[M-GI29] Data assimilation: A fundamental approach in geosciences

[MGI29-P04] Ensemble-based iterative variational data assimilation approach and its extension for count data

*Shin ya Nakano^1,2,3 (1.The Institute of Statistical Mathematics, 2.Center for Data Assimilation Research and Applications, Joint Support Center for Data Science Research, 3.School of Multidisciplinary Science, SOKENDAI)

Keywords:data assimilation, ensemble variational method, Gauss-Newton method, Poisson distribution

The 4-dimensional variational data assimilation, which is one of the typical approaches to data assimilation, can be regarded as a method for finding the maximum a posteriori (MAP) estimate that maximizes posterior probabilities given a time series of observed data. The 4-dimensional variational data assimilation problem is normally solved by the adjoint method. However, the adjoint method requires an adjoint model which takes time and effort to develop.

To avoid the problems with the adjoint model, there exist some ensemble-based approaches for solving the problem of the 4-dimensional variational method without the adjoint model. The ensemble-based approaches employ results of ensemble simulation runs with various initial conditions and parameter settings. These ensemble-based variational methods are much easier to implement than the conventional adjoint method because a simulation model is treated as a black-box. However, since the existing ensemble-based methods are derived under the assumption that observations obey Gaussian distributions, they can not immediately be applied when observations obey other distributions. In this paper, we propose an ensemble-based algorithm for observations obeying Poisson distributions, which can be used for data assimilation into a black-box simulation model. We also conduct a simple experiment with a one-dimensional fluid model to confirm the performance proposed approach.