> Every time i've actually tried something like this it has not outperformed statistical process control.

There are clear counterexamples to your experience, most notably in maintaining plasma stability in tokamak reactors: https://www.nature.com/articles/s41586-021-04301-9