I know which way to move — but where do I stop?
A two-level factorial answers the first question well: raise the temperature, yield goes up. It cannot answer the second, and the reason is geometric — two points define a line, and a line has no top. Push far enough along it and you will walk straight past the optimum without the experiment ever warning you.
What a response surface adds. First, repeated runs at the centre. If the response there sits well off the line joining the corners, the surface is curved — and that is testable before you commit to anything else. Second, axial runs: each factor is pushed beyond its usual high and low, giving it a third and fourth level. Only then can a squared term be estimated, and only a squared term can bend.
CCD or Box-Behnken? The central composite design includes the corner runs, so it also gives you the plain factorial information. Its axial points sit outside ±1 — good statistically, impossible if a factor is already at a physical limit; then use the face-centred version. Box-Behnken never sets every factor to an extreme in the same run, which is what you want when that combination would be unsafe.
Reading the answer. Solving the fitted surface gives a stationary point — the place where the slope is zero in every direction. It is not automatically the peak. The eigenvalues decide: all negative is a maximum, all positive a minimum, mixed signs a saddle. A saddle rises one way and falls another, so there is no optimum inside the region at all; the honest move is to pick a ridge, follow it, and run a new design further along. SenSight reports the classification rather than just handing you coordinates.
One more trap. If the stationary point falls outside the range you actually ran, the model is extrapolating. It was only fitted where you collected data. Treat such a point as a direction worth exploring, never as a setting to adopt.
What a response surface adds. First, repeated runs at the centre. If the response there sits well off the line joining the corners, the surface is curved — and that is testable before you commit to anything else. Second, axial runs: each factor is pushed beyond its usual high and low, giving it a third and fourth level. Only then can a squared term be estimated, and only a squared term can bend.
CCD or Box-Behnken? The central composite design includes the corner runs, so it also gives you the plain factorial information. Its axial points sit outside ±1 — good statistically, impossible if a factor is already at a physical limit; then use the face-centred version. Box-Behnken never sets every factor to an extreme in the same run, which is what you want when that combination would be unsafe.
Reading the answer. Solving the fitted surface gives a stationary point — the place where the slope is zero in every direction. It is not automatically the peak. The eigenvalues decide: all negative is a maximum, all positive a minimum, mixed signs a saddle. A saddle rises one way and falls another, so there is no optimum inside the region at all; the honest move is to pick a ridge, follow it, and run a new design further along. SenSight reports the classification rather than just handing you coordinates.
One more trap. If the stationary point falls outside the range you actually ran, the model is extrapolating. It was only fitted where you collected data. Treat such a point as a direction worth exploring, never as a setting to adopt.
Try it in the app
Try: DOE tab, Response Surface, design a 2-factor CCD and note the axial runs at ±1.41. Run it (or load a dataset with centre points), analyse, and read the curvature test before the peak.