Prediction is not understanding
The knowledge you want is already in the trained model: it encodes how the output moves as each input changes, over every combination your data covers. Sensitivity analysis reads that structure out. Without it, the model hits its validation targets and you still cannot say which of a dozen process parameters the prediction hangs on.
You stop treating the surrogate as an oracle that returns numbers and start interrogating it. The answers come in two forms: a global ranking of what drives the output anywhere in the space, and a local attribution that explains one prediction at a time.
Global ranks the space, local explains the case
The two views answer different questions, and you will usually use both.
Which inputs drive the output, in general?
The Sobol total-order index measures each input's share of the output variation across the whole space, interactions with other inputs included.
What you do: focus control and monitoring on the top of the ranking, and settle arguments about the bottom.
Why did this specific case come out the way it did?
SHAP splits one prediction's distance from the model's baseline across the inputs, showing which pushed it up and which pulled it down.
What you do: explain an unusual batch or design, and check the model's reasoning against what you know about the process.
Both methods ship in STOCHOS and run on the trained model. The sensitivity analysis page covers the wider method set.
The ranking is a test the model has to pass
Hold the global ranking against your process understanding. If it matches what experience says drives the process, the model has earned some trust. If it does not, either the data samples the space badly or the process behaves differently from what you assumed, and both are worth finding out before the next decision.
Local attributions run the same test case by case. An explanation that contradicts process knowledge flags a problem earlier than prediction error will, and one that holds up is something you can put in front of colleagues who never saw the model.
Common questions
01Can I trust indices computed on a surrogate rather than on the real process?
The indices describe the model, so validate the fit first. A surrogate that predicts held-out results well has captured the input-output structure the indices summarise. Where the model is uncertain, read the ranking with the same caution.
02Can the global and local views disagree?
Yes, and the disagreement is information. An input can rank low across the whole space yet dominate one prediction taken near the edge of your data. Global averages over everything you sampled; local looks at a single point.
03What does SHAP compare a prediction against?
A baseline, the model's average prediction over the data. Each input's contribution measures how far knowing its value moved this prediction away from that baseline, and the contributions sum to the full gap.
Sensitivity Analysis
The method page: what each method measures and when to reach for it.
How STOCHOS works
The DIM-GP model behind the predictions the indices explain.
Which parameters actually matter
The planning side: rank inputs before you design the next campaign.
News and Guides
All the short reads, collected on one page.
