Table of Contents
Ore grade prediction for grade control estimates the grade of each small block in a blasted bench so that every truckload goes to the right destination: mill, stockpile or waste dump. Most mines use kriging on blasthole assays. Machine learning can add drilling and sensor data, but published comparisons show it does not beat kriging by default.
Key takeaways
- Grade control decides where mined material goes, using all available data. Misclassifying a block causes either ore loss or dilution, and both cost money (Zacche da Silva et al., 2020).
- Blasthole samples are dense but hard to take correctly: cuttings segregate, subdrill material is included and the fractured top of the bench yields few cuttings (Holmes, 2010).
- In published comparisons, machine learning alone did not reliably beat ordinary kriging (Potakey, 2024); a hybrid that fed a machine learning trend into kriging did better in one copper case study.
- Random cross-validation flatters spatial models. Test by holding out whole blasts or spatial blocks (Akpabio and Minnitt, 2026).
- Reconciliation with mill results, not model accuracy alone, shows whether a new method adds value (Ebrahimi, 2025).
What is grade control, and what does misclassification cost?
In an open pit, the long-term resource model is built from widely spaced exploration drillholes. It is too coarse to decide where to dig. Grade control fills the gap: as each bench is drilled for blasting, samples are taken, a short-term grade model is built and dig lines are drawn between ore and waste before loading. One study states the goal as predicting "the destination of extracted material based on all available data collected" (Zacche da Silva et al., 2020).
Each block can be wrong in two ways:
| Error | What happens | Cost to the mine |
|---|---|---|
| Ore loss | A block above cutoff is classified as waste and dumped | Metal never recovered; lower revenue |
| Dilution | A block below cutoff is classified as ore and milled | Mill capacity, energy and reagents spent on material that does not pay; lower head grade |
The two costs are rarely equal, which matters when deciding how to use a model. The gap between plan and outcome can be large: in reconciliation work at six mines, an SRK consultant observed reported tonnages deviating from plan by -4% to +12% and grades by -7% to +20% (Ebrahimi, 2025). Not all of that comes from grade control, but grade control is one stage where the gap can be measured and reduced. For the wider picture of AI in exploration and mining, see our guide to AI in mining and AI-powered geological analysis.
What data feeds a grade prediction model?
| Data source | What it provides | Main limitations |
|---|---|---|
| Blasthole cone samples | One assay per hole on a dense pattern (5 m spacing in one copper study) | One sample for the whole hole, for example 10 to 15 m deep; segregation and subdrill bias the sample |
| Reverse circulation (RC) grade control drilling | Smaller samples that are easier to collect; several benches drilled at once | Considerably higher cost |
| Measure-while-drilling (MWD) | Penetration rate, torque, rotation and air pressure for every hole, at 0.1 m depth resolution | Indirect: reflects rock properties, not chemistry |
| Geological logging and the long-term model | Lithology, alteration, domain boundaries | Logging is subjective; boundaries can be misplaced |
| Sensors at the face (for example XRF on shovel buckets) | A grade estimate for each bucket during loading | Performance figures come mainly from vendors |
Blasthole samples are the default because the holes are drilled for blasting anyway. They are also the weak link. Holmes describes segregation of cuttings in the cone and the inclusion of subdrill material from below the bench. Because the top of the bench is already fractured, few cuttings come from the top few metres, so the sample is biased towards the middle and lower bench. He cites separate RC drilling as the best solution, at considerably higher cost, and warns that "bias cannot be eliminated once it is present" (Holmes, 2010). No model can correct a biased sample, so a sampling audit comes before any modelling. Our post on industrial data quality problems covers the general version of this problem.
MWD data are logged by modern drill rigs for every hole. In a pilot at an iron ore mine in the Pilbara, with MWD from almost 7,000 holes, models predicted iron content with a correlation of 0.79 against lab assays in one region but only 0.64 in another, where the grade distribution was different (Khushaba et al., 2022). MWD is a useful extra input, not a replacement for assays.
Sensors at the face are newer. MineSense reports that its shovel buckets fitted with XRF sensors, with machine learning to characterise each bucket, re-routed on average 5.9% of measured trucks from ore to waste and 7.2% from waste to ore at nine open-pit copper mines over part of 2023 (CIM Magazine, 2024). These are the vendor's figures, not an independent study.
Kriging, simulation or machine learning: which method?
Ordinary kriging is the standard. A survey of Pilbara iron ore operations calls it the "method of choice" in geostatistics, usually paired with sequential Gaussian simulation to compensate for kriging's smoothing and to estimate grade uncertainty. Gaussian processes, a probabilistic machine learning method, have been adapted to treat blasthole assays as intervals rather than points (Leung et al., 2023).
General-purpose machine learning, such as random forests, support vector machines and neural networks, can combine many inputs and does not need kriging's stationarity assumption within a domain. But it "does not guarantee data reproduction at sampled locations" and does not explicitly account for spatial correlation (Zacche da Silva et al., 2020).
What do head-to-head comparisons show?
- Two open-pit projects: ordinary kriging misclassified about 8.5% of blocks and a support vector classifier about 10.5%. Adding sampling error reduced the performance of both by about 4% (Potakey, 2024).
- Platinum, coordinates only: k-nearest neighbours and random forests looked excellent under random cross-validation (R² of 0.92 and 0.86) but often scored below zero under spatial block cross-validation, while kriging stayed more stable. The authors conclude that coordinate-only models "behave as interpolators" (Akpabio and Minnitt, 2026).
- Copper, hybrid: at the Carmen de Andacollo mine in Chile, using a neural network trend inside collocated co-kriging misclassified 5.4% less material than ordinary kriging over 10 blasts. The benchmark was simulated, because true reconciliation data were not available (Zacche da Silva et al., 2020).
| Method | Strengths | Weaknesses | When to use |
|---|---|---|---|
| Ordinary kriging | Standard, auditable, honours the data | Smooths; one estimate per block; assumes stationarity within a domain | Baseline for every project |
| Conditional simulation | Quantifies uncertainty; supports expected-profit decisions | More computation and expertise | High misclassification costs or uncertain boundaries |
| Machine learning on coordinates only | Quick to try | Acts as an interpolator; weak under spatial validation | Rarely on its own |
| Machine learning with drilling and geological inputs | Uses MWD, logging and sensor data | Needs assays as labels; can fail in new areas | When extra data exist for every hole |
| Hybrid (machine learning trend with kriging) | Combines extra inputs with spatial correlation | More complex workflow | Strong trends within a domain |
Machine learning earns its place when it brings information kriging does not use, such as drilling response, logged geology or sensor readings, not when it reinterpolates the same assays. The same logic underlies soft sensors in process plants: cheap measurements estimate a costly one, and the lab stays the reference.
How should uncertainty change ore and waste decisions?
Comparing a single estimate with the cutoff ignores that the two errors cost different amounts. A better approach uses many simulated realisations of grade. Datamine's CSOWOPT process computes, for each block, the expected loss of sending it to the mill or to waste, using mill recovery, metal price, grade and the cost of mining and milling. A block is classified as ore only when its probability of exceeding the cutoff is above a breakeven probability (Datamine Studio RM documentation). Magri and Norrena described the same idea as choosing the classification that "maximizes expected profit" across realisations (Magri and Norrena, 2000).
This applies to machine learning too. A model that returns a probability or a distribution, not just a grade, can feed the same decision rule.
How do you validate a grade model before trucks follow it?
- Hold out whole blasts or spatial blocks. Khushaba et al. used leave-one-blast-out testing, which mimics predicting a new blast, and note that random splits can give over-optimistic errors because of spatial autocorrelation (Khushaba et al., 2022).
- Score decisions, not only grade error. Count tonnes misclassified as ore and as waste, and value each at current prices and costs.
- Compare with the current method on the same blasts. Kriging on the same data is the minimum baseline.
- Check each domain and area separately. Performance can drop in a new part of the pit, as the 0.79 and 0.64 correlations above show.
- Run in parallel. Let the model propose dig lines alongside the existing process before it changes truck destinations.
- Reconcile. SRK's F1 compares the short-term grade control model with the long-term model, and F2 compares what the mill receives with what the mine reports delivering (Ebrahimi, 2025). A better model should narrow the gap between grade control predictions and mill results.
After go-live, grade models drift as mining moves into new domains. Our guide to model drift monitoring covers what to track.
What we have seen at KABAS AI
KABAS AI has not published a grade control project. What follows is how our data science team approaches this kind of work on industrial data:
- Audit sampling first. Check how blastholes are sampled, whether subdrill is included and how sample IDs, coordinates, depths and assay dates are linked. Fix bias before modelling.
- Build the kriging baseline. A machine learning model has to beat what the mine geologist already produces on the same blasts.
- Add inputs that exist for every hole. MWD, logged lithology and sensor data are where machine learning can add information.
- Keep the data on site. Drillhole, assay and grade data are commercially sensitive, and many pits have weak connectivity. Our mining solutions run on-premise without an internet connection, and our post on AI at remote sites explains the trade-offs.
If you want to test whether your blasthole and drilling data can support better grade predictions, contact us.
Sources
- Zacche da Silva, Nisenson and Boisvert, Grade control with Ensembled ML: A comparative case study at the Carmen de Andacollo copper mine, CCG Annual Report 22, 2020
- Holmes, Sampling mineral commodities: the good, the bad, and the ugly, Journal of the Southern African Institute of Mining and Metallurgy 110, 2010
- Akpabio and Minnitt, Improving grade estimation using machine learning: a comparative study of ordinary kriging against machine learning algorithms, SAIMM, 2026
- Potakey, Grade Control Optimization in Open Pit Mining using Supervised Machine Learning Approach, Master's thesis, Queen's University, 2024
- Khushaba, Melkumyan and Hill, A Machine Learning Approach for Material Type Logging and Chemical Assaying from Autonomous Measure-While-Drilling (MWD) Data, Mathematical Geosciences, 2022 (arXiv 2202.02959)
- Leung, Hill and Melkumyan, Automation and AI Technology in Surface Mining With a Brief Introduction to Open-Pit Operations in the Pilbara, arXiv 2301.09771, 2023
- Ebrahimi, Reconciliation Explained and Solutions Recommended, SRK Consulting, 2025
- Datamine, CSOWOPT process, Studio RM documentation
- Magri and Norrena, Optimal grade control using geostatistics and economics: methodology and examples, SME, 2000
- CIM Magazine, Industry partner spotlights 2024: MineSense
Frequently asked questions
What is the difference between a resource model and a grade control model?
The long-term resource model is built from widely spaced exploration drillholes and is used for reserves and long-term planning. The grade control model, also called the short-term model, is built bench by bench from denser data such as blasthole or grade control drilling assays. It is the model used to draw dig lines and send each truck to the mill, a stockpile or the waste dump.
Can machine learning replace kriging for grade control?
Not by default. In published comparisons, machine learning models that use only sample coordinates and assays did not beat ordinary kriging, and some failed under spatial validation. Machine learning helps when it adds information kriging does not use, such as drilling data, logged geology or sensor readings, or when its output is combined with kriging.
Can measure-while-drilling data replace blasthole assays?
No. MWD data such as penetration rate, torque and air pressure are recorded for every hole and can be correlated with grade, but the link is indirect and varies between areas of the same mine. Assays remain the reference used to train, check and correct any model that uses MWD.
What are ore loss and dilution?
Ore loss is ore that is sent to the waste dump because it was classified as waste, so its metal is never recovered. Dilution is waste that is sent to the mill because it was classified as ore, which consumes mill capacity, energy and reagents and lowers the head grade. Grade control tries to minimise the combined cost of both errors.
How do we know whether a new grade control method is working?
Run it in parallel with the current method on the same blasts, compare the tonnes it would misclassify, and then track reconciliation over time. If the new method is better, the gap between what the grade control model predicts and what the mill receives should narrow. Model accuracy figures alone are not enough.



