Why grade prediction ML errs on ore streams
An ore body is almost never homogeneous in grade. Ore grade prediction is built on a geological model that smooths values between drillholes, while a block with a local anomaly — an oxidized zone, a veinlet, a waste inclusion — arrives at the face.
A model trained on such labeling confidently errs precisely in transition zones, where the cost of error is highest.
The second reason is sampling lag. Between sample collection and the laboratory analysis result, time passes during which the ore has already gone through the crusher and ended up in the blend.
The label is tied not to the interval where the material actually was, and grade prediction ML begins to learn on shifted labeling. The longer the sample preparation chain, the more “grade in the model” and “grade in the stream” diverge.
Then the mixing of ore types intervenes. Several faces, stockpiles, and different zones of the deposit converge into one stream, and the online analyzer sees an averaged signal.
Unstable metal grade creates a class of errors that cannot be written off as noise: overgrading sends metal to tailings, undergrading drives excess reagent and breaks concentrate specification.
And on top of everything is manual adjustment of the regime by the operator. It pulls the process along, but at the same time changes the conditions in which the model was supposed to work: we do not see what the grade would have been without intervention.
This is the root of ore classification errors at most processing plants — we examine how this is addressed.
What a controllable metal grade prediction provides
Metal grade prediction becomes controllable when the model outputs not a single number but an interval with an uncertainty estimate.
The operator sees where the prediction is stable and where sample verification is needed, and blend decisions stop being a bet on an average estimate. What follows are effects that accumulate in the very first weeks of operation.
Reducing metal losses in tailings
The main metal leak is not accidents but a slow shift in the regime: flotation drifts off, and the valuable component leaks into tailings unnoticed until the next sampling.
The grade prediction model reads geology, sampling data, and process parameters and catches this shift earlier than it appears in laboratory analysis.
The early signal gives time to correct the reagent regime and pulp density — the share of metal in tailings falls, and losses become a predictable value rather than a month-end result.
Fewer false mis-sorts in ore sorting
The accuracy of ore sorting is limited not by the sensor but by the decision threshold. A rigid threshold based on one feature sends pieces with borderline grade to the dump and at the same time lets waste rock into the mill feed.
A prediction with an uncertainty interval makes it possible to shift the threshold to where the error is cheaper — the share of erroneous decisions is reduced by about a third.
A side effect is stability of the processing regime: smoother feed, less variation in particle size and grade, fewer manual adjustments.
It is the borderline classes that account for most mis-sorts; industry estimates from the International Council on Mining and Metals agree that managing this area delivers the largest gain in recovery.
Options for building grade prediction: comparison of modes
The prediction mode is chosen before model configuration: it determines the set of input data, the recalculation frequency, and who makes the decision on a disputed sample. Below are three schemes used at mining operations and their limitations.
| Option | When it fits | Limitations |
|---|---|---|
| Batch prediction by samples | Exploration, shift planning, block mining | The answer arrives with a delay and is not tied to the current conveyor belt |
| Streaming prediction on the conveyor | Real-time sorting and blend control | Requires working sensors and regular calibration |
| Hybrid scheme with geologist rules | Transition zones, mixed ore types | More settings and maintenance, requires a conflict resolution procedure |
Batch prediction by samples
Samples arrive in batches; the model assigns each to a grade and returns the result to the geologist before the shift starts. The decision can be rechecked against the original sample, so the mode is convenient for reporting and discrepancy analysis.
Streaming prediction on the conveyor
Sensor readings are read continuously, the estimate updates as the ore moves — the blend route can be changed before the material goes to the dump. The limitation is sensitivity to signal interruptions: when calibration drifts, accuracy falls unnoticed by the operator.
The hybrid scheme removes some of these limitations: the model provides an estimate, while geologist rules cover disputed cases by ore class.
We fix the applicability boundaries before launch — where the prediction is reliable and where the operator receives a warning; results remain reproducible according to JORC requirements.
Stages of implementing a grade prediction model at a plant
The workflow goes from historical sampling data to a model embedded in the plant loop. Each stage is closed with an artifact that can be verified — otherwise we do not start the next step. Below is the sequence and what will be required on your side.
Audit of sampling data and labeling
- Sampling data audit: we reconcile sampling logs, sample certificates, and laboratory records, looking for gaps and discrepancies.
- Sample labeling: we tie the analysis result to the batch, sampling time, and point in the process flowsheet.
- Training on historical data: we split the sample by time, not randomly — this keeps validation honest.
- Metric alignment: we fix the acceptable share of classification and grade errors by which we judge.
| Stage | What we do | What is needed from the customer |
|---|---|---|
| Sampling data audit | We reconcile logs, sample certificates, laboratory records | Exports for the available period, reference data |
| Sample labeling | We tie the analysis to the batch and flowsheet point | Confirmation of the sampling methodology |
| Training on historical data | We assemble the sample and separate validation by time | Agreement on the period and list of grades |
| Prediction acceptance | We compare with the laboratory on a holdout period | Access to control samples |
Pilot operation and acceptance
The model outputs a prediction in parallel with the current decision and does not replace anything. Acceptance is based on a holdout period: we look at the share of classification errors overall and separately for transition zones and “difficult” grades.
After acceptance, we transfer the model to operation: retraining procedure, drift control, responsible persons at the plant.
Integration schemes, configurations, and retraining procedures
The customer receives not only the trained model but also everything for independent operation: configurations, interfaces with the accounting system, access policies, and procedures. We agree on the list of artifacts with the plant’s technical service before production launch.
Integration scheme with the plant accounting system
- Flow map: which reference data (block, level, shift, ore type) comes from where and who owns the source.
- Interface through an industrial gateway: read-only exchange, buffering during communication loss, retry without duplicates.
- Rules for matching keys between the accounting system and the model storage, a discrepancy log for the dispatcher.
- Data access policies: roles and visibility scopes, separation of industrial and personal data.
- Model documentation: a model card with the list of features, history window, metrics, and applicability boundaries.
Retraining and drift control procedure
The procedure specifies who launches retraining and on what signal so that the model does not degrade unnoticed.
- Triggers: shift in the feature distribution, drop in the metric on the holdout sample, change in the sampling scheme.
- Data drift control: per-shift checks of gaps and ranges, weekly reconciliation of prediction against fact.
- Acceptance procedure for a new version before entry into the ore quality control loop.
- Rollback procedure: the previous version of the model and configuration are stored, return in one step, a responsible person is assigned.
Artifacts remain in the customer’s repositories and systems — operation does not depend on us.
Case study: how we reduced ore classification errors by 30%
What it was like before implementation
Ore grade was determined from a laboratory sample: collection, delivery, analysis — the verdict came several hours later, when the stream had already moved further through the flowsheet.
During this window, the operator worked from indirect signs — crusher load, particle size, pulp color — so decisions from different shifts diverged.
At the Rudny processing plant, customer measurements showed 21% erroneous grade assignments and a spread between shifts of up to 8 percentage points on the same type of ore.
The transition took place without stopping production. Mill telemetry, laboratory protocols, and shift logs were brought to a single time scale so that each sample was matched to the operating regime at that moment.
The grade prediction model was launched in advisory mode: the operator confirmed or rejected the estimate, and every discrepancy was returned to the training set. The prediction was moved into the working loop when the share of confirmations stopped growing.
Result in measurable terms
The share of erroneous grade assignments fell by 30% against the baseline measurement, and the spread between shifts decreased almost threefold — the decision stopped depending on the experience of a particular operator.
The prediction is recalculated from current telemetry every few minutes, so the stream is switched to another grade before processing, not after the laboratory verdict.
Metrics were calculated on control samples that did not enter training: once per shift, the model’s verdict was reconciled with the laboratory. Such a measurement shows the result of implementation, not fitting to history.
How ML grade prediction differs from a geological block model
A geological block model answers the question “what is the average grade in this volume of subsurface.” It is built by interpolating sampling data — drillholes, channel and sludge samples — and produces a smoothed field: peaks and troughs are averaged over a block that is an order of magnitude larger than a bucket.
For resource estimation this is correct, but at the cut-off grade boundary, smoothing starts to work against you.
ML grade prediction answers a different question: what exactly is in this bucket, truck, or ore stream right now.
The model learns on the same sampling data, but together with operational features — ore type, oxidation, particle-size distribution, hardness, drilling rate, spectral measurements. It captures nonlinear relationships and interactions that geostatistics does not describe.
The difference between the approaches is in the horizon and in the carrier of the prediction. The block model lives in subsurface coordinates and predicts an average over volume; the machine prediction lives in process coordinates and predicts a specific batch of material.
That is why the metrics are different too: for the block — resource reproducibility, for ML — the share of correctly assigned batches at the cut-off grade threshold and the cost of error in each class.
Combine them, do not oppose them. The block prediction is fed into the model as a spatial feature, while ML is further trained on actual sampling data — ore assays, shipment certificates, plant results — and removes the systematic shift that sends ore to the dump.
One condition: unified sampling methodologies and labeling of discrepancies between the laboratory and online sensors. Otherwise the prediction will inherit assay noise and confidently err.
Technical concerns before launching the model into the control loop
Engineers rightly ask what the model does between sample collection and the arrival of the analysis result, and whether its output can be sent to setpoints.
Below are five concerns that we resolve together with process engineers and the process control system service before the prediction enters the control loop.
Laboratory analysis arrives with a delay of several hours — is the prediction “blind” all that time?
The model is trained on the pair “current process indicators → delayed analysis result,” so in each cycle it relies on live equipment measurements, not on the latest sample.
The laboratory is pulled in as a target variable and refines the weights during recalculation, but it does not block the output.
For the interval between samples, we produce a prediction with a widened uncertainty interval — the operator sees how much the model can be trusted right now.
Ore composition changes — does the model “drift apart” over time?
We calculate drift continuously: we compare the distribution of input features and prediction residuals against a reference window. If a threshold is exceeded, recalculation on fresh data is launched, and if quality degrades, a rollback to the previous saved version occurs.
This way, a change in level or raw material supplier does not turn into quietly incorrect setpoints.
How is the prediction interfaced with dispatching?
We perform the interface through the standard exchange interfaces of the dispatching system — tags, value archive, task tables — without intermediary layers that add extra delay.
The model publishes a value with a confidence label, while the interlock logic remains in the controller: it still decides whether to accept the recommendation.
Who is responsible for retraining and when should it be launched?
Retraining follows a procedure with two triggers: calendar-based and metric-degradation-based. Each model version is versioned together with the dataset and settings, so any prediction is reproducible, and a disputed section can be recalculated anew.
Responsibility for launching is assigned to a specific role, not “spread” between the contractor and the process control system service.
Is “average” accuracy enough for control?
No, and we do not treat it that way. We look at the error in the working grade range and at transitions between grades — that is exactly where the cost of a miss is highest.
Plus a confidence rule: with a wide interval, the model gives advice to the operator rather than replacing the setpoint, so disputed periods do not go into automation silently.
Let us discuss your ore grade prediction task
To come back with an assessment and a plan rather than general words, we need to understand the physics of your processing operation and what data is already recorded at the site.
Below is what we clarify in the first conversation: it is usually enough to name the class of task, the expected prediction accuracy, and the main risks.
- Which indicator we are predicting: metal grade, grade by class, or particle-size distribution — and how it is measured today.
- What is already accumulating in history: conveyor scale and online analyzer readings, grinding and flotation regimes, batch certificates.
- Where the delay occurs between sample collection and the analysis result — it is most often what breaks prediction accuracy.
- How the grade assignment error is calculated and what it changes in blending and recovery.
- Who uses the prediction: an operator prompt or automatic regime correction.
- Where we deploy the computations: the plant loop, access to historical data, the procedure for updating reference data.
- Who on your side accepts the model and provides feedback on prediction quality.
After that, we propose a pilot at one site: we assemble a sample for past periods, train the model, reconcile the prediction with laboratory results, and calculate the share of correct grade assignments.
If the metrics hold on the holdout period, we move to transferring the model into operation: retraining procedure, data shift control, monitoring, and instructions for the shift. Describe the task — we will come back with an assessment and a work plan.
