The Development of the GFS Hybrid 4DEnVar System with Multiresolution Ensembles

GSI-based 4DEnVar is further developed to have the capability to ingest ensemble at multi-resolutions. Due to the constraints of limited computational resources, ensemble background is run at a lower resolution in the operational GSI–based 4DEnVar in GFS system. In the GSI-based operational 4DEnVar update, analysis increments including both the flow dependent and static components are generated at a reduced low resolution (hereafter, Single Resolution-Low, SR-Low). In the mean time, current GSI-based 4DEnVar has a capability to generate the analysis increment at high resolution using high-resolution static background error covariance (BEC) and low-resolution ensemble BEC (hereafter, Dual Resolution, DR). Although these SR-Low and DR configurations can save computational costs, neither ensemble or static BEC contributes to the highly flow-dependent analysis increment at smaller scales (SR-Low); only static BEC contributes to this increment (DR). In addition, small-scale information in high resolution observations such as radar and satellite need to be thinned or superobbed to be consistent with the reduced resolution in BEC. To remedy these problems, we introduce a new term associated with high-resolution ensemble into the current GSI-based 4DEnVar cost function. Therefore, data assimilation can be with a mixture of low-resolution and high resolution ensemble BEC (hereafter, Multi Resolution Ensemble, MR-ENS). The new capability allows 4DEnVar to ingest ensembles from various sources at various resolutions.

The newly developed MR-ENS 4DEnVar system is examined during a 5-week period, 2400 UTC July – 0000 UTC 30 August 2013. Our results show the MR-ENS has better fit of 6hr background to the observations compared to the DR and SR-Low. Power spectrum of analysis increment shows that the MR-ENS analysis increment has larger power than the DR and SR-Low due to the contribution from the high-resolution ensemble BEC at small scales. The evaluation for forecast shows that the MR-ENS improves the forecast at short-to-mid lead times (6-72hr) compared to the DR and SR-Low. More detailed diagnostics and the theoretical description of the new method will be presented at the conference.