DOE Design Generator

What are DOE design matrices and when should you use each type?

Design of Experiments (DOE) uses structured design matrices to systematically explore the effect of multiple factors on a response. Each row represents an experimental run and each column represents a factor set to a specific level. The choice of design type depends on the number of factors, available resources, and experimental objectives.

Full Factorial (2^k) tests all combinations but grows exponentially. Fractional Factorial (2^(k-p)) uses a fraction of runs while maintaining estimability of main effects. Plackett-Burman designs are screening designs for identifying significant factors among many candidates. CCD and Box-Behnken add center and axial points for response surface modeling.

Taguchi orthogonal arrays (L4, L8, L9, etc.) are pre-built fractional designs focused on robustness. Definitive Screening Designs (DSD) are modern alternatives that screen factors and estimate quadratic effects in a minimal number of runs. All designs use coded levels (−1, 0, +1) for standardized analysis.

Formula: Full Factorial: Runs = 2^k (or 3^k for 3-level) Fractional Factorial: Runs = 2^(k-p) CCD: Runs = 2^k + 2k + center points Box-Behnken: Runs ≈ 2k(k-1) + center points Plackett-Burman: Runs = multiple of 4, N ≥ k+1

Example Calculation

For 4 factors: Full Factorial = 2⁴ = 16 runs. Half-fraction (p=1) = 2³ = 8 runs. CCD = 16 + 8 + 5 center = 29 runs. Box-Behnken = 24 + 3 center = 27 runs. A Plackett-Burman with 4 factors needs 8 runs. Choose based on available resources and whether interactions or curvature need estimation.

When to Use This Calculator

• At the start of a DOE project to generate the correct design matrix for your chosen experimental design type and factor count
• When comparing design options — generate matrices for Full Factorial, Fractional Factorial, and Plackett-Burman to compare run counts and resolution
• When a response surface design (CCD or Box-Behnken) is needed for optimization after screening has identified the important factors
• When using Taguchi methods and need to select the appropriate orthogonal array for the number of factors and levels
• During experiment documentation to produce a clean, randomized design matrix for the lab or shop floor team to execute

Common Mistakes to Avoid

• Using a full factorial when you have 6+ factors — 2^6 = 64 runs is often impractical; use a fractional factorial or screening design for initial investigation
• Choosing a fractional design with too low resolution — Resolution III designs alias main effects with 2-factor interactions; use Resolution IV or higher if interactions may be important
• Not randomizing run order — the generated matrix should be randomized before execution to protect against lurking variables like time, temperature drift, or operator fatigue
• Using CCD or Box-Behnken when only screening is needed — these designs are for response surface modeling; for screening, simpler designs (Plackett-Burman, Definitive Screening) are more efficient
• Selecting the wrong Taguchi array size — an L8 can handle up to 7 factors at 2 levels; overloading the array by adding more factors than columns leads to complete aliasing

How to Interpret Results

• Each row of the design matrix is one experimental run and each column is a factor in coded units: −1 is the low level, +1 the high level, 0 a center point — convert to real units with actual = center + coded × half-range
• Rows of all zeros are center points — they estimate pure error and detect curvature; CCD and Box-Behnken designs include them by default
• In a Rotatable CCD, axial-point values beyond ±1 mean those runs are set outside the factorial range — verify those settings are physically safe and achievable; choose Face-Centered or Inscribed if levels must stay within bounds
• Compare the Number of Runs against your experimental budget — if the run count is unaffordable, increase the generators (more fractionation) or switch to a screening design (Plackett-Burman, Definitive Screening) and accept reduced interaction information
• Randomize the run order before executing — the matrix is generated in standard order, and running it as-listed confounds factor effects with time-related drift

Related Standards & References

• NIST/SEMATECH e-Handbook of Statistical Methods — Comprehensive reference for DOE design selection and analysis methodologies
• AIAG APQP 3rd Edition — Advanced Product Quality Planning, recommending DOE for process optimization during product development
• ISO 3534-3 — Statistical vocabulary related to design of experiments
• JIS Z 9090 — Japanese standard for Taguchi orthogonal array experiment methodology

Frequently Asked Questions

How do I choose between Full Factorial and Fractional Factorial?

Use Full Factorial when you have ≤ 5 factors and can afford 2^k runs, as it estimates all interactions. Use Fractional Factorial when resources are limited or you have 6+ factors — it assumes higher-order interactions are negligible (sparsity of effects principle). Start with a screening fraction, then augment to a full design for significant factors.

When should I use CCD vs Box-Behnken designs?

Use CCD (Central Composite Design) when you need to estimate a full quadratic model and can run axial points outside the factorial cube (e.g., factor levels beyond ±1). Use Box-Behnken when factor levels must stay within bounds (no axial points), or when you have exactly 3-7 factors. Box-Behnken never tests all factors at extreme levels simultaneously.