Computer model interpolation and validation with gaussian process emulators


Christopher E. Coleman-Smith

Interesting physical questions frequently require the creation of complex computational models, furthermore really interesting questions require models of such complexity that sampling their output over the full phase space of the problem would require semi-infinite amounts of computer time. A further fundamental problem with computational models is to ascertain how well the model adequately represents reality?

I will discuss a method to mitigate the former issue by the construction of an emulator, or surrogate for the computational model which provides a probability distribution for the model's output at all locations in the parameter space. I outline a method for using the emulator, along with field data, to validate the model and it's departure from "reality". The emulator may also be used to globally estimate the sensitivity of the underlying model upon it's parameters.

Back to seminar 2010