Ensemble Generation for Seamless Prediction in the GEOS-S2S Forecast System

Abstract. Improving the quality of short term climate (subseasonal to seasonal) forecasts depends on improving both the quality of the forecast model and the quality of the initial conditions, with the latter typically consisting of an ensemble of states that are equally likely estimates of the true initial state. In practice, due to our limited knowledge of the true initial errors, an alternative goal is to insure that the initial perturbations project onto the relevant fastest growing modes. With that goal in mind, we present here a relatively simple to implement, yet effective, strategy for generating initial perturbations that are particularly relevant to the short-term climate prediction problem. The strategy, referred to as the Synchronized Multiple Time-lagged (SMT) approach, uses the information about the temporal coherence of nearby analysis states to generate multiple perturbations that are imposed at a specified initial time, with pre-specified amplitudes determined as a fraction of the climatological variance. We show that the perturbations so generated consist of a rich array of physically realistic atmosphere and ocean modes of variability that appear to have some correspondence with the fastest growing modes determined from a singular value decomposition of the model’s linear propagator. Furthermore, recognizing the conflicting goals of increasing ensemble size and increasing model complexity, we outline a strategy for reducing, after a specified lead time, an initially large forecast ensemble, which involves performing a stratified sampling of the early larger ensemble in a way that accounts for the emerging directions of error growth.