When to use RSM (after screening)
Response surface methodology doe is an optimisation technique: it fits a curved (second-order) model to a small number of continuous factors so you can predict the settings that maximise titer, yield, or purity. It belongs at the end of an experimental campaign, after a screening design has already told you which factors matter. Running it first wastes expensive curved designs on factors that may turn out to be inert.
Response surface methodology is the right tool when you have 2–5 critical factors, you suspect the response has an interior optimum (curvature), and you need a precise prediction of the best operating point rather than just a ranking. The classic workflow is screen, then optimise: a Plackett-Burman or fractional factorial design narrows the field, then a central composite design or Box-Behnken design maps the surface around the promising region.
Response surface methodology was introduced by Box and Wilson in 1951 to find optimum operating conditions in chemical processes, and the same logic now drives most bioprocess media and process optimisation. The reason curvature matters is biological: titer-versus-pH or titer-versus-temperature relationships almost always peak at an interior value, so a two-level design that only fits a straight line will point you toward an extreme that is actually past the optimum.
Central composite design (CCD)
A central composite design is the most widely used response-surface design. It takes a two-level factorial cube, adds a set of axial (star) points that stick out beyond the cube faces, and replicates the centre point several times. Those axial points are what let the design estimate the pure quadratic (curvature) term for each factor, because they place each factor at a fifth level beyond its factorial high and low.
For three factors, a central composite design typically has 8 factorial (cube) points, 6 axial points, and around 6 centre replicates — about 20 runs. The centre replicates do double duty: they give a model-free estimate of pure experimental error and they test for lack of fit. The axial distance, often written α, is usually set near 1.68 for three factors to make the design rotatable, meaning prediction variance depends only on distance from the centre, not direction.
The strength of a central composite design is coverage: because the axial points reach beyond the cube, the design models the full region including the extreme combinations. The cost is exactly that — it asks you to run the most extreme settings, which in a bioreactor can mean a combination of high temperature and high pH that the cells may not tolerate.
Box-Behnken design
A box-behnken design is a three-level response-surface design that avoids the extreme corners entirely. Instead of cube vertices and axial points, it places experimental points at the midpoints of the edges of the factor region, plus centre replicates. No run combines all factors at their extreme high or all at their extreme low, which is the property that makes Box-Behnken attractive for biology.
For three factors a box-behnken design uses 12 edge points plus 3 centre replicates — 15 runs, slightly fewer than the equivalent central composite design. It was introduced by Box and Behnken in 1960 specifically to study quantitative variables efficiently at three levels. Because every factor takes only three levels (−1, 0, +1), it is also simpler to set up on equipment that can hold only a few discrete setpoints.
The trade-off is that a Box-Behnken design does not sample the corner region at all, so it cannot predict the response there. If your true optimum sits near an extreme corner, a Box-Behnken design will model the interior well but miss the corner — a situation where a central composite design is the safer choice. For most bioprocess work, though, the optimum is interior and the corners are operationally irrelevant, which is why Box-Behnken is so popular in media optimisation.
CCD vs Box-Behnken (run counts, levels, corners)
The choice between a central composite design and a box-behnken design comes down to three things: how many levels your equipment can hold, whether the extreme corner combinations are safe to run, and how many runs you can afford. Both fit the same second-order model and both give you a contour plot at the end; they differ in where they place the experimental points.
| Property | Central composite (CCD) | Box-Behnken (BBD) |
|---|---|---|
| Levels per factor | 5 (−α, −1, 0, +1, +α) | 3 (−1, 0, +1) |
| Runs (3 factors) | ~20 | 15 |
| Runs (4 factors) | ~30 | 27 |
| Samples extreme corners? | Yes (+ beyond, via axial points) | No |
| Rotatable? | Yes (with α ≈ 1.68 for k=3) | Near-rotatable |
| Best when | Optimum may be near a corner; full region matters | Extreme combinations unsafe/impossible; 3 setpoints only |
Run counts grow quickly with factor number, which is the practical reason response surface methodology is reserved for the 2–5 factors that survived screening. The chart below shows how a central composite design and a Box-Behnken design scale, with a full three-level factorial included to show why nobody runs one for optimisation.
Box-Behnken needs at least three factors — it is not defined for two — whereas a central composite design works from two factors up. For two-factor optimisation, a central composite design (or a small full factorial with centre and axial points) is the natural choice.
Reading a contour plot
A contour plot is how you actually read a response-surface model. It is a 2-D map of two factors with lines joining all combinations that give the same predicted response, exactly like elevation lines on a topographic map. The optimum sits at the centre of the innermost closed contour; tightly packed contours mean the response is sensitive to that factor, widely spaced contours mean it is forgiving.
Because a quadratic model has many terms, you read it one slice at a time: pick two factors for the axes and hold the others at their best values, then draw the contour plot for that slice. A ridge that runs diagonally across the plot is the visual signature of an interaction between the two factors — the optimum of one depends on the level of the other.
Worked example: 3-factor CHO Box-Behnken
You have screened a CHO feed process down to three critical factors and want the titer optimum: A = temperature (33–37 °C), B = pH (6.6–7.2), C = feed rate (low–high). You run a 15-run box-behnken design (12 edge points + 3 centre replicates) and fit a quadratic model. A representative result:
- Titer (g/L) = 3.8 − 0.45·A − 0.30·A² + 0.22·B − 0.38·B² + 0.18·C + 0.15·A·B
- The negative A² and B² terms (p < 0.05) confirm interior optima for both temperature and pH — the surface peaks inside the region.
- Setting the partial derivatives to zero gives an optimum near 34.5 °C, pH 7.0, high feed, with a predicted titer of ~4.0 g/L.
- The significant A·B interaction (the diagonal ridge on the contour plot) means the best temperature shifts slightly with pH — you cannot optimise them independently.
Confirmation: run two or three replicate bioreactors at the predicted optimum. If the measured titer falls within the model's prediction interval, the response surface is validated; if not, the optimum likely sits outside the tested region and the design must be shifted toward it.
Don't extrapolate beyond the design region
The single most common response surface methodology doe mistake is trusting a prediction outside the factor ranges you actually tested. A fitted quadratic is mathematically happy to report an optimum just past the edge of your design, complete with a confident-looking predicted value — but there is no data there, and real biological responses rarely remain quadratic outside the tested window.
If the optimum your model reports sits at or beyond the boundary of the design (for example, "best titer at the maximum tested feed rate"), that is a signal, not an answer. It means the true optimum is probably outside your region. The correct response is to run a new central composite design or Box-Behnken design shifted toward the indicated direction — the classic method of steepest ascent — not to extrapolate the current model. Always sanity-check the predicted optimum against process and biological limits before committing a confirmation run.
Build an RSM design free
You do not need a commercial statistics package to run response surface methodology. A free design of experiments calculator builds both a central composite design and a Box-Behnken design for 2–5 factors, sets the coded levels and axial distance, randomises the run order to protect against time trends, and exports a run sheet ready for the bench.
Enter your factor names and their low and high values, choose CCD or Box-Behnken, and the tool returns the full design. After the runs, fit the quadratic model and read the optimum from the contour plot. Because the same generator also builds the screening designs, you can run a Plackett-Burman screen and then promote the survivors into a response-surface design without changing tools — the full DOE for bioprocess optimization workflow in one place.
Build a CCD or Box-Behnken design now
A free, no-install DOE generator: central composite and Box-Behnken designs, coded levels, randomised run order, and run-sheet export.
Frequently Asked Questions
What is response surface methodology in DOE?
Response surface methodology (RSM) is a collection of DOE techniques for modelling a curved relationship between a few continuous factors and a response, so you can find the settings that maximise or minimise that response. It fits a second-order (quadratic) polynomial using a design like a central composite design or a Box-Behnken design, then uses the fitted model to predict the optimum and draw contour plots of the response surface.
What is the difference between a central composite design and a Box-Behnken design?
A central composite design adds axial (star) points outside the factorial cube, so it samples five levels per factor and reaches the extreme corner-plus-axial region; a Box-Behnken design places points at the edge midpoints and uses only three levels, never visiting the extreme corners. For three factors a CCD needs about 20 runs and a Box-Behnken needs 15. Choose Box-Behnken when the extreme combinations are unsafe or impossible to run, and a central composite design when you need to model the full region including corners.
How many runs does a response surface design need?
For three factors, a central composite design needs about 20 runs (8 cube points, 6 axial points, and 6 centre replicates) and a Box-Behnken design needs 15 runs (12 edge points and 3 centre replicates). Run counts grow with the number of factors, which is why response surface methodology is normally applied to only 2 to 5 critical factors identified by an earlier screening design.
What is a contour plot in response surface methodology?
A contour plot is a 2-D map of the fitted response surface, where lines connect factor combinations that give the same predicted response — like elevation lines on a topographic map. The optimum sits at the centre of the innermost closed contour. Contour plots are the main way engineers read an RSM model, because they show both the location of the optimum and how sensitive the response is to small changes in each factor.
When should I use response surface methodology?
Use response surface methodology after a screening design has reduced your factors to the 2 to 5 that genuinely matter and you suspect the response is curved (has an interior optimum). RSM is an optimisation tool, not a screening tool — running it on too many factors wastes runs, and running it before screening risks spending an expensive curved design on factors that turn out to be inert.
Can I extrapolate beyond the response surface design region?
No. A response-surface model is only valid inside the factor ranges you actually tested. The fitted quadratic can predict a plausible-looking optimum just outside the design region, but that prediction is unsupported by data and frequently wrong because real biological responses rarely stay quadratic outside the tested window. If the optimum sits at the edge of your design, run a new design shifted toward it rather than trusting the extrapolation.
Related Tools
- DOE Experiment Generator — Build central composite, Box-Behnken, and screening designs free in the browser with randomized run order.
- CHO Process Troubleshooter — Diagnose the temperature, pH, DO, and feed factors you would optimise in a response-surface design.
- Media & Feed Estimator — Cost the media and feed regimes you compare across a response-surface study.
References
- 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. DOI: 10.1111/j.2517-6161.1951.tb00067.x
- Box, G.E.P. & Behnken, D.W. (1960). Some new three level designs for the study of quantitative variables. Technometrics, 2(4), 455–475. DOI: 10.1080/00401706.1960.10489912
- Mandenius, C.F. & Brundin, A. (2008). Bioprocess optimization using design-of-experiments methodology. Biotechnology Progress, 24(6), 1191–1203. DOI: 10.1002/btpr.67