Comparison of Three Surrogate Modeling Techniques: Datascaper, Kriging, and Second Order Regression

Using surrogate models in place of high fidelity engineering simulations can help reduce design cycle times and cost by enabling rapid analysis of alternative designs. Surrogate models can also be used in a deliverable product as an efficient replacement for large lookup tables or as a soft sensor to predict quantities than cannot be directly measured. Many different surrogate modeling techniques exist, including new commercial technologies, each with different capabilities and pitfalls. The goal of this research is to aid the designer in selecting the appropriate surrogate model by comparing two popular techniques, second order regression and kriging, along with a new commercial application called Datascape. The three different modeling techniques are compared on model accuracy, computational efficiency, robustness, transparency, and ease of use. The comparisons were done using three test problems: an Earth-Mars transfer orbit problem, the analytic Shekel function, and a low Earth orbit three-satellite constellation design problem. It was found that kriging models performed the best when the sample data used to build the models was sparse, when larger sample sets were used Datascape produced more accurate models.