Development of a GSI-based Global Data Assimilation System for the Model for Prediction Across Scales

In numerical weather prediction (NWP) systems, data assimilation plays a crucial role. In determining the optimal weight between the background and observation error covariances, the ingestion of observations can have a positive impact on the analysis and initial conditions for the next assimilation cycle, not to mention the longer-term forecast.

For frameworks such as the Weather Research and Forecasting (WRF) model and NCEP’s Global Forecasting System (GFS), variational minimization using the Gridpoint Statistical Interpolation (GSI) assimilation software is performed on regular or Gaussian grids. For models which operate on unstructured meshes, such as the Model for Prediction Across Scales (MPAS), minimization with GSI becomes problematic, primarily due to the lack of background error covariance information (and its generation) for those types of meshes.

The Weather Company has chosen MPAS as its next-generation global model, which will serve as the eventual replacement for the current WRF-based forecasting system (known as “RPM”) and will be run at high resolution in a cycled assimilation framework with hourly updates. To date, MPAS has been running quasi-operationally at 15-km uniform resolution and initialized (cold start) from the 0.25-degree GFS analyses at 00Z, 06Z, 12Z, and 18Z each day, with the addition of the NASA SPoRT 2-km SST and 4-km VIIRS green vegetation fraction data. Work is also underway to process, QC, and assimilate cell-phone pressure observations, which IBM has been retrieving and archiving at one-minute intervals, totaling approximately 350 million per day.

In order to transition the GSI-based data assimilation techniques from regional WRF to global MPAS, the analysis variables on the 15-km unstructured mesh must first be reprojected to a Gaussian grid using interpolation weight files, which are generated via Earth System Modeling Framework (ESMF) regridding techniques. The ESMF functions take into account the specifications of both the unstructured mesh and Gaussian grid in order to determine the number of links between each, and output a file that is read by first-order interpolation routines (based on an NCAR utility). The GSI is then run in a similar manner to NCEP’s Global Data Assimilation System (GDAS), and the analysis increments (Ps, U, V, T, Qv, and density) are interpolated back to the MPAS mesh.

An overview of the MPAS and GSI implementations will be discussed, along with details of the background error covariance matrix and results from single-observation tests. The preliminary impacts of real observations, including conventional PrepBUFR and cell-phone pressures in a 3DVAR framework, will also be reviewed and discussed. Plans are in place for the enablement of an hourly-updating hybrid-EnVAR system, which will be outlined as well.