Skip to content
Back to search results

Central Composite Design — Response Surface Experimental Design

Introduced by
George E. P. Box and K. B. Wilson
Year
1951
Type
Response surface experimental design
Family
Engineering methods
Font size

What does it do?

Central Composite Design (CCD) is a second-order response surface design that allows researchers to efficiently fit a full quadratic model relating multiple continuous input factors to one or more response variables.

Introduced by Box and Wilson in 1951, it combines a factorial (or fractional factorial) core, axial (star) points, and center-point replicates into a single unified design, making it the most widely used design for process optimization in engineering, chemistry, and manufacturing.

The intuition behind the method

Suppose you are optimizing the yield of a chemical reaction by varying temperature and pressure. A simple factorial design tells you about main effects and interactions at just two levels per factor, but misses the curved surface that often exists near an optimum. CCD adds two extra points beyond the factorial corners along each factor axis — the axial or star points — so the experiment reaches out far enough in each direction to estimate curvature. A few center-point runs are added to measure pure experimental error and check that the relationship really is curved rather than linear. Together these three sets of points let you map the entire response surface with far fewer runs than a full three-level factorial would require.

History

Central Composite Design was introduced by George E. P. Box and K. B. Wilson in their landmark 1951 paper in the Journal of the Royal Statistical Society, which simultaneously introduced the concept of response surface methodology. Box and Wilson showed that combining a factorial design with axial points allowed the curvature of the response surface to be estimated and an optimum to be located with far fewer experiments than existing methods required. The design became central to industrial process optimization throughout the 1950s and 1960s and was popularized by Box and Draper's textbook Empirical Model-Building and Response Surfaces (1987). Today CCD remains the most widely cited and applied second-order experimental design across engineering, chemistry, and the life sciences.

When to use it

Use CCD when you need to fit a second-order (quadratic) model to map and optimize a continuous response over a region of k quantitative factors (typically k = 2–6). It is the standard choice in process optimization studies in chemical engineering, pharmaceutical development, food science, and manufacturing when you expect curvature in the factor–response relationship or when a first-order model has already shown significant lack of fit. CCD is preferred over Box-Behnken design when you can afford (or need) factor settings at the extreme axial points beyond the factorial range, and over a 3^k full factorial when run economy matters. Do NOT use CCD with categorical factors, binary or count response variables, or when the experimental region is geometrically constrained such that the axial points fall outside feasible operating ranges (in which case a Box-Behnken or D-optimal design is more appropriate). Also avoid CCD if a first-order model with a steepest-ascent strategy has not yet been used to move the design to the vicinity of the optimum — fitting a quadratic far from the optimum wastes runs.

Limitations and considerations

  • Requires all factors to be continuous and quantitative; cannot handle categorical factors without modification.
  • Axial points at distance alpha > 1 extend beyond the factorial cube, which may be infeasible when factor levels are physically constrained.
  • As k grows (k > 5), the total run count increases rapidly and resource demands become substantial.
  • The quadratic model is a local approximation; if the true response surface has higher-order nonlinearity, the CCD model may still show lack of fit near the optimum.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Efficiently estimates all coefficients of a full second-order polynomial with fewer runs than a 3^k factorial design.
  • Rotatability: the face-centered or rotatable CCD provides equal prediction variance at all points equidistant from the design center, giving uniform predictive precision.
  • Center-point replicates provide an independent estimate of pure error, enabling a valid lack-of-fit test.
  • Flexible: can be built sequentially — run the factorial core first (as a screening design), then augment with axial and center points if curvature is detected.
  • Well supported in standard statistical software (JMP, Minitab, R rsm package, SAS PROC RSREG).

Applications

  • Pharmaceutical process development: optimizing tablet compression force, granulation moisture, and binder concentration to maximize dissolution rate and hardness simultaneously.
  • Chemical engineering: mapping yield and selectivity as functions of reaction temperature, pressure, and catalyst loading to find operating optima.
  • Food science: optimizing baking or fermentation conditions (time, temperature, ingredient ratios) for texture and sensory quality.
  • Environmental engineering: modeling pollutant removal efficiency as a function of pH, contact time, and adsorbent dose in water treatment studies.
  • Manufacturing: optimizing machining parameters (cutting speed, feed rate, depth of cut) to minimize surface roughness and tool wear.
  • Materials science: fitting strength or conductivity as a function of composition and processing conditions to guide alloy or polymer formulation.

How it works

How it works

1

Identify the k continuous quantitative factors to be studied and set their low and high factorial levels (coded -1 and +1). Choose the axial distance alpha — for rotatability alpha = (2^k)^(1/4); for face-centered CCD alpha = 1. Determine the number of center-point replicates n0 (typically 3–6) to estimate pure error and stabilize variance.

2

Assemble the three components: (1) a 2^k full factorial or a suitable 2^(k-p) fractional factorial core providing 2^k (or fewer) corner runs; (2) 2k axial (star) points placed at distance alpha from the center along each factor axis; and (3) n0 center-point runs at the origin. The total run count is N = 2^k + 2k + n0. For k=2 factors this is typically 13 runs; for k=3, about 20 runs.

Formula
N = 2^k + 2k + n_0
3

Conduct the N experimental runs in a randomized order to prevent systematic bias from time trends or equipment drift. Record the response value(s) for each run. If resources allow, blocking the factorial and axial portions into separate experimental sessions preserves rotatability while controlling for day-to-day variation.

4

Use least squares regression to fit the full second-order polynomial: y = beta0 + sum(beta_i * x_i) + sum(beta_ii * x_i^2) + sum(beta_ij * x_i * x_j) + epsilon. This model captures main effects, pure quadratic (curvature) effects, and two-factor interactions. The CCD point structure ensures all second-order coefficients are estimable.

Formula
y = \beta_0 + \sum_{i}\beta_i x_i + \sum_{i}\beta_{ii}x_i^2 + \sum_{i<j}\beta_{ij}x_ix_j + \varepsilon
5

Evaluate model adequacy through analysis of variance (ANOVA), the coefficient of determination (R^2 and adjusted R^2), residual plots, and a lack-of-fit test (enabled by the center-point replicates). Confirm that the model meets regression assumptions — normality and constant variance of residuals — before drawing conclusions.

6

Use the fitted response surface to locate the optimum — a maximum, minimum, or saddle point — via canonical analysis, contour plots, or numerical desirability-function optimization. Verify the predicted optimum with one or more confirmation runs at the optimal factor settings before committing to operational changes.

Common pitfalls

  • Running the design far from the true optimum — CCD is a second-order design intended for fine-tuning near an optimum, not for broad exploration; use screening designs and steepest-ascent first.
  • Ignoring the lack-of-fit test — center points are there precisely to test whether the quadratic model fits; a significant lack-of-fit signal means the model is inadequate and optimization conclusions are unreliable.
  • Failing to randomize run order — systematic execution of the factorial block followed by axial and center runs without randomization confounds time effects with factor effects.
  • Choosing alpha without checking feasibility — a rotatable CCD (alpha = (2^k)^(1/4)) places axial points well beyond the factorial range; always verify that the resulting factor levels are physically achievable before finalizing the design.
  • Over-interpreting the surface far from the design region — the fitted quadratic is reliable only within the experimental region; extrapolating to find a global optimum outside that region is statistically unjustified.

Frequently asked questions

What is the difference between a face-centered CCD and a rotatable CCD?

In a face-centered CCD (alpha = 1) the axial points lie on the faces of the factorial cube, so all factor settings stay within the [-1, +1] range — useful when axial excursions are not feasible. In a rotatable CCD alpha = (2^k)^(1/4), placing axial points farther out; this gives equal prediction variance at all points the same distance from the center (rotatability) but requires factor levels beyond the factorial range. Choose face-centered when physical constraints prevent extreme settings; choose rotatable when uniform predictive precision across the design sphere is the priority.

How many center-point replicates should I use?

The standard recommendation is 3–6 center-point replicates. Three is the practical minimum to get a meaningful lack-of-fit test and a stable pure-error estimate. Using 5–6 center points improves the stability of the variance estimate and is advisable when experimental noise is expected to be high. More than six center points rarely add proportionate value and increase total run count.

Can I use CCD when I have more than six factors?

Yes, but run economy deteriorates. For k > 5 the full 2^k factorial core becomes expensive, so practitioners typically substitute a resolution V fractional factorial, creating a fractional central composite design. Software such as JMP or Minitab automates this construction. For k > 6 a D-optimal or I-optimal computer-generated design often provides a more run-efficient alternative to any classical CCD.

How does CCD compare to Box-Behnken design?

Both are second-order response surface designs for k >= 3 factors. CCD includes corner (factorial) points plus axial points plus center points; Box-Behnken uses only edge midpoints and center points, so it never requires factor combinations at the extreme corners of the design space. Box-Behnken is preferred when corner settings are physically dangerous or infeasible and when alpha > 1 axial points cannot be achieved. CCD is preferred when corner settings are feasible and when the sequential (augmentation) strategy is desired, because the factorial block can serve double duty as an earlier screening experiment.

What software can I use to set up and analyze a CCD?

All major statistical packages support CCD: Minitab (Response Surface > Central Composite), JMP (Custom Design or Response Surface platform), SAS PROC RSREG, and R with the rsm package (ccd() function). The rsm package is freely available and provides full design generation, model fitting, and surface plotting in one workflow.

Related methods

Sources

  1. Box, G. E. P., & Wilson, K. B. (1951). On the experimental attainment of optimum conditions. Journal of the Royal Statistical Society: Series B, 13(1), 1–45.

    Open DOI reference
    Copied.
  2. Montgomery, D. C. (2017). Design and Analysis of Experiments (9th ed.). Wiley.

    ISBN: 978-1119492443
    Copied.