A statistical approach to ocean model testing and tuning

Abstract

For many purposes such as the forecasting of oceanic conditions or the prediction of climate changes, it is important to have realistic models of the ocean, that is models which are able to reproduce sufficiently well the main features of interest, and their time variability. This is particularly critical in the design of coupled ocean-atmosphere models where model errors are generally exacerbated by the coupling, so that artificial “flux-corrections” (Sausen et al, 1988) are often used to prevent a drift toward unreasonable climate conditions, even though they may distort the dynamics. Although more complex models should represent reality more correctly than simpler ones, they do not necessarily do so, as illustrated by the El Niño-Southern Oscillation (ENSO) phenomenon where simple ocean-atmosphere models have so far provided forecasts as good as general circulation models (GCMs), in part because of the drift of the latter (Latif et al, 1993). Provided models are consistent with known physics within the tolerance allowed by the approximations made, model adequacy should thus be judged by the ability at reproducing the relevant observations.