Box-Behnken
Edge midpoints, never the extreme corners.
Box-Behnken fits a full quadratic response surface with fewer runs than a central composite design, and it never sets every factor to its extreme at once. For processes where the all-corners combination is unsafe or impossible, that makes it the safer, cheaper choice.
What it measures
Box-Behnken is a three-level response-surface design built from a specific, efficient geometry:
- The point placement: it combines two-level factorial points on pairs of factors with the remaining factors held at their center, which places the runs at the midpoints of the cube's edges, plus replicated center points. It never places a run at a corner, where every factor is at its extreme simultaneously.
- What it fits: the full quadratic model, main effects, two-factor interactions, and pure quadratic (curvature) terms, the same surface a central composite design fits, so you can find a curved optimum.
- The efficiency: for three and four factors it typically needs fewer runs than a comparable CCD, and being a three-level design (rather than CCD's five levels with axial points outside the cube) it keeps every run inside the original factor box.
How to read the output
The defining property to weigh is that no run combines all factors at their extremes. That is exactly what you want when the corner combination is dangerous, wasteful, or physically impossible, maximum temperature and maximum pressure and maximum time all at once might ruin the sample or the equipment, and Box-Behnken simply never asks for it. The trade-off is at the corners themselves: because it does not sample them, Box-Behnken predicts least well in the extreme corner regions, so if your optimum might live in a corner, a central composite (which has axial points reaching out there) is the better fit. Read the design choice as a question about where the answer is likely to be and whether the extremes are safe to run.
A real use case
A sintering study for a ceramic separator varies temperature, dwell time, and heating rate, and the team knows from experience that maximum temperature at maximum dwell at maximum rate cracks the parts, so an all-corners design would waste runs on guaranteed scrap. A Box-Behnken design fits the full quadratic in 13 to 15 runs without ever commanding that destructive corner combination: every run sits at an edge midpoint, with two factors pushed while the third stays central. The fitted surface locates a curved optimum, moderate temperature with longer dwell, that no two-level design would have seen, and not a single run was lost to the corner that everyone already knew would fail.
Common mistakes
- Using Box-Behnken when the optimum may sit in an extreme corner. It samples edge midpoints, not corners, and predicts worst exactly there.
- Expecting it to handle two factors. Box-Behnken needs at least three factors by construction; for two, use a central composite or a small factorial with center points.
- Treating it as interchangeable with CCD without considering run safety. Its signature advantage is avoiding the all-extreme combination; if that combination is fine, CCD's corner coverage may be worth more.
- Skipping the replicated center points, which provide the pure-error estimate and the curvature check the design depends on.
- Fitting a quadratic and then extrapolating outside the design region, where a response surface has no support.
Quadratic surfaces without the dangerous corner
Niobia builds the Box-Behnken design for three or more factors, runs the analysis to fit the full quadratic (main effects, interactions, and curvature), and reports the fitted response surface with its optimum. It weighs Box-Behnken against a central composite design by what the process needs, the run budget, whether the extreme-corner combination is safe to run, and where the optimum is likely to sit, rather than defaulting to one design. The fitted model carries into a prediction profiler for exploring the window, and into Bayesian optimization when the next run should be chosen adaptively.
Frequently asked
When should I choose Box-Behnken over a central composite design?
When the combination of all factors at their extremes is unsafe, wasteful, or impossible, and when you want to keep every run inside the original factor box. Box-Behnken never samples corners and uses three levels. Choose CCD instead when the optimum might be in a corner, since CCD's axial points reach out there.
How many factors does Box-Behnken need?
At least three, by construction, since it pairs factorial points on two factors while holding the rest at center. For two factors, a central composite or a factorial with center points is the appropriate response-surface choice.
Does it fit curvature like a CCD?
Yes. Box-Behnken is a response-surface design that fits the full quadratic model, including the pure-quadratic curvature terms, so it can locate a curved optimum, typically in fewer runs than a comparable CCD for three or four factors.