We compare climatological differences across the three historical WRF experiments from 1980 through 2010, using ERA5-Land data at 9 km horizontal resolution as a reference. Reanalysis data are remapped to each WRF simulation's resolution using conservative regridding. Differences are calculated as simple differences between simulated and ERA5-Land values for temperature and as percent differences from climatological values for precipitation; precipitation differences are also computed as differences in climatological precipitation rate (mm d⁻¹). Seasonal differences are assessed for winter (DJF) and summer (JJA) only, for brevity.

It is important to note that ERA5 and ERA5-Land are not direct observations, and these reanalysis products carry their own biases relative to station and satellite-based estimates . Direct comparison to point observations is also complicated by the scale mismatch between gridded model output and station data. We therefore use the term “difference” throughout the manuscript when referring to comparisons against ERA5-Land, reserving “bias” for cases where it is substantially more concise (e.g., “cold bias” rather than “negative temperature difference relative to ERA5-Land”). Additionally, to provide context about how the estimated bias can differ depending on the choice of reference dataset, we have included climatological differences of the WRF simulations from PRISM and Daymet.

We estimate diurnal precipitation peaks following , fitting a second-order harmonic function (Eq. ) to hourly precipitation from 1 June through 31 August for each year from 1980 through 2010, comparing the three WRF simulations against GPM IMERG V07B . 1Pt=P+A1×cos⁡2πt24+B1×sin⁡2πt24+A2×cos⁡4πt12+B2×sin⁡4πt12

Here Pt is precipitation at hour t, P is total daily precipitation, A1 and B1 control the 24 h cycle, and A2 and B2 control the 12 h cycle. Fourier transform estimates provide both the phase (hour of peak precipitation) and magnitude.

We also evaluate the mean Spring (March through May) diurnal temperature range, a period characterized by large differences in daily minimum and maximum temperatures.

We define the North American Monsoon region as northern Mexico (115–105° W and 25–32° N) and the southern US (115–105° W and 32–40° N) monsoon extent, comparing daily climatological precipitation against IMERG.

We assess snow biases across the three WRF simulations against the University of Arizona SWE dataset , examining gridded differences in SWE for late winter and early spring (JFMAM) in the Western US. We additionally compare SWE at SNOTEL station locations to the closest gridpoint from each WRF simulation, aggregating results by hydrologic unit code-8 (HUC-8) watershed regions to assess the seasonality of the snow cycle, including the accumulation, peak, and melt phases.

We compare the three simulations using the “perfect model framework”, treating the results from C404 as the goal-post to assess the efficacy of NAC6 and NAC5. Comparisons to reanalysis data are also provided.

We evaluate precipitation extremes at three temporal scales: hourly, 6-hourly, and daily accumulated precipitation. To reduce data volumes and memory requirements, we compute the maximum precipitation value at each gridpoint for each calendar month and temporal scale, then flatten these gridpoint-level monthly maxima across space into a single distribution for each dataset. Histograms and extreme percentiles (90, 95, 99, 99.9, and 99.99th) of this flattened distribution are used to characterize the extreme tail of precipitation over the CONUS domain (130–60° W, 30–50° N).

For tropical cyclones (TCs), we select Hurricane Ivan as a case study, evaluating storm intensity and impact metrics (minimum surface pressure, maximum precipitation rate, and maximum wind speed) within a 150 km radius of the storm center, defined as the gridpoint with lowest surface pressure. We extend this into a climatology using the same metrics for every TC that tracked into the Gulf of Mexico (20–40° N, 105–75° W) and was labeled a hurricane by the National Hurricane Center. We use IBTrACS data for an initial guess at storm location, with the actual storm center in each WRF simulation identified as the minimum surface pressure within 50 km of the IBTrACS position; TC category is also taken from IBTrACS/National Hurricane Center classification.

For the mesoscale convective system case study, we examine the 22 July 2003 “Mid-South Derecho” using 750 hPa horizontal moisture flux (|Q|=(qu)2+(qv)2) into the Midwest, moisture flux convergence (MFC; Eq. ), and 850 hPa temperature advection (Eq. ). 2MFC=-∫925hPa700hPa∂(qu)∂x+∂(qv)∂ydp3-V⋅∇T=-u∂T∂x+v∂T∂y

In these equations, q is specific humidity, u and v represent zonal and meridional winds, p is pressure, and T is temperature. Grid coordinates are represented by x and y.

Analysis of precipitation biases and the diurnal cycle of precipitation has suggested that the North American Monsoon in these WRF simulations is wet biased in the southwest US (Figs. , ), and Fig.  confirms this. In northern Mexico, the seasonal evolution of the monsoon in all three historical simulations matches IMERG quite well. Notably, the NAC6 simulation has the largest wet bias, though the magnitude of the simulated difference is quite small day-to-day (approximately 0.5 mm d⁻¹). When integrated over the July through September months, NAC5 has the smallest difference from IMERG (+9.8 % of IMERG's climatological magnitude).

In contrast, for the northern extent of the monsoon region in the southern US (Fig. b), there is much more divergence in the simulated monsoon precipitation across the WRF simulations. All simulations produce more precipitation than IMERG in the cold season (November through April), and then begin to diverge in Spring (April through June), at which point C404 most closely matches observations. The NAC6 run is the best at capturing the timing of the monsoon precipitation peak in August (note how both C404 and NAC5 produce a secondary peak near 25 August), but it produces about 28. 8 % more precipitation than IMERG from July through September. This is a marked improvement from NAC5 (+42.5 % precipitation in July through September) for the southern US monsoon region.

As estimated by the University of Arizona SWE dataset, climatological values are notably higher at high elevation throughout the western United States, with the prominent mountain ranges in the region experiencing normal SWE values well above 400 kg m⁻² for the January through April mean (Fig. S12). We highlight this map to show that there are wide swaths of the region that also experience trace amounts of snow water equivalent below 50 kg m⁻². The subsequent analysis focuses on the gridcells that have climatological values higher than 50 kg m⁻² so that the bias assessment is not heavily skewed by these smaller SWE values.

The spatial maps (Fig. a, c, e) display very heterogeneous differences in January through April climatological SWE, but they highlight the granularity of SWE values for the higher resolution NAC6 and C404. A mean dry signal is evident in all simulations and at all elevation bands (Fig. b, d, f), and in general the magnitude of the difference from the reference SWE decreases with increasing spatial resolution. From the 4 km C404 to the 25 km NAC5, the mean dry differences substantially increase: 12 % (C404) to 35 % (NAC5) for 0–500 m, 33 % (C404) to 65 % (NAC5) for 500–1500 m, 29 % (C404) to 45 % (NAC5) for 1500–2500 m, and 25 % (C404) to 50 % (NAC5) for elevations greater than 2500 m. NAC6 has a higher mean bias in SWE than C404 at all elevation bands, but the differences are small: 15 % at the lowest elevation band, and only 1 %–3 % in the three elevation bands above 500 m.

When comparing the gridded WRF data with station data, we expect WRF SWE to be lower than the SNOTEL sites because the gridpoint values represent an integration over a large area. In all simulations, the seasonal peak in SWE is earlier than the maximum seasonal peak from the SNOTEL stations, indicating that the snowpack accumulation phase is too short and snowmelt begins generally too early in the year. When all stations are integrated together, the peak SWE timing is 29 d too early in C404, 27 d too early in NAC6, and 41 d too early in NAC5 (Fig. S12). The seasonal cycle is improved from NAC5 to NAC6 and C404, with a notable elongation of the maximum SWE at all stations (Fig. b) and in the HUC-8 watershed regions (Fig. c–e). This early peak in SWE is consistent across the diverse watershed regions in the western US. showed that early season melt events accelerate snowmelt processes in WRF and Noah-MP simulations of the western US, likely due to poorly-represented snow albedo processes, excessive downward surface shortwave radiation, and strong surface winds that enhance surface heat exchange.

These improvements in the representation of the snow seasonal cycle and climatology likely come from the improved representation of snow-relevant orographic processes, including terrain-induced circulation features and local ridge shadowing that reduces snowmelt and sublimation . The land-surface model also plays a large role in the representation of snow and frozen water storage, which has been substantially improved in Noah-MP (used in NAC6 and CONUS404) by updates to radiative transfer and snow compaction .

Figure compares precipitation extremes by displaying histograms of the maximum 1-hourly, 6-hourly, and daily precipitation rates in each simulation over the CONUS domain (130–60° W and 30–50° N). In any precipitation dataset or simulation, the frequency and intensity of extreme precipitation rates are dependent on the spatial resolution, with precipitation intensity decreasing as gridpoint area becomes larger . While spatially integrating datasets or simulations with different horizontal resolution may provide more equitable precipitation rates, it is critically important to capture localized extreme rates too.

showed that the C404 dataset simulates extreme precipitation rates that closely match station-based observational estimates, rendering this simulation a quality benchmark for our precipitation comparisons . Figure  shows that the probability density of extreme precipitation rates in NAC6 is very similar to C404 at all temporal frequencies (especially 1 h), even without resolved convection. Precipitation rates exceeding 224 mm h⁻¹ are not found in the 12 km NAC6, but the overall distribution of extreme rainfall events is a marked improvement over NAC5 which more noticeably deviates from the extreme precipitation distribution at 1 h timescales.

Quantitative comparison of extreme precipitation percentiles (Fig. ) confirms that this agreement between NAC6 and C404 holds across the tail of the distribution, not just in bulk. At 1 h resolution, NAC6 percentiles closely bracket C404 from the 99th through the 99.99th percentile (45.1 vs. 47.8 mm h⁻¹ and 111.6 vs. 109.5 mm h⁻¹, respectively), differing by only a few percent even at the most extreme thresholds. At the coarser 6 h and daily accumulations, NAC6 remains close to C404 through the 99th percentile but modestly exceeds it at the 99.9th and 99.99th percentiles (248.7 vs. 220.7 mm per 6 h at the 99.99th percentile, a 13 % difference), suggesting that parameterized convection at 12 km may produce proportionally larger accumulated totals over longer windows than explicitly resolved convection, even where the two simulations agree closely on sub-daily rates. NAC5, in contrast, underestimates extreme precipitation relative to C404 at every percentile and frequency, and the underestimate grows with percentile, widening from roughly 30 %–55 % at the 99th percentile to 35 %–61 % at the 99.99th percentile depending on frequency. This indicates that the improvement from NAC5 to NAC6 is not confined to the bulk of the precipitation distribution but extends into its extreme tail.

The following analysis examines processes historically difficult to simulate at coarser model resolutions, including a case study and climatology of (TC) characteristics and a case study of a severe summertime convective storm. Hurricane Ivan is chosen as the TC case study because it was a strong TC within the temporal and spatial range of all datasets. Ivan reached Category 5 status on 9 September 2004 and then made landfall as a Category 3 hurricane in Orange Beach, Alabama on 16 September 2004, producing a record-breaking tornado outbreak and up to 43 cm of rainfall with widespread flooding . While these simulations cannot represent the tornado outbreak, we evaluate storm intensity and impact metrics: minimum surface pressure, maximum precipitation rate, and maximum wind speed within a 150 km radius of the storm center (the gridpoint with lowest surface pressure).

Figure a–c shows the downward shortwave radiation at the surface for the hurricane in each simulation. While outgoing longwave radiation would provide a better visualization of the TC structure, this variable is not output at the needed frequency in NAC5 to enable a comparison between the three datasets. C404 clearly resolves the inner eye and spiral rainbands, with a notably symmetric pre-landfall structure indicating low wind shear, closely resembling observed Hurricane Ivan at this time (15 September 2004, 18:00 UTC) . NAC6 also resolves important storm characteristics, including a well-defined eye, coarse spiral banding, and symmetric structure (Fig. b). At 25 km, NAC5 still resolves a central eye but fails to depict realistic spiral banding or storm symmetry (Fig. c).

NAC6 also reliably produces metrics describing storm intensity and potential impacts (Fig. d–f). Both C404 and NAC6 show very similar evolutions of minimum surface pressure, maximum precipitation rate, and maximum wind speed while the TC was in the Gulf of Mexico (14 through 17 September). TC metrics from NAC5 are comparatively very weak.

We note that part of the improvement in simulated TC structure and intensity from NAC5 to NAC6 (Figs. , ) may be attributable to the driving reanalysis rather than WRF resolution or physics alone. ERA5 represents tropical cyclones with greater intensity and structural realism than ERA-Interim, owing to its finer resolution and more advanced assimilation of observations . Since NAC6 is nudged toward ERA5 and NAC5 toward ERA-Interim, some of the TC improvement documented here likely reflects this difference in boundary conditions. Figure S13 highlights that the updates to ERA5 result in a much more realistic looking TC, as the ERA-Interim representation of Hurricane Ivan completely lacks an eye and has very weak TC intensity metrics throughout the life of the storm (minimum surface pressure of 975 hPa, a maximum precipitation rate of 20 mm h⁻¹ and a maximum wind speed of about 31 m s⁻¹).

This distinction also matters for interpreting these results in the context of the planned NA-CORDEX-CMIP6 ensemble, which will be driven by CMIP6 ESMs with resolution coarser than either reanalysis; TC statistics from that ensemble may therefore resemble NAC5 more closely than the ERA5-driven NAC6 evaluation run presented here.

We extend this analysis into a climatology by evaluating the same metrics–minimum pressure, maximum precipitation rate, and maximum 10 m wind speed – for every TC that tracked into the Gulf of Mexico within 20–40° N and 105–75° W, and was labeled a hurricane by the National Hurricane Center. We use the IBTrACS TC data for an initial guess at the storm location (the actual storm center in each WRF simulation is obtained by finding the minimum surface pressure within 50 km of the IBTrACS storm location) and for the TC category as defined by the US National Hurricane Center.

Violin and box plots of these metrics are shown for the three simulations in Fig. . For all three intensity metrics, the C404 and NAC6 runs have similar distributions associated with TCs, with the strongest similarity shown for minimum pressure (Fig. a) and maximum precipitation (Fig. b). The wind speed distributions do clearly decrease with the resolution (Fig. c), but NAC6 still performs well compared to C404. These results suggest that the NAC6 dataset is suitable for analyzing larger-scale statistics and impacts of tropical cyclones. Research focused on TC dynamics and storm structure (for example, eyewall updrafts and mesovortices) requires a much higher resolution than even 4 km .

Convectively generated mid-latitude storms pose substantial threats to human life and cause significant economic damage each year , especially in central North America. The most impactful are often mesoscale convective systems (MCSs), in which localized convection organizes into larger propagating complexes of intense thunderstorm activity, sometimes producing powerful windstorms or derechos.

On 22 July 2003, a powerful MCS known as the “Mid-South Derecho” propagated from Arkansas through Northern Alabama with wind gusts exceeding 40 m s⁻¹ . We compare the severe storm environment across the three runs using 750 hPa horizontal moisture flux (|Q|=(qu)2+(qv)2) into the Midwest on 22 July 2003 (Fig. ). Moisture flux at this level indicates increased atmospheric instability and potential for heavier precipitation . The large-scale circulation (500 hPa) is similar across simulations, but notable differences emerge in the moisture flux fields, which are not nudged in C404 and NAC6. We do not expect parameterized convection to adequately simulate an MCS, but highlight differences in the resolved storm environment across simulations.

Figure shows substantially different mesoscale storm structures in the three WRF runs. All datasets produce an elongated longitudinal band of enhanced moisture flux aligned with low-level winds, but smaller-scale convective features are only present in C404, visible as localized areas (< 20 km) of enhanced moisture flux, particularly in the northeast quadrant of Fig. a. Both NAC6 and NAC5 resolve a more contiguous moisture flux band lacking these smaller-scale convective features (Fig. b–c).

Figure d–e highlights the importance of convection-permitting scales for representing distributions of moisture flux convergence (MFC; Eq. ; Fig. d) and temperature advection (Eq. ; Fig. e), two metrics associated with intense localized instability and extreme precipitation rates (spatial plots in Fig. S14). All simulations exhibit distributions centered near zero with similar overall structure, but both CORDEX simulations have substantially narrower distributions than the convection-resolving dataset, with far fewer high-magnitude convergence and advection events. C404 shows substantially larger variability than NAC6 and NAC5 in both moisture flux convergence and temperature advection, with 95th percentile magnitudes 3.0–4.0 times higher (1.72×10-3 vs. 5.68×10-4 and 4.35×10-4 kg kg⁻¹ s⁻¹; 1.79 vs. 0.54 and 0.59 K h⁻¹) than the CORDEX simulations. NAC6 and NAC5 show comparable magnitudes to each other in both metrics, with NAC6 only modestly exceeding NAC5 (e.g., 95th percentile moisture convergence of 5.68×10-4 vs. 4.35×10-4 kg kg⁻¹ s⁻¹), underscoring that resolved convection, not resolution alone, drives the sharper gradients captured by C404.

Both reanalyses show even narrower distributions than their driving simulations, consistent with their coarser native resolution (∼ 0.25° for ERA5, ∼ 0.7° for ERA-Interim) relative to the 12 and 25 km WRF grids (Fig. S15). Notably, NAC6's 95th percentile magnitudes closely track ERA5 (5.68×10-4 vs. 5.01×10-4 kg kg⁻¹ s⁻¹ for moisture convergence; 0.54 vs. 0.55 K h⁻¹ for temperature advection), while NAC5 exceeds ERA-Interim by a larger margin, particularly for moisture convergence (4.35×10-4 vs. 2.39×10-4 kg kg⁻¹ s⁻¹, roughly 1.8 times higher). This suggests that part of the apparent gap between NAC6 and NAC5 reflects the difference in resolution and structure between their respective driving reanalyses, rather than the WRF grid spacing alone. Disentangling these reanalysis-driven contributions would require another simulation with the same forcing reanalysis at a different regional resolution, which is out of the scope of the current case study.