How to decide what data to collect: statistical thinking, the DOE toolkit for screening and optimisation, and how to turn a table of runs into a defensible answer.
Design first. The design step decides what you can learn — if you can't say what model, what runs, and what factors, don't start running yet.
Measure your ruler. A measurement systems analysis is not optional — if your ruler is worse than the effect, no design saves you, and bias isn't fixable by replication.
Use structure. One-factor-at-a-time freezes every other factor, so interactions — common in chemistry — stay invisible; factorial designs vary factors jointly to reveal them.
Close the loop. Plot, fit, check, reduce, interpret, and verify with a confirmation run — the analysis is what turns a table of numbers into a defensible answer.