Bayesian Inference for Complex Computer Models and Large Multivariate Spatial Data for Climate Science

• Activity Number: 406
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): K. Sham Bhat*+ and Murali Haran and Klaus Keller
• Companies: Los Alamos National Laboratory and Penn State University and Penn State University
• Keywords: computer experiments ; hierarchical Bayes ; Gaussian processes ; multivariate spatial data ; uncertainty quantification ; climate change

Abstract

Computer model calibration involves combining information from a complex computer model with physical observations of the process. Computer model output is often in the form of multiple spatial fields, particularly in climate science. We study an effective inferential approach by using Gaussian processes to emulate the computer model, linking the calibration parameters with the multivariate spatial observations. Then we infer the calibration parameters using Bayesian methods, while incorporating more flexible approaches allowing for non-linear relationships among spatial fields and non-separable covariance functions. In addition, we incorporate the uncertainty due to model discrepancy and measurement error into our inference and predictions, which usually results in more accurate and sharper inference of the calibration parameter and improved characterization of uncertainties. We utilize kernel mixing and matrix identities in order to make computations tractable for large spatial data sets. We apply our approach to infer vertical diffusivity, a climate model parameter from which we obtain projections of the Atlantic Meridional Overturning Circulation (MOC).