According to the numerous tests described in [20], we have chosen to impose the boundary conditions of the nested model as follows :

- For the barotropic part, the transport of water normally to the boundary is interpolated from the corresponding interface in the main model and imposed at the boundaries of the nested model. This insures conservation of the amount of water present in the nested model and the corresponding part of the main model, thus conserving the mean value of free surface elevation on the corresponding areas.
- For the baroclinic part, the component of velocity normal to the boundaries of the nested model has been imposed in each cell of those boundaries. Thoses values are of course interpolated from the corresponding values obtained in the coarse model. This insures the main dynamical forcing of the nested model to be coherent with the main model's dynamics.
- For the physical scalar variables, i. e. temperature, salinity and turbulent kinetic energy, what we called the "continuous scheme" has been adopted : interpolated values of the variables are imposed at the boundaries of the nested model. This scheme has been adopted preferentially to a "conservative" one (imposing fluxes at the boundaries) for numerical stability reasons, as the tests showed. Of course we don't have conservation of heat or salt quantities in the nested model and the corresponding part of the main model, but values of the scalars do not differ very much from one grid point to the next, so that the adopted method isn't a problem for short-term integrations.
- For the other scalar variables, i.e. biological model variables or passive tracers,
variables whose effects on the physics can be neglected, the conservative scheme is used.
At entrance boundaries of the nested model (boundary points where water enters the fine model),
fluxes of the variables interpolated from the coarse model are imposed.
At exit boundary points, the information on the variable evolution comes with the flow, i.e.
comes from the inner part of the fine model, and so nothing coming from the main model is imposed.
In this last case a null horizontal gradient is used at the boundary point.

Wed Jun 10 10:53:44 DFT 1998