Two goals that pull in opposite directions
A response surface finds the peak of one response. No real process has one response. You want the yield up and the scrap down, the hardness on target and the ductility not ruined — and the setting that maximises the first usually damages the second. Optimising them one at a time hides exactly the thing you need to see.
Putting different things on one scale. Yield in percent and scrap in kilograms cannot be added together. So each response is first converted to a desirability: a number between 0 and 1, where 0 means completely unacceptable and 1 means the target was reached. You supply two or three numbers per response — where it becomes unacceptable, and where it is good enough — and everything after that is comparable.
Why the geometric mean, and why it matters. The individual scores are combined by multiplying them and taking the root, not by averaging. The consequence is the whole point: if any single response scores zero, the combined score is zero. Under an ordinary average, a setting that produces out-of-spec product could score 0.65 because everything else was excellent, and would be recommended to you as a decent compromise. Multiplication makes that impossible. One failed requirement vetoes the solution no matter how good the rest look.
Weights say what you actually care about. Equal weights treat every requirement as equally important, which is rarely true. Raising the weight on scrap pulls the answer towards low scrap and accepts a worse yield. This is where engineering judgement belongs — not hidden inside the arithmetic.
When there is no answer. Sometimes nothing works: the combined score is zero everywhere. That is a genuine result, not a bug. It means at least one requirement is beyond what this process can deliver in the range you explored. SenSight names which requirement is unreachable, so you argue with the right one instead of chasing a setting that does not exist.
The last step is always a real run. The optimum comes from fitted models, and models are wrong at the edges. Treat the answer as the best place to run a confirmation trial, never as a setting to put straight into production.
Putting different things on one scale. Yield in percent and scrap in kilograms cannot be added together. So each response is first converted to a desirability: a number between 0 and 1, where 0 means completely unacceptable and 1 means the target was reached. You supply two or three numbers per response — where it becomes unacceptable, and where it is good enough — and everything after that is comparable.
Why the geometric mean, and why it matters. The individual scores are combined by multiplying them and taking the root, not by averaging. The consequence is the whole point: if any single response scores zero, the combined score is zero. Under an ordinary average, a setting that produces out-of-spec product could score 0.65 because everything else was excellent, and would be recommended to you as a decent compromise. Multiplication makes that impossible. One failed requirement vetoes the solution no matter how good the rest look.
Weights say what you actually care about. Equal weights treat every requirement as equally important, which is rarely true. Raising the weight on scrap pulls the answer towards low scrap and accepts a worse yield. This is where engineering judgement belongs — not hidden inside the arithmetic.
When there is no answer. Sometimes nothing works: the combined score is zero everywhere. That is a genuine result, not a bug. It means at least one requirement is beyond what this process can deliver in the range you explored. SenSight names which requirement is unreachable, so you argue with the right one instead of chasing a setting that does not exist.
The last step is always a real run. The optimum comes from fitted models, and models are wrong at the edges. Treat the answer as the best place to run a confirmation trial, never as a setting to put straight into production.
Try it in the app
Try: DOE tab, Multi-Response Optimisation. Add two responses that fight each other, set one weight to 5 and re-run — watch the recommended setting move towards the response you weighted.