Climate Model Calibration Using High-Dimensional and Non-Gaussian Spatial Data

Open Access
Author:
Chang, Won
Graduate Program:
Statistics
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
May 13, 2014
Committee Members:
  • Murali Haran, Dissertation Advisor
  • Murali Haran, Committee Chair
  • Klaus Keller, Committee Member
  • Bing Li, Committee Member
  • Donald Richards, Committee Member
Keywords:
  • Climate Model Calibration
  • Gaussian Process
  • High-dimensional Spatial Data
  • Non-Gaussian Spatial Data
Abstract:
This thesis focuses on statistical methods to calibrate complex computer models using high-dimensional spatial data sets. This work is motivated by important research problems in climate science where such computer models are frequently used. Climate models play a central role in generating projections of future climate. An important source of uncertainty about future projections from these models is due to uncertainty about input parameters that are key drivers of the resulting hindcasts and projections. Climate model calibration is a statistical framework for inferring the input parameters by combining information from climate model runs and observational data. When the data are in the form of high-dimensional spatial fields, climate model emulation (approximation) and calibration can pose significant modeling and computational challenges. The goal of this research is to develop new approaches to computer model calibration that are computationally efficient, accurate, and carefully account for uncertainties. The main contributions of this thesis are three-fold: (1) to develop a highly efficient reduced-dimensional climate model calibration approach that enables the use of high-dimensional spatial data; (2) to formulate a novel calibration method based on block composite likelihood and study the asymptotic properties of the resulting estimates for input parameters; and (3) to introduce a calibration framework that generalizes the existing method to the one-dimensional exponential family and formulate a reduced-dimensional approach that can efficiently handle the high-dimensional non-Gaussian spatial data. Our methods provide insights about current and future climate. In our first application we make projections of the North Atlantic Meridional Overturning Circulation (AMOC), an ocean circulation that transports heat from low- to high-latitude areas in the Atlantic and contributes to the mild climate in Northern and Western Europe. AMOC changes are projected to impact human and natural systems. We demonstrate that utilizing information from high-dimensional spatial data reduces parametric uncertainty and thus results in an AMOC projection with reduced uncertainties. In the second case study, we demonstrate an application of our approach for non-Gaussian spatial data to calibration of a Greenland ice sheet model and show that our approach can improve upon current methods for projections of sea level rise contributed by the Greenland ice sheet.