Más información:
Información de contacto del evento:
Semillero de Investigación en Asimilación de Datos, siad.eafit@gmail.com
TRANSMISIÓN EN VIVO
Objective
Data Assimilation is the process by which an imperfect numerical forecast is adjusted according to real-noisy observations. In general, two well-known family of methods are employed under operational DA scenarios: sequential and variational formulations. In sequential DA methods, observations are assimilated one at a time while variational methods, typically, seek for the initial condition which best fit a given assimilation window. The ensemble Kalman filter (EnKF) is a well-established and -recognized filter into the sequential DA context. The EnKF is usually exploited for parameter and state estimation in highly non-linear numerical models, its popularity obeys to his simple formulation and relatively ease implementation. In the EnKF, an ensemble of model realizations is employed to estimate moments of background error distributions and through observations, to obtain samples of posterior (analysis) distributions. Stochastic and deterministic formulations of such filter are widely found in the current literature. Regardless his nature, ensemble sizes are bounded by the hundreds while model resolutions are in the order of the millions.
Consequently, sampling errors can impact the quality of posterior samples which can lead to poor estimates of error dynamics. Localization methods are commonly employed to counteract the effects of sampling noise, for instance, some of them are covariance matrix localization (B-localization), observation localization (R-localization), and local domain decomposition. Despite these methods have been tested under operational DA settings, their practical implementations are an active field into the DA community wherein, for instance, scientific efforts are typically centered in proposing matrix-free EnKF implementations. In the variational context, Four Dimensional Variational (4D-Var) methods are commonly utilized for the estimation of initial conditions in numerical models for a given set of temporally-spaced observations. These methods rely on adjoint models which are labor-intensive to develop and computationally expensive to run. Yet another family of methods are the hybrid ones which have proven to properly work under operational settings. In these methods, an ensemble of model realizations is employed to build surrogate models of the 4D-Var cost function onto ensemble sub-spaces wherein analysis innovations are estimated and via their projections onto model spaces, initial analysis ensembles (whose model trajectories best fit observations) can be estimated.