Four hundred and fifty generated batches from a three-stage fermentation, with a known cause-and-effect structure planted inside. Left: what a correlation screen ranks highest. Right: the structure a causal model learns when it is told the order of the process. Below: what happens to yield if you actually set a variable.
Rank process variables by how strongly they move with yield and test top against bottom batches. A variable that merely shares a cause with yield passes; an optimum in the middle of the range fails.
Later stages cannot cause earlier ones, and variables in the same stage are not linked. That rules out more than half the possible edges before learning; nonlinear structural equations are fitted on what remains.
With the structure and equations in hand, set a variable — do(X = x) — and propagate through its descendants. That is the in-silico experiment process engineers use to pick conditions worth trying on a real line.