The benefits of a hybrid‐3DEnVar data assimilation scheme for the convection‐permitting NWP model AROME–Austria

Abstract We apply a hybrid three‐dimensional ensemble variational (Hybrid‐3DEnVar) data assimilation (DA) method to the convection‐permitting numerical weather prediction (NWP) model Application of Research to Operations at Mesoscale over Austria (AROME–Austria) and evaluate whether the resulting analysis outperforms GeoSphere Austria's operational three‐dimensional variational DA (3DVar) system. Whereas 3DVar uses static model‐error covariances computed from a climatology of past forecasts, Hybrid‐3DEnVar includes flow‐dependent error covariances sampled from a 50‐member ensemble generated by the Convection‐permitting Limited‐Area Ensemble Forecasting (CLAEF) system of GeoSphere Austria. To investigate how the hybrid weighting of static versus flow‐dependent covariances impacts the resulting analysis, we conduct experiments using evenly spaced weighting values, ranging from 0 (3DVar) to 1 (3DEnVar), with multiple Hybrid‐3DEnVar values in between. We assimilate Mode‐S aircraft observations for 12 different test cases and verify the resulting analyses against independent Aircraft Meteorological Data Relay (AMDAR), wind profiler, and radiosonde observations. Although 3DEnVar reveals mixed performance, Hybrid‐3DEnVar with a 10%–70% weight assigned to the ensemble‐error covariance matrix consistently outperforms 3DVar in almost all test cases for both temperature and wind. However, the exact value that results in the most accurate analysis varies with weather conditions, the assimilated variable, and the test case. Furthermore, we confirm that the chosen localization length‐scale affects which hybrid weighting performs best. Smaller localization scales are most effective when model errors are determined primarily by the ensemble, whereas larger localization scales perform better with a larger weight for the static component.