5. Future Work
- Some evidence suggests the UGA-WRF configuration handles other southeastern CAD events with similar success, but the next step for this research is more statistical verification across a greater sample of CAD cases using a greater number of variables and vertical levels.
- Test simulations are also planned for the UGA-WRF configuration that look to further optimize and find the best physics schemes and dynamical settings for southeastern CAD events.
- Additionally, a UGA configuration of the MPAS-A model is also in development and early operational testing, which could serve as a test to determine if these CAD-retention strategies hold true across unstructured grids and newer dynamical cores.
Success in Simulating CAD: High-Resolution UGA-WRF Simulations of the January 23-25th Southeast U.S. Ice Storm
Scout M.B. Carlson 1
John A. Knox 2
1 ATSC Undergraduate, University of Georgia, scout.carlson@uga.edu, 2Josiah Meigs Distinguished Teaching Professor, Department of Geography, University of Georgia, johnknox@uga.edu
1. Introduction
The CAD Forecast Challenge
- Cold Air Damming (CAD) is a phenomenon that occurs when a shallow area of dense, relatively colder air becomes trapped or “wedged” against the lee-side of a mountain range, funneling it down to the south (in the NH). CAD is an important weather phenomenon for locations east of the Appalachian Mountains most of the year.
- Operational mesoscale models like the HRRR and NAM along with global models like the GFS struggle to maintain CAD. Frequently, the models will both underestimate the strength and duration of the wedge, mixing it out prematurely.
- This premature surface warming from models leads to frequent forecast headaches and large errors regarding temperatures, precipitation location and amount, and – during winter weather – precipitation type.
- This project is a case study of the performance of the UGA-WRF weather model, a regional, customized, ultra-high-resolution configuration of the WRF-ARW model during the January 23-25, 2026 Southeast Ice Storm. This crippling ice storm was accompanied by the shallow cold air provided by a particularly strong CAD event across Northeast Georgia.
2. Model Setup
UGA-WRF Model Configuration & Domain Setup
- Grid Resolution: ~1.5km (highest resolution operational model over GA)
- Domain: Centered around Atlanta, slightly extending into the 5 neighboring states
- Physics Suite: Modified version of NCAR’s convective-permitting “CONUS” suite [defaults in brackets]:
- Boundary and Surface Layer: Mellor-Yamada-Janjic [Same as Default]
- Microphysics: Eta (Ferrier) [Thompson]
- Cumulus: Disabled
- Longwave/Shortwave Radiation: RRTM [RRTMG]
- Land Surface: NOAH LSM [Same as Default]
Methodology
- In order to isolate why the UGA-WRF tends to perform well at CAD forecasting, simulations were run using identical forcing data and initialization times across three different configurations:
- Operational Data: Standard UGA-WRF Configuration at 1.5km
- Control Comparison: Standard WRF CONUS suite at 1.5km
- Resolution Comparison: Standard UGA-WRF Configuration degraded to 3km
- All models were run out to 48 hours, with initial forcing data and boundary conditions from the HRRR model. Initialization times in this study are:
- The WRF configurations were then compared against data from the HRRR, NAM, and GFS models at the same initialization times.
3. Results
- Promising results were found when first comparing our custom configurations versus the NCAR-provided defaults in the CONUS physics suite and our configuration downgraded to 3km, as illustrated in the freezing rain forecasts above.
- Our configurations held on to the CAD slightly longer than the CONUS configuration and much longer than other operational models, which allowed for greater accuracy in precipitation type forecasting.
Fig. 2: 48-hour freezing rain totals from each configuration, initialized at 12z.
L-to-R: UGA-WRF 1.5k, UGA-WRF 3k, CONUS 1.5k
- As shown in Figure 4, the WRF configurations appear to be outperformed by the other operational models. However, as shown in Figure 3 and 5, the WRF configurations exhibited a colder bias and maintained the wedge better, whereas the GFS, NAM, and HRRR erroneously warmed.
- During the “peak CAD phase”, when the swath of subfreezing temperatures was at its greatest in Figure 6, the UGA-WRF outperformed all configurations except for the NAM 3k. However, while the NAM 3k possessed low RMSE, it had a pronounced warm bias in the CAD area, indicating that despite a low overall error, it struggled to retain the CAD properly.
- In Figure 3, this is very pronounced in the second comparison image, with the NAM running +2.7°F warmer than observed at CNI, +4.5°F at AHN, and +6.9°F at GSP. For comparison, the UGA-WRF only ran +1.1°F, -0.9°F, and +2.0°F respectively.
- It is important to note that the NAM and the UGA-WRF configurations share the same microphysics setup, Ferrier, while the HRRR uses the Thompson scheme.
Fig. 4: Domain-Wide RMSE for WRF configurations along with HRRR, NAM 3k, and GFS 0.25deg, initialized 1/24 12z.
Fig. 5: 2m Temp vs Obs for WRF configurations along with HRRR, NAM 3k, and GFS 0.25deg, initialized 1/24 12z.
Fig. 6: RMSE and Mean Bias for WRF configurations and operational models over peak CAD phase of 1/25 12z - 1/26 00z, initialized 1/24 00z and 1/24 12z.
4. Conclusion and Implications
- Consistently over this case, our higher resolution configuration prevailed over the degraded resolution configuration and provided slight improvements over the CONUS configuration.
- While the NAM 3k achieved the lowest overall error, the custom WRF configurations successfully outperformed the HRRR and GFS in terms of CAD strength and duration. The WRF configurations also outperformed the NAM in some of the deepest areas of the CAD where it underestimated the wedge and exhibited strong warm biases.
- From this case study, it would be reasonable to conclude that the Ferrier microphysics scheme used in both the UGA-WRF and NAM models potentially provides an edge over the Thompson microphysics scheme - the default used by the CONUS setup and the HRRR - in terms of CAD retention.
- It would also be reasonable to conclude that an increase in resolution correlates to slight improvements in CAD forecasting, presumably due to improvements in resolving the topography of the Appalachian Mountains. This is very apparent when comparing the failure of the GFS model to all the high-resolution models, whose coarse 0.25-degree grids cannot neatly resolve the complexities of the terrain.
- However, there do appear to be diminishing returns, as while improvements were noted over the UGA-WRF 1.5k vs 3k, they were slight and hint that at such fine scales (<3km), parameterizations and initial/boundary conditions matter more than increases in raw resolution.
- The tendency of national operational models to prematurely erode wedges creates significant blind spots during Southeastern winter storms. With local high resolution tuned models, supplementary guidance could be provided to mitigate public forecast failures.
Fig. 1: Total ice accumulation from NWS Peachtree City.
Special thanks to Joseph Cristea for aiding in providing some custom Python plotting code.
Fig. 3: Temperature forecasts from WRF configurations, NAM, HRRR, and GFS vs OBS/RTMA analysis.
Valid 1/25 04z (CAD onset) and 1/25 15z (CAD peak swath).