Description:
Technology
Overview
Software for
Reduced-Dimension Data Assimilation (RDDA) uses fast model surrogates to allow
rapid data assimilation in complex data sets, including those modeling weather
and oceanographic currents. Model
surrogates are statistical models that are trained to approximate the dynamics
of traditional computational models at a fraction of their computational
cost. This method of data
assimilation reduces computational run times by several orders of magnitudes.
The RDDA
software consists of three components:
1)
A dimension reduction toolbox
(EOF-toolbox). EOF-toolbox applies Empirical Orthogonal Function (EOF) dimension
reduction technique to large spatial-temporal datasets, such as outputs of
oceanographic circulation models. Algorithms used in the EOF-toolbox are
described in (Frolov 2007).
2)
A model surrogate training code
(model-surrogate toolbox). Model-surrogate toolbox extends the functionality of
an existing neural network training toolbox (Netlab) to training of statistical
models with time-lagged inputs and outputs. Algorithms used in the
model-surrogate toolbox are described in (Frolov
2007; Frolov et al. 2009; van der Merwe et al. 2007).
3)
A set of driver routines that implement
the RDDA machinery (RDDA drivers). RDDA drivers plumb together the EOF-toolbox,
the model surrogate toolbox, the ReBel toolbox for state estimation, routines
for preparation of the observational data specific to the Columbia River
observing system, and routines for post processing of model outputs. Estimation algorithms used by the RDDA
drivers are described in (Frolov 2007; Frolov et
al. 2009).
References
used:
Frolov S (2007) Enabling technologies for data assimilation in a
coastal-margin observatory. Ph.D. thesis, Oregon Health & Science University
Portland, OR
Frolov S, Baptista AM, Leen TK, Lu Z, van der Merwe R (2009) Fast
data assimilation using a nonlinear Kalman filter and a model surrogate: an
application to the Columbia River estuary. Dynamics of Atmosphere and Oceans 48
(1-3):16-45.
doi:doi:10.1016/j.dynatmoce.2008.10.004
van der Merwe R, Leen TK, Lu Z, Frolov S, Baptista AM (2007) Fast
Neural Network Surrogates for Very High Dimensional Physics-based Models in
Computational Oceanography. Neural Networks.
doi:doi:10.1016/j.neunet.2007.04.023