A combined array approach to minimise expected prediction errors in experimentation involving mixture and process variables

This paper proposes experimental planning methods for experiments involving mixture variables. These methods generally foster relatively accurate predictions using fewer experimental runs. Methods are based on a criterion that is the expected squared value of these errors, taking into account model bias and the application of alternative candidate sets. We compare the proposed methods with alternatives, using case studies from the literature. Further, we show that computer generated combined arrays based on new, combined models can be expected to yield lower prediction errors compared with relevant alternatives including design methods combining central composites and simplex centroids and D-optimal block designs.