A three-dimensional scale-dependent view of atmospheric predictability

This contribution provides a scale-dependent analysis of atmospheric predictability in two global models, the operational ensemble prediction system (EPS) of ECMWF and the forecast ensemble based on the ensemble Kalman filter DART/CAM. The loss of predictability is presented by using the three-dimensionally orthogonal normal modes that provide interpretation of balance and predictability in terms of quasi-geostrophic and inertio-gravity (IG) modes. The normal-mode framework is particularly suitable for the tropics where the IG circulation dominates on all scales and where the short-range forecast errors are largest.

New characteristics of the ensemble spread revealed by the projection of the 3D forecast fields onto normal-mode functions include quantification of the initial spread among the balanced and IG contributions and its time evolution as a function of the horizontal and vertical scales. In particular, both the ECMWF EPS and DART/CAM forecast ensemble are characterized by around half of their initial spread in unbalanced modes in relation to the largest analysis uncertainties in the tropics.

The growth of spread in the short range is substantially different in the balanced, eastward-propagating and westward-propagating IG modes. The eastward IG spread in the ECMWF system grows much faster then the westward IG spread primarily due to the impact of Kelvin modes in the tropics. In ECMWF EPS the spread growth is dominated by the increase of balanced spread on large scales while the subsynoptic spread increases little after the first 24-hr forecast. The ensemble reliability analyzed in the modal space reveals that distribution and evolution of the ensemble spread and the root-mean-square-error of ensemble mean appear very similar on the synoptic and subsynoptic scales, while a large underestimation of the spread develops in the zonally-averaged state after the second day of forecast.

These results from ECMWF EPS are contrasted with the outputs from a perfect-model ensemble data assimilation and forecast experiment which assimilated globally homogeneous observations of dynamical variables within the DART/CAM system.