Results showed that the environmental conditions including moisture, instability and convective inhibition were highly predictable at the spatiotemporal scale explored herein while a limited intrinsic predictability was observed at the storm scale such as the location, morphology and development of individual storms. It was found that initial errors that had a standard deviation of one magnitude smaller than current observational and analysis errors, although not being able to modify the environmental convective conditions prior of convection initiation (CI), produced large uncertainties at storm scale in terms of ensemble probabilities of simulated radar reflectivity, precipitation and updraft helicity. These uncertainties were not reduced even when the initial errors were cut down by as much as 90%, indicating that predictability of this thunderstorm event was inherently limited that improvement in initial conditions may not lead to expected improvement in forecast skills. Scale decomposition of errors showed that deep moist convections facilitated the growth of initial errors firstly in meso-γ scale and became saturated shortly after CI. Errors in meso-γ scale transferred upscale into meso-β scale and grew continuously until the entire convective system began to weaken. These error-growth mechanisms were consistent with previous studies in larger scales.
To understand how the simulated thunderstorms became distinct from each other in the ensemble, two specific simulations were further inspected. Results showed that during the CI stage, large variability occurred in the location and intensity of updrafts, and vertical wind shear (due to pre-existing convective rolls and surface outflow boundaries of surrounding storms), making it hard to predict whether initiated storms would be long-lived or not. For the long-lived thunderstorms, randomly shifted updraft maxima during early CI might lead to changes in both strength and extent of storm reflectivity and updrafts in tens of minutes. After CI, surface cold pools would form and influence the development and organization of storms with their interaction with the environment. More complex situations might occur when splitting and merging processes happened: the juxtaposition of two thunderstorms before their merging was found to be essential for the later development of the merger storm, leading to bifurcated development of maintenance versus dissipation of thunderstorms in the two simulations, even though their initial differences were almost negligible.