Abstract:  We have developed extremely fast surrogates that mimic the dynamics
of a non-linear, high-dimensional (107 degrees of freedom) finite element model
of the Columbia River Estuary and near ocean region.  We developed the surrogate
technology as a component in a portable Bayesian model/data fusion system (called
data assimilation in oceanography and meteorology).  The surrogates use neural
networks trained to respond to tidal, atmospheric and river flux forcings as do
the full finite element models.  They accelerate the modeling by factors between
1,000 and 12,000. We expect such surrogates to be useful in other domains that
use high-dimensional numerical simulations (such as meteorology, hydrology, 
aerodynamics, and computational fluid dynamics) and for a range of applications
including parameter optimization and ensemble probabilistic prediction.