Ensemble Assimilation

Welcome to the SAMA's Ensemble Assimilation group page.
This theme is coordinated by Frédéric Chevallier ()

The data processed with statistical assimilation methods are random variables, represented by probability distributions. In some cases, the distributions can be described with analytical functions, like normal or Poisson laws. For ensemble assimilations methods, these distributions are represented with a discrete population. The number of members varies, depending on the application, from about ten to some hundreds. Such an approach is motivated by empirical (when the analytical law is not known) or practical reasons (when computation is simpler on a discrete population as on an analytical law).
Ensemble Kalman Filter (EnsKF), first described by Geir Evensen in 1992 (see. http://www.nersc.no/~geir/) is an archetypic example of these methods. Compared to a classical Kalman Filter, the EnsKF replaces the analytical computation of the probability distribution with a rebuilding using the finite population. The filter itself remains analytical. This approach allows to take into account a possible nonlinearity of the governing equations of the system, even if the optimality of the EnsKF is only demonstrated for linear equations.
Particle Filter is another strong example of these methods. Here, the goal is to apply the Bayes theorem to each of the ensemble members, for each successive collection of observations.
Ensemble methods are developing in an always more parallel than vectorial computing strategy. Unlike variational methods, they do not require heavy developpments of tangent linear and adjoint codes. On the other hand, the sampling results in a loss of precision, especially for ensembles with few members.
Some research domains connected to ensemble assimilation: members choice, link with predictability and robustness of members ensembles (degeneracy of the cycled ensembles is observed).