Geophysical phenomena occur from
the interaction between atmosphere, ocean, land, and solar radiation.
Understanding the phenomena requires us to grasp their feedback properties
and nonlinear processes. For the grasp of these characteristics we conventionally
have the only means: computer simulation. However, recent development
of the earth observation systems, spacecraft observation for example,
enables us to obtain a dataset so global as we could merely estimate via
computer simulations so far. On the other hand, it is a fact that conventional
simulations have been executed under the parameter setting considerably
idealized and simplified compared with the real world, due to a computational
limit and a lack of available observation to refer for. Data assimilation
technique integrates the global dataset with simulation results, aiming
at more precise prediction of the geophysical phenomena and production
of a new synthesized dataset that will be used to examine the prediction.
Data assimilation in effect fits models to observations. A model fitting might recall us to a line fitting with least squares. When a simulation model is used instead of a line model, we call the fitting Data assimilation. While simulations originally proceed with time integration by themselves once we give the initial and boundary conditions, data assimilation modifies the simulation result at each time step with the help of data, and goes to the next step calculation with the modified result. Thanks to the modification process included, even a simple simulation model can predict more realistic state. Figure 1 shows an example of data assimilation: sea surface height data are assimilated into a simulation model reproducing the interaction between atmosphere and ocean.
As mentioned above, we can regard data assimilation as simulation taking in data. Execution of data assimilation then requires not only a simulation run but also a modification process with data, so the computational load exceeds 100 times of that for a pure simulation. This institute has a parallel computer of SGI Japan, Ltd. shown in Figure 2 (256 CPU and 1920GB shared memory at the maximum) by which we can conduct data assimilation experiments without minding heavy computational load.
Data assimilation in effect fits models to observations. A model fitting might recall us to a line fitting with least squares. When a simulation model is used instead of a line model, we call the fitting Data assimilation. While simulations originally proceed with time integration by themselves once we give the initial and boundary conditions, data assimilation modifies the simulation result at each time step with the help of data, and goes to the next step calculation with the modified result. Thanks to the modification process included, even a simple simulation model can predict more realistic state. Figure 1 shows an example of data assimilation: sea surface height data are assimilated into a simulation model reproducing the interaction between atmosphere and ocean.
As mentioned above, we can regard data assimilation as simulation taking in data. Execution of data assimilation then requires not only a simulation run but also a modification process with data, so the computational load exceeds 100 times of that for a pure simulation. This institute has a parallel computer of SGI Japan, Ltd. shown in Figure 2 (256 CPU and 1920GB shared memory at the maximum) by which we can conduct data assimilation experiments without minding heavy computational load.
Members
Genta Ueno , Tomoyuki Higuchi
Top: Sea surface height observation when El Nino
event occurs,
Bottom: a simulation result which the data are assimilated into. SSH means
Sea Surface Height.
Parallel computer for statistical science (SGI Altix3700 Super Cluster).
