this complexity, a clear non-monotonic grid spacing dependence emerges in the core monsoon region. The 40 km run shows the best performance with a minimal median (mean) early bias of -2 (-2) d. On the contrary, both the 10 and 80 km runs exhibit a delayed onset, with biases of 3 (3) and 2 (4) d, respectively. This result suggests that while coarse resolution (80 km) is insufficient to capture key onset dynamics, the highest resolution (10 km) may potentially introduce its own biases that require further investigation. Overall, the mean biases for NAmerM are notably smaller in magnitude than those for the SAsiaM and WAfriM domains.

In summary, results reveal that ICON reliably captures the large-scale onset characteristics but exhibits region-dependent biases that are modulated by model grid spacing in non-systematic ways. For the WAfriM, finer grid spacing yields marginal improvement. However, for the SAsiaM and NAmerM, the relationship is non-monotonic. The overall magnitude of onset timing biases is smallest for the NAmerM, followed by the WAfriM. These results remain qualitatively consistent when validated against CMORPH, ERA5 and MSWEP (shown in Fig. S9). All ICON simulations use monthly SST data linearly interpolated to daily resolution. Sensitivity experiments with a daily high-resolution SST product reveal that monsoon onset over SAsiaM and WAfriM is sensitive to SST bias and update frequency. Nevertheless, since all grid spacing experiments (10–80 km) use identical SST boundary conditions, our conclusions about grid spacing remain robust.

Accurate representation of the spatial patterns and magnitude of monsoon precipitation has been a long-standing challenge for weather/climate models. Here, we investigate the precipitation features of the SAsiaM, WAfriM, and NAmerM domains and their associated biases.

Figure a–d show monsoon precipitation over the SAsiaM during local summer. IMERG shows peak precipitation (> 15 mm d⁻¹) over the head Bay of Bengal, the Himalayan foothills and Western Ghats, and over the south China/west Philippines Sea. Over the agriculturally critical central Indian core region, precipitation reaches ≈ 11 mm d⁻¹. A characteristic coastal maximum with subdued inland rainfall is also evident in the EAsiaM sub-region. All simulations reproduce the observed spatial patterns of these precipitation peaks. However, they exhibit region specific differences indicating resolution sensitivity. Most notably, only the 10 km simulation reproduces the observed IMERG precipitation peak over the Himalayan foothills, underscoring the need of a fine scale orography for realistic orography rainfall. Notably, all the runs capture the precipitation peak over the SAsiaM core region, indicating a reliable process representation in ICON such as the inland moisture transport, monsoon LPS, active/break oscillations and local land-atmosphere coupling. Across the EAsiaM region, the 10 km run also better captures the coastal to inland precipitation gradient, indicating added value in capturing local atmosphere–land–sea interactions.

Over the WAfriM domain, IMERG reveals distinct precipitation features characterized by sharp meridional gradients. An intense precipitation maxima along the Guinea coast and adjacent eastern Atlantic (> 15 mm d⁻¹) followed by a secondary continental maximum over the Sudanian region (∼ 9–12° N), creating a characteristic meridional profile with a relative minimum in the intervening Sahelian zone. A localized precipitation peak over the western slopes of the Ethiopian Highlands (40° E) is also evident. All runs capture the primary precipitation coastal maxima and the orographic induced precipitation over the Ethiopian Highlands. However, the simulations fail to capture the sharp meridional gradient observed in the IMERG over the Sudanian and Sahel regions due to an overestimation of precipitation in these regions. Furthermore, the orographic precipitation over the Ethiopian Highlands is also overestimated in all the simulations.

Over the NAmerM domain, IMERG reveals a defining land-ocean precipitation structure. The primary feature is the continental precipitation peak (> 6 mm d⁻¹) on the windward slopes of Sierra Madre Occidental, with a sharp rain shadow on its leeward side. This orographic peak is fed by moisture transport from the eastern Pacific Ocean, southwest of Mexico. All simulations capture the continental maximum over the Sierra Madre Occidental, with the 10 km run showing a sharper and more realistic orographic gradient. However, a major systematic overestimation of precipitation (> 16 mm d⁻¹) over the eastern Pacific source region is present in all simulations.

All simulations reproduce observed precipitation patterns well, but region-specific differences remain. We quantify these biases in Fig.  by showing the difference between the simulations and IMERG, followed by an investigating into the monsoon core regions using violin plot. Over the SAsiaM domain (Fig. a–d), all runs exhibit a coherent dipole bias pattern, a systematic dry bias (> 2 mm d⁻¹) over the southern Indian peninsula and adjacent equatorial Indian Ocean coupled with a wet bias (≈ 2 mm d⁻¹) over central and northern India. Additional biases include a dry bias over the foothills of the Himalayas, and a wet bias over Tibetan Plateau. Fine grid spacing (10 km) reduces the dry bias over the Himalaya and wet bias over the Tibetan Plateau, but it amplifies the wet-dry dipole bias over the Indian sub-continent. Over the monsoon core region, a median (mean) bias over central India in 10, 40, 80 km of 0.9 (1.2), 1.3 (1.0), 0.1 (0.1) mm d⁻¹ exists respectively. A bimodal distribution of bias in 10 km run further shows that finer grid spacing enhances the dipole bias pattern over India. This points out that finer grid spacing, does not necessarily improve biases, rather it might further activate regional processes over SAsiaM domain, such as orography moisture blocking, local land-atmosphere interactions or intensification of synoptic systems .

Over the WAfriM domain (Fig. e–h), all runs show a wet bias (≈ 1.5 mm d⁻¹) in the continental central Sahel, Sudanian regions and coastal Atlantic, which intensifies with finer grid spacing. Over the monsoon core region, the 10, 40, and 80 km simulations show a median (mean) bias of 1.7 (1.7), 1.2 (1.6), and 1.1 (1.3) mm d⁻¹, respectively, further indicating intensification of biases. These results are also consistent with the findings in SAsiaM core region. Over the NAmerM domain, systematic slight dry bias (≈ 0.5 mm d⁻¹) is revealed in all simulations over the Sierra Madre Occidental region, and a large wet bias over the eastern Pacific ocean (> 8 mm d⁻¹). This large oceanic bias shows a deficiency in ICON in representing tropical eastern Pacific convection, air-sea interactions, or the subtropical high-pressure system that normally suppresses rainfall in this area in summers . Finer grid spacing (10 and 40 km) shows a slight improvement in the magnitude of dry bias (0.8 mm d⁻¹) over the monsoon core region compared to 80 km (1.2 mm d⁻¹) run.

Overall, ICON reliably simulates the large-scale patterns of SAsiaM, NAfriM, and NAmerM monsoon precipitation, but it exhibits systematic regional biases. Importantly, finer grid spacing improves some features (e.g., orographic gradients) but amplifies others, such as continental wet-dry dipoles and excessive oceanic rainfall. This demonstrates that increased resolution alone does not necessarily improve performance; it can instead more vigorously activate both realistic and biased physical processes. Furthermore, our findings remain qualitatively consistent when validated against CMORPH, ERA5 and MSWEP (shown in Fig. S10)

Here, we assess the ICON performance across various timescales to investigate the representation and sensitivity of monsoon variability to grid spacing. Figure  shows the spatial patterns of variance ratio (ICON/IMERG) associated with 30–90 d bandpass filtered daily precipitation data over the SAsiaM (a–d), WAfriM (e–h) and NAmerM (i–l) domains.

Over the SAsiaM domain, the 10 km simulation shows an overestimation of intraseasonal variability, with a peak variance ratio > 1.5 over the central Indian monsoon core region, Bangladesh and eastern Tibetan Plateau. The 40 km simulation also overestimates the variance over the monsoon core region, but the peak pattern is confined more towards the east coast of India. In contrast, the ICON 80 km variance ratio is ≈ 1 over most of the domain. Over the monsoon core region, the median (mean) variance ratio is 1.9 (1.8) for 10 km, 1.4 (1.1) for 40 km, and 0.8 (0.7) for 80 km simulation. This systematic resolution dependence, where the fine grid spacings produce greater intraseasonal variance, explains the earlier finding of enhanced extreme-precipitation frequency. The 10 km simulation exhibits approximately 58 % greater variance than the 80 km run. Since the 30–90 d variability in this domain is largely governed by the boreal summer intraseasonal oscillation and tropical waves, these results suggest that the simulation of these large-scale convective phenomena is highly sensitive to model grid spacing.

In the WAfriM domain, all the simulations systematically overestimate the variance, with most of the overestimation occurring in the Sahel region. The spatial patterns of this variance overestimation exhibit a clear north-south gradient, closely mirroring the mean precipitation bias shown in Fig. . The magnitude and the spatial extent of the overestimation increase with fine grid spacing. The 10 km simulation shows the highest variance > 2, and also covers a larger swath over the Sahel compared to the coarse grid spacing simulations. Quantitatively, over the monsoon core region, the median (mean) variance ratio is 3.2 (2.4), 2.1 (1.4), and 1.7 (1) for ICON 10, 40, and 80 km simulation, respectively. This represents an 88 % increase in mean core variance from 80 km run to 10 km, the largest resolution-induced amplification among all three monsoon domains. Over the continental NAmerM domain, the spatial patterns of variance ratio in ICON simulations look almost identical, but with a slightly enhanced variance over the Sierra Madre Occidental in 10 km simulation compared to coarser grid spacing runs. Over the monsoon core region, the median (mean) variance ratio is 1 (0.9) for 10 km, 0.9 (0.7) for 40 km, and 0.8 (0.8) for 80 km simulation. This indicates a slight underestimation of 30–90 d variability across all simulations, with minimal sensitivity to model grid spacing. These results qualitatively agree when simulations are compared to MSWEP, CMORPH and ERA5 observations (Figs. S13–S15)

Next, we investigated the 2–30 d high frequency variability (Figs. S17–S20) across the domains. The results mirror those of the 30–90 d band, i.e., the finer grid spacing systematically amplifies variance, with ICON 10 km showing the highest variance over the SAsiaM and WAfriM domains. The percentage increase over the SAsiaM core monsoon region between the ICON 10 and 80 km is ≈ 54 % and 42 % over WAfriM. In these regions, the high-frequency 2–30 d oscillations are generally attributed to synoptic systems, wave activity, which contributes to active/break monsoon periods. Often, these systems are attributed to the extreme precipitation over these regions, causing massive flooding. Key examples include low pressure systems (LPS) over of the Bay of Bengal that propagate northwestward into the Indian core region, and African Easterly Wave activity that triggers organized mesoscale convective systems over the Sahel. We investigate these processes explicitly in the following section.

In contrast, over the NAmerM domain, variance is suppressed (ratio < 1) in all the simulations, suggesting that the model does not fully resolve the key drivers of high-frequency variability in this region, such as Gulf of California moisture surges. Analysis of interannual variability (Fig. S22) shows similar resolution-dependent patterns, but we do not emphasize these results, as a simulation period longer than our 10-year sample is required for robust statistics at these longer timescales. Overall, the analysis of 30–90 and 2–30 d bands demonstrates that finer grid spacing systematically amplifies precipitation variance across the SAsiaM and WAfriM monsoons, with the strongest effect occurring at the finest resolution (ICON 10 km).

To determine if intense precipitation in fine grid spacing simulations disproportionately influence the variance ratios, we applied a log⁡(1+P) transformation to all precipitation (P) datasets prior to band-pass filtering. This transformation reduces the statistical leverage of intense rainfall, while preserving the timing of precipitation events. After transformation, the systematic increase of variance persists across both the 30–90 and 2–30 d bands over SAsiaM and WAfriM domains (Figs. S16 and S21). Crucially, the transformed variance ratios are below unity for 2–30 d, revealing that all simulations underestimate organized precipitation variability when intense precipitation is statistically controlled, while the ratio of 30–90 d in fine grid spacing remains near unity. This indicates that the variance ratios exceeding 1 in raw data primarily reflect the fine-grid simulations over production of intense precipitation (> 10 mm d⁻¹). These findings are robust when seasonal cycle is removed in both raw and transformed datasets. The persistence of grid spacing dependence on variance ratios after transformation confirms they represent real model sensitivities rather than statistical artifacts from intense precipitation.

Finally in Fig. , we examine the spatial patterns of the local hour of maximum precipitation (peak diurnal timing) in local solar time (LST) compared against IMERG during the local summer monsoon season.

Over the SAsiaM domain Fig. a–d, IMERG shows a wide spread afternoon peak (≈ 15:00–18:00 LST) over continental India, and the Tibetan Plateau, driven by strong daytime land-surface heating, and a nocturnal peak over the adjacent oceans, likely associated with nighttime cloud-top cooling and local air–sea interactions . Notable exceptions include a nocturnal peak over the Himalayas and its adjacent foothills associated with the interaction of katabatic winds with the monsoonal flow, and a patchy nocturnal to afternoon transitions over the open oceans. All simulations capture the diurnal signal over the continental afternoon/oceanic nocturnal signal. However, ICON 10 km marginally better agrees in the timing of the peak signal (≈ 15:00–18:00 LST), while 40 and 80 km simulations exhibit slight delay, with a peak occurring mostly at ≈ 18:00 LST. All simulations are unable to reproduce the nocturnal peak over the Himalayas, instead they unrealistically produce the peak towards the south of the Himalayan foothills. Over the oceans, the simulated diurnal phase patterns align well with observations across all grid spacings.

Over the continental WAfriM domain Fig. e–h, IMERG shows a wide spread late afternoon maxima (≈ 16:00–18:00 LST) over the interior, primarily driven by land surface heating, and a coastal zone with nocturnal maxima along the Guinea coast, eastern Atlantic and coastal Sahel indicative of land-sea breeze convergence. Inland nocturnal patches are observed, particularly over the central Sahel, likely associated with the propagation of mesoscale convective systems initiated during the afternoon. All the simulations capture the broad inland afternoon peak. The 10 km simulation marginally better reproduces the timing, while the coarse grid spacing show a slight delay. The coastal early morning peak near the Sahelian coast is also well captured in all the simulations, indicating that the local land-sea breeze is reasonably well simulated.

Over the NAmerM domain Fig. i–l, IMERG reveals the expected afternoon maximum (≈ 15:00–18:00 LST) over the orographically complex regions of western North America (Sierra Madre Occidental, Rocky Mountains) and Central America, driven by slope heating, as well as over the eastern US and coastal regions, driven by land-surface heating. A distinct early morning peak (≈ 03:00–07:00 LST) over the Great plains resulting from the nocturnal eastward propagation of mesoscale convective systems initiated over the mountains. A nocturnal peak dominates open oceans, while patchy afternoon peaks are observed near coast lines. All simulation capture the broad pattern of afternoon peak over both, the eastern and western North American regions. The 10 km simulation shows the correct peak time, while the coarser grid spacing runs exhibit a slight delay, consistent with the behavior seen over SAsiaM and WAfriM continents. The nocturnal maximum over the central plains of America is also simulated well in all simulations. Overall, the diurnal phase is surprisingly well simulated in all the simulations, with 10 km run providing only marginal improvement in timing over land.

These results demonstrate that while finer grid spacing improves some aspects of monsoon simulation (diurnal timing, variance structure) depending upon the region, it exacerbates others (extreme overproduction, mean biases), revealing complex trade-offs in fine grid spacing monsoon modeling.

The previous section analysis showed that ICON's fine grid spacing systematically amplified the variance in the 2–30 d band over the SAsiaM and WAfriM domains. Here we investigate the characteristics of the processes responsible for this variability, namely the rain carrying synoptic low pressure systems over SAsiaM domain and African Easterly wave activity. While the 30–90 d mode shows similar behavioral trends, its dynamics involve more complex large scale ocean–atmosphere interactions that extend beyond the scope of this regional process evaluation.