Estimation of input parameters in complex simulation using a Gaussian process metamodel

Abstract The unknown input parameters of a simulation code are usually adjusted by the nonlinear least squares estimation (NLSE) method which minimizes the sum of differences between computer responses and real observations. However, when a simulation program is very complex and takes several hours for one execution, the NLSE method may not be computationally feasible. In this case, one may build a statistical metamodel which approximates the complex simulation code. Then this metamodel is used as if it is the true simulation code in the NLSE method, which makes the problem computationally feasible. This ‘approximated’ NLSE method is described in this article. A Gaussian process model is used as a metamodel of complex simulation code. The proposed method is validated through a toy-model study where the true parameters are known a priori. An application to nuclear fusion device is presented.