Impact of Ensemble Design in the Rapid Refresh Forecast System Using Initial Condition and Stochastic Physics Perturbations

As part of the transition to the Unified Forecast System (UFS), development has begun on the future convection-allowing Rapid Refresh Forecast System (RRFS) ensemble that uses a single dynamic core (dycore) and eventually a single physics suite. This system is intended to replace the multi-dycore, multi-physics High Resolution Ensemble Forecast (HREFv3), which is NOAA’s current operational convection-allowing ensemble forecast system. To account for model uncertainty, it is important that an ensemble system adequatedly sample the space of potential outcomes and thus provide forecast spread representative of the variability of the atmosphere. Although the HREFv3 provides enough spread, due to its multi-dycore and multi-physics nature its members tend to cluster by dycore/physics pairing, resulting in multi-modal statistics. An ensemble based on a single dycore and single physics suite can remedy this problem, but can lack sufficient spread. As part of a Developmental Testbed Center (DTC) project on ensemble design, we test configurations of time-lagging, neighborhood probabilities, initial condition (IC) perturbations, and stochastic physics perturbations to evaluate their potential to provide sufficient spread to represent forecast errors. This analysis focuses on the impacts of IC perturbations from the Global Ensemble Forecast System (GEFS) and stochastic physics including the Stochastically Perturbed Parameterizations (SPP), wherein the spread of ensemble members is assessed as a function of lead time. While IC perturbations drive spread from the initialization of the ensemble, stochastic physics perturbations maintain it later in the forecast. In an operationally-similar environment to the RRFS, we assess the impact of IC and stochastic physics perturbations on selected cases, including comparisons of variability between members within the experiment and the need to address magnitude changes to balance spread contributions from each technique. We focus on qualitative evaluations of storm mode, structure, intensity, and convective evolution, as well as quantitative assessments of ensemble-based and neighborhood probabilistic metrics using the enhanced Model Evaluation Tools (METplus) verification framework.