Monday, 30 July 2001

Impact of Diabatic Processes on 4-Dimensional Variation with the Global Spectral Model at NCEP/NOAA

The impact of diabatic processes on 4-dimensional variational data
assimilation (4DVAR) was studied using the global spectral model
(1995 version) with full physics at National Centers for Environmental
Prediction (NCEP) and the developed adjoint. With a cost function
measuring spectral errors of 6-hour forecasts, the adaibatic and
diabatic 4DVAR experiments were carried out based on the same NCEP
re-analysis data set by employing the limited memory quasi-Newton
algorithm (L-BFGS). Comparisons on results found that the decrease
of the diabatic cost function remains stuck after some iterations in
minimization due to discontinuities in parameterizations of the model
while the adiabatic cost function continuously decreases. In order to
seek a sounder performance of the diabatic minimization, the decrease
in minimization of the defined cost function in different spectral
domains was first investigated. Results found that at the beginning
stage (before iteration 10) the minimization process mainly adjusts
large-scale flows, in which discontinuous parameterizations in
diabatic model have almost no hurt on minimization. As iterations
increase, the adjustment gradually moves to relatively small scales
through around 20-iteration for transition. During the transition
period, discontinuities in parameterizations may start to damage
minimization to cause the diabatic cost function to be stuck instead
of continuously decreasing as an adiabatic one does. Then a mixed
4DVAR experiment scheme was designed. The scheme first uses the
adiabatic model and adjoint to adjust large-scale flows and then
the diabatic model and adjoint is invoked to adjust relatively small
scales. The switch occurs once the adjustment goes through some
critical period (iteration 10 to 20, in this case) where discontinuous
parameterizations may have the most impact on minimization. Experiment
results showed that the new mixed scheme causes 29% more reduction on
the cost function and 35% on the gradient norm for 60 iterations and
saves 21% CPU time on the NCAR Cray cluster. The resulted optimal
initial conditions from the new scheme improve the short-range
forecast skills with 48-hour statistics.

The results suggest that in solving a diabatic 4DVAR problem, ones may choose an economical differentiable optimization algorithm if the addressed is a relatively large-scale problem. And also, The mixed 4DVAR scheme, to some degree, is able to avoid the hurt on minimization due to discontinuities in parameterizations of a diabatic model with a cheaper computation cost and it is therefore a practicable choice in operational applications.

Supplementary URL: