Development of a global data assimilation system for climate model study
Tadashi Tsuyuki*, Yoshiaki Takeuchi
Numerical Prediction Division, Japan Meteorological Agency
Nobuhiro Ishida
Research Organization for Information Science and Technology
*e-mail:
Data assimilation is a technique for producing accurate initial conditions for numerical weather prediction from observational data and a numerical weather prediction (NWP) model. Product of data assimilation, which is called analysis, has also been widely used as verifying data for examining performance of a climate model. Comparison of long-term simulation of a climate model and climatology obtained from analyzed data is a usual method for examining performance of the climate model. For this purpose, however, it is desirable to use the same model in data assimilation itself, because the analysis is more or less affected by model bias and a different model has different bias. A data assimilation system for a climate model could provide another method for examining the climate model. The model would be run in NWP model and then checked using the wide range of NWP verification technique such as synoptic forecast quality, objective scores, weather parameters, etc., as well as diagnostics of hydrological budgets, surface fluxes, cloud amount, diabatic heating, etc. Since observational data and short-range model forecasts are directly and intensively compared in a data assimilation system, the system itself could also be an effective tool for model verification.
The Numerical Prediction Division of the Japan Meteorological Agency (NPD/JMA) and the Research Organization for Information Science and Technology (RIST) have been cooperated to construct an advanced global data assimilation system on a distributed-memory parallel computer since 1998. The adopted assimilation method is a 4-dimensional variational (4D-Var) technique with time-evolving background error covariance matrix a Kalman filtering technique. The model used is basically the same as the unified global spectral model for NWP and climate study,which is being developed at JMA and the Meteorological Research Institute (see Sugi (2001) and Kuma (2001) in this workshop). The current operational data assimilation systems at JMA are based on the conventional optimal interpolation (OI). JMA plans to introduce a global 4D-Var assimilation system to the operational NWP on a Hitachi SR8000/E1 early 2004, based on the cooperative work between NPD/JMA and RIST. Incidentally, JMA has also been developing a 4D-Var system for mesoscale NWP since 1997, and the mesoscale 4D-Var system will be operational early 2002.
The development of the global 4D-Var system consists of two parts: development of a 3-dimentional variational (3D-Var) assimilation system and that of the tangent linear and adjoint models of the global spectral model. The basic design of the background error covariance matrix in the 3D-Var system is similar to that in the variational assimilation systems at ECMWF, NCEP and CMC. Background error standard deviations of vorticity, divergence, surface pressure, temperature and specific humidity are assumed to be zonal symmetric, while those of unbalanced divergence and unbalanced surface pressure and temperature are assumed to be horizontally homogeneous. Partial decoupling between mass and wind fields in the tropics are taken into account by utilizing the singular value decomposition (Daley, 1996). The background error statistics is obtained by the NMC method (Parrish and Derber, 1992) and the maximum likelihood method (Dee et al., 1999). It is found that the 3D-Var system shows better performance in the southern hemisphere than the JMA operational OI system. Positive or neutral impacts are also seen in the northern hemisphere. The tangent linear and adjoint codes of the dynamical process of the model were constructed by partially utilizing the automatic adjoint code generator, TAMC, by Giering and Kaminski (1995). The 4D-Var system with the constant background error covariance matrix is currently being developed by combining the 3D-Var system and the adjoint model.
The 3D-Var system was optimized for the Hitachi SR8000/E1 by using the following methods: parallel processing of the wave-grid transformation according to the method used in the global spectral model and node allotment based on the type and number of observational data. Figure 1 compares the CPU times between single node and four nodes for the T213L30 version of the model with an inner-loop model of the T63L30 model. For the case of four nodes, the CPU time is about one fourth of that for single node. However, Input and output processes are not easy to parallelize. Figure 2 shows the job flow of the currently developed 4D-Var system for a distributed-memory machine. There are a couple of methods for predicting the background error covariance matrix in the 4D-Var assimilation system: a reduced-rank Kalman filter approach and an ensemble Kalman filter approach. Which method is to be taken is not yet decided. It is also to be noted that availability of a massive parallel computers provides an opportunity for investigating the feasibility of ensemble data assimilation, though it is a more NWP-oriented work.