Modeling of Computer Experiments with Different Mesh Densities

This article considers deterministic computer experiments with real-valued tuning parameters which determine the accuracy of the numerical algorithm. The aim is to integrate computer outputs with different tuning parameters. Finite element method with mesh density as the tuning parameter is an important case. Novel nonstationary Gaussian process models are proposed to establish a framework consistent with the results in numerical analysis. Maximum entropy designs are considered for experimental planning. Numerical studies show the advantages of the proposed method. The methodology is illustrated with a problem in casting simulation.