Ensemble initialization of the oceanic component of a coupled model through bred vectors at seasonal-to-interannual timescales

Abstract

Abstract. We evaluate the ensemble spread at seasonal-to-interannual timescales for two perturbation techniques implemented in the ocean component of a coupled model: (1) lagged initial conditions as commonly used for decadal predictions; (2) bred vectors as commonly used for weather and seasonal forecasting. We show that relative to an uninitialized reference simulation the implementation for bred vectors can improve the ensemble spread compared to lagged initialization at timescales from one month up to three years. As bred vectors have so far mostly been used at short timescales, we initially focus on the implementation of the bred vectors in the ocean component. We introduce a depth-dependent vertical rescaling norm, accounting for the vertical dependence of the variability, and extending the commonly used upper-ocean rescaling norm to the full water column. We further show that it is sufficient for the (sub-surface) ocean to breed temperature and salinity (i.e., scalar quantities), and rely on the governing physics to carry the temperature and salinity perturbations to the flow field. Using these bred vectors with a rescaling interval of 12 months, we initialize hindcast simulations and compare them to hindcast simulations initialized with lagged initial conditions. We quantify the ensemble spread by analyzing Talagrand diagrams and spread–error ratios. For both temperature and salinity, the lagged initialized ensemble is particularly under-dispersive for the first few months of predictable lead time. The ensemble initialized with bred vectors improves the spread for temperature and salinity for the 0–700 m and 1000–3500 m means, compared to the lagged ensemble at lead times of several months to one year. As the lead time increases to years, the differences between the two ensemble initialization techniques become more difficult to discern. While the results need to be confirmed in an initialized framework, the present analysis represents a first step towards improved ensemble generation at the transition from seasonal to interannual timescales, in particular at lead times up to one year.