Application of uncertainty quantification, benchmarking, and model diagnostic tools to inform climate risk assessments
Open Access
- Author:
- Ye, Haochen
- Graduate Program:
- Geosciences
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 25, 2023
- Committee Members:
- Donald Fisher, Program Head/Chair
Murali Haran, Outside Unit & Field Member
Antonia Hadjimichael, Major Field Member
Klaus Keller, Special Member
James Kasting, Chair & Dissertation Advisor
Robert Nicholas, Dissertation Advisor - Keywords:
- Climate risk
Uncertainty analysis
Sensitivity analysis
Benchmark - Abstract:
- Complex models provide inputs to climate risks assessments. Projected climate risks present challenges in multiple sectors. Climate risk-informed decision-making can require the use of models with complex dynamics. In these models, uncertainties can arise from multiple sources including observational data, model structures, model parameters, and projected forcings. Neglecting these uncertainties may lead to biased climate risk projections and poor decisions. However, quantifying these uncertainties and their impacts can be nontrivial. This dissertation applies and benchmarks selected existing model diagnostic tools for uncertainty and sensitivity analysis using case studies. The first study assesses uncertainties in maize yield projections through a weather-yield statistical model. This study addresses two questions: (1) How does the incorporation of parametric and climate forcing uncertainties change maize yield projections? (2) What is the relative importance of these uncertainties on the uncertainties surrounding maize yield projections? This study first adopts a pre-calibration method to sample the parametric uncertainties and then applies a cumulative sensitivity decomposition method to quantify the impacts of parametric uncertainty and climate forcing uncertainty on maize yield projection variance. Considering both uncertainties expands the tail of yield projections. Our results show that parametric uncertainty is more important than climate forcing uncertainty. The second encompasses ongoing work that seeks to quantify uncertainties of soil and crop-related parameters in soil moisture simulation hindcasts and projections from a hydrological model. This study focuses on how to sample these uncertainties using limited available soil moisture observations. This study applies a pre-calibration method similar to the first study to sample plausible soil moisture simulations. This is a step towards more comprehensive uncertainty and sensitivity analyses for crop yields and economic projections. The third study benchmarks the performance of different global sensitivity analysis methods for a wide range of model complexities and evaluation times. This study addresses two main questions: (1) How do model space dimensions and evaluation times influence the choice of a global sensitivity analysis method? (2) How do model space dimensions and evaluation times change the estimated computational cost of the fastest sensitivity analysis method? This study compares the results and computational costs of four existing global sensitivity analysis methods using polynomial test models with varying model evaluation times and parameter space dimensions. The emulation and adaptive sampling methods are faster than the standard Sobol’ methods for high-dimensional or slow models, and the Bayesian adaptive spline surface method is fast for most scenarios. The pre-calibration method is a straightforward approach to sample uncertainties but can have a low sampling efficiency for high-dimensional problems. The choice of global sensitivity analysis methods depends on nontrivial interactions between the dimensionality of the model’s parameter space and model evaluation time. These studies can guide the choice of uncertainty and sensitivity analysis methods in climate risk assessments. Proper selection and use of these methods can help sample and quantify important model uncertainties with lower computational costs. Future work can expand these studies by considering more types of uncertainties, more types of models, and more uncertainty and sensitivity analysis methods. This can improve the understanding and application of different model diagnostic tools in climate risk assessments and relevant decision-making.