Key Takeaways

  • Throughput, recovery, flotation kinetics, and concentrate grade can each be modeled as a “soft sensor” — an ML model that continuously predicts plant performance from live sensor/historian data instead of a static simulation.
  • JKSimFloat and METSIM remain the standard purpose-built flotation simulators for offline circuit design and what-if analysis.
  • Imubit is the clearest AI-native product in this space — closed-loop reinforcement-learning optimization already deployed across process industries, including flotation-adjacent applications.
  • Published plant studies use NSGA-II multi-objective genetic optimization combined with ML feature selection to lift recovery beyond what static simulators achieve alone.
  • Concentrate-grade ML prediction is the least mature of the four (Emerging) — grade modeling still runs mostly inside the simulators rather than as a standalone real-time model.

TL;DR

Use JKSimFloat/METSIM for offline flotation circuit design, then layer a continuously-retrained ML model (or a closed-loop platform like Imubit) on top of live plant sensor/historian data to predict and optimize throughput, recovery, and grade in real time rather than relying on a static simulation.

How Do I Predict Flotation Performance With AI?

Flotation performance depends on dozens of interacting variables — reagent dosage, pH, residence time, air rate, feed mineralogy — that shift throughout a shift in ways a static process simulation can’t fully capture. The traditional tool for flotation modeling is JKSimFloat, the industry-benchmark simulator from JKTech, used alongside JKSimMet for combined comminution-flotation circuit modeling; METSIM covers similar ground across the broader flowsheet. These remain the right tools for offline circuit design, debottlenecking studies, and what-if scenario testing — nothing here replaces them for that purpose.

Where AI changes the picture is in real-time prediction and control. Instead of a fixed simulation, a “soft sensor” is a regression model — typically scikit-learn/XGBoost-class gradient boosting, sometimes a neural network — trained on your plant historian data (SCADA tags: reagent flows, pH, froth camera features, feed grade) against lab-measured recovery and grade results, then run continuously to predict throughput, recovery, and grade minute-by-minute instead of only at the cadence of your metallurgical sampling. Published plant studies go further, using NSGA-II (a multi-objective genetic algorithm) combined with ML feature selection to search for reagent/operating-condition combinations that lift recovery beyond what the static simulator’s baseline recommends — this is where real production gains get reported, not just prediction accuracy.

Imubit represents the most AI-native end of this spectrum: a closed-loop reinforcement-learning platform (their “Optimizing Brain” product) that doesn’t just predict plant behavior but actively adjusts setpoints in a continuous control loop, already deployed across more than 90 industrial applications including flotation-adjacent process control. This is meaningfully different from a soft sensor that only predicts — it’s a system making (or recommending) real-time operating decisions.

Concentrate grade prediction (Cu, Mo, Au, Ag) is the weakest link of the four right now — it’s still primarily built inside the process simulators (JKSimFloat, METSIM) rather than existing as its own real-time ML model, so treat grade-prediction claims from any vendor with more scrutiny than throughput or recovery claims, which have more published validation behind them.

Try It With Geocluster

Building a soft-sensor model that’s actually trustworthy requires reconciling historian tag data against lab assay ground truth over a long enough window to capture real operating variation — exactly the kind of applied research Geocluster is designed to help with. If you’re standing up flotation soft sensors, Geocluster can help you build and validate the training pipeline behind them.