Causal analysis of an industrial iron-ore flotation circuit: Which setpoint effects can be identified from plant data

Machine-learning soft sensors predict concentrate grade in froth flotation with good accuracy, and SHAP or partial dependence are then read as guidance for changing a setpoint. Those readings answer an associational question. A setpoint decision is interventional, and the two answers diverge whenever the plant runs under closed-loop control. We treat a public record from an industrial iron-ore reverse flotation plant (737,453 rows at a 20 s cadence, March to September 2017) as a causal inference problem. Causal discovery under process-knowledge constraints recovers a directed acyclic graph of the circuit. Interventional effects on the silica content of the iron concentrate are estimated by backdoor adjustment, by g-computation and by double machine learning, each with a 24 h block-bootstrap interval, and compared with a gradient-boosted predictor. The predictor reports pulp pH as inert, with a partial dependence of − 0.01 percentage points per standard deviation and an interval that includes zero. The flexible causal estimators return − 0.08 and − 0.10 percentage points with intervals that exclude zero, carrying the sign that reverse flotation practice would predict at this operating pH. The paired difference is negative in all 200 resamples, and three regime shifts with balanced feed assays confirm the causal sign. A time-lagged model shows that the pH effect appears about one hour after the change, which is why a same-hour attribution misses it. The amine collector effect is not identifiable from these data, because the dosing decision responds to ore mineralogy that the routine assays do not measure. Aeration produces the largest identified response, and it is carried by one of the seven flotation columns rather than by the bank mean. Causal modeling does not replace predictive machine learning. The two answer different questions, and the causal structure improves the predictor by exposing an assay that leaks the outcome.