Sensitivity Analysis Methods And Applications Of Complex Geophysical Models

Computer-based system models often contain many parameters, which exert great influenceonmodels’performance. Numerousstudieshaveshownthatcorrectparametervaluescangreatlyimprovemodels’performance. However,howtospecifymodelparametersproperlyisacomplexscientific problem. It involves how one can correctly identify and ultimately optimize the mostimportant model parameters. Sensitivity analysis (SA) is a commonly used approach foridentifying important parameters that dominate model behaviors, and is also an important meansto fully understand how the model parameters affect the model’s performance. SA provides thescientific basis for model parameter optimization. Through a comprehensive evaluation andcomparison of various global SA methods, we designed an uncertainty quantification (UQ)framework for two-stage SA, and conducted sensitivity analyses for several geophysical modelsof different complexities and different variable types. The ultimate objective of this dissertationis to provide practical guidance to researchers on how to choose appropriate SA methods foranalyzing complex system models. The main contributions of this dissertation are as follows:(1) This study performed a comprehensive review of various SA methods, including thecommonly used local SA methods and the global SA methods that have become popularin recent years. The effectiveness and efficiency of nine widely used global SA methods,including six qualitative and three quantitative ones, were then evaluated. All SAmethods were tested using a variety of sampling techniques to screen out the leastsensitive parameters from the sensitive (i.e., important) ones. The Sacramento SoilMoisture Accounting (SAC-SMA) model, which has thirteen tunable parameters, wasused for illustration. By evaluating different SA methods, we have demonstrated theirstrengths and limitations as well as applicability, and have provided guidance on how tochoose appropriate SA methods for other applications.(2) Based on the merits of both qualitative and quantitative global SA methods, an UQframework was further developed to examine parameter sensitivities of a newlydeveloped Conjunctive Surface-Subsurface Process (CSSP) land surface model forterrestrial hydrological modeling. This seamless stepwise SA was conducted using a multi-objective two-stage approach, in which the first stage is a qualitative SA using theLatin Hypercube-based OAT (LH-OAT) method for screening out insensitive parameters,and the second stage is a quantitative SA using the MARS-based Sobol’sensitivityindicesfor apportioning total response variance to the contribution of each sensitive parameter.Parameter behaviors were understood and sensitive parameters were identified. Bydoing this, we aim to reduce parameter dimensionality and simplify the complex modelfor parameter optimization. This study serves to provide an effective and efficient wayfor parameter SA of complex system models.(3) Statistical methods were adopted to analyze the sensitivities of physical processes of acomplex numerical weather prediction model—Weather Research and Forecasting (WRF)model. The goal of this study is to employ SA methods to help identify whatcombinations of WRF model parameterization schemes are acceptable. This problem isbasically a combinatorial optimization problem of discrete categorical variables, whichare different from the continuous variables (such as model parameters) or discrete integervariables (which assume integer values). Discrete categorical variables cannot beassigned to any numerical values, as they represent only symbols or categories. For theWRF model, a parameterization scheme of a physical process represents a category of acategorical variable. The combinatorial problem for the WRF model is especiallychallenging, as the number of combinations of the WRF model parameterization schemesis very large (more than106). To analyze sensitivities of the physical processes, analysisof variance (ANOVA) method was used in this study to test whether there are significantdifferences among the performances of different parameterization schemes of the samephysical process, for the modeling of air temperature and precipitation in Beijing. Tukeymultiple comparison test was then used to find the parameterization schemes that performsignificantly different from others. We also tried to explore if it is possible to identifyand then discard the worst parameterization schemes step by step for each physicalprocess, as to reduce the number of scheme combinations and find a few optimal schemecombinations that perform without significant differences. This study puts forward anew idea for SA and combinatorial optimization for the discrete categorical problems.