Sequential design with a visible rationale
Bound the question
Specify the response, maximize or minimize objective, factor bounds, experiment count, and a per-experiment cost budget. Bounds must remain within the supported observed range.
Model the response
For supported continuous problems, fit a Gaussian process and quantify its conditional uncertainty. The initial workflow supports up to three factors with enough usable observations.
Rank useful candidates
Expected improvement balances the predicted response against model uncertainty. Candidates are suggestions to review, never a guarantee of experimental improvement.
Close the loop
Record new measurements in a new dataset version, repeat analysis, and compare the next iteration with earlier experiments and model runs.
Complex mixture constraints, multi-objective Pareto optimization, categorical optimization, and unvalidated extrapolation are outside the initial optimizer. Generated candidates require laboratory feasibility and safety review.