Baird Langenbrunner and J. David Neelin
Journal of Advances in Modeling Earth Systems,submitted Sept. 2016.
Supplement (6.4 MB).
Abstract Global climate model (GCM) integrations exhibit notable uncertainty in simulating the hydrological cycle, arising substantially from sensitivity to sub-grid scale physics or parameterizations. The dependence of a climate field on parameter values can be significantly nonlinear, though its explicit form has not been thoroughly addressed in ensemble studies. Here, we investigate the parameter dependence of global seasonal precipitation in an existing perturbed physics ensemble, in which four parameters were sampled in the deep convection scheme of a fully-coupled GCM. This ensemble is used to train a metamodel or model emulator that allows for a reconstruction of parameter space at a fraction of the computational cost relative to brute-force sampling. A quadratic metamodel reproduces seasonal root-mean-square error (rmse) relative to observations with decent skill, though it fails over certain highly sensitive parameter ranges. Narrowing to a trust region along the parameter axis works well at the cost of excluding these most sensitive ranges. Using an approach known as high-dimensional model representation (HDMR), an alternative metamodel is constructed from empirical orthogonal functions across parameter axes, yielding principal uncertainty patterns (PUPs). This PUP-HDMR metamodel performs well at reproducing seasonal rmse measures on and near parameter axes, even in highly sensitive ranges. This metamodel is then used to explore questions of parameter optimization by analyzing rmse surfaces as a function of multiple parameters. We introduce a set of iterative steps that a modeler can follow in the process of characterizing the parameter sensitivity of a GCM and constraining it against observations or reanalyses.
Citation B. Langenbrunner and J. D. Neelin, 2015: Multiobjective constraints for climate model parameter choices, Part I: High-dimensional model representation strategies in CESM1. Journal of Advances in Modeling Earth Systems, submitted Sept. 2016.