Operating-zone identification of an auxiliary electrode electrostatic roll separator for ilmenite recovery using explainable machine learning

Ilmenite recovery from heavy mineral sand processing rejects was investigated using an auxiliary electrode electrostatic roll separator (AEERS). The AEERS incorporates an independently positioned lower electrode beneath the grounded rotating drum, which is intended to extend electric field control into the particle discharge region relative to a conventional high-tension roll separator. A representative reject sample containing 11.97% ilmenite and 66.20% garnet, with minor amounts of rutile, zircon and sillimanite, was characterised and used as feed. A six-factor Central Composite Design (CCD) comprising 30 experimental runs was employed to evaluate the effects of feed rate, feed-roll speed, separation-roll speed, feed temperature, splitter position and applied voltage on ilmenite recovery. A quadratic RSM model was developed and validated by analysis of variance (ANOVA). The model was statistically significant (F = 58.19, p < 0.00001, R 2 = 0.9987). Maximum ilmenite recovery of 52.3% at a concentrate grade of 90.4% was achieved at a feed rate of 0.20 TPH, feed-roll speed of 25 RPM, separation-roll speed of 165 RPM, feed temperature of 110 °C, splitter position of 60° and an applied voltage of 24 kV. The experimental data were further analysed using decision tree regression (DTR) and SHapley Additive exPlanations (SHAP) to classify the operating space into zones and quantify the contribution of individual operating variables. Analysis of variance identified applied voltage as the variable with the largest effect on recovery, followed by separation-roll speed and splitter position, with feed rate and feed-roll speed contributing less and feed temperature not significant. The DTR model classified the operating space into optimal ( ≥ 48% recovery), transitional (40%–48%) and sub-optimal ( < 40%) zones. The resulting operating-zone map complements the RSM model by defining an indicative operating window instead of a single optimum operating point. The combined RSM and explainable machine learning framework provides a practical approach for operating-zone identification and process optimisation in the electrostatic separation of ilmenite from heavy mineral sand processing rejects, with potential application to other mineral separation processes. All experiments were performed with the auxiliary electrode energised. Its isolated contribution relative to a conventional single-electrode configuration was not quantified.