1. Supercell Predictability Studies in Support of NOAA Warn-on-Forecast
Corey K. Potvin1,2,3, Montgomery L. Flora3, Elisa M. Murillo3, and Louis J. Wicker2
1Cooperative Institute for Mesoscale Meteorological Studies, Norman, OK
2NOAA/OAR National Severe Storms Lab, Norman, OK
3School of Meteorology, University of Oklahoma, Norman, OK

2. Motivation: Warn-on-Forecast (WoF)
Goal: convection-allowing NWP ensembles that provide probabilistic guidance for thunderstorm hazard forecasts & warnings
Tornadoes, hail, wind, flash flooding
Challenge: real-time operation requires ensemble design compromises
Ensemble size, resolution, physics schemes
Question: how to optimize ensemble design?

3. Approach
Optimizing ensemble design requires knowledge of impacts from different forecast error sources
Construct sensitivity experiments that allow isolation and systematic exploration of errors
For some studies, combine idealized and real-data frameworks
Retain complete knowledge of “truth” but with full physics and observationally-constrained IC’s
Initial focus on supercells

4. Sensitivity to radar-to-storm geometry (Potvin and Wicker 2013, WAF)
Idealized OSSEs
3 TRUTH supercells
Assimilate pseudo-radar data with EnKF, then ensemble forecast
Radar #1 > 100 km away
Radar #2 repositioned to vary radar-storm distance, cross-beam angles (CBAs)

5. Excellent Cross-Beam Angles Poor Cross-Beam Angles
Poor radar-to-storm geometry does not unduly degrade low-level rotation forecasts
Excellent Cross-Beam Angles
Poor Cross-Beam Angles
Both radars > 100 km away
Radar #2 much closer
CBA = 70-90°
CBA = 20-30°
CBA = 0-30°
Neighborhood ensemble probability of strong low-level rotation; Red = TRUTH

6. Sensitivity to forecast grid spacing (Δx) Potvin and Flora (2015, MWR)
Idealized simulations with Δx = 333 m (TRUTH), 1-4 km
Range of environments (figure)
Compare to TRUTH and to TRUTH with scales < 2Δx filtered out
Identify errors arising from unresolved upscale effects

7. Time-height composites of ζ 2-h forecasts of 1-km AGL dBZ
WK82 Snd
4-km Δx too coarse
Delayed storm development
Premature demise
3-km Δx much better
Del City Snd
El Reno Snd
Time-height composites of ζ
TRUTH
Middle row: TRUTH, upscaled
Bottom row: coarse sims
2-h forecasts of 1-km AGL dBZ
1-km Δx needed to resolve low-level mesocyclone
Even finer Δx needed to forecasting its timing

8. Sensitivity to IC resolution Potvin et al. (2017, JAS)
Select member of a full-physics, 3-km EnKF analysis of a real supercell and downscale to Δx ≈ 300 m
Integrate for 2 h (CNTL) and compare to simulations with spatially filtered ICs
Cutoff wavelengths = 2, 4, 8, or 16 km
Add noise to IC’s to generate ensembles
Since deterministic framework insufficient to reveal systemic impact of IC resolution (and some other types of error; see extra slide)

9. Surprising insensitivity to IC resolution!
Probability-matched ensemble mean dBZ at t = 2 h, z = 2 km AGL

10. Largely since missing scales regenerate within 5-10 min of forecast
w spectra at t = 0 min
w spectra at t = 5 min
Similar results for other cases & variables
In organized convection, larger scales (intra-storm and environment) strongly determine evolution

11. Sensitivity to IC spread (Flora et al., in prep)
Evaluate practical & intrinsic predictability of supercells using feature-oriented framework
Downscale real-data 3-km ensemble analyses to 1-km, integrate for 3 h
Repeat with IC spread reduced to 50%, 25%, 10% of original spread
SEE POSTER 4 (session 2)
Left: spread in forecast-maximum updraft helicity (UH)
Right: neighborhood probability UH > 300 m2s-2

12. Conclusions
Typical (poor) radar-storm geometries do not preclude accurate supercell forecasts
Forecast ∆x must be ≤ 3 km to reliably predict general storm evolution
Major additional improvements require ∆x ≤ 1 km
However, reducing analysis ∆x below 3 km may be unnecessary
Finite computing resources are better spent on reducing forecast ∆x than analysis ∆x
Option 1: Perform coarse (e.g., 3-km) DA, then interpolate to fine forecast grid (e.g., 1 km)
Option 2: Mixed-resolution DA

13. Conclusions (cont.)
Intrinsic predictability limit < 3 h in some cases, but practical predictability limit longer
Improving supercell prediction at > 1-2 h lead times may require much better analyses of mesoscale environment
Real-world 1-km ensemble forecasts from 3-km EnKF analyses have severe biases (ongoing WoF work)  physics schemes remain a leading error source

14. Ongoing/Future Work Acknowledgments
Real-world 1-km EnKF DA and/or forecasts
If these improve upon 3-km, then explore dual-resolution EnKF methods
Extend to mesoscale convective systems
Acknowledgments
Warn-on-Forecast team: Thomas Jones, Kent Knopfmeier, Patrick Skinner, Dusty Wheatley, Nusrat Yussouf, Gerry Creager (and others)
National Weather Center REU program

15. Extra Slides

16. Why an ensemble framework?
If storm evolution is very sensitive to IC, then IC resolution impact could vary wildly between two nearly identical storms (or moment to moment in the same storm), making it impossible to use a single pair of sims to generalize the effect of IC resolution error
IC resolution error >> IC perturbation error  deterministic framework is sufficient
IC resolution error ≈ IC perturbation error  need ensemble framework!