Every percentage point of scrap in a flotation circuit is revenue that left the plant floor and never came back. Not latent capacity. Not recoverable rework. Concentrate that went to the tailings dam because the froth condition shifted twenty minutes before the operator's walk-through, and the assay confirmed the grade failure three hours later when the stockpile was already built. Plant executives running rougher-scavenger-cleaner circuits in 2026 are no longer asking whether predictive scrap analytics works. They are asking why their Cpk is still oscillating between 1.1 and 1.4 when peer operations are sustaining 1.67+ across every ore zone transition. The difference is not in the ore. It is in what the monitoring layer can predict before the scrap event matures.
What Is Predictive Scrap Analytics for Mining Flotation — and Why It Determines Your Cpk Ceiling
Predictive scrap analytics in mining flotation is the application of multivariate machine learning models to the continuous forecast of concentrate grade excursions, recovery losses, and off-spec events before they materialise as scrap. Where conventional SPC monitors what has already happened against static control limits, predictive scrap analytics learns the subtle, multi-variable precursor patterns — froth vision features, reagent dosage drift, pulp density shifts, airflow distribution changes — that statistical process control charts are not designed to detect until the signal crosses a fixed threshold.
The Cpk connection is structural. Process capability index for a flotation circuit is determined by two things: how tightly the process output clusters around the target grade, and how centrally that cluster sits relative to the specification limits. Static SPC, calibrated to annual average feed conditions, produces control limits that are too wide for stable ore zones and too narrow for transition zones — creating a Cpk profile that oscillates between 1.1 and 1.4 as the circuit moves through the mine plan. Predictive scrap analytics replaces those static limits with self-tuning control bands that contract when the process is stable and expand appropriately when ore zone chemistry shifts, holding the circuit in the 1.67+ Cpk band continuously rather than recovering to it after every transition.
The result is a scrap profile that does not cluster at ore zone boundaries. Off-spec events go from 6-12 per quarter to 1-2, and the remaining events are intercepted by the predictive model's early warning system before the concentrate reaches the stockpile. The scrap that does occur is documented with ML-ranked root cause evidence, producing the ISO 9001 corrective action record at the moment of the event rather than three weeks after the investigation.
Conventional SPC tells you your Cpk dropped below 1.33 after the assay results arrive. Predictive scrap analytics tells you your Cpk trajectory is diverging from target three to six hours before the grade failure — and gives you the specific process variable adjustment that will bring it back. The 12-18% scrap reduction in the first quarter is not a function of new equipment. It is a function of information arriving at the intervention decision point with enough lead time for the operator to act.
The Five Scrap Precursor Signals That Static SPC Misses — and Predictive ML Catches
Flotation scrap does not announce itself with a single sensor crossing a red line. It builds through the interaction of five precursor signals, each of which is individually innocuous and collectively predictive of an imminent grade excursion. Static SPC, monitoring each variable independently against fixed limits, detects none of these interactions until the combined effect produces a limit breach on the final output — at which point the scrap is already in the stockpile. Predictive scrap analytics, using multivariate ML models trained on your circuit's historical scrap events, identifies the precursor pattern and escalates it as a developing event.
How Predictive Scrap Analytics Integrates With SPC: The Architecture That Sustains Cpk 1.67+
Predictive scrap analytics does not replace SPC. It replaces the static control limit paradigm that forces SPC to operate reactively. The integration stack layers ML scrap forecasting, self-tuning UCL and LCL, Western Electric rule monitoring, and root cause classification on top of the existing SPC foundation — producing a quality monitoring layer that predicts, detects, diagnoses, and documents scrap events in a single workflow rather than across four disconnected systems.
Scrap Reduction by Circuit Stage: Where Predictive Analytics Closes Each Gap
Scrap in flotation does not originate from a single stage. It accumulates because each stage's specific loss mechanism — froth condition degradation in the rougher, feed quality mis-match in the scavenger, grade sensitivity in the cleaner — operates below the detection threshold of static SPC long enough for the combined effect to produce an off-spec concentrate event. Predictive scrap analytics addresses each stage's mechanism with a stage-specific model that understands the different precursor signals and intervention windows of rougher, scavenger, and cleaner circuits.
The rougher circuit processes the full feed stream and sets the recovery ceiling for every downstream stage. Scrap in the rougher — underrecovered mineral sent to the scavenger — is the highest-impact loss because it reduces recovery across the entire circuit, not just in a single stage. The predictive model monitors rougher froth condition, pulp density, and reagent dosage at per-minute frequency, detecting the froth vision drift and dosage-recovery curve displacement that indicate the rougher is losing recovery before the tailings grade confirms it. At ore zone transitions — the highest-risk window — the model pre-loads zone-specific starting setpoints from the last seven transitions of the same boundary, compressing the recovery stabilisation window from two shifts to under one hour.
Scavenger scrap is typically a rougher propagation problem. When the rougher underperforms for even a single shift, the scavenger receives higher-than-expected mineral loading at lower-than-expected grade, and fixed reagent setpoints calibrated to average feed conditions become immediately suboptimal. The predictive model detects the feed quality change from the rougher's froth vision and grade estimate output and propagates the dosage adjustment recommendation to the scavenger circuit before the loading imbalance materialises as a recovery loss. Scavenger dosage tracks the incoming rougher output in real time rather than the mine plan average — eliminating the shift-long delay between a rougher performance change and the scavenger's response.
Cleaner scrap has the most direct commercial consequence because it determines final concentrate grade. A single off-spec cleaner event can produce an entire shift's output that falls outside the off-taker's tolerance window. The predictive model in the cleaner uses froth colour-texture grade estimates, density-airflow interaction maps, and the zone transition classifier to generate a continuous Cpk trajectory forecast for the cleaner output. When the forecasted Cpk drops below the 1.67 threshold, the model fires an alert with the specific variable adjustment — usually airflow distribution across cleaner cells or recirculation load balance — that has the highest causal weight in restoring the grade trajectory. The alert arrives three to six hours before the LIMS assay would confirm the failure, and the intervention recommendation is specific enough that the operator can act without escalation.
Our Cpk was oscillating between 1.1 and 1.4 for eighteen months. Every quarterly review we explained it as feed variability. After iFactory's predictive scrap analytics went live, the Cpk stabilised at 1.72 within the first six weeks and has held there through three ore zone transitions that would have generated off-spec events under our old SPC. The scrap events dropped from nine per quarter to one — and that one was intercepted by the predictive alert before it reached the stockpile. Our quality manager's pre-audit preparation time went from three weeks to four days.
— Concentrator Operations Manager, Copper-Gold Operation, 22,000 tpdWhat Changes on the Plant Executive Dashboard When Predictive Scrap Analytics Goes Live
Plant executives do not manage ML models. They manage the outputs those models protect. When predictive scrap analytics goes live, the change is visible in the three metrics that define the commercial performance of the operation — and those three metrics move in a specific, predictable sequence that reflects the model maturation curve.
The Cpk trajectory in a flotation circuit running predictive scrap analytics stops oscillating with the mine plan. Instead of dropping below 1.33 at every ore zone transition and recovering over two to three shifts, Cpk holds at 1.67+ continuously because the self-tuning control limits contract and expand with the zone, and the ML forecaster pre-loads the zone-specific starting parameters from the historical performance database. The dashboard shows Cpk not as a single number but as a trend per ore zone — with the predictive forecast line extending three to six hours ahead, showing where Cpk is heading before the grade data confirms it.
Scrap events in a predictive-monitored circuit do not disappear immediately. They decline in a characteristic curve: an initial drop of 50-60% in the first four to six weeks as the model eliminates the false alarm volume and operators begin responding to every genuine alert, followed by a sustained reduction to 80-90% below baseline by the end of the second quarter as the ML root cause classifier refines its intervention recommendations from confirmed outcomes. The dashboard tracks scrap event count and financial impact per event, with the trend line compared against the pre-deployment baseline and the current ore zone's predicted scrap risk.
ISO 9001 compliance in a flotation circuit running predictive scrap analytics is no longer a pre-audit exercise. Every scrap event and every intercepted near-event generates its own quality record automatically — froth condition state at detection, process variable readings, ML-ranked root cause, recommended intervention, and actual corrective action taken — at the moment the event closes. Clause 8.7 nonconforming output records and clause 10.2 corrective action records are produced as standard operating output rather than reconstructed from DCS logs after the event. The three weeks of pre-audit preparation that consumes quality team capacity every cycle compresses to three to five days because the documentation is complete and audit-ready at the close of each event.
Deployment: Read-Only Integration, No DCS Changes, Live in 4-8 Weeks
The most common question plant executives ask about predictive scrap analytics deployment is whether it introduces operational risk to a running circuit. The answer is that iFactory connects to existing infrastructure through read-only interfaces. Camera feeds integrate through standard industrial video protocols. Process historian data is accessed with no write permissions. The LIMS remains the laboratory record of truth, and its assay results feed as lagged validation inputs into the ML scrap forecaster to improve prediction accuracy over time. No DCS modification. No SCADA schema change. No operational exposure at any point in the deployment.
Conclusion: The Cpk Gap Is a Monitoring Gap
Sustaining Cpk 1.67+ in a mining flotation circuit is not a control problem. The DCS can adjust airflow, level, and reagent dosage with precision. The operators understand the process and respond to alerts when they trust them. The ore variability is real and will not reduce. The gap between actual and achievable Cpk is a monitoring problem — specifically, a problem of detection latency, false alarm volume, and the absence of a prediction layer that can see the scrap event before the limit breach.
Predictive scrap analytics closes that gap by replacing static SPC with a self-tuning, ML-powered quality monitoring layer that continuously learns your circuit's scrap precursor patterns, forecasts the Cpk trajectory three to six hours ahead, and delivers a ranked root cause and recommended intervention with every alert. The 50-60% scrap reduction in the first quarter, the Cpk stabilisation at 1.67+ across ore zone transitions, and the quality management overhead that shrinks because the documentation writes itself — these outcomes come not from new equipment or additional instrumentation, but from a fundamental shift in when the monitoring layer detects the developing condition and how precisely it tells the team what to do about it.
The ore will continue to change at every zone boundary. The mine plan will not reduce the number of transitions. The off-taker's grade specifications will not widen. The only variable a plant executive can change is the speed and precision with which the operation detects, diagnoses, and corrects the conditions that separate actual Cpk from achievable Cpk. Predictive scrap analytics is how that variable is changed — available now, deployable in four to eight weeks, and operational without a single modification to the infrastructure you have already built.
