|Title||Evaluating ecosystem model complexity for the northwest North Atlantic through surrogate-based optimization|
|Publication Type||Journal Article|
|Year of Publication||2019|
|Authors||Kuhn A.M, Fennel K.|
|Type of Article||Article|
|Keywords||biogeochemical models; chlorophyll-a; circulation; Data assimilation; dynamics; marine; Meteorology & Atmospheric Sciences; model complexity; North Atlantic; ocean model; oceanography; optimization; parameter-identification; performance; Plankton; sensitivity|
Objectively determining the level of ecosystem model complexity necessary to achieve meaningful representations of biogeochemical cycles at different spatial and temporal scales is an outstanding issue in marine ecosystem modeling. As part of the development of a three-dimensional (3D) Regional Ocean Modelling System (ROMS) application for the northwest North Atlantic Ocean, we compare model results from three alternative ecosystem model versions in which ecological complexity was increased in a step-wise fashion. In order to ensure an objective comparison, the models were optimized to replicate observations of satellite surface chlorophyll, and in situ chlorophyll and nitrate profiles. To overcome the high computational cost of optimizing 3D models, we use a surrogate-based optimization method; that is, an ensemble of one-dimensional (1D) models is used as a proxy of the ecosystem model behavior in the 3D setting. The 1D models were configured at locations where in situ profiles are available. A total of 17 optimization experiments aim to evaluate different aspects of the comparison between the ecosystem models. We show that for all ecosystem model versions the optimized model performance degrades when the optimization includes all observed variables at all locations instead of individual locations only. Moreover, the choice of parameters to be optimized can significantly affect the behavior of the optimized models and is most noticeable when multiple phytoplankton and zooplankton groups are included. Additionally, evaluation of spatial patterns in optimal parameter values at individual locations allows us to assess geographical model portability. In general, an optimized complex model can achieve lower model-data misfits against assimilated data than simple models, but is also more prone to generating unintended trophic relations. The more complex model also had decreased performance when applied to locations different than those used for calibration (i.e., "portability experiments"), which is discussed in the context of the design of the cost function and selection of parameters to optimize.