reduce runs in RPD

[paulprew]

Hello, I'm working on a Robust Parameter Design, 5 control factors and 2 noise factors. The factors are being treated as nominal 2-level, except one 3-level. Following Montgomery's Design text, a D-optimal design is used to model a combined array. This results in 44 runs.

Assumptions
•Model complexity is at the level of 2-factor interactions. Pure quadratic effects are not examined (no term for effect of Nozzle.press^2).
•To reduce the number of experiments, no interaction term is modeled for the ‘noise’ factors Press.inlet*Temp.inlet. It’s not necessary to understand the noise variables at that level of detail.

Factor Name Function Units Type Subtype Low Actual Mid Actual High Actual
A Screen.size Control opi Categoric Nominal 4x4 10x10 Levels: 2
B Flow.reg Control psi Categoric Ordinal 2 3.5 Levels: 2
C Nozzle.type Control gpm/angle Categoric Nominal 0.39/84 0.50/65 0.56/90 Levels: 3
D Inlet.OD Control inch Categoric Nominal 3/8 1/2 Levels: 2
E Nozzle.press Control psi Categoric Nominal 5 15 Levels: 2
F Temp.inlet Noise C Categoric Nominal 90 140 Levels: 2
G Press.inlet Noise psi Categoric Nominal 20 100 Levels: 2


3 questions:
1) Can the number of runs be reduced in an intelligent manner? I added 4 replicates to the 5 lack-of-fit runs suggested by DE. The power is quite high, over 90% for effects >= 2 standard deviations (except the interactions with Nozzle.type, the 3-level factor, are < 80%).

2) Should I be using the standard deviation attribute for POE? In Montgomery's text, the noise variable is typically set at standard deviation = 1. How does that fit in with the fact that e.g. noise variable Press.inlet is being tested at levels = 20, 100?

3) The engineer wonders if the Flow.reg variable should be tested at 3 levels (2, 3, 3.5), and changed to a continuous variable. Any advice on whether this is advisable?

thanks, Paul