PT Unknown AU Prieß, M TI Surrogate-Based Optimization for Marine Ecosystem Models PY 2012 PU Christian-Albrechts-Universität zu Kiel WP https://macau.uni-kiel.de/receive/diss_mods_00007782 LA en DE marine ecosystem models; parameter identification; optimization; data-assimilation; model calibration; computationally efficient optimization; accelerated optimization; surrogate-based optimization; response correction; aggressive space mapping; space mapping; low-fidelity models; coarse models; numerical stability; marine Ökosystem-Modelle; Parameteridentifikation; Optimierung; Daten-Assimilation; Modellkalibrierung; rechenzeiteffiziente Optimierung; beschleunigte Optimierung; surrogat-basierte Optimierung; Modellkorrektur; grobe Modelle; numerische Stabilität AB Marine ecosystem models are of great importance for understanding the oceanic uptake of carbon dioxide and for projections of the marine ecosystem’s responses to climate change. The applicability of a marine ecosystem model for prognostic simulations crucially depends on its ability to resemble the actually observed physical and biogeochemical processes. An assessment of the quality of a given model is typically based on its calibration against observed quantities. This calibration or optimization process is intrinsically linked to an adjustment of typically poorly known model parameters. Straightforward calibration attempts by direct adjustment of the model parameters using conventional optimization algorithms are often tedious or even beyond the capabilities of modern computer power as they normally require a large number of simulations. This typically results in prohibitively high computational cost, particularly if already a single model evaluation involves time-consuming computer simulations. The optimization of coupled hydrodynamical marine ecosystem models simulating biogeochemical processes in the ocean is here a representative example. Computing times of hours up to several days already for a single model evaluation are not uncommon. A computationally efficient optimization of expensive simulation models can be realized using for example surrogate-based optimization. Therein, the optimization of the expensive, so-called high-fidelity (or fine) model is carried out by means of a surrogate – a fine model’s fast but yet reasonably accurate representation. This work comprises an investigation and application of surrogate-based optimization methodologies employing physics-based low-fidelity (or coarse) models. Seeking a computationally efficient calibration of marine ecosystem models serves as the fundamental aim. As a case study, two illustrative marine ecosystem models are considered. Here, coarse models obtained by a coarser temporal resolution and by a truncated model spin-up are investigated. The accuracy of these computationally cheaper coarse models is typically not sufficient to directly exploit them in the optimization loop in lieu of the fine model. I investigate suitable correction techniques to ensure that the corrected coarse model (the surrogate) provides a reliable prediction of the fine model optimum. Firstly, I focus on Aggressive Space Mapping as one of the original Space Mapping approaches. It will be shown that this optimization method allows to achieve a reasonable reduction in the optimization costs, provided that the considered coarse and fine model are sufficiently “similar”. A multiplicative response correction approach, subsequently investigated, turned out to be very suitable for the considered marine ecosystem models. A reliable surrogate can be obtained. Exploiting the latter in a surrogate-based optimization algorithm, a computationally cheap but yet accurate solution is achieved. The optimization costs can be significantly reduced compared to what is achieved by the Aggressive Space Mapping algorithm. The proposed methodologies, particularly the multiplicative response correction approach, serve as initial parts of a set of tools for a computationally efficient calibration of marine ecosystem models. The investigation of further enhancements of the presented algorithms as well as other promising approaches in the framework of surrogate-based optimization will be highly valuable. PI Kiel ER