Multi-scale Localization Using SABER Spectral Filters with MPAS

In data assimilation, background error covariances (encoded in the B-matrix) can be estimated on-the-fly from an ensemble of perturbed forecasts. Due to computational limitations, only a limited number of ensemble “members” can be produced and used to estimate a B-matrix. So, the resulting estimated B-matrix is prone to sampling noise, typically in the form of spurious covariances at large spatial separations. The application of a localization matrix through an element-wise Schur product can preserve covariances up to a certain length-scale while removing the generally spurious covariances beyond a chosen length-scale. However, background covariances can occur at a variety of scales, from convective (short) to synoptic (long), so the ability to localize ensemble covariances at multiple scales allows for an improved analysis. The UK Met Office originally implemented generic multi-scale localization with spectral filtering methods in the SABER (System Agnostic Background Error Representation) package for use in their global DA system. SABER’s generic covariance models are developed with the intention to allow for application into new systems with minimal effort. This presentation will show the progress towards applying the generic spectral filtering methods for multi-scale ensemble localization to the MPAS (Model Prediction Across Scales) model.