DoE Planner
Design of Experiments: Full Factorial, Fractional, Plackett-Burman, CCD, Box-Behnken, Taguchi
Overview
Design of Experiments (DoE) plans experiments so that maximum insight about several factors is gained from as few runs as possible. While the classical "one factor at a time" approach needs many runs and misses interactions, DoE delivers structured designs that efficiently separate main effects and interactions.
Factor: An input deliberately varied — e.g. temperature, pressure, time, material type. Factors can be continuous (temperature in °C) or categorical (supplier A/B).
Level: The concrete values a factor is set to in the experiment. 2 levels (−1 / +1) are standard for screening, 3+ levels for optimization (e.g. response surface).
Response: The output to optimize or understand — e.g. yield, strength, cycle time. Measured once per run.
Full factorial: All combinations of all levels are tested. With k factors at 2 levels: 2ᵏ runs. Exact but quickly expensive (6 factors = 64 runs).
Fractional factorial: A deliberately chosen fraction of the full design — e.g. a half (2ᵏ⁻¹) or quarter. Efficient for screening many factors, but specific interactions become confounded (aliased).
Interaction: Two factors do not act independently — the effect of A depends on the level of B. Detecting interactions is one of DoE's main benefits; they are entirely missed in the "one factor at a time" world.
Center point: Run(s) at the center of the design region (all factors at 0). Checks for curvature and estimates pure variation.
Dispersion DoE (dual response): For designs with at least two replicates per point, the planner additionally fits a model for ln(s²). This identifies factor settings that reduce process variability independently of the mean — shown in the analysis section below the design quality summary.
DoE is the hardest weapon of the Improve phase — if you don't know which factors matter, a screening design answers in a few runs. If you already know your factors and want to optimize, use response-surface designs.
Approach
- Clearly define the response — what should be optimized or understood?
- Pick factors and sensible levels — with process experts, not at the desk.
- Identify noise variables to be held constant or randomized.
- Choose a design type: screening (Plackett-Burman, fractional 2ᵏ⁻ᵖ), optimization (response surface, central composite), mixtures.
- Randomize — run order random, not systematic.
- Run the experiments, measure the response, record data carefully.
- Fit the model, assess main effects and interactions.
- Confirm with runs at the predicted best settings.
Pitfalls
No randomization: Running experiments in the standard-table order confounds factor effects with drift over time (tool wear, temperature change). Randomization is mandatory, not optional.
Levels too narrow: When levels sit close together, the effect hides in the noise. Be bold with wide levels — they need only stay meaningful, not "safe".
Too many factors, too little knowledge: Studying 10 factors at once yields a huge design and little time per effect. Rule of thumb: prior knowledge and prioritization first, DoE second — not the other way around.
Interactions ignored: A half-fractional design confounds certain main effects with interactions. Before analyzing, inspect the alias structure, otherwise conclusions will be wrong.
One data point per run: A single value per condition provides no noise estimate. Use at least two replicates or center points, otherwise pure variation remains unknown.
Model without confirmation: The found optimum is a prediction — until it is run on the real process, it is just an equation. Confirmation runs are the decisive final step.
Extrapolation: The model applies inside the studied region. Predictions outside (e.g. even higher temperature than ever tested) are pure speculation.
Examples
This module ships with the following example datasets — load any of them in the app with a single click.
Available in the following cycles
- DMAIC: Improve
- DMADV: Design
- 8D: D5 — Corrective Actions