Michael Goldstein (University of Durham), Daniel Williamson (University of Exeter), Peter Young (University of Lancaster)
Thu 11 Apr 2019, 09:00 - 12:30
JCMB 5328

If you have a question about this talk, please contact: Stephen Catterall (v1scatte)

This is a series of three seminars on emulation and calibration of complex environmental system simulators, jointly organised by Biomathematics & Statistics Scotland and the University of Edinburgh. All are welcome to attend, no need to register.


9:00 Welcome by Professor Glenn Marion, Head of Research, BioSS

9:15 Professor Michael Goldstein, University of Durham, The Bayes linear approach to emulation and history matching for complex computer simulators

Abstract. This talk gives an introductory overview of the Bayes linear approach for handling uncertainty for complex computer models. The approach is based around careful structural model discrepancy analysis and Bayes linear emulation as a basis for history matching against real system data. The approach will be illustrated in the context of environmental system simulators.

10:10 Coffee

10:40 Dr. Daniel Williamson, University of Exeter, Uncertainty Quantification for Calibrating Spatio-temporal Models using Basis Methods

Abstract. When building emulators for climate models, often model output is in the form of spatio-temporal fields. The two main approaches to emulating these so that the model can be calibrated with observations are either to emulate every point in space and time independently using simple Gaussian processes, or to use a basis to describe the spatio-temporal signal, project the model runs onto this and to emulate the coefficients. Conceptually, the attractions of each approach are clear. Basis methods allow us to capture large scale spatio- temporal correlation structures that have a basis in the underlying model physics and allow us to build fewer emulators. Single grid box methods are conceptually simpler and, for any given grid box, the uncertainty may be lower as there is no loss of information due to discarding low-signal basis vectors. In this talk I will present our novel approaches with basis methods and compare their performance in calibration for emulating climate model fields with the single grid box approach.

11:35 Professor Peter Young, University of Lancaster, Nominal or Full Dynamic Model Emulation?

Abstract. The approach to Dynamic Model Emulation (DME) considered in this presentation is part of the Data-Based Mechanistic (DBM) modelling philosophy. The aim is to exploit Dominant Mode Analysis (DMA) to construct a stand-alone emulation of a high order computer simulation model in one of two possible forms: Nominal Emulation, which is quick and simple but restricted to models with specific parameter values; and Full Emulation which is much more complex but provides for emulation over a whole, user-defined, range of parameter values. The presentation reviews this two-stage approach to emulation and discusses the merits of the two emulation model forms using a model for the transport and dispersion of a solute in a river as an illustration. If time allows, the nominal emulation of large (AOGCM) global climate models from step response data will also be considered.