Key Takeaways
- A flotation soft sensor is an ML model that continuously predicts throughput, recovery, flotation kinetics or concentrate grade from live sensor and 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 optimisation already deployed across process industries, including flotation-adjacent applications.
- Published plant studies combine NSGA-II multi-objective genetic optimisation 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
Keep JKSimFloat or METSIM for offline flotation circuit design. Then train a flotation soft sensor on live plant sensor and historian data to predict throughput, recovery and grade in real time. For the next step, a closed-loop platform like Imubit acts on those predictions instead of only reporting them.
What Is a Flotation Soft Sensor?
A soft sensor is a model that estimates a value you cannot measure directly in real time. In flotation, the values that matter most are recovery and concentrate grade. Both normally come from lab assays, which arrive hours after the sample was taken. By then, the plant has already moved on.
A flotation soft sensor fills that gap. It learns the relationship between what you can measure continuously and what the lab reports later. Continuous inputs include reagent flows, pH, air rate, feed grade and froth camera features. Once trained, the model predicts recovery and grade minute by minute. Operators, or a controller, can then react before an off-spec shift shows up in the assay results.
How Do I Predict Flotation Performance With AI?
Flotation performance depends on dozens of interacting variables: reagent dosage, pH, residence time, air rate and feed mineralogy. They shift throughout a shift in ways a static simulation cannot fully capture. The workflow below keeps the simulators for what they do well and adds AI where it helps.
Keep the simulators for circuit design
The traditional tool for flotation modeling is JKSimFloat, the industry-benchmark simulator from JKTech. It is used alongside JKSimMet for combined comminution and 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 scenarios.
Train the flotation soft sensor on historian data
Where AI changes the picture is real-time prediction. A flotation soft sensor is usually a regression model, often gradient boosting from scikit-learn or XGBoost and sometimes a neural network. You train it on plant historian data (SCADA tags for reagent flows, pH, froth camera features and feed grade) against lab-measured recovery and grade. Then it runs continuously.
Froth images add a lot of signal. Research systems extract bubble size, colour and texture from camera video and feed those features into the model. The main difficulty is labelled data. Lab assays are far less frequent than sensor readings, so each model learns from relatively few ground-truth points.
Optimise, not just predict
Published plant studies go a step further. They use NSGA-II, a multi-objective genetic algorithm, together with ML feature selection. The search looks for reagent and operating-condition combinations that lift recovery beyond the simulator’s baseline. This is where real production gains get reported, not just prediction accuracy.
Imubit sits at the most AI-native end of this spectrum. Its closed-loop reinforcement-learning platform does not just predict plant behaviour. It actively adjusts setpoints in a continuous control loop, and it is deployed across more than 90 industrial applications. Closed-loop AI flotation control is meaningfully different from a soft sensor that only predicts.
Where Flotation Soft Sensor Predictions Are Weakest
Concentrate grade prediction (Cu, Mo, Au, Ag) is the weakest link right now. It is still built mainly inside the process simulators rather than as its own real-time ML model. Treat grade-prediction claims from any vendor with more scrutiny than throughput or recovery claims, which have more published validation behind them.
Soft sensors also drift. When ore feed or circuit configuration changes, the relationships the model learned can stop holding. Plan for regular retraining and track prediction error against new assays. The same discipline applies to managing plant variability more broadly.
FAQs
What is a soft sensor in mineral processing?
A soft sensor is a software model that estimates a process value from other measurements, instead of measuring it directly with a physical instrument. In mineral processing it typically predicts values like recovery, grade or particle size, which otherwise need lab analysis or expensive online analyzers.
What data do you need to build a flotation soft sensor?
You need continuous plant data from your historian, such as reagent flows, pH, air rate, feed grade and froth camera features. You also need matching lab assays for recovery and grade over a long enough period to capture real operating variation. Aligning timestamps between the two is often the hardest part.
Can a flotation soft sensor predict concentrate grade?
It can, but grade is the least mature target. Recovery and throughput predictions have more published validation. Grade models are harder because assays are infrequent and grade depends on subtle changes in mineralogy. Validate any grade model carefully against new lab results before relying on it.
Is a soft sensor the same as closed-loop control?
No. A soft sensor only predicts. Closed-loop control uses predictions like these to change setpoints automatically. Many plants start with a soft sensor for operator decision support and move to closed-loop control once they trust the model.
Try It With Geocluster
A trustworthy soft sensor needs historian tag data reconciled against lab assay ground truth, over a window long enough to capture real operating variation. That applied research is what Geocluster is designed to help with. If you are standing up flotation soft sensors, it can help you build and validate the training pipeline behind them.