DoE Advisor

Overview

The DoE Advisor helps you choose the right experimental design for a Six Sigma project. The choice of design decides whether a limited number of runs yields robust insight into main effects, interactions and optima — or whether effort is wasted.

Overview: Side-by-side table of all designs relevant in everyday Six Sigma (full factorial 2^k, general full factorial, fractional, Plackett-Burman, Definitive Screening Design, CCD rotatable and face-centered, Box-Behnken, Taguchi, dispersion DoE, mixture designs, D-optimal, EVOP) with factor range, levels, effort and key advantage.

Advisor: Guided question wizard. From DMAIC phase, objective and constraints (factor limits, allowed corner points, curvature, mixtures, robustness) it derives a concrete design recommendation with reasoning.

CCD comparison: Direct comparison of the two most common response-surface designs: rotatable (CCC) versus face-centered (CCF) — star point location, α value, number of levels, rotatability, typical applications.

This module does not replace planning in the DoE Planner — it comes before: at the selection step. Once the design is chosen, the DoE Planner generates the concrete plan and the Worksheet evaluates it.

Approach

• Clarify the DMAIC phase — Define/Measure carries no DoE yet; Analyze means screening, Improve means optimization, Control means EVOP.
• Sharpen the objective — screening (which factors?), understanding interactions, finding the optimum, robustness against noise, or formulating a mixture.
• Count factors — few (≤4) allow full factorial; many (>10) call for Plackett-Burman.
• Decide on factor levels — pure 2-level factors fit 2^k; multi-level or categorical factors (material A/B/C, 3 pressure levels) need the general full factorial.
• Pick the robustness path — for variance reduction, ask whether explicit noise factors are available (Taguchi inner-outer array) or whether dispersion should be modeled directly from replicates (dispersion DoE / dual response).
• Check factor limits — are they expandable (CCD rotatable) or hard due to safety/spec limits (CCD face-centered)?
• Check corner points — are extreme combinations of all factors physically allowed (full factorial / CCD)? If not: Box-Behnken or D-optimal.
• Account for curvature — if quadratic effects are suspected, choose DSD or an RSM design (CCD / Box-Behnken).
• Implement the recommendation in the DoE Planner, transfer data to the Worksheet, analyze via regression / ANOVA.

Pitfalls

Full factorial with too many factors: A full factorial with 8 factors means 256 runs — too much even with a generous budget. From 5–6 factors onward, always consider fractional, Plackett-Burman or DSD.

Resolution ignored: A 2^(7-3) plan is not automatically a good choice — resolution (III, IV, V) decides which effects are aliased with which. Resolution III for screening only; for interactions at least IV, better V.

CCD outside factor limits: In a rotatable CCD the star points lie OUTSIDE the original factor range (α > 1). With hard limits (pressure, temperature, safety) this is not allowed — pick face-centered (α = 1) instead.

Box-Behnken with too few factors: Box-Behnken needs at least 3 factors. It is undefined for two factors; use CCD or a full factorial with center points instead.

Taguchi instead of interactions: Taguchi arrays are primarily optimized for robustness (signal-to-noise); interactions are hard to separate from main effects. To understand interactions, use 2^k or CCD.

Mixtures treated as independent factors: In mixtures the proportions sum to 100 %, so components cannot vary independently. Full factorial or CCD are mathematically wrong — use Simplex-Lattice / Simplex-Centroid.

Multi-level factors squeezed into a 2^k plan: Categorical factors with three or more levels (material A/B/C) cannot be meaningfully reduced to "high/low". Use the general full factorial instead — effort equals the product of the level counts.

Dispersion DoE without true replicates: A dispersion design requires true, independent replicates per run — only then is log(s²) a valid response. Multiple measurements on the same run (pseudo-replicates) capture only measurement uncertainty, not process variability.

Available in the following cycles

• DMAIC: Improve
• DMADV: Design
• 8D: D5 — Corrective Actions