Ensemble Data Assimilation: Challenges and New Ideas

Estimates of the state of a physical system can be inferred by blending information from both observations and model predictions of the system itself. These estimates should reflect uncertainties arising from both errors in the observations and predictions, and the sub-optimality of the inference procedure. Ensemble data assimilation (EnDA) proposes an attractive solution to the inference problem in the context of large geophysical systems such as the atmosphere. While promising, EnDA is a relatively new discipline requiring further exploration and development. Some aspects of the EnDA ripe for advancements include the forward operator for observations, treatment of model errors caused by model inadequacies, and localization procedures. These outstanding issues, along with the potential of particle filters, will be discussed to outline possible avenues for the future advancement of EnDA. I will show examples of Bayesian inference and stochastic sampling as solutions of the inverse problem of determining atmospheric sources that can be extended to account for errors generated by model parameter uncertainties. I will also show results of experiments using ensembles and Kalman filtering in the context of air quality predictions, along with a new method based on an analog concept to reduce both bias and random prediction errors in surface wind predictions.
I will discuss how the analog concept can be extended to improve the forward operator, as well to populate an ensemble to alleviate sampling issues and to account for the need of localization in EnDA.