Reinforcement‑learning froth‑depth control for flotation recovery.
A deep reinforcement‑learning controller that drives the flotation drain valves directly and learns feed‑forward action from the upstream valve, holding froth depth on setpoint and breaking the disturbance cascade that propagates tank to tank. It keeps valuable metal in the float instead of the tails, and runs closed‑loop on the plant's existing Allen‑Bradley PLC over native EtherNet/IP, with no new instrumentation.
A flotation cell concentrates target metals by floating particle‑laden froth off the top of each tank while gangue drains out the bottom, and recovery depends on holding froth depth steady on setpoint. But the disturbances start upstream: when the lead tank's valve moves to defend its own froth depth, it sends a wave down every downstream tank. The downstream froth‑depth meter doesn't register it for roughly ten seconds, by which point the PID loop is already chasing setpoint, swinging, and the swing propagates down the bank, carrying valuable metal to tails.
Ai‑OPs deployed a Koios deep reinforcement‑learning (DRL) controller that drives the drain valves directly and learns a feed‑forward response, taking the upstream valve position as an input so it acts on the cause before the effect reaches the downstream meter. Validated against roughly three days of plant data across four sets, it held froth depth nearly exactly on setpoint, an estimated >80% reduction in error, running closed‑loop on the existing Allen‑Bradley PLC.
Ai-OPs
·Flotation Froth-Depth DRL Control01
Background & objective
The objective was to break the disturbance cascade: stop the froth‑depth swing that begins at the head of the bank and grows tank by tank, and hold each cell on setpoint so valuable metal floats instead of dropping to tails.
In a flotation cell, slurry feeds in and air bubbles carry target‑metal particles up into a stabilized froth that overflows to concentrate, while gangue drains out the bottom. Each tank holds its froth depth with a drain valve on a single froth‑ball measurement. This was a cascaded flotation unit: the cells run in series and step down progressively in size, rougher, scavenger, cleaner, so every tank's outlet is the next tank's disturbance.
The control problem
- Upstream origin: an upset begins at the lead tank; as its valve responds to hold froth depth, it creates a domino effect of disturbances down every downstream tank.
- Detection lag: there is a ~10‑second delay before the disturbance reaches the downstream froth‑depth meter, by which time it is already too late to mitigate.
- Reactive PID: the loop only starts chasing the level setpoint after the swing has landed, so it never catches up to the moving disturbance.
- Lost recovery: each swing costs efficiency and production, and carries valuable metal to tails as it propagates down the cell.
A faster reactive loop still has to wait for the froth meter to see the disturbance. The win is to act on its cause, the upstream valve move, before its effect ever reaches the downstream meter.
FIG. 1: CASCADED FLOTATION UNIT. The cells run in series and step down progressively in size; slurry passes cell to cell while froth overflows to concentrate. A valve move at the lead tank propagates downstream as a froth‑depth swing that each downstream froth meter sees only after a ~10s transport lag, so a reactive PID is always behind it.
Ai-OPs
·Flotation Froth-Depth DRL Control02
Closed-loop control architecture
A deep reinforcement‑learning agent drives the drain valve directly and watches the upstream valve, so it functions like a feed‑forward loop: information passes from the upstream valve to the controller, and it moves before the disturbance reaches the local froth meter.
FIG. 2: CONTROL & CONNECTIVITY ARCHITECTURE. The agent takes the upstream valve position as a feed‑forward input and drives the local drain valve, with the native PID loop held ready for clean fallback. Koios reaches the loop directly over native EtherNet/IP, across both controller‑scoped and program‑scoped tags, no OPC server in between.
What the agent watches
Froth depth setpoint and PV on the controlled tank, the local valve opening, and, critically, the upstream valve opening. It is data‑driven, no hydraulic model or simulator, and inferences once per second on the prior 10 samples of runtime so it acts on a trend, not a single noisy reading.
What it controls
It outputs one positional command, the first‑cleaner drain‑valve opening, limited to a 2% maximum step per move. Because it already knows what the upstream valve did, it pre‑empts the incoming disturbance instead of reacting to it after the froth meter finally sees it.
Ai-OPs
·Flotation Froth-Depth DRL Control2.3
Digital twin & control model
Breaking the valve correlation
In the raw data the upstream and downstream valves were so highly correlated that the system dynamics could not be modeled directly. The fix was a single engineered feature, the difference between the two valves (Valve_01‑02_diff). It broke the correlation and let Koios build a dynamic time‑series forecast, a digital twin (AiPV) of froth depth, by machine learning.
The twin was bump‑tested by shifting the valve difference between −2% and +2%; froth depth responded as a clean sine‑wave, confirming the dynamics had been captured before any control model touched the loop.
Control objective
The control model is trained to hold froth depth on setpoint with smooth, low‑amplitude valve motion, leaning on the feed‑forward signal from the upstream valve so it pre‑empts disturbances rather than chasing them. Precise, stable tracking is the priority the policy optimizes toward.
| Output tag | First‑cleaner valve opening % (Unit 02) |
| Output type | Positional |
| Max step | 2.0% |
| Scan rate | 1 s · train & execute |
| History depth | 10 samples |
| Froth depth SP (in) | 4 – 18 |
| Froth depth PV (in) | 0 – 20 |
| Valve opening, Unit 01 | 0 – 34% |
| Valve opening, Unit 02 | 0 – 34% |
TABLE 1: MODEL CONFIGURATION. The digital twin predicts froth depth from the valve difference; the control model outputs the Unit 02 valve opening, normalized and clamped to its valid range.
03
Implementation & integration
No new hardware
The controller reads existing froth‑depth and valve tags and writes the existing drain‑valve loop directly on the Allen‑Bradley PLC. The improvement reached the field entirely in software, with no new instrumentation and no capital build‑out.
Train, validate, deploy
Roughly 70% of the supplied data trained the models; the remainder drove simulated bump tests and validation. Models run in monitoring first, then promote to closed‑loop control, with PID fallback available at any time.
The model built for one tank can be replicated for each tank regardless of size, so a single proven controller scales across the rougher‑scavenger‑cleaner bank rather than being re‑engineered cell by cell.
Ai-OPs
·Flotation Froth-Depth DRL Control04
Results
In validation the agent learned to rely heavily on the feed‑forward signal from the upstream valve, moving instantly to hold froth depth nearly exactly on setpoint, while the old reactive control kept swinging through the same disturbances.
FIG. 3: CONTROL VALIDATION. Recreated from the validation report. The old froth depth (green) swings widely through the bank's disturbances; the Ai‑OPs froth depth (orange, dashed) tracks the set point (blue) almost exactly. Across the four validation sets, error sums fell consistently, an estimated >80% reduction in froth‑depth error.
"The RL algorithm has learned to rely heavily on the feed‑forward information from the upstream valve. It moves instantly to hold the froth depth nearly exactly on set point. We expect an error‑from‑setpoint reduction of ~80% when deployed."
Ai‑OPs Model Validation: Froth Level Stabilization for Recovery Optimization
Ai-OPs
·Flotation Froth-Depth DRL Control05
Where this applies
The same pattern, a reinforcement‑learning agent that learns feed‑forward action across coupled loops and runs closed‑loop on the existing control system, generalizes across mineral processing and the wider plant. Applications available today on Koios:
About Koios
Koios is Ai‑OPs' self‑hosted machine‑learning inferencing platform for deploying scalable AI models on industrial control systems. It is hardware‑agnostic, runs fully on‑premises and air‑gapped, and writes optimized control back to PLCs, SCADA and smart instruments in real time. Your models and data stay inside your plant; you own every model Koios trains.
It installs as a single Docker image from a Docker Verified Publisher, or as an offline ISO / virtual‑appliance image, and is licensed as an annual per‑instance subscription.
| Protocols | EtherNet/IP · OPC‑UA · Modbus TCP · BACnet & more |
| Deployment | On‑prem · air‑gapped |
| Install | Single Docker image / ISO |
| Min. hardware | 4‑core 64‑bit · 8 GB RAM · 128 GB SSD · 1 GbE |
| Per license | Unlimited devices, tags & models · historian & diagnostics · support |