Our inability to accurately model marine food webs severely limits the prognostic capabilities of current generation marine biogeochemistry models. To address this problem we examine the use of data assimilation and mesocosm experiments to facilitate the development of food web models. The components of the data assimilation demonstrated include the construction of measurement models, the adjoint technique to obtain gradient information on the objective function, the use of parameter constraints, incorporation of discrete measurements and assessing parameter observability. We also examine the effectiveness of classic and contemporary optimization routines used in data assimilation. A standard compartment-type food web model is employed with an emphasis on organic matter production and consumption. Mesocosm experiments designed to examine the interaction of inorganic nitrogen with organic matter provide the data used to constrain the model. Although we are able to obtain reasonable fits between the mesocosm data and food web model, the model lacks the robustness to be applicable across trophic gradients, such as those occurring in coastal environments. The robustness problem is due to inherent structural problems that render the model extremely sensitive to parameter values. Furthermore, parameters governing actual ecosystems are not constants, but rather vary as a function of environmental conditions and species abundance, which increases the sensitivity problem. We conclude by briefly discussing possible improvements in food web models and the need for rigorous comparisons between models and data (a modeling workbench) so that performance of competing models can be assessed. Such a workbench should facilitate systematic improvements in prognostic marine food web models.