Single-Suite Stochasticity for Thunderstorms: Can It Beat a Mixed-Physics Suite?

In ensemble forecasting systems, an optimal distribution of members must approximate the true uncertainty of the atmosphere. In operational systems, suitable ensemble dispersion is approached through intra-member variation in the parameterisation suite; however, this yields an ensemble wherein each member is not equally as likely, while the multitude of physics suites may pose issues for the integration and stability of data-assimilation systems. To remedy this, others have proposed the introduction of ensemble perturbations via so-called stochastic methods: a sort of curated randomness. Despite many studies into the efficacy of stochastic perturbations in limited-area numerical weather prediction (NWP) models, there is less focus regarding stochasticity at the thunderstorm scale, specifically (a) the impact of such perturbations on the thunderstorm’s structure, and (b) development of novel schemes that specifically address a lack of diversity in state-of-the-art models. Further, many concerns remain in the scientific community regarding the validity of stochastic techniques at convective scales. Many of these concerns are addressed herein.

We document the impact—both meteorologically and statistically—of five stochastic perturbation schemes on 0–3-h supercell WRF simulations. These experiments (comprising both novel and established schemes) address a hypothesised ensemble overconfidence in, e.g., soil moisture, boundary-layer turbulence and variables, microphysical parameters, vegetative processes in the surface layer, etc. From the perspective of individual thunderstorms, results confirm the importance of constraining (1) the distribution from which to sample stochasticity, perhaps with observations, and (2) the magnitude of such a “shock” to the simulated atmospheric state, to preserve realism. Preliminary work suggests the combined impact of all stochastic schemes within a single-physics suite (SP-Stoch) results in generally higher skill scores and more reliable spread than in the control ensemble (no stochasticity). Surprisingly, SP-Stoch also outperformed a mixed-physics configuration (no stochasticity).

Herein, we present results from all permutations of the five stochastic schemes, and give meteorological examples of the impact of each scheme. The experiments use novel verification techniques to understand the impacts of each scheme in more detail. While stochastic perturbations are potentially a stop-gap measure to account for deficient representations of sub-scale processes, it is preferable to a mixed-physics suite for reliability and efficiency reasons. We anticipate the results will inform immediate NWP developers in designing the next iteration of convection-allowing-scale ensemble systems.