High anisotropy space exploration with co-Kriging method

For a black box data-space exploration, classical DoE method prescribes a cluster of quasi-isotropic design points following a certain space-infill criterion, however the objective function behaves. Adding new points often disturbs the original spatial-infill properties, which restricts dimension-augmentation and reusability of data. Surrogate model based sequential design considers not only the spatial features of the design space but also the characteristic of the objective function. The latter may favor a highly anisotropic design for the sake of computational effort effectiveness, especially when highly informative samples are used. This work presents a high anisotropy design solution when derivative-assisted co-Kriging method is employed. Adaptation of this method has been proposed for the special case where only one single sample appears at certain dimensions. A 2-dimensional test function with distinctive spatial properties at each dimension is composed and used to assess the validity of the proposed method.