Progression of Data Assimilation for IBM GRAF: Towards a JEDI-Driven System

Data assimilation (DA) is an integral component of any numerical weather prediction system, as it determines the optimal weight between the background (B) and observation (R) error covariances, characterizing the uncertainty in the analysis. This reduces analysis error and brings the initial conditions closer to reality. IBM has been continuously developing its DA capabilities, beginning with regional WRF implementations at IBM Research and The Weather Company (TWC).

The IBM Global high-Resolution Atmospheric Forecasting (GRAF) system has extended these capabilities to the Model for Prediction Across Scales (MPAS), the underlying unstructured-mesh model for all current and future operations. To date, the Grid-point Statistical Interpolation software has been utilized to assimilate surface weather observations (operationally) and radiosondes (experimentally) in a partially cycled 3D-variational approach. The analysis increments are computed on a Gaussian grid and interpolated back to a native MPAS mesh at 15/4-km. A cloud analysis utility has also been developed, which ingests site-specific radar reflectivity and operates directly on the mesh, after the DA.

The static B matrix was derived from 180 retrospective cold-start forecasts across a continuous 3-month period, initialized at 00Z and 12Z from the 0.25-degree GFS, 2.5-km global SST analyses, and daily updated NESDIS 4-km green vegetation fraction data. The resolution and duration of each forecast was 15/3-km and 36 hours, respectively. The output at 12- and 36-hours was used as input to the NMC method, to compute the perturbations for regression on a Gaussian grid.

Moving forward, GRAF is employing more advanced DA techniques using the JCSDA’s Joint Effort for Data assimilation Integration (JEDI) software. TWC’s objective is to develop a rapidly updating, fully cycled 3D/4D-EnVar capability with JEDI, driven by an in-house MPAS ensemble (EnKF/LETKF) and higher-resolution deterministic forecast. A summary of the ongoing JEDI work will be discussed, including updates to the static B matrix, observation conversions and priorities, and recent experiments. The latest GRAF roadmap/workflow will be outlined as well.