Parameter identification in climate models using surrogate-based optimization

摘要

We present initial steps and first results of a surrogate-based optimization (SBO) approach for parameter optimization in climate models. In SBO, a computationally cheap, but yet reasonably accurate representation of the original high-fidelity (or fine) model, the so-called surrogate, replaces the fine model in the optimization process. We choose two representatives, namely two marine ecosystem models, to verify our approach. We present two ways to obtain a physics-based low-fidelity (or coarse) model, one based on a coarser time discretization, the other on an inaccurate fixed point iteration. Since in both cases, the low-fidelity model is less accurate, we use a multiplicative response correction technique, aligning the low- and the high-fidelity model output to obtain a reliable surrogate at the current iterate in the optimization process. We verify the approach by using model generated target data. We show that the proposed SBO method leads to a very satisfactory solution at the cost of a few evaluations of the high-fidelity model only.