Resources What's New Support Download Contact Sales: Arthur-Industrial@outlook.com
Home › Resources › Design of experiments

Design of experiments

Why one-factor-at-a-time fails, factorial designs, fractional designs and resolution, interactions, response surfaces and robust design.

Design of experiments (DOE) changes several input factors together in a planned pattern so that their effects, and how they interact, can be estimated with few runs.

Why not one factor at a time?

Changing one factor at a time needs more runs, cannot see interactions (the effect of temperature depends on pressure), and can miss the best settings entirely.

Factorial designs

  • A full factorial tests k factors at a low and high level in every combination: 3 factors in 8 runs, 4 in 16.
  • Main effect: average response at the high level minus at the low level.
  • Interaction: half the difference between a factor's effect at the high and low levels of another factor.
  • Fractional factorials () run a planned fraction when there are many factors. Some effects become aliased (confounded).
  • Resolution: III (main effects aliased with two-factor interactions; screening only), IV (main effects clear), V (main effects and two-factor interactions clear).
  • Plackett-Burman and definitive screening designs screen many factors efficiently.

Good practice

  • Choose factor levels far enough apart to see an effect, but within safe operation.
  • Randomize run order to protect against drift.
  • Replicate runs to estimate pure error, or add center points to check for curvature.
  • Block runs when conditions change (days, material lots).
  • Confirm the result with confirmation runs at the chosen settings.

Response surface methods

When curvature matters and you want an optimum, use central composite or Box-Behnken designs and fit a quadratic model. Contour plots and an optimizer find the best settings for one or several responses.

Robust design

Taguchi's idea: choose settings that make the product insensitive to noise factors you cannot control (material variation, ambient temperature), and measure success with the variation of the response, not just its mean.

Questions and answers

How many factors should we test?

Screen many factors (6 to 15) with a small fractional or definitive screening design, then study the 2 to 4 that matter in more detail.

The ANOVA shows no significant effects. Was the experiment wasted?

Not if the design had enough power: you learned that those factors, in that range, do not matter much. Check that the levels were wide enough and the measurement system good enough.

In RS-1: DOE tab: screening, definitive screening, response surface, mixture and Taguchi designs, and their analysis.

← All articles