Calibration of computer models with multivariate output

The problem of calibrating computer models that produce multivariate output is considered, with a particular emphasis on the situation where the model is computationally demanding. The proposed methodology builds on Gaussian process-based response-surface approximations to each of the components of the output of the computer model to produce an emulator of the multivariate output. This emulator is then combined in a statistical model involving field observations, which is then used to produce calibration strategies for the parameters of the computer model. The results of applying this methodology to a simulated example and to a real application are presented.