Neill Bowler, Met Office

The development of an ensemble of 4D-ensemble variational assimilations

Coauthors
A. M. Clayton, M. Jardak, E. Lee, A. C. Lorenc, C. Piccolo, S. R. Pring, M. A. Wlasak, D. M. Barker, G. W. Inverarity, R. Swinbank

Abstract:

A key aspect for a hybrid data assimilation system is to improve the quality of the flow-dependent covariance information it receives from an ensemble. A large change has been tested to the Met Office's ensemble prediction system - moving from an ETKF-based ensemble to one based on an ensemble of data assimilations. The data assimilation method used is four-dimensional ensemble-variational (4DEnVar) minimisation. This has the advantage that it is close to the operational hybrid 4DVar method used for data assimilation while being considerably cheaper - using this system it has been possible to test a 200 member ensemble.

The quality of the ensemble prediction system is strongly affected by the method used to inflate the ensemble spread. To simulate the effects of model error we have supplemented our stochastic physics schemes with additive inflation - randomly sampling an archive of analysis increments to perturb the model's trajectory. It has previously been shown that analysis increments contain a component of model error. The relaxation to prior perturbations inflation method is effective at maintaining the spread in the mid-latitudes, but works less well in tropical regions and produces perturbations which are highly balanced. The relaxation to prior spread method is able to maintain the spread across different regions of the globe, but we find that very large relaxation factors are required to produce adequate performance. Acceptable performance has been found by combining these two inflation methods, and this change to the ensemble prediction system gives a substantial improvement to the Met Office’s deterministic hybrid data assimilation.

File available here