The SWANFAR^® system is built from the spectral wave model SWAN and can assimilate frequency-directional wave spectra in arbitrary nearshore domains. When provided with initializing boundary spectra, bathymetry, and observed spectral data (at individual locations within the domain), the system produces a "best estimate" of spectra throughout the domain by minimizing the model-data error at the observation locations.

The structure of the SWANFAR^® system is based on the structure of 4DVAR NCOM. The main program calls forward, adjoint, and representer-perturbation (RP) versions of SWAN, utilizing their output together with observed spectra to compute a cost function. In an iterative process, the cost function is minimized using a conjugate gradient technique and repeated calls to adjoint and RP SWAN.

The adjoint SWAN code is the fundamental core of the SWANFAR^® data assimilation system. When provided with the model-data error (or innovation) spectra at one or more "observation" locations in the domain, the adjoint propagates the perturbations "backward" throughout the domain, determining the sensitivity of spectral energy levels at all locations to the perturbations from the assimilation points.

The numerical adjoint is composed of a collection of adjoint subroutines, each individually constructed from a corresponding subroutine in forward SWAN. Every subroutine in the forward model that acts on the wave action is matched by a complementary subroutine in the adjoint model. The current version of adjoint SWAN includes adjoints to all source and sink subroutines in the forward model, and it also allows for nonstationary conditions and wave-current interaction.

System Flowchart (click chart for larger image)