Show simple item record

dc.contributor.authorVolodina, V
dc.date.accessioned2020-06-08T09:21:41Z
dc.date.issued2020-06-01
dc.description.abstractIn this thesis, we provide the Uncertainty Quantification (UQ) tools to assist automatic and robust calibration of complex computer models. Our tools allow users to construct a cheap (statistical) surrogate, a Gaussian process (GP) emulator, based on a small number of climate model runs. History matching (HM), the calibration process of removing parameter space for which computer model outputs are inconsistent with the observations, is combined with an emulator. The remaining subset of parameter space is termed the Not Ruled Out Yet (NROY). A weakly stationary GP with a covariance function that depends on the distance between two input points is the principal tool in UQ. However, the stationarity assumption is inadequate when we operate with a heterogeneous model response. In this thesis, we develop diagnostic-led nonstationary GP emulators with a kernel mixture. We employ diagnostics from a stationary GP fit to identify input regions with distinct model behaviour and obtain mixing functions for a kernel mixture. The result is a continuous emulator in parameter space that adapts to changes in model response behaviour. History matching has proven to be more effective when performed in waves. At each wave of HM, a new ensemble is obtained to update an emulator before finding an NROY space. In this thesis, we propose a Bayesian experimental design with a loss function that compares the volume of the NROY space obtained with an updated emulator to the volume of the “true” NROY space obtained using a “perfect” emulator. We combine Bayesian Design Criterion with our proposed nonstationary GP emulator to perform calibration of climate model.en_GB
dc.identifier.urihttp://hdl.handle.net/10871/121314
dc.language.isoenen_GB
dc.publisherUniversity of Exeteren_GB
dc.subjectUncertainty Quantificationen_GB
dc.subjectHistory matchingen_GB
dc.subjectBayesian optimal designen_GB
dc.subjectKernel mixtureen_GB
dc.subjectNonstationary Gaussian Processesen_GB
dc.subjectEmulationen_GB
dc.titleUncertainty Quantification for complex computer models with nonstationary output. Bayesian optimal design for iterative refocussingen_GB
dc.typeThesis or dissertationen_GB
dc.date.available2020-06-08T09:21:41Z
dc.contributor.advisorWilliamson, Den_GB
dc.publisher.departmentMathematicsen_GB
dc.rights.urihttp://www.rioxx.net/licenses/all-rights-reserveden_GB
dc.type.degreetitlePhD in Mathematicsen_GB
dc.type.qualificationlevelDoctoralen_GB
dc.type.qualificationnameDoctoral Thesisen_GB
rioxxterms.versionNAen_GB
rioxxterms.licenseref.startdate2020-06-01
rioxxterms.typeThesisen_GB
refterms.dateFOA2020-06-08T09:21:45Z


Files in this item

This item appears in the following Collection(s)

Show simple item record