7B.2 Application of the randomized
incremental optimal technique (RIOT) for parallelization of 4D-Var in
WRFDA-Chem
Guerrette, Jonathan, National Oceanic and
Atmospheric Administration, Nicolas Bousserez,
and Daven Henze, University of Colorado at Boulder
Atmospheric pollutants (e.g., carbon dioxide, methane,
aerosols) have quantifiable impacts on weather, climate, and human health,
often on the time scale of a few days.
While operational weather prediction improves significantly by
applying data assimilation (DA) to initial state, the current chemical state
also depends strongly on surface fluxes, whose relationship to large-scale
observations can only be approximated by a 4D model. WRFDA-Chem uses incremental
four-dimensional variational data assimilation
(4D-Var) to quantify emissions of atmospheric chemical tracers in limited
area domains for research applications on single-week timescales. 4D-Var is computationally expensive
for real-time forecasting, partially due to its iterative minimization
procedure, which is divided into inner and outer loops. The recent proliferation of randomized
methods for matrix decomposition enables the parallelization of the inner
loop. We utilize an ensemble of adjoint integrations in the randomized incremental
optimal technique (RIOT) in WRFDA-Chem.
We achieve wall-time reductions of nearly an order of magnitude, with
some increase in the required core-hours. This new development can be extended
to any application of WRFDA 4D-Var, independent of observation and control
variable combinations. Together,
WRFDA-Chem and RIOT might enable simultaneous meteorological and chemical DA
in order to improve both types of forecasts. |