Application of covariance structures to improve the fit of response surfaces to simulation data

This paper presents a covariance-based interpolating procedure for constructing response surfaces that pass through all the data points created by computer-based experiments. The justification for this approach is that simulation-based experiments, in general, do not suffer from random errors. The paper presents the method in the context of fitting response surfaces to TCAD results for VLSI processes. It is shown that there can be a real benefit from using these covariance-based models when polynomial-based response surfaces have difficulty accurately representing the simulated variation. A number of examples are presented, one of which is a plateau region next to a steep cliff, which is well known to present a severe challenge to the fitting of response surfaces.