On the use of supersaturated designs in DOE

Fractional factorial designs are popular in screening investigations because they allow the investigator to map the influence of many factors in relatively few runs. Sometimes, however, the number of runs encoded by a fractional factorial design might be prohibitive, e.g. due to time or economical constraints. In such a situation a so called supersaturated design might be a viable alternative. A supersaturated design is a reduced design in which the number of investigated factors exceeds the number of runs. This means that the run size is not large enough for estimating all the main effects of the factors, and so there will be complex confounding among the model terms. The objective of this webinar is to give an introduction to supersaturated designs, what they are and how they can be used.

A central concept in screening DOE is the design resolution. Resolution V and V+ can be said to correspond to the highest level of information retrievable from fractional factorial designs. This high resolution is accomplished when N >> K (runs >> factors) and corresponds to regression models in which two-factor interactions are unconfounded with each other. It should be realized, however, that for screening purposes resolution V/V+ designs encode unnecessarily many experiments. Resolution IV designs, on the other hand, are more apt for screening since they demand fewer runs and still enable two-factor interactions to be computed clear of one another. Roughly, resolution IV designs have N ≈ 2K (for up to eight factors). Resolution III designs encode still fewer runs (N ≈ K), but now main effects and two-factor interactions are confounded. Finally, we have designs of resolution II where N < K, i.e., the so called supersaturated designs, for which complicated confounding patterns arise.

Topics for this webinar:

- Introduction to supersaturated design
- Implementation in the MODDE 12
- Criticality of Effects sparsity assumption
- Design resolution and its consequences for model estimation
- Data analytics using PLS
- Data analytics using MLR