With recent advances in exascale high-performance computing (HPC) architectures, the computational throughput of various weather and climate models has substantially improved over the last decade. Especially energy-efficient heterogeneous architectures, leveraging GPU accelerators, have been developed and can make climate model integrations very efficient . However, substantial investments in enabling km-scale weather and climate model codes are needed to port them and run them efficiently on new HPC systems . Rapidly emerging technologies in hybrid architectures require agile, architecture-agnostic climate model codes. The EXCLAIM (EXtreme scale Computing and data platform for cLoud-resolving weAther and clImate Modeling) project at ETH Zürich addresses this issue by refactoring the ICON code with Domain Specific Language (DSL) with a Python front-end using GridTools for Python, GT4Py, while enabling backend flexibility across architectures .

We use the version of ICON Model developed within the EXCLAIM project , which features a complete refactoring of the numerical core into GT4Py (icon-exclaim v0.2.0). Two global experiments were conducted on the Swiss National Computing Center's new HPC infrastructure, ALPS. A ten-year (2006–2016) spin-up simulation was first performed at a 10 km horizontal grid spacing. A 2.5 km grid spacing simulation following the DYAMOND protocol is conducted for a span of 4 years (20 January 2020 to 1 April 2024), with its soil state taken from the 10 km course grid simulation integrated for 10 years . The 2.5 km simulation uses 240 GH200 GPU nodes (960 GPUs in total), achieving a throughput of 0.25 simulated years per day. This setting was chosen for computational efficiency, and throughput can be increased by increasing the number of nodes. More information on the computational aspects of the specific model used for this simulation can be found in .

The 2.5 km simulation uses lower boundary conditions from daily updated sea surface temperatures and the sea-ice fraction from the ESA Climate Change Initiative (ESA-CCI) product at a horizontal grid spacing of 1/20° . Furthermore, the setup includes 120 vertical levels extending up to 85 km with a terrain-following hybrid setup and the smooth level vertical (SLEVE) coordinate . The single-moment cloud bulk microphysics scheme is used to simulate cloud water, ice, snow, graupel, and rain. The second-order turbulent kinetic energy-based surface transfer and planetary boundary layer parameterization are used to represent turbulence . Land processes are simulated using the TERRA Multi Layer (TERRA-ML) model , which employs eight soil layers and governs the exchange of heat, moisture, and momentum. The TERRA-ML model employs a tile approach to accurately estimate cell-averaged surface fluxes and account for deviations in sub-grid-scale surface characteristics. The shallow and deep convection schemes are turned off in the 2.5 km simulation, allowing an explicit treatment of convection. Additionally, subgrid-scale orographic drag and non-orographic gravity wave drag schemes are disabled. We used a numerical time step of 20 s, updated cloud cover and the microphysics schemes every time step, and applied the radiation scheme every 6 min. The external parameters required by ICON, namely the topographic height of the Earth's surface, surface roughness, vegetation cover, and land/sea/lake distribution, were prepared using the External Parameters for Numerical Weather Prediction and Climate Application software tool . Finally, we use the global Max-Planck-Institute Aerosol Climatology version 2 (MAC-v2) dataset with 1° grid spacing at a monthly temporal frequency .

The 2.5 km model configuration is close to the ICON-CLM (Climate Limited-area Modelling) community setup and benefits from the vast experience of the community's participation in CORDEX (COordinated Regional Climate Downscaling EXperiment) activities across multiple CORDEX regional domains. We did not perform any additional model tuning for this simulation. We saved output variables following the DYAMOND III protocol .

We use a range of observational datasets alongside reanalysis to assess our simulation's performance.

The global average pattern of T2M is well simulated with a Spearman pattern correlation coefficient of 0.998 and a root mean squared error of 1.48 °C for land grid points (Fig. ). As expected, ocean regions exhibit particularly small differences due to the pre-described sea surface temperature. Striking are cold differences in ICON during polar winters that might be related to the representation of sea ice and snow in addition to differences stemming from uncertainties in ERA5 T2M in these regions (Fig. d, f). Larger differences also occur during boreal winters over land areas with up to 3 °C negative differences in eastern Canada and large positive differences in Siberia and Alaska.

During boreal summers, temperatures exceeding 3 °C warm differences emerge in continental regions such as the central U.S. and Eurasia (Fig. f). Part of these differences is likely related to a misrepresentation of land-atmosphere interactions, such as the lack of lateral groundwater flow and the corresponding reduced evapotranspiration in land-atmosphere coupling hot-spots such as the central U.S., which can cool down these regions by several degrees Celsius during the summer season . Additionally, misrepresented cloud radiative forcings frequently (Fig. S1 in the Supplement) contribute to surface temperature biases in models with convection parameterizations e.g., and km-scale models e.g.,.

It is important to note that the differences shown here and in the following sections are partly related to the short simulation period of four years. Climate internal variability can cause large differences at local scales even though observed sea surface temperatures are used.

Figure  shows the daily mean surface incoming shortwave radiation (SW_in) for (a) ICON, (b) the EWEMBI observational dataset, and (c) their difference. Overall, SW_in is underestimated over oceans, especially along the Intertropical Convergence Zone, and overestimated over land, particularly in mountainous regions. This overestimation over land is also evident when comparing the simulation to flux tower observations (Fig. d, e). The median difference in land mean SW_in is approximately 6.8 W m⁻², with a median difference of around 28.2 W m⁻². The overall median difference is consistent with the mean bias error of approximately 7.8 W m⁻² reported for CMIP6 models by . Previous studies have shown that shallow clouds are frequently a main contributor to continental surface energy biases in km-scale models e.g.,.

Figure d and e show the spatial distribution of station median total and noon differences in longwave downward radiation (LW_in), latent heat flux (LE), and sensible heat flux (H). LW_in is slightly underestimated. LE shows a median difference close to zero across stations, with regionally over- and underestimations of more than 10 W m⁻² that cancel each other on average. This spatial compensation suggests region-dependent limitations in the representation of land-surface controls on evaporation, rather than a uniform model bias. In contrast, H is more consistently overestimated, particularly during noon, agreeing with the overestimation in SW_in.

The spatial pattern of station-based mean noon differences in LE and H during JJA shows a pronounced underestimation of LE and an overestimation of H at most stations in ICON (Fig. f, h) in agreement with the T2M differences shown in Fig. . We suspect that parts of the systematic dry-summer biases are due to the absence of a lateral subsurface water-flux scheme in TERRA. Previous studies have shown that the lack of such lateral groundwater redistribution can lead to substantial surface flux and temperature biases at km-scale grid spacing . In contrast, there is more station-to-station variability for winter biases. The spatial patterns and characteristics of surface flux differences are consistent with biases in temperature and precipitation.

Annual and seasonal differences between ICON and GPM-IMERG are shown in Fig. . The Pearson pattern correlation of annual average precipitation is 0.91 (0.87 over land cells) and the corresponding relative RMSE is 41 %. Most noticeable are wet differences over most tropical ocean regions except for the west Maritime Continent (Fig. c). This is consistent with the negative SW_in differences shown in Fig. c. Land precipitation is relatively well simulated overall, although dry differences emerge in several continental regions during boreal summer (Fig. g). In some of these regions, including parts of the central U.S. and Eurasia, the dry differences coincide with warm T2M differences (Fig. f), higher SW_in, lower LE, and higher H, suggesting a contribution from land-atmosphere coupling and cloud radiation biases.

Zonal average precipitation agrees very well between GPM-IMERG and ICON, partly because of error cancellation effects (Fig. d). The differences are typically smaller than the inter-annual variability (shading in Fig. d) except for high-latitudes where IMERG is known to underestimate precipitation accumulations due to issues in remote sensing solid precipitation complicating a reliable evaluation. The location of the wet regions in the tropics (defined as the 80th percentile of tropical average precipitation; Fig. h) is remarkably well captured with a single tropical rainband.

This is a big achievement since simulating a realistic rainband is challenging, even at a km grid spacing, as found a double ITCZ in the Indo-Pacific using km-scale ICON simulations with model physics different from ours. They found a lack a surface fluxes (caused by too little surface winds) in this region, which they resolved by increasing the minimum surface wind speed in the lowest model level from 1 to 4 m s⁻¹ . Our model settings, which are similar to ICON's NWP setup, do not feature this issue. While it is unclear exactly what causes the differences in model performance among the ICON setups, a key difference lies in the turbulence parameterizations and surface-layer schemes, while found large sensitivities to the formulation of microphysics. The location of mid-latitude precipitation bands associated with the storm tracks is reasonably well simulated (Fig. a, b).

Comparing peak hourly precipitation over the four-year evaluation period between GPM-IMERG and ICON shows that our simulation yields substantially larger values, particularly in the tropics and subtropics (Fig. a, b). This is predominantly caused by too low heavy hourly precipitation in GPM-IMERG since comparisons with hourly precipitation gauge observations from the HadISD dataset show that ICON reliably reproduces observed intense precipitation rates across various climate regions. At the same time, GPM-IMERG fails to capture intense rates, particularly at lower latitudes (Fig. c–f), which is in agreement with previous research . However, a proper evaluation of hourly precipitation using station data is not possible in many parts of the globe due to insufficient data coverage and a lack of data sharing e.g.,. Our presented analysis is primarily based on data from the U.S. and Europe (see Fig. S2). Results for additional climate regions, often based on only a few stations, are shown in Fig. S3.

Using the same station network, we assess the June–July–August (JJA) season-average precipitation diurnal cycle (Fig. ; we use 0.1 mm h⁻¹ as the threshold for precipitation intensity and frequency statistics). The performance of ICON is regionally dependent and shows good agreement for capturing precipitation amount, frequency, and intensity in the North Central America (NCA) region that is heavily influenced by the North American Monsoon (Fig. g–i). In Eastern North America (ENA), precipitation frequency is well captured compared to station data. Still, precipitation intensity is overestimated, resulting in an overly amplified evening and nocturnal peak in precipitation amplitudes (Fig. a–c). Western Central Europe (WCE) shows a too-early peak in precipitation that decays too quickly. The precipitation amount is also low-biased due to the too low precipitation frequency. Generally, ICON shows greater fidelity in capturing the diurnal cycle of convective precipitation in southern regions (ENA and NCA) than GPM-IMERG.

From Fig. , we see that the peak timing of convective precipitation amount is relatively well captured by GPM-IMERG, consistent with previous findings e.g., although IMERG has been reported to have a phase lag relative to Ku-band radar in many regions which should be kept in mind . This allows us to evaluate ICON on a near-global scale (Fig. ). We perform this analysis during seasons dominated by convective precipitation, namely JJA at > 30° N, December–January–February (DJF) at < 30° S, and annually in the tropics. GPM-IMERG and ICON feature afternoon precipitation peaks over land. In contrast, tropical ocean regions feature predominantly nighttime and early-morning peaks. Some land regions also feature nocturnal precipitation peaks, particularly in the lee of mountain ranges (e.g., the central U.S., the La Plata basin, eastern China, and the Sahel and Western Africa), which are related to nocturnal mesoscale convective system activity. ICON can capture these nocturnal peaks reasonably well, but the spatial extent of nocturnal regions is often too small, particularly in the Amazon Basin and the central U.S. The relative amplitude of the convective precipitation diurnal cycle (diurnal amplitude divided by mean precipitation) is also relatively similar, with a tendency for larger values in ICON compared to GPM-IMERG.

In summary, we recommend using GPM-IMERGv7 hourly precipitation with caution in model evaluation studies, given its tendency to overestimate precipitation frequency and underestimate precipitation intensity. Where possible, in-situ hourly rain gauge observations should be used.

Tropical cyclones cause the most intense UV10 extremes during the evaluation period (Fig. a). Storm tracks of extratropical cyclones are also clearly visible in both hemispheres. Over land, mountain regions and plateaus such as the Tibetan Plateau and the Andes stand out, as do mid-latitude plains that feature straight-line windstorms (e.g., the central U.S., the La Plata Basin, and the Sahel). Also, the coastal regions of Greenland and Antarctica feature peak wind speeds exceeding 40 m s⁻¹.

Evaluations against UV10 HadISD station observations show overall good agreement with a tendency of heavy wind speeds (>20 m s⁻¹) to be underestimated in ICON (Fig. b). However, regional differences are observed, with lower differences at high wind speeds, mostly in Mid-Latitude regions such as CNA and WCE (Fig. c, d). The region with the most pronounced underestimation of UV10 extremes is Eastern North America (Fig. S5), which might be partly related to a low difference in tropical cyclone frequency as we will see in Sect. , but regional differences in the simulation of extra-tropical cyclones could also contribute. Intense winds are well simulated in most tropical and high latitude areas such as East Asia (EAS) and Northern Australia (NAU) (Fig. e, f). The station density that is available for comparison in HadISD is shown in Fig. S4, and statistics for additional IPCC regions are shown in Fig. S6. Also, ICON reproduces the temporal autocorrelation of hourly winds well (Fig. S5), indicating that it captures the persistence and day-to-day temporal variability of near-surface wind fluctuations, including both the rapid decorrelation at short lags and the diurnal recurrence seen in the observations.

The wind speed evaluation should be interpreted with caution, as we compare station data with 2.5 km grid-cell averages. Additionally, short-term average wind speeds (e.g., 2 min averages at U.S. stations, 10 min averages at many WMO stations) are compared with hourly instantaneous model winds. Instantaneous model snapshots tend to produce broader, higher-tailed wind-speed PDFs than temporally averaged station observations, while spatial averaging over 2.5 km grid cells damps small-scale wind maxima. The combined effect partially compensates for differences. Km-scale modes are generally more comparable to station observations because the scale differences between the simulated processes on the grid and point-based observations are smaller than in coarser-resolution models.

Tropical cyclone tracks compare well in frequency and location between ICON and IBTrACS, except for the North Atlantic Basin, which features a low bias (Fig. a, b). Note that the longer tracks in IBTrACS are caused by identifying TCs already as tropical depressions and following them after the extra-tropical transition. In contrast, MOAAP only identifies the main TC phase. The inter-annual variability of basin total tropical cyclone frequencies is also well captured in ICON, except for the Southwest Pacific and South Indian Ocean, where the variability is smaller (Fig. b).

The TC pressure wind relationship is relatively well captured, but ICON wind speeds are too low (Fig. c), consistent with the UV10 analysis in the previous section and simulations with the NICAM model at km-scale . The low difference is also reflected in the probability histogram of tropical cyclone lifetime peak wind speeds, which drops off too rapidly at high wind speeds (Fig. d). The IBTrACS distribution shows a bi-modal distribution with a minimum around 40 m s⁻¹ and a primary peak at low wind speeds. The reason for this primary peak is that the tropical depression and tropical storm phases of TC are included in IBTrACS. In contrast, the MOAAP tracker is calibrated to identify only storms of tropical storm intensity. Similar to the comparison to station-based wind observations in the previous chapter, comparing modeled instantaneous wind fields to IBTrACS observations is challenging. The simulation provides gridded 10 m near-surface winds, while IBTrACS reports storm-level maximum sustained winds from observational best-track analyses. They are therefore not directly equivalent, since the model resolves spatial wind structure and average wind speeds within grid cells, whereas IBTrACS summarizes cyclone intensity. This could explain parts of the low difference in wind speed seen in ICON.

Figure  shows Hoffmöller diagrams of tropical average Tb in ICON and GPM-MERGIR, with labels indicating equatorial waves from MOAAP output for May 2020. While Rossby waves are well simulated in ICON, other waves that are strongly coupled to deep convection are underestimated in both frequency and amplitude. Kelvin and inertia-gravity waves (IGWs) are very infrequent, and Mixed Rossby-Gravity Waves (MRGs) are only regularly observed over Africa.

Global kilometer-scale (storm-resolving) models substantially improve the simulation of convectively coupled equatorial waves by explicitly representing deep convection, leading to more realistic precipitation spectra, wave propagation, and dynamical structures compared to models with parameterized convection . However, despite these advances, biases remain in wave–convection coupling strength, phase speed, and intermittency, indicating that the representation of equatorial waves at km-scale resolution is improved but still model- and configuration-dependent rather than fully resolved .

Wheeler–Kiladis space–time spectra computed from the four-year evaluation period confirm these results (Fig. ). We follow the approach of and , and apply it to outgoing longwave radiation (OLR) from the ICON simulation and the NOAA-interpolated OLR dataset. Overall, ICON reproduces the observed spectral characteristics of equatorial waves reasonably well. Rossby wave activity shows the closest agreement with observations, while other modes – such as Kelvin waves, the Madden–Julian Oscillation (MJO), and IGW – are captured but exhibit lower amplitudes than in the NOAA OLR data.

Equatorial waves often organize convection into clusters of MCSs. The latent heating from these MCSs can, in turn, amplify and sustain the wave, creating a feedback loop between the large-scale disturbance and convection . The general pattern of MCS initiation frequency is well captured in ICON, but initiation is underestimated over tropical oceans and overestimated over tropical land regions (Fig. a–c). The relative underestimation of oceanic MCS frequency is largest in the northern hemisphere, followed by the Southern Hemisphere, and tropics (Fig. g–i). In the tropics, oceanic MCS frequencies are well simulated between December and March but have a systematic low bias in the other months of the year (Fig. h).

Simulated land-based MCSs frequencies are most similar to observations in the northern hemisphere. Frequencies are overestimated in the Southern Hemisphere during summer and in the tropics year-round, except in June and July.

These frequency biases result in a systematic underestimation of the MCS to total precipitation ratio over almost all ocean regions, while this ratio is overestimated over Africa (Fig. d–f). The reasons for the MCS frequency biases are unclear, but they might be related to biases in tropical waves and the thermodynamic-convection coupling , which will be the topic of future research.

In addition to frequency statistics, we also performed lifetime analysis of oceanic and land-based MCSs in the tropics, northern, and Southern Hemispheres (Figs.  and S7 and S8 respectively). We discuss only the results from the tropics here, since they are similar in both hemispheres. Observed MCS sizes are similar over tropical oceans and land (Fig. a). Simulated MCSs are generally smaller than observed systems. The simulated MCS movement speed is systematically higher for most quintiles (Fig. b), while the mean precipitation under the MCS cloud shields is lower (Fig. c), particularly for land-based MCSs. Finally, modeled MCS durations are too short (2–4 h median difference), with larger differences over land than over ocean (Fig. d).

To compare the temporal evolution of MCSs, we differentiate between short-lived (4–9 h) and long-lived (>14 h) tropical MCSs, given lifetime-dependent model biases. The growth and decay of MCS cloud shields are well simulated for oceanic and land-based tropical MCSs (Fig. e, f). However, long-lived storms do not grow fast enough and reach a peak size that is about 20 %–30 % too small. On the other side, long-lived storms feature more accurate movement speeds while small storms move about 20 % too fast (Fig. g, h).

Mean MCS precipitation starts high during the initiation phase of MCSs and steadily decreases during their lifetime (Fig. i, j). This is because most MCSs start from a burst of small-scale convection and develop stratiform precipitation areas as they grow upscale until the convective areas die off in the decaying phase, resulting in mostly stratiform precipitation e.g.,. ICON simulates mean precipitation within ±20 % of observed values during the growth phase, with the exception of land-based long-lived systems, which show a high difference of up to 40 %. Long-lived MCSs maintain high mean precipitation rates during the growth phase of the storm (especially over tropical land), while simulated MCSs show a decline in mean precipitation even early on during their lifetime. This results in simulated mean MCS precipitation becoming increasingly low-biased over the life cycle of storms ending at low differences of up to -60 % during the decaying phase. This indicates difficulties in simulating MCS stratiform precipitation, which is a common issue in many km-scale models e.g.,.

Finally, the combined differences in mean precipitation rates and MCS size are affecting the differences in MCS precipitation volume, which is better simulated during the growth phase but becomes increasingly negatively biased over their lifetime, especially for short-lived systems (Fig. k, l). The dominant driver of this bias is the low mean precipitation difference for short-lived systems, as well as the low size difference for long-lived storms. We refrain from analyzing differences in the development of heavy precipitation rates due to the deficiencies in capturing these in GPM-IMERG (see Fig. .)

We found some of the largest model wet biases in tropical ocean ITCZ precipitation (Fig. ). At the same time, MCSs are too infrequent in this region. To better understand the causes of this bias we also tracked smaller precipitating cloud systems that have cold cloud tops in addition to MCSs (see Sect.  and Fig. a–c). As previous studies have shown, cloud sizes (i.e., cold cloud top areas) follow a power-law frequency distribution e.g.,. Observations show that cold cloud shields that exceed our MCS threshold (>40 000 km²) contribute the most to tropical precipitation with a maximum contribution from clouds with ∼75 000 km² (Fig. a). This maximum is shifted towards smaller cloud sizes and lower volumes in ICON (Fig. b) with a low contribution bias for large clouds and an overestimation from small clouds (Fig. c). Not visible in Fig.  are the low-frequency biases in MCSs over land and ocean, and in non-MCS precipitating cold clouds over land (Fig. S9), since only normalized distributions are shown. These results are broadly consistent with , who showed that a large fraction of tropical precipitation originates from congestus clouds. GPM-IMERGv7 is not well suited to robustly quantify precipitation specifically from shallow or congestus clouds due to deficiencies in detecting precipitation from warm rain clouds .

A key question is: which clouds cause the excess simulated precipitation in the tropics if deep convective clouds are too infrequent and too small? Figure d–f show that in observations, most tropical precipitation originates beneath cold cloud tops, while ICON simulations have a too large fraction of rainfall coming from warm clouds. This issue is particularly evident in tropical land and ocean regions, but is also evident to a lesser extent in the Northern and Southern Hemispheres (Figs. S9 and S10). Excess precipitation from clouds with warm cloud tops indicates overly active warm-cloud processes, too slow glaciation in tropical clouds in the ICON microphysics scheme, or errors in GPM-IMERGv7 in capturing warm rain. This is consistent with the underestimated surface shortwave radiation over tropical oceans (Fig. c) that could be caused by overly reflective or overly abundant water clouds.

The output from the MOAAP algorithm allows us to analyze the interactions among atmospheric phenomena in creating hourly heavy precipitation events in observations and in the ICON model (Fig. ). We decided to exclude equatorial waves and fronts from the analysis in Fig. a–f since equatorial waves dominate tropical heavy precipitation statistics due to their large size, and including them obscures details from smaller-scale phenomena (Fig. S12). We decided to exclude fronts because they are more frequent in ICON, likely due to its higher model resolution, which allows sharper gradients to be simulated. Results including fronts and equatorial waves are shown in Fig. S12.

ICON can effectively simulate the processes that cause extreme precipitation across different climate regions, though it has some notable deficiencies. In observations, MCSs are more often the most frequent producers of heavy precipitation in tropical and subtropical ocean regions (Fig. a, b). This agrees with our previous analysis, which showed low-frequency and size differences in simulated MCSs. In contrast, simulated MCSs are more dominant over tropical and subtropical land regions, where observations highlight non-MCS cold clouds as most important. At the same time, moisture streams (areas of low-level strong and coherent horizontal moisture transport) are frequently more relevant in ICON than in observations, particularly in subtropical ocean regions.

Figure g–i show frequency statistics that visualize how often a phenomenon was co-located with hourly heavy precipitation in the Northern European (NEU), Southeast South America (SES), and the Equatorial Pacific Ocean (EPO) region. In the high-latitude NEU region, observations and ICON agree that surface cyclones and mid-level cyclones are the most frequent drivers with secondary contributions from jet streams and moisture streams (Fig. g). ICON features more heavy precipitation events from non-MCS clouds than observed and a larger contribution of frontal systems. Cut-off lows are less frequently involved in simulated heavy precipitation events than in observations.

In the SES region, ICON agrees with observations that MCSs and non-MCS clouds are the dominant heavy precipitation-producing phenomena (Fig. h). ICON also captures the importance of moisture streams and jet streams well, while frontal involvement is much higher than in ERA5.

We find a relatively poor agreement in the tropical EPO region (Fig. i) where ICON substantially underestimates the importance of MCSs and all equatorial waves except for Equatorial Rossby waves, in agreement with previously shown results. However, ICON can capture the importance of non-MCS cold clouds and moisture streams in the production of heavy precipitation.