Thursday 13 Oct 2016: Statistical Science Seminar: Calibrating climate models with spatial data: beyond principal components
James Salter and Daniel Williamson - University of Exeter
In the majority of climate applications, spatial fields are calibrated by projecting the data onto a basis derived from the ensemble, before constructing emulators and calibrating based on these projections (Higdon et al 2008). However, our work has suggested that this approach may be flawed when the ensemble, and hence the basis, does not contain patterns that are also found in the observed data that we are intending to calibrate our model to. Ignoring patterns that may be important can lead to concluding that the computer model does not have parameter choices that are consistent with the observations, or worse, highlighting regions of parameter space that lead to poor reconstructions, when the model may actually be able to more accurately represent the observations.
We present a methodology for calibration using spatial observations that overcomes this flaw in current approaches. Our method allows us to search along elicited physically important directions (or spatial patterns) and develops a procedure for optimal projection-based computer model calibration for spatial model output. We demonstrate an application of our method using the Canadian climate model.