Modern semiconductor experiments are constrained by more than statistics. Lot sizes, review cycles, and operational risk all limit how much experimentation can realistically happen in production. At the same time, practitioners are often asked to choose between two powerful – but imperfect – approaches: design of experiments (DOE), which is efficient but model-dependent, and Bayesian optimization, which is flexible but often resource-hungry.

This talk explores what happens when you stop choosing and start combining. This work presents a hybrid experimentation strategy that uses definitive screening designs (DSDs) as high-quality training data for Bayesian optimization. The idea is simple: start with a DOE that provides broad, structured coverage of the factor space, then let Bayesian optimization build on that foundation to efficiently refine and optimize.

Using simulated and published examples from dry etch, PECVD, and CMP processes, we show how this hybrid approach improves learning speed, reduces experimental cost, and, in some cases, identifies better operating conditions than either method alone. Compared to Bayesian optimization with space-filling training data, the DSD-based approach converges faster and with significantly fewer runs – an important advantage in resource-constrained fab environments.

Attendees learn how modern DOE concepts can directly enhance active learning workflows in JMP, when this approach is most effective, and how to think practically about deploying it in real manufacturing settings. If you’ve been curious about Bayesian optimization but hesitant to apply it in production, this session aims to close that gap.