| abstract: |
Data Assimilation (DA) is a technique for a synthesis of information from a dynamic (numerical) model and observation data. It is an emerging area in earth sciences, particularly oceanography, stimulated by recent improvements in computational and modeling capabilities and the increase in the amount of available observations. In statistical methodology, DA can be formulated in the generalized state space model that draws much interest of the researchers in various domains such as the time series analysis, signal processing, and control theory. As a result, DA can be applicable to any scientific domains involving numerical simulation models. The four-dimensional variational analysis (4DVAR), in particular, the adjoint method is usually employed as DA in an oceanography and meteorology. 4DVAR is an optimization technique to find the optimal values for specifying the initial and boundary conditions, and then its general framework can be easily accepted by geophysical researchers. However, there remains difficulty in applying the adjoint method to a wide variety of problems in any scientific domain, because the adjoint method requires much human resources to develop an original code for each numerical simulation. On the other hand, an ensemble-based sequential DA (EnSDA), which is not the optimization technique but a procedure for a statistical inference, has an advantage in terms of less human resources which is achieved by plugging into the existing “forward” simulation codes. We address pros and cons between the non-sequential DA (4DVAR) and EnSDA for the practitioners who intend to introduce DA. Most of issues are laid on human and computational resources. We briefly explain a recent advancement in EnSDA. Finally, we discuss on what and how we develop the EnSDA methods for realizing DA on HPC.
|
