A Dime a Hundred: Cost Effective Large Ensembles through Probit-space Ensemble Size Expansion for Gaussian Copulas (PESE-GC)

Physics-based numerical weather prediction (NWP) models are computationally costly. This cost limits most NWP ensembles to less than 100 members, thus resulting in sampling errors in the ensembles’ statistics. These errors limit the performance of socioeconomically important applications of ensemble statistics (e.g., ensemble data assimilation (DA), probabilistic NWP and risk assessment). As such, cost-effective approaches to increase ensemble size are likely to improve the performance of those applications.

To improve those important applications, especially ensemble DA, this study proposes a novel cost-effective approach to construct large ensembles from a small number of NWP runs: Probit-space Ensemble Size Expansion for Gaussian Copulas (PESE-GC). PESE-GC constructs additional members by combining the user’s knowledge of univariate forecast statistics with a scalable covariance-preserving non-Gaussian multivariate resampling scheme. In fact, PESE-GC is so economical that merely 8 CPU core-hours is sufficient to construct 100 additional members (10⁷ state vector elements) from a 10-member Navy Global Environmental Model (NAVGEM) ensemble. Finally, PESE-GC can be extended to incorporate expertise on important diagnostic relationships (e.g., geostrophic balance). Future work will investigate PESE-GC’s impacts on the performance of the Ensemble Navy Aerosol Analysis Prediction System (ENAAPS) and NAVGEM.

In this talk, we will introduce and illustrate PESE-GC through applications with NAVGEM ensembles from June 2023 (a period of severe haze and severe weather). The realism and multivariate statistics of the additional members will also be examined.