Grant Hutchings
Title: Computationally Efficient Methods for Emulation and Bayesian Calibration of Computer Models
Date: November 26th, 2025
Time: 2:00pm
Location: LIB 2020
Supervised by: Derek Bingham
Abstract:
The design and analysis of computer experiments encompass a broad range of statistical methodologies. This thesis presents 3 research projects in the field with a unifying theme; all projects are motivated by Bayesian model calibration and leverage Gaussian process models. Chapter 2 develops a novel approach for fast emulation, calibration, and active learning for computer models with functional response, focusing on computational efficiency when very large ensembles of runs are available for analysis. Careful implementation of local Gaussian process regression with scaled inputs facilitates this efficiency. In Chapter 3, a new approach for efficient uncertainty propagation in modular and cut-Bayesian inference problems is introduced. As in Chapter 2, computational efficiency is of utmost importance, and the proposed methodology is shown to provide significant savings over traditional approaches. Chapters 2 and 3 are closely related, both presenting novel methodology for efficient modular Bayesian calibration targeting different applications. Chapter 4 tackles an important statistical challenge in computational wildfire science, the generation of representative mid-story fuels for wildfire simulations. We propose a novel generative model based on a spatial point process. Bayesian model calibration is again used to tune the generative model using sparse survey data so that representative fuels can be defined over a large domain.