A convection‐permitting dynamically downscaled dataset over the Midwestern United States

Climate change is expected to have far‐reaching effects at both the global and regional scale, but local effects are difficult to determine from coarse‐resolution climate studies. Dynamical downscaling can provide insight into future climate projections on local scales. Here, we present a new dynamically downscaled dataset for Indiana and the surrounding regions. Output from the Community Earth System Model (CESM) version 1 is downscaled using the Weather Research and Forecasting model (WRF). Simulations are run with a 24‐hr reinitialization strategy and a 12‐hr spin‐up window. WRF output is bias corrected to the National Centers for Environmental Protection/National Center for Atmospheric Research 40‐year Reanalysis project (NCEP) using a modified quantile mapping method. Bias‐corrected 2‐m air temperature and accumulated precipitation are the initial focus, with additional variables planned for future releases. Regional climate change signals agree well with larger global studies, and local fine‐scaled features are visible in the resulting dataset, such as urban heat islands, frontal passages, and orographic temperature gradients. This high‐resolution climate dataset could be used for down‐stream applications focused on impacts across the domain, such as urban planning, energy usage, water resources, agriculture and public health.


| INTRODUCTION
Global climate studies are important for understanding large-scale patterns and trends in the climate record.A commonly used type of tool, Global Climate Models (GCMs), can provide future projections of climate change under a variety of scenarios, as well as hindcasts to understand different contributions to past climate change (Flato et al., 2013).GCMs, which often have a spatial resolution between 50 and 400 km, are commonly employed in studies of large-scale climate responses (Taylor et al., 2012).At this resolution, many atmospheric processes, such as clouds and convective precipitation, are highly parameterized.These parameterizations are designed to represent processes that occur across the globe, and therefore, will not be optimized for a specific region.
Additionally, anthropogenic climate change is expected to affect different locations in drastically different ways, requiring specific regional studies to uncover local trends and make conclusions about local impacts that are otherwise lost in broader studies (e.g., Zhai et al., 2005).The results of these studies can then be applied in downstream applications to explore specific local, regional, or state-level effects.For example, climate change impacts on infrastructure within the state of Indiana are currently being studied, including green infrastructure and energy supply and demand (Raymond et al., 2020;Reynolds et al., 2020;Wachs & Singh, 2020).These regional climate studies may also shed light on important impacts on ecological systems (Höök et al., 2020;Phillips et al., 2020) and industries such as agriculture (Bowling et al., 2020) and public health (Filippelli et al., 2020).
Such regionally focused impact studies often require climate data at finer resolutions than those generated by GCMs.Those data are created from GCM output by downscaling.Downscaling methods are broadly categorized into either statistical or dynamical approaches, each with their own strengths and weaknesses.Statistical downscaling involves deriving empirical historical relationships between large and small scales to predict local-scale weather features based on future large-scale atmospheric conditions (Ekström et al., 2015).Statistical downscaling often draws upon an ensemble of GCMs for large-scale information, thereby producing a range of projections (and therefore uncertainties), but the methods usually assume that the historical empirical relationships persist in a future climate.Additionally, since statistical downscaling is often performed on a single variable point-by-point, results within the final product can lose spatial coherence or physical consistency, although multi-variate biascorrection techniques have been developed to address this issue (e.g.Cannon, 2018).
Dynamical downscaling involves using a Regional Climate Model (RCM) that ingests GCM output and computes local weather features at a finer resolution.Unsurprisingly, dynamical downscaling is computationally demanding, requires additional post-processing to correct RCM biases, and often uses input from a single source (like a GCM), which can increase uncertainty due to model bias (Ekström et al., 2015).However, dynamical downscaling is often able to resolve local scale weather features missing or smeared out by GCMs, such as fronts and extreme weather systems, as well as to resolve processes that are parameterized in GCMs, such as cloud and precipitation formation (Mearns et al., 2003) and atmospheric deep convection (Gutowski et al., 2020), and to represent still unresolved processes with parameterizations tailored to the region being studied.Furthermore, dynamically downscaled data are coherent and consistent by construction.Because some of the main concerns of future climate change in Indiana are flooding and extreme events -events which can benefit from a continuous and physically consistent dataset -we decided that dynamical downscaling was the appropriate method for our research.
Dynamical downscaling requires numerous decisions to be made when designing simulations.Many past studies have explored the impacts of these different approaches and identified limitations of dynamical downscaling.Lo et al. (2008) found that RCM simulations performed better when reinitialized weekly rather than monthly or not at all.Additionally, Lo et al. (2008) determined that 3-D nudging techniques were able to reproduce regional-scaled patterns that were not resolved by the usual boundary-condition updating methods.Liu et al. (2012) varied the downscaling ratio (the resolution of the largest domain compared to the resolution of the next largest domain) and found that the often-used 1:3 ratio underperformed at reproducing some storm systems when compared to 1:5 and 1:7 ratios.Other studies found that a single RCM poorly represented lake temperatures, especially when examining ice cover and temperature gradients, and they, therefore, coupled additional models to the RCM to overcome these limitations (Gula & Peltier, 2012;Hostetler et al., 1993;Mallard et al., 2014).Additionally, multimodel downscaling efforts, such as the Coordinated Regional Downscaling Experiment (CORDEX; Giorgi et al., 2008), have offered common downscaling protocols and frameworks as well as common domains to facilitate ensemble comparisons (Giorgi & Gutowski, 2015).The sheer variety of these different studies highlights the multitude of ways that downscaling methods can be applied, and demonstrates that downscaling approaches are not one-size-fits-all.
While research into the specific limitations of dynamical downscaling continues, the value added by an RCM to a coarser GCM is well-established.The Weather Research and Forecasting (WRF; Skamarock et al., 2008) RCM is well-suited for both forecast and research purposes.Soares et al. (2012), Caldwell et al. (2009), and Heikkilä et al. (2011) each used WRF to improve the local-scale representation of meteorological output from coarser data.Soares et al. (2012) improved spatial representation of precipitation -especially precipitation extremes -in Portugal by capturing orographic features that the European Centre for Medium-Range Weather Forecasts (ECMWF) Interim Reanalysis (ERA-Interim) dataset did not resolve.Caldwell et al. (2009) was able to better represent spatial distributions of surface temperatures and resolve precipitation extremes that the Community Climate System Model version 3 (CCSM3) could not.Similarly, Heikkilä et al. (2011) improved representation of precipitation frequency and extremes over the complex topography of Norway by downscaling the 40-year ECMWF reanalysis (ERA-40).Conversely, Racherla et al. (2012) found that WRF showed no improvement in representing seasonally and regionally averaged historical temperature and precipitation when downscaling the Goddard Institute for Space Studies (GISS) ModelE2 over the continental United States.These contrasting results reinforce the need for a tailored approach to dynamical downscaling.
Indiana has been included within the domain of other dynamical downscaling studies (e.g., Gao et al., 2012;Gula & Peltier, 2012;Mallard et al., 2014), but few downscaling projects have focused on the future climate of the state.Gula and Peltier (2012) and Mallard et al. (2014) both focused on exploring the representation of the Great Lakes' temperatures.Prein et al. (2017) dynamically downscaled ERA-Interim across the continental United States and focused on the effect of rising temperatures on precipitation extremes.Gao et al. (2012) included Indiana in their domain, which covered the eastern United States, and downscaled the Community Earth System Model (CESM), using WRF, to a horizontal resolution of 4 km.Along with a marginal increase in resolution (3 km vs. 4 km), our downscaling efforts use a much new version of WRF (3.9.1 vs. 3.2.1)which includes numerous important updates, such as those to cumulus, microphysics, boundary layer, and ocean physics suites (NCAR, 2021b).Hamlet et al. (2019) performed statistical downscaling for Indiana based on an ensemble of models from Phase 5 of the Coupled Model Intercomparison Project (CMIP5; Taylor et al., 2012).Additionally, Schoof and Pryor (2001) used Indianapolis as the focus of a comparison between different statistical downscaling methods.While Hamlet et al. (2019) and Schoof and Pryor (2001) provide valuable information about potential changes in Indiana's future climate, they are limited in scope to the geographic borders of the state.This spatial limitation restricts the ability to perform studies that include surrounding regions, such as hydrology studies of the Wabash River basin that extends beyond the borders of Indiana.Moreover, none of the past Indiana-focused studies has conducted convectionpermitting downscaling, thus limiting their usefulness in simulating some of the most impactful events, such as extreme precipitation and flooding.
Here, we introduce and describe a new dynamically downscaled dataset covering Indiana and some of the surrounding regions at a 3 km horizontal resolution.The current dataset includes a historical period from 1951 to 2005 and a no-policy scenario (RCP8.5)from 2006-2100, although additional scenarios (RCP6.0 and RCP4.5) are planned to be released in the future.The rest of this article is arranged as follows.Section 2 details the data and models used for the downscaling simulations, as well as the bias-correction method adapted specifically for this dataset.Section 3 provides examples of value added compared to the driving GCM as well as a high-level overview of the climatic results seen in the dataset.

| DATA AND METHODS
The dataset featured in this project was produced through two major efforts: 1. Dynamical downscaling of CESM using WRF. 2. Bias correcting the WRF output against reanalysis data.Sections 2.1.1 and 2.1.2describe the GCM and RCM, respectively, used in the downscaling simulations.Section 2.2 describes the bias-correction method that is developed and applied to the WRF simulation output.Currently, only 2-m air temperature and accumulated precipitation are presented for the historical period and RCP8.5.We are currently planning future efforts to add additional data variables, such as minimum and maximum temperature, humidity, and wind, as well as additional future scenarios (RCPs 6.0 and 4.5).

| Global climate models
We initialize downscaling with simulations from version 1 of NCAR's CESM (Hurrell et al., 2013), used to support CMIP5 (Taylor et al., 2012).We choose CESM because it is a well-characterized, widely used state-of-the-science climate model that falls well within the range of climates simulated by other models participating in CMIP5.Rather than process CESM output directly, we use a pre-processed dataset (Monaghan et al., 2014); this dataset (freely available from NCAR) was designed for dynamical downscaling using WRF and has the necessary variables to initialize the WRF atmosphere and land models (i.e., air temperature, boundary layer winds, geopotential height, humidity, sea ice concentration, sea level pressure, sea surface temperature, skin temperature, snow water equivalent, soil moisture/water content, soil temperature, surface pressure, surface winds, upper air temperature, and upper level winds) in the Intermediate File Format specific to WRF (Monaghan et al., 2014).Additionally, the data have already been bias-corrected following the method from Bruyere et al. (2015) against the ERA-Interim reanalysis to remove the mean bias from the state fields in the CESM output.The dataset has a spatial resolution of approximately 1° and provides six-hourly output.The dataset includes a 20th century simulation from 1951 to 2005, as well as three Representative Concentration Pathway (RCP) scenarios from 2006 to 2100 (Monaghan et al., 2014).The three scenarios, RCP4.5, RCP6.0 and RCP8.5, represent middle-of-the-road climate change, low mitigation and no-policy (often called "business-as-usual") future scenarios, respectively.

| Regional climate model
To downscale output from the CESM dataset, we configure WRF 3.9.1 using the NCAR convection permitting suite, which has been used extensively for real-time forecasting of convective weather across the Continental United States (NCAR, 2021a).We run the model with two nested domains, both centered on 39.872° N and 86.25° W -close to the center of Indiana.The outer domain has a resolution of 9 km, and the inner domain a resolution of 3 km.The inner domain measures approximately 1,056 km on a side and the outer domain extends an additional 462 km in all directions (Figure 1).
There are two main dynamical downscaling approaches: continuous downscaling and reinitialization downscaling.In a continuous scheme, RCM simulations are initialized once from the driving dataset and then run for the entire length of the study in a single simulation.The boundary conditions are updated as frequently as allowed by the driving dataset, but the features within the simulation domain are allowed to develop and evolve freely.Conversely, reinitialization downscaling consists of many short-term simulation runs that are reinitialized from the driving dataset across the entire simulation domain.Numerous studies (e.g., Lo et al., 2008;Lucas-Picher et al., 2013;Qian et al., 2003) have compared reinitialization downscaling to continuous downscaling and found that reinitialization methods reduce the systematic error that can be generated from longer continuous runs.Additionally, shorter simulation lengths tend to result in lower systematic error when compared to observations (Lo et al., 2008;Qian et al., 2003).Therefore, we choose to employ the reinitialization downscaling approach with a short simulation length.
We initialize the WRF simulations from CESM output every 12 hr and update boundary conditions every 6 hr.WRF outputs simulation data every 3 hr.We run each simulation for a total of 24 hr, with the first 12 hr of simulation time excluded from the processed dataset.We test different spin-up lengths at different times of year and compare WRF output to interpolated CESM output for the same simulation time.WRF and CESM mean and median 2-m air temperature differences show small differences between the spin-up times (Table 1).Additionally, we find no appreciable difference in spatial patterns between different spin-up lengths (Figure 2).Thus, we move forward with 12 hr of spin-up time in order to decrease computational requirements.

| Bias correction
The existence of systematic biases in RCMs and their downscaled output has been well documented (Christensen et al., 2008).There are various biascorrection methods, such as linear scaling, local intensity scaling, delta-change methods, distribution derived transformations, parametric transformations and nonparametric transformations (Gudmundsson et al., 2012;Teutschbein & Seibert, 2012), that can be used to adjust RCM output.Many studies (e.g., Berg et al., 2012;Gudmundsson et al., 2012;Teutschbein & Seibert, 2012) comparing bias-correction methods have found that non-parametric quantile mapping performs well at reducing a wide range of systematic biases, although Berg et al. (2012) notes that no single method outperforms all other methods in every test.We apply quantile-quantile mapping, similar to those described by Wood et al. (2004) and Boé et al. (2007).
We bias-correct WRF simulation output against output from the NCEP/NCAR 40-year Reanalysis project (NCEP; Kalnay et al., 1996) provided by NOAA/OAR/ ESRL PSL, Boulder, CO, USA and freely available from their website (http://psl.noaa.gov/).NCEP reanalysis data are available at 1.9° resolution every 6 hr (00Z, 06Z, 12Z and 18Z).We recognize that different reanalysis sets will yield different results, and we choose to use NCEP as the initial reanalysis for bias correction because its temporal range includes a long historical period wellsuited for the downscaling simulations as well as for its sub-daily temporal scale.We apply bias-correction methods in order to correct systematic differences in the empirical probability distributions of the WRF simulations, as compared to those of the NCEP reanalysis data.For example, WRF consistently overestimates June 2-m air temperature at 00Z and 18Z and underestimates at 12Z (See Figure A1).In order to correct for these biases, the climatological distributions from the WRF simulations during the historical period were compared to those of the NCEP reanalysis.This correction occurs in two parts described in more detail below: first we calculate the bias, and then we apply the bias correction (Figure 3).
We develop and employ a modified quantile-quantile mapping technique to perform the bias correction.In this method, we compare quantiles of the WRF simulations to the same quantile of the NCEP reanalysis (Figure 4a,e).In this way, we compare high temperatures to high temperatures, low temperatures to low temperatures, and so on.We calculate biases separately for each time step in each month (e.g., January 00Z biases were calculated independently from January 06Z biases).To calculate the biases, we first bin NCEP reanalysis and WRF simulation data into quantiles for each gridpoint.We initially applied an average-based bias correction for each quantile directly to WRF values, but this method allowed extreme outliers in either dataset to strongly influence the highest or lowest quantiles, creating "gaps" in the climatological distribution (Figure 4b,f).We also see these effects from extreme values when using the mean of the values in each quantile to calculate bias (Figure 4c,g).Therefore, we use the median of the observations in each bin (hereafter, "quantile median") to  reduce the impact of extreme outliers in the highest or lowest quantiles and to retain continuity in the climatological distribution (Figure 4d,h).
For each quantile, we then apply a bilinear interpolation to the NCEP quantile median values from the coarse NCEP grid to the fine-resolution WRF grid.Because interpolation can often result in the loss of extreme values, we first calculate the quantile medians in order to retain the effects of observed extreme events.We then bilinearly interpolate directly to the quantile medians of each bin.We calculate the biases at each grid point for each quantile by taking the difference between the interpolated NCEP quantile medians and the WRF quantile medians such that for a given time step t (e.g., January 00Z), and a given quantile q, the bias B is where over-tilde terms are the quantile medians, as described above.In order to calculate biases for 'intermediate'time steps with no matching NCEP reanalysis observations (03Z, 09Z, 15Z, and 21Z), we take the arithmetic mean of the biases from the previous and next time steps (e.g.B 03Z,q = B 00Z,q + B 06Z,q 2 for each quantile q).Because we group data within discrete quantile bins, a choice must be made about the number of bins used.We investigate the effect of the number of bins on the quality of the bias correction (Figure A2) and use the Kolmogorov-Smirnov test to qualitatively measure the quality of the bias correction.The number of bins has no discernible effect on the bias correction within the range that we selected, and the KS test shows that the biascorrected distributions are statistically indistinguishable from the NCEP distributions for all bias-correction methods.We therefore choose 40 quantile bins in the interest of balancing downscaling quality and computational speed.
We then apply the bias correction separately to the historical and future WRF simulation output, first binning WRF historical and future simulation data into separate bins, as detailed above.We next apply the corresponding bias B to the simulated data such that for any value X at a time step t and a quantile q, the bias-corrected value X ′ is We apply the bias-correction method to temperature and precipitation data separately.Temperature was processed and bias corrected as described above, with no changes needed to the method.On the other hand, precipitation spans a wide range of values and is highly skewed, so the bias-correction method described above could not simultaneously span the range of precipitation values and resolve the dry-end of the precipitation histogram.Thus, we first calculate the logarithm of the positive precipitation amounts before applying the bias correction.In order to maintain the number of rain-days simulated by WRF, we did not bias correct for dry-day frequency.
However, when bias correcting the rain-days, we found that WRF has a banded drizzle effect (See Figure 5) which, in turn, caused our bias correction to result in banded precipitation totals.Upon further examination, this banded (1) B t,q = X NCEP,t,q − X WRF,t,q (2) X � = X + B t,q .
F I G U R E 4 Histograms of NCEP (red) and WRF (blue) 2-m air temperature resulting from different bias-correction methods.Histograms show (a,e) uncorrected WRF data, (b,f) bias-corrected WRF data using direct quantile comparison, (c,g) mean of quantile edges, and (d,h) median of values between the quantile edges (the method used here).WRF data shown for (a-d) historical time period , and (e-h) future time period .All the data are for August at 12Z.Note that Direct Quantile Comparison and Mean Method (b,c and f,g, respectively) create gaps in the tails of the corrected WRF distributions (in blue) due to outliers in the NCEP record.effect from WRF appeared to be due to floating point rounding errors when WRF writes output, whereby zero values were represented by values on the order of 1 × 10 −15 mm.Rather than attempt to troubleshoot a minor WRF bug, we restrict precipitation data to totals greater than 1 × 10 −6 mm, treating any smaller values as if they were zero.This restriction is implemented before creating and applying the bias correction, which side-steps the issue.
On the opposite end of the distribution, extreme precipitation projections that are outside of the range of historical observations are corrected using the bias correction from the highest quantile.These implausibly high extreme events occur in the historical as well as the future period and appear to be an artefact of WRF rather than an effect of the climate change signal.Therefore, no extrapolation is applied to the bias correction for extremely high precipitation totals, nor is any limit set to cap extreme precipitation.

| RESULTS
The downscaled final dataset (see Figure 6), as expected, shows notable improvement in fine-scale feature representation over the CESM input data.Figure 7 shows a direct comparison between the 2-m air temperature for the evening of 3 May 2025 from CESM output data and WRF simulation output after 0 hr (when CESM data has been consumed and interpolated, but no simulation has been run) and after 12 hr of runtime.Note that a variety of small-scale meteorological features are visible in the WRF output (Figure 7c), such as frontal activity (northwestern corner of the domain), urban heat islands (e.g., Indianapolis in the center of Indiana), and orographic temperature changes (southeastern corner of the domain).Additionally, temperature over water is seen to be lower than the surrounding land, as is expected in early May.This is particularly apparent over the Great Lakes, but also visible over smaller lakes, such as Kentucky Lake near the southwestern corner of Kentucky.
To further demonstrate the simulation of frontal systems, Figure 8 shows a cold front in early June moving across the domain from 21Z to 09Z the following day.Note the strength of the temperature gradient, which at times reaches 8°C in the span of 9 km.This gradient (approximately 1°C km −1 ), could never be resolved on a typical GCM resolution of 100 km, as this would result in a temperature difference of nearly 100°C between adjacent grid cells.Figure 8 also shows the precipitation that is generated by the passage of the front.Note the banded structure of the precipitation, as well as the correlation between the placement of the precipitation in relation to the temperature gradient, with the precipitation following behind the cold front.Additionally, WRF is able to resolve fine-scale, heavy precipitation totals that would be impossible to simulate through a GCM (See Figure A3).Representation of such extreme precipitation events can be valuable for impact studies, especially those focusing on flooding and water availability.
Analysis of the bias corrected downscaled product also shows climate changes that are consistent with other climate change projections (Hamlet et al., 2019;Hurrell et al., 2013).Comparing average temperature between three 20-year periods from the beginning (1951-1970), middle (2021-2040), and end (2081-2100) of the simulation period (Figure 9), annual mean temperature is seen to increase by 2°C across the domain by the middle of the simulation period.March-April-May (MAM) and September-October-November (SON) are seen to have F I G U R E 5 Percent of time steps in a single year with precipitation totals (a) greater than 1 × 10 −6 mm and (b) less than 1 × 10 −6 mm but greater than 0 mm.slightly lower temperature increases (1.6°C) in the first half of the simulation period, while December-January-February (DJF) sees the largest temperature increase (2.7°C) up until the middle 20-year period (See Figure A4 for seasonal temperature changes).June-July-August (JJA) sees a temperature increase of 2.1°C, nearly that of the annual mean temperature increase.
Temperature increases accelerate later into the century, with annual mean temperature increasing an additional 3.6°C between 2050 and the end of the century.MAM, JJA and DJF all see slightly smaller temperature increases (3.3°C-3.4°C) in the second half of the simulation period.SON sees the largest growth in temperature increase with a warming of 4.2°C during the second half of the century.Overall, annual mean temperature increases by 5.6°C from 1951-1970 to 2081-2100, which is consistent with the driving dataset (Hurrell et al., 2013).The largest total seasonal warming is seen in DJF (6.1°C) and the smallest seasonal temperature increase is found in MAM (5.0°C; See Table 2).These results are comparable to those found by Hamlet et al. (2019), who found a similar annual mean warming of 5.6°C across Indiana.Seasonal results show mixed agreement; MAM shows the least amount of warming in both studies, but Hamlet et al. (2019) shows the largest warming in JJA rather than in DJF.
These projected temperature changes fit well within the existing climate change projections.The IPCC reports a similar temperature increase across the domain of our study by the end of the century under RCP8.5 simulations (Romero-Lankao et al., 2014).This annual mean warming varies from 4°C in the southeast to 6°C in the north of our domain, which correlates well with the spatial patterns seen in our results (Figure 9).The National Climate Assessment shows similar temperature increases and spatial patterns (Hayhoe et al., 2018).Additionally, multimodel ensemble comparisons from CMIP5 note that the greatest projected warming is seen at higher latitudes in the wintertime (Maloney et al., 2014), which is well represented in our results (See Figure A4).
While temperature changes throughout the downscaled dataset are relatively uniform and monotonic, precipitation changes vary, both seasonally and spatially (Figure 10).Regionally averaged annual precipitation sees a slight decrease in the first half of the simulation, but this trend is reversed in the second half of the simulation period to result in no net change in annual precipitation through the entire simulation period.MAM and DJF both exhibit a general increase in the total precipitation, with MAM showing a stronger increase in the first half of the simulation period than the second, while DJF exhibits a fairly consistent change between all three periods.JJA shows a very strong drying trend in the first half of the simulation, with a much smaller precipitation decrease towards the end of the century.SON exhibits similar trends to JJA, although the precipitation decreases in the first half of the simulation are much smaller than seen in JJA.Table 3 shows the spatially averaged precipitation totals and changes for the different time periods.
While the precipitation results presented here do not completely align with other projections, they do not contradict past studies either.Many large-scale studies (e.g., Hayhoe et al., 2018;Maloney et al., 2014;Romero-Lankao et al., 2014) predict a general increase in annual total precipitation across the midwestern U.S., which is not seen in our results.However, we note that precipitation projections are less straightforward than temperature changes and that not all models agree on the sign of future precipitation changes.
While our dataset does not show an increase in annual total precipitation, seasonal trends show agreement with other studies.For example, seasonal projections from Hamlet et al. (2019) agree on the sign of the change with our results.Additionally, both Hamlet et al. (2019) and Hayhoe et al. (2018) project the strongest increases in DJF and MAM, while changes in JJA and SON show large levels of uncertainty.Indeed, an assessment of the CMIP5 results shows substantial model disagreement on the sign of the expected precipitation change in JJA and SON (Maloney et al., 2014).
Despite being focused on Indiana, the results from Hamlet et al. (2019) still do not provide a direct comparison to the results presented here, as the two studies use different domains and timescales.A qualitative analysis shows strong agreement between the two datasets.Both studies project an increase in DJF and MAM precipitation and a general drying in JJA and SON.The seasonal trends from Hamlet et al. (2019) are all slightly wetter than the results presented above, which accounts for the discrepancy in annual precipitation trends.Additional comparisons between our dataset and Hamlet et al. (2019) are planned in order to investigate the differences between statistical and dynamical downscaling, and perhaps resolve the annual precipitation total discrepancies.
While regionally averaged precipitation trends displayed in Table 3 show a general drying in JJA and SON along with an increase in total precipitation during MAM and DJF, spatial averages mask important geographic variability in our precipitation projections.Annual precipitation averages across the entire region decrease and then return to historical values, but specific areas within the domain show different patterns.For example, Illinois, Indiana, Michigan and Ohio are shown to have a general drying through the entire simulation period, while orographic regions in the southeastern region of the domain see an increase in precipitation.Similarly, SON shows a general decrease in precipitation, but most areas of Wisconsin, Kentucky, and Tennessee see an increase in SON precipitation.With such spatial variations seen in   high level precipitation metrics, future research is needed to thoroughly examine expected changes in rainfall and water availability across the domain.

| CONCLUSION
Here, we have presented a new dynamically downscaled dataset for the state of Indiana and the surrounding regions.In addition to the downscaled temperature and precipitation data for RCP8.5, we have generated a workflow that can be applied to other future scenarios or other variables.We have presented bias corrected temperature and precipitation data, as these variables are most commonly found in other studies and provide the most comparisons.However, one of the advantages of dynamical downscaling is that the regional model outputs hundreds of variables, all of which are internally consistent.Pending an appropriate bias correction, these too can be explored to gain deeper insights into future regional changes at fine scales.Future efforts are planned to explore additional variables and climate indices, such as humidity, wind, daily maximum and minimum temperatures, water availability, and potential evapotranspiration.All of these data can be applied to downstream applications focusing on the impacts of climate change, such as water resources (Dierauer & Zhu, 2020), urban energy demand (Raymond et al., 2020;Wachs & Singh, 2020), agriculture (Bowling et al., 2020) and public health (Filippelli et al., 2020).While this new dataset is promising, additional work is necessary to fully characterize the potential of this dataset.Hamlet et al. (2019) provided a statistically downscaled dataset over a similar region to that described here; future work could involve a comparison between the two to understand the added value provided by (and potential limitations of) dynamical downscaling, particularly for precipitation where statistically based products typically oversmooth spatial fields and underestimate extremes (Ensor & Robeson, 2008).Another point of uncertainty in our dataset is that it was created by bias correcting WRF output to the NCAR/NCEP Reanalysis.Since the output of our driving GCM was corrected to ERA-Interim, a part of our bias correction is working to "correct away" the differences between the ERA-Interim and NCAR/NCEP reanalyses.Different datasets, such as different reanalyses or syntheses of observations (e.g.PRISM; Daly et al., 2008) may yield slightly different bias corrections, which is a form of uncertainty in our results.Although it is beyond the scope of our study and dataset, in principle one could drive the regional model with boundary conditions different from those provided by CESM; our workflow establishes this possibility.
One of the key advantages of dynamical downscaling is its ability to better resolve extreme events than statistical downscaling.Our dataset enables research into these events, such as extreme precipitation (weather leading to floods or droughts) or heat waves.Understanding the changing probabilities of these events in the future is of critical importance to citizens and decision makers.Although there is no one-size-fits-all method for dynamical downscaling, by providing a documented workflow, we have enabled the possibility for others to adapt and test our modelling setup in other locations.

F
Weather Research and Forecasting model simulation domains employed for this study.The outer domain (blue) has a 9 km resolution.The inner domain (red) has a 3 km resolution.T A B L E 1 Mean and median absolute 2-m air temperature difference between interpolated CESM input and WRF output with different spin-up lengths.

F
Difference between WRF output after 12, 24, 48 and 72 hr of runtime and CESM output interpolated to WRF grid for representative timesteps in January, April, July and October.No appreciable difference is seen between the different runtime lengths within each month, showing that a 12-hr spin-up length is effective.FI G U R E 3 Schematic of the steps involved in the bias-correction process.

F
I G U R E 6 Domain-averaged 2-m temperature (top) and precipitation (bottom) from WRF over the entire simulated time series.Values shown are for annual averaged (ANN) and averaged over each season (MAM, JJA, SON, DJF).F I G U R E 7 Example of fine-scale features added by WRF.Comparison of (a) CCSM 2-m temperature, (b) WRF 2-m temperature after 0 hr of simulation, and (c) WRF 2-m temperature after 12 hr of simulation.All plots show model output at May 3, 2025 at 00Z.

F
Weather Research and Forecasting model simulation output showing passage of a cold front, with associated 2-m temperature change and accumulated precipitation amounts over a 15-hr period from June 2nd, 2025 at 21Z to June 3rd, 2025 at 09Z.

T A B L E 2
Seasonal and annual mean 2-m temperatures over the spatial domain shown in Figure9for three time periods(1951-1970, 2021-2040 and 2081-2100), as well as the amount of warming seen between the first two periods (Early warming), the last two periods (Late warming), and between the first and the last period (Total warming).F I G U R 1 0 Precipitation means from bias corrected WRF output.Annual precipitation means for (a) 1951 to 1970, (b) 2021 to 2040, and (c) 2081 to 2100.(d-o) As in (a-c), but seasonal precipitation means for (d-f) spring, (g-i), summer, (j-l) autumn, and (m-o) winter.
Averages of seasonal and annual precipitation totals over the spatial domain shown in Figure9for three time periods(1951-1970, 2021-2040  and 2081-2100), as well as the amount of precipitation change between the first two periods (Early precipitation change), the last two periods (Late precipitation change), and between the first and the last period (Total precipitation change).F I G U RA 3 Examples of fine-heavy precipitation totals from WRF simulation output.3-hr accumulated precipitation totals from (a) 7 August 2006 at 03Z, (b) 25 July 2049 at 21Z, (c) 14 October 2079 at 21Z, and (d) 24 June 2098 at 06Z.

Spin-up time (hr) Mean difference Median difference
Values are calculated by comparing every time step in 12-hr intervals for an entire year of simulations. Note: