Design of experiments
Pick the design family for the question you have. Screening (Plackett–Burman) screens for potentially important main effects among many factors, in N runs for up to N − 1 factors, assuming interactions are negligible or tolerable as confounding. A two-level factorial measures main effects and interactions, with center points to check for curvature. A response surface design fits a full quadratic so you can find a maximum, a minimum, or a setting that hits a target, and read it on a contour plot. A Taguchi orthogonal array with a signal-to-noise ratio finds settings that stay good when noise you cannot control changes.
In a factorial, effect = contrast / (N/2), using factorial points only; center points estimate curvature and are not part of that contrast. A half or quarter fraction aliases effects, so do not read an interaction as free unless the generators say it is. The response surface sheet reports coefficients, ANOVA with lack of fit and pure error, R² adjusted and predicted, the stationary point with its eigenvalues, and the best setting inside the design region. The response surface example is practice data made up for this sheet, and its results are checked against a separate calculation in NumPy.
Whatever the design: randomize the run order, hold everything you are not studying constant and record it, and confirm the chosen setting with a few new runs before you change the process.