Supercell Predictability: Exploring Ensemble Sensitivity to Initial Condition Spread

As convection-allowing ensembles are now routinely used to explicitly forecast the evolution of severe thunderstorms, developing an understanding of the limits of storm-scale predictability is warranted. The motivation for this study largely stems from the NOAA Warn-on-Forecast (WoF) program and the need for a suitable full-physics/real-data numerical weather prediction (NWP) framework for studying storm-scale predictability. Using the aforementioned framework, the sensitivity of ensemble forecasts of supercells to initial condition (IC) uncertainty is investigated. Three sets of ICs for our simulations were generated from the real-time NSSL Experimental WoF System for Ensembles (NEWS-e) during the 2016 NOAA Hazardous Weather Testbed Spring Forecasting Experiment. Simulations were initialized with developing thunderstorms and integrated for 3 h. The forecast sensitivity to IC uncertainty was assessed by repeating the simulations with the initial ensemble perturbations reduced to 50 %, 25 %, and 10% of their original magnitudes. Supercell features examined include mid- and low-level mesocyclone, updraft, downdraft, severe surface winds, and rainfall.

Forecast spread was substantially reduced with decreasing IC spread for all examined supercell features in all three cases. In cases where storm lifetime is only 3 - 4 h an intrinsic predictability limit (IPL; i.e., the lead time beyond which forecast uncertainty can no longer be reduced by decreasing IC uncertainty) may not exist. Upon reaching the IPL, the forecast spread can remain modest, allowing the practical predictability limit (i.e., the lead time beyond which the forecast becomes too uncertain to be useful) to exceed the IPL. Comparing to previous work, initializing post-convective initiation (CI) rather than pre-CI substantially improves predictability. Similar to tropical cyclones, storm location was far more predictable than a given feature’s intensity. The practical predictability of strong low-level rotation (a tornado potential proxy) increased from 30-90 min to 3 h when IC spread was reduced by 50%.

Impacts of IC uncertainty within vs. outside of the storm were also explored. Forecast spread for all supercell features in all three cases benefited greatly early on from reducing intra-storm ensemble perturbations to zero (e.g., storm certainty). Forecasts benefited more from eliminating uncertainty in the storm environment (e.g., environment certainty) later in the simulations, but the results were more case-dependent. Storm certainty reduced forecast location spread for lead times of up to 2-3 h. Low-level rotation was more sensitive to intra-storm perturbations rather than environment perturbations suggesting that tornado forecasts out to lead times of 1.5 – 2 h would benefit more from improving initial storm state certainty.