The design of most modern systems requires the tight integration of multiple disciplines. In practice, these multiple disciplines are often optimized independently, given only fixed values or targets for their interactions with other disciplines. The result is a system that may not represent the optimal system-level design. It may also not be a robust design in the sense that small changes in each subsystem’s performance may have a large impact on the system-level performance. The use of kriging models to represent the response surfaces of subsystems that are then combined to estimate system-level performance can be used as a method to provide collaboration between design teams. The difficulty with this method is the creation of the models given potentially large number of dimensions or observations. This paper presents a method to reduce the dimensionality of the input space for kriging models used for designing of complex systems. The input dimensionality of the kriging model is reduced to only includes the most important factors needed for the prediction of the observed output. A result of using these reduced dimensionality models is the need to no longer force interpolation of all of the observations used to create the models.