First, we look at the overall atmospheric results over the CONUS (20–50∘ N, 65–125∘ W) by comparing the spatial RMSEs of the historical ensemble means between LR (blue triangles) and NARRM (red triangles) in Fig. . The same metric is used in Fig.  for the global results, but we adjust the variables to be more relevant to the CONUS. NARRM generally produces better (as quantified by smaller RMSE numbers) results than LR for these annual and seasonal climatologies, such as SW CRE (Fig. a), precipitation (Fig. c), and 200 hPa zonal wind (Fig. f). Because we have not retuned the physics of NARRM, some deteriorations are expected, for example longwave (LW) CRE in DJF and March–April–May (MAM) (Fig. b).

Precipitation and clouds, which are obviously important quantities for the water cycle, are largely improved in NARRM compared to LR. The precipitation patterns are better captured by NARRM as observed at the Sierra Madre Occidental in JJA (Fig.  left column) and in the western US in DJF (Fig.  right column) due to the better-resolved topography in NARRM (see Fig. ). The poor representation of marine stratocumulus clouds is a long-standing problem that plagues many ESMs e.g.,. The underestimation of summertime low clouds (manifested as the excessive TOA shortwave CRE) in the California stratocumulus region is substantially improved with NARRM (Fig. ). The improvement in this bias is likely due to a reduced bias in the simulated sea surface temperature (SST) in the coupled RRM. To constrain the impact of the RRM on the SST bias, we conduct two additional experiments: one where the LR ocean is coupled to the RRM atmosphere, and another where the RRM ocean is coupled to the LR atmosphere. The simulated SST bias averaged over years 51–100 of the piControl is shown in Fig. . Comparison of Fig. a and b shows a clear reduction in bias for the NARRM simulation. Figure c and d illustrate that the regional refinement in the atmosphere (c) is primarily responsible for the reduction in SST bias; however, comparing to the NARRM result, we see that the bias is further reduced when regional refinement is included in both components, highlighting the advantage of coupled RRM over a single-component RRM. Further details will be described in a future paper by Van Roekel et al. (Luke P. Van Roekel, personal communication, 2023).

Another well-known issue of global ESMs is the poorly captured diurnal propagation of organized mesoscale convective systems (MCSs). Over CONUS, such MCSs originate from the front range of the Rockies in the afternoon and propagate eastward, manifesting as a nocturnal precipitation peak in the central US and contributing as much as half of the summertime rainfall in that region . Both LR and NARRM simulate the summertime nocturnal rain peak in the central US (Fig. ) because of the new convective trigger method for deep convection . However, the magnitude is weaker and the area of nocturnal peak extends much larger (almost the whole eastern half of the US) than the observation, which could be caused by remaining propagation or convective trigger deficiencies. Nevertheless, NARRM reduces the underestimation of maximum diurnal cycle magnitude with an 80 % greater value (5.61 mm d-1 vs. 3.10 mm d-1) than LR on the same 100 km grids. The NARRM maximum can be as high as 6.71 mm d-1 on the 25 km grids but still biases low compared to the observation (10.89 mm d-1). This result suggests that a resolution of ∼ 25 km is not adequate to capture the physics driving propagating MCSs, which probably require convection-permitting atmospheric simulations to achieve a good agreement with observations .

The E3SMv2 LR model simulates too strong aerosol-related forcing, which has been identified as the primary cause of the underestimated warming in the later portion of historical period in Fig. . We will examine here if the NARRM configuration helps bring down the biases in aerosols and anthropogenic forcing by better resolving the meteorological and climate fields. In addition, since NARRM employs the hybrid time step approach that eliminates retuning the scale-dependent aerosol parameters used in LR, e.g., the global scaling factor used to constrain the total emission fluxes of natural aerosols (dust and sea salt), which depend non-linearly on the model-resolved small-scale surface winds, we will also discuss the impact of increasing model horizontal resolution on the natural aerosols and total aerosol optical depth (AOD) in NARRM.

Impact on anthropogenic aerosols.

Aerosols in the NARRM configuration are represented in the same manner as in the LR with the enhanced MAM4 and improved dust aerosol properties . Anthropogenic and wildfire emissions used in the LR and NARRM experiments are also from the same input datasets. However, cloud microphysical processes, horizontal advection, and convection can affect aerosol loading within the NARRM high-resolution domain if wet deposition and/or transport are substantially different from the LR configuration . We first compare modeled surface mass concentrations of SO2, sulfate, black carbon, and organic carbon between the ensemble means of LR and NARRM historical simulations, and evaluate them against ground-based observations from the Clean Air Status and Trends Network (CASTNET) and the Interagency Monitoring of Protected Visual Environments (IMPROVE). The results are shown in Fig. . In general, both sets of simulations show a strong correlation with measurements, and biases are very similar between the two for the anthropogenic aerosol species. NARRM simulations slightly increase (less than 7 %) the concentrations of the four species shown here, compared to LR simulations, which may result from less wet removal or vertical transport in the refined mesh.

Here we examine the impact of increased horizontal resolution over NA on the simulated clouds and their radiative effects with the LR and NARRM historical simulations and on cloud feedback changes with the quadrupling 4xCO2 simulations.

Figure  compares the cloud cover between the E3SM CALIPSO (Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation) simulator output and the GCM-Oriented CALIPSO Cloud Product (CALIPSO-GOCCP) . Cloud cover and cloud thermodynamic phase are diagnosed with the same algorithm in the CALIPSO simulator and CALIPSO-GOCCP data, facilitating consistent model–observation comparisons . Observed total cloud cover is larger over the NA polar region than the CONUS in the CALIPSO-GOCCP data. Total cloud cover is larger than 60 % over the eastern Pacific Ocean, northern Atlantic Ocean, and Arctic Ocean. Strong land–ocean contrast is observed – liquid phase clouds dominate over the ocean, while ice phase clouds prevail over the land in areas such as Greenland and CONUS. Compared to CALIPSO-GOCCP, LR overestimates total cloud cover at NA high latitudes and the western CONUS and underestimates it in Greenland, particularly over Baffin Bay and near the Greenland coast. The underestimated cloud cover over the western coast of the CONUS is also notable, which is consistent with the previous discussion on Fig. . The excessive modeled total cloud cover is primarily attributed to the positive biases in the liquid cloud over the polar region. The positive biases of ice cloud cover contributes to the biases over the mountainous regions in the western NA. On the other hand, ice clouds are underestimated over Greenland and northern Canada.

Over land, NARRM displays improvements relative to LR in western NA and the Arctic for both cloud phases. For instance, the ice cloud biases are significantly reduced from Alaska to the western CONUS (Fig. f, i), and the liquid cloud deficiencies over Alaska and Greenland are generally decreased (Fig. e, h). The better represented topography in NARRM (Fig. ) is probably the key factor of these NARRM improvements. The impact of increased horizontal resolution on simulated cloud phase is also noted in the E3SMv1 model with the CALIPSO simulator , where increased horizontal resolution also slightly decreases simulated liquid and ice clouds at temperatures warmer than -40 ∘C in the Arctic region.

Over ocean, NARRM substantially improves the stratocumulus clouds to the west of coastal regions. NARRM also moderately outperforms LR in representing liquid clouds over the North Atlantic to the west of Greenland. This is related to the warmer (∼ 1.5 ∘C) NARRM surface air temperature over the Labrador Sea. This warmer NARRM surface air temperature is consistent with the decreased sea ice concentration in that region (not shown), which is somewhat expected as an advantage of refining grids for both atmosphere and ocean and sea ice (see Fig. ). Further process-level analysis is necessary to fully understand this LR-NARRM model behavior change and will be reported in separate papers.

Figure compares the simulated SW and LW CRE with the CERES-EBAF Ed4.1 observations. Large negative SW CRE biases and positive LW CRE biases are shown over the Arctic land area and western coast of NA (i.e., Alaska to Oregon) in LR, which mainly result from the overestimated cloud cover. These biases are substantially reduced in NARRM, primarily owing to the improved cloud cover (Fig. ). Given the reduced negative bias of marine stratocumulus clouds in NARRM near the western coasts of the CONUS, simulated SW CRE is also largely improved. As discussed by , sea ice concentration is largely overestimated over the North Atlantic Ocean in LR. The too large sea ice extent leads to weaker SW and LW CRE than observed in the Labrador Sea. This is primarily because of the brighter and colder sea ice surface in LR that reflects more SW radiative fluxes and emits less LW radiative fluxes than the observations under clear-sky conditions (not shown). Compared to LR, the maximum positive bias in NARRM sea ice extent in Labrador Sea is greatly alleviated. The better simulated sea ice extent reduces the biases of overly reflective clear-sky SW radiation and the insufficient clear-sky outgoing LW radiation. With generally comparable all-sky SW and LW radiative fluxes between LR and NARRM, those reduced clear-sky biases thus lead to a better CRE in NARRM over Labrador Sea.

Given the improved historical cloud cover and cloud radiative effects over NA, we further examine the regional climate feedbacks over this region in Fig. . Relative to the global mean value, the total climate feedback over NA is more negative from NARRM than from LR. This mainly results from the more negative Planck feedback and less positive SW cloud feedbacks. The more negative Planck feedback is related to the stronger surface warming over the northeastern Pacific (not shown).

Figure shows the spatial distribution of cloud feedback of LR and NARRM in the NA region. Due to the different cloud types over land and ocean, we report their regional averages separately. Notably, the total land cloud feedback is 0.41 W m-2 K-1 smaller in NARRM (0.26 W m-2 K-1) than in LR (0.67 W m-2 K-1), which is dominated by the reduced SW cloud feedback over the northeastern US (Fig. e). Further examination indicates this reduction is mainly related to the weaker reduction in low cloud cover under warming in NARRM. Figure g shows that the overestimated cloud cover is slightly alleviated over the Arctic land region in NARRM, implying that the lower mean state cloud cover might contribute to a weaker cloud reduction under warming there. Over ocean, NARRM presents a stronger SW cloud feedback and a weaker LW cloud feedback over the marine low cloud regime, leading to a small change in total cloud feedback. Across the CSS/WGNE Pacific Cross-Section Intercomparison (GPCI) transect , NARRM tends to show a weaker positive cloud feedback near the coast and more positive cloud feedback off the coast. These factors suggest that the regional refinement can significantly affect the regional cloud responses under warming and the predictability of regional climate.

One of the primary motivations for pushing climate simulation resolution is to potentially better capture extremes. Extratropical cyclones (ETCs) are a major weather extreme phenomenon at middle and high latitudes, bringing with them strong winds and precipitation that can exert substantial societal impacts along their pathways over days and over hundreds to thousands of kilometers. Climate changes are likely to induce changes to the dynamical and physical characteristics of ETCs as well as their geospatial distribution (e.g., ). Projections of such future changes rely heavily on numerical climate and ESM models. Their skills in simulating major weather systems like ETC have been carefully scrutinized by modeling centers and by the climate science community in conjunction with the major intercomparison campaigns . While the conventional climate models with grid resolution around 100 km show reasonable skill in producing ETC frequency and spatial track density, it has also been found that higher-resolution models are better capable of capturing more intense ETCs e.g.,, which is critical for using ESM to project future climates, as growing evidence shows that global warming tends to shift the weather spectrum to the more extreme end . Here, we will demonstrate the benefits of higher resolution in simulating ETCs in a regionally refined setting by comparing the results from NARRM with those from LR simulations against the ETC activities derived from the ERA5 reanalysis.

The ETC tracks and statistics can be obtained using automated identification and tracking algorithms (e.g., ). The objective identification and tracking also make it suitable to compare ETC activities and statistics derived from different data sources in particular for model evaluations. The algorithms usually identify and track the spatial features of a meteorological variable, such as mean sea level pressure (MSLP) or 850 hPa vorticity, that can characterize the structure of cyclones and their movements. In this work, we use a community feature detection and tracking framework, TempestExtremes , to derive ETC activities from 6-hourly MSLP data during the period of 1985–2014 from the E3SM simulations and the ERA5 reanalysis. Considering higher-resolution data can more accurately identify the storms and their tracks , all the model and reanalysis data are placed on 1∘×1∘ grids to feed the tracking software. The algorithm takes two steps. First, a candidate cyclone is detected when a minimum MSLP feature is enclosed by a contour of 200 Pa interval within 6∘ of the center. Candidates within 6∘ of one another are merged, with the lower center pressure taking precedence. The candidates are then stitched together to define the tracks if the features persist for at least 60 h with a maximum gap of at most 18 h. From the start to the end, a candidate cyclone must travel at least 12∘ great circle distance to qualify as an ETC.

Over the NARRM high-resolution domain, ETCs are most active during winter. Figure  shows the mean DJF track density for the models and the analysis derived by casting the computed ETC track data onto 5∘×5∘ grids. The tracks are mostly concentrated over the northeastern Pacific and northwestern Atlantic that form the well-known storm tracks. Both LR and NARRM simulations capture these main features to a large extent. There are, however, notable differences between LR and NARRM over these oceanic storm tracks. The track densities are clearly underestimated in the LR simulations inside the refined region, except for the Atlantic storm track in the coupled mode. NARRM clearly produces higher ETC track density than LR does for both sections of the oceanic storm tracks and mostly agrees better with the ERA5 data, although in the coupled mode the track density tends to be overestimated. The shape and orientation of both the Pacific and Atlantic storm tracks are much better produced by NARRM in the coupled mode. Given the significant differences from both coupled LR (Fig. d) and uncoupled NARRM (Fig. c), it is reasonable to believe that the better captured storm track shapes in the coupled NARRM are due to interactions with the refined ocean (Fig. ). Several secondary centers of active ETCs in the ERA5 over land are also reproduced by the models, including the active regions over the Great Lakes and Hudson Bay, although the densities are overestimated in the models (more so in the NARRM). It is worthwhile mentioning that the NARRM simulations are able to produce the chain of secondary centers to the east of the Rocky Mountains that are also present in the ERA5 reanalysis but are largely missing in the LR simulations. This is presumably due to the NARRM's better-resolved mountainous terrain features, a benefit by design.

The benefit of grid refinement can be further seen in Fig. , which shows the histograms of the ETC as a function of the minimum center pressure and the maximum deepening rate during its lifetime. All events within the refined region bounded by the dashed black lines as shown in Fig.  are used to compute these statistics. The maximum deepening rate is defined as the maximum 6-hourly center pressure drop, normalized by sin⁡(φref)/sin⁡(φ), with φ being the latitude and φref the reference latitude at 45∘ see also in. Clearly the NARRM very closely reproduces the number of intense cyclones (minimum center MSLP <960 hPa), while unsurprisingly the LR model underproduces. This is true in coupled and uncoupled modes. Both LR and NARRM simulations overestimate the number of weaker ETCs. On a similar note, the observed number of rapid-growth cyclones is closely reproduced by the NARRM simulations but is clearly underestimated by a large margin in the LR simulations.

Natural storage provided by mountain snowpack is central to water supply reliability in the western US . To evaluate model skill in representing this critical hydroclimate benchmark variable, intra-annual snowpack dynamics are evaluated using the methodology known as the snow water equivalent (SWE) triangle . The seven metrics that make up the SWE triangle attempt to distill management-relevant aspects of the accumulation and ablation of snowpack (e.g., peak water volume and snowmelt rate) for any arbitrary gridded SWE dataset. Five HUC2 basins of the mountainous western US are used to derive five-member ensemble and basin average evaluations of LR and NARRM fully coupled historical simulations and are compared with ERA5. All datasets are bi-linearly regridded using the Earth System Modeling Framework (ESMF) to 0.25∘ resolution prior to masking and computing the basin-average SWE triangle metrics.

NARRM provides enhanced winter (DJF) climatological representation of the spatial variability of SWE across the CONUS relative to LR (Fig. a, b). This is seen through higher SWE magnitudes and more granular spatial structures in NARRM compared with LR, particularly in coastal mountain ranges such as the Cascades and Sierra Nevada, and corroborates a long history of ESM studies that highlight the critical importance of horizontal resolution (≤ 0.25∘) in properly representing the mountainous hydrologic cycle . As shown through the more granular intra-seasonal perspective of the SWE triangle metrics, certain aspects in the snowpack dynamics are improved with NARRM (e.g., peak water volume), namely in the Pacific Northwest and California (Fig. c). With that said, some E3SM SWE biases are not ameliorated with horizontal resolution and may arise due to the combination of higher winter season precipitation (Fig. ) and a general cool bias (Fig. ) in both the LR and NARRM fully coupled historical simulations.

NARRM better captures spatial variability in land hydrologic processes, as indicated by two most important land hydrologic variables, total runoff (Fig. ) and evapotranspiration (ET) (Fig. ). For instance, in the coastal Pacific regions, the Rocky Mountains block atmospheric moisture from ocean to inland areas and lead to two distinct hydrologic regimes: a wet regime in the western mountains and a dry regime in the eastern mountains. This abrupt spatial shift from a wet to dry hydrologic regime can be clearly seen in the composite runoff map from the Global Runoff Data Center (GRDC) , Fig. a, and the observed evapotranspiration map from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite observations , Fig. a. Note that the GRDC runoff map is not completely based on the observational data since runoff measurements are not available at the regional or global scales due to technical and economic limitations. It is nevertheless a more realistic estimate than any model simulations because it was first generated with a monthly hydrologic model (hence producing spatiotemporal variability) and then corrected for bias against discharge measurements at thousands of river gauges . This abrupt shift of hydrologic regime around the Rocky Mountains, along with the other spatial variations, is much better resolved in the NARRM simulation than LR, which is the case for both simulated runoff and evapotranspiration, as shown in Figs. b, c and b, c. The NARRM simulated spatial patterns are thus more realistic than the LR ones over NA. Over the remaining regions of the globe, the NARRM and LR simulated spatial patterns are quite similar to each other in terms of both runoff and evapotranspiration (not shown).

The simulation biases in runoff and evapotranspiration are further examined in terms of absolute biases, i.e., the absolute difference between the simulated and “benchmark” values. Here the GRDC runoff and MODIS ET data are used as the benchmark data. Figures d and d show the maps of absolute bias difference, i.e., the difference between the absolute biases in the LR simulation and those in the NARRM simulation (former subtracting latter), for annual mean runoff and evapotranspiration, respectively. For a specific grid cell in these two maps, a positive difference means the absolute bias in the LR simulation is larger than that in the NARRM and vice versa. It appears that there are more absolute biases in LR than NARRM over both the western and eastern US. Using the longitude 100∘ W as the divide, the average absolute bias differences (positive indicates NARRM has less overall absolute bias than LR) are 22.8 and 0.9 mm yr-1 over the western and eastern US, respectively, for annual mean runoff, and are 21.6 and 18.5 mm yr-1 over the western and eastern US, respectively, for annual mean evapotranspiration. When compared to LR, NARRM can thus help reduce simulation biases in hydrologic variables.

Streamflow simulations are typically affected by multiple sources of uncertainties, such as the biases in the simulated runoff, the uncertainties in the river model parameters (e.g., river network topology, channel geometry, Manning's roughness coefficients), and water demand data . For river network topology, the 0.5∘ and 0.125∘ resolution river network data are used for the LR and NARRM simulations, respectively, as shown in Fig. a, b. It is expected that a higher-resolution river network data can represent rivers more smoothly and hence more realistically. Another benefit of higher resolution river network data is to enable more extensive streamflow validation. Terrestrial water fluxes, particularly surface runoff and streamflow, are dominated by gravity and controlled by topography and hence mostly follow irregular watershed boundaries. In most land surface and ESMs, including E3SM, regular lat–long grids are used to resolve spatial heterogeneity for both runoff and river processes to be compatible with the other land and atmospheric components . Both the magnitude and timing of streamflow at each river gauge are dominated by the corresponding upstream drainage area. Streamflow simulations are thus largely affected by the discrepancies between the watershed boundaries and regular lat–long grids. These discrepancies can be significantly reduced with higher-resolution river network data. For example, in this study a 10 % discrepancy threshold is used to select the river gauges for validating streamflow simulations; i.e., the relative difference between the real upstream drainage area of a river gauge and that estimated from a lat–long grid-based river network should not exceed 10 %. Over the NA domain, 615 river gauges satisfy the requirement for the 0.5∘ resolution river network, whilst 2924 river gauges satisfy the requirement for the 0.125∘ resolution river network. There are 563 river gauges that simultaneously satisfy the requirement for both resolutions.

Figure c shows the comparison between the annual mean observed and simulated streamflow over these 563 river gauges. Overall, both simulations produce the long-term average streamflow reasonably well across these gauges. NARRM performs noticeably better (closer to the red 1:1 line) for the top four gauges with the largest discharges. An additional analysis (figure not shown) indicates that LR produces greater absolute bias than NARRM in 330 out of 563 gauges (about 60 %) in the streamflow simulation. Interestingly, it appears that the overestimation and underestimation of JJA and DJF streamflow are concentrated in the western and eastern US, respectively, for both LR and NARRM, as shown in Fig. a–d. Figure e, f displays the difference in absolute biases between LR and NARRM (the former subtracted from the latter) at individual river gauges for the JJA and DJF seasons. Positive differences (indicating greater bias in LR than in NARRM, purple color) dominate over most gauges in the eastern US during JJA and over the CONUS during DJF. Taken together, Figs.  and suggest an overall better performance of NARRM despite all the uncertainties.