[Mining](https://bfcs.ai/industry/mining/) · [Faster](https://bfcs.ai/faster/) # Can AI Speed Up Ore Grading? Open-pit grade control has long waited on the laboratory: drill, sample, assay, then decide where the dig line goes. Sensors on conveyors and shovels, their developers say, now read grade in seconds, and machine learning turns drilling data into estimates. The measurement is faster. Whether the decision gets faster, and better, depends on the orebody and on who signs off. Updated **October 2026** · 7 min read · Evidence-graded · 8 sources ## The BFCS verdict Every article scored on the same four dimensions | Dimension | Verdict | Why | | --- | --- | --- | | Better | Partly · Low to moderate confidence | Finer decisions where the orebody varies · Sensors can classify ore in tonnes rather than whole dig blocks, but the strongest site results come from the vendor MineSense and the developer NextOre. In a peer-reviewed Pilbara study, machine learning performed similarly to an expert-calibrated method rather than beating it. | | Faster | Partly · Moderate confidence | Seconds at the sensor; the sign-off is unchanged · Bulk sensors report grade in seconds against a laboratory loop that, in an AusIMM paper, miners want closed within two or three days. The speed figures are developer-reported (CSIRO and NextOre), and the models still learn from assays. | | Cheaper | Conditional · Low confidence | Only where grade varies at sensor scale · Diverting waste before the mill saves processing cost only if the orebody has a heterogeneity signature. The cost and copper figures we found are a CSIRO researcher's estimate and MineSense's own extrapolation. | | Safer | Conditional · Governance-gated | Reporting integrity, not injury rates · We found no study measuring injury outcomes from sensor-based grade control. The safety question here is whether automated ore and waste calls stay traceable to estimates a Competent Person can stand behind. | ## The problem, in numbers In an open pit, grade control decides which rock goes to the mill and which goes to the waste dump. The traditional loop is slow. A blast pattern is drilled, the cuttings are sampled, the samples are prepared and assayed, a geologist draws dig lines, and only then does the shovel move. In an AusIMM conference paper, the sampling specialist F. F. Pitard wrote that the miner wants the grade control geologist to decide within two or three days at most, which he argued leaves too little time for sampling, preparation and the laboratory to do a good job. (Source 1) That is his assessment of practice, not a measured survey. He also judged the resulting misclassification of ore and waste responsible for severe reconciliation problems and large financial losses each year. (Source 1) The loop is also coarse. An International Mining feature on CRC ORE says a dig block may be about 15,000 tonnes, a truck 100 to 300 tonnes and a shovel 50 to 100. (Source 5) In our reading, a call made per block sends the whole block one way, even when the grade inside it varies. ### The evidence, in four numbers 2 to 3 days · the most, in one sampling expert's account, a miner wants to wait for a grade control call · author's assessment of blasthole practice, not a measured survey (AusIMM, 2008) 16.7 t vs 15,000 t · a sensor diversion unit against a dig block · calculated example in an International Mining feature on CRC ORE, at 2,000 t/h with a 30-second reading 2,800 t/h · sorting rate claimed at the Kansanshi mine · developer-reported by NextOre, the CSIRO joint venture, 2022 5.9% and 7.2% · of measured trucks rerouted to waste and to ore, on average · nine copper mines, part of 2023; vendor-reported by MineSense, not independently verified ## What improves, and why Sensing changes both the speed and the scale of the decision. CSIRO developed a magnetic resonance sensor that, its developers say, reads all the ore on a conveyor rather than only the surface. (Source 4) In a trial on the 1,300 t/h Ridgeway portal conveyor at Cadia, it measured chalcopyrite, and by inference copper grade, down to a one-tonne scale for months at a time. (Source 3) The developers say the commercial version reports grade in seconds. (Source 4) Scale matters as much as speed. In the same article's worked example, a conveyor running at 2,000 t/h with a sensor that integrates over 30 seconds can divert as little as 16.7 tonnes at a time, and an X-ray fluorescence (XRF) sensor reading every 10 seconds 5.6 tonnes. (Source 5) The vendor MineSense describes shovel-mounted XRF that moves the call earlier still, to the bucket, with machine learning aggregating each bucket into a truck destination. (Source 6) Machine learning works on the drilling side too. Measure-while-drilling (MWD) records penetration rate, torque and other signals from blast holes. A 2024 peer-reviewed review notes that this data is cheap and plentiful compared with sparse exploration holes, and that machine learning allows real-time or near real-time analysis of it. (Source 8) > A sensor can read grade in seconds. It cannot create variability the orebody does not have, and it does not sign the resource estimate. ## What the evidence actually says The physical case for sorting has some peer-reviewed support. In a journal paper, CSIRO researchers found fractal ordering of copper grade in bore core from the Cadia Ridgeway porphyry, then monitored mine output every 20 seconds for a month to test whether it survived mining. (Source 3) The abstracts we could read do not state that result. They argue that such distributions, with many low-grade pods and large-scale clustering of grade, favour separating ore from waste. (Source 3) The operating results are mostly reported by the people selling the systems. NextOre, the CSIRO joint venture that commercialised the sensor, claims in its own release a sorting rate of 2,800 t/h at the Kansanshi mine and calls it the highest-capacity ore sorting plant in the world. (Source 4) A CSIRO researcher estimated in 2017 that processing costs could fall by as much as 20% in some mines, a projection from the team's own numbers. (Source 4) MineSense reports that across nine copper mines in part of 2023, its shovel system sent on average 5.9% of measured trucks to waste that had been marked as ore, and 7.2% to ore that had been marked as waste. (Source 6) At one unnamed site, shifts with 80 to 100% coverage showed about 4% better recovery than shifts under 20%; MineSense does not say whether that is relative or in percentage points. (Source 6) These are vendor figures without named sites or sample sizes, and the shift comparison is observational. On grade estimation itself, the evidence is promising but modest. A systematic review of 31 studies from 1990 to 2019 found machine learning outperformed conventional estimation methods. (Source 8) In our reading, a published review of published studies is likely to favour positive results. The most useful Australian test we found came from a Pilbara iron ore deposit: researchers supported by the Australian Centre for Field Robotics and the Rio Tinto Centre for Mine Automation used machine learning to update geological boundaries with blasthole assays, training on 35,733 exploration samples (arXiv author version). (Source 7) They found it performed similarly to the expert-calibrated method, with a good prospect that stronger models could match or exceed it. (Source 7) In our reading, that is automation of a manual step, not a leap in accuracy, and it still runs on assays. ## Where it falls short - **No variability, no value.** CRC ORE's principal geologist says bulk ore sorting implementation "is not a plug and play opportunity" and cannot be implemented without a heterogeneity signature. (Source 5) - **Block models hide the signal.** Predicting sorting performance needs the grade-tonnage curve at sensor scale; an estimator that smooths grade at that scale will not predict the separation correctly, and kriging is described as less suited to it. (Source 5) - **Laboratory results are not plant results.** One multispectral sorting study classified samples by arsenic level with 91.9% accuracy or more, on 14 quartz-diluted samples under idealised laboratory lighting, and named field trials as future work. (Source 8) - **The clock that matters is unmeasured.** We found no independent study timing the full loop from drilling to dig decision before and after sensors. ## Where governance decides the outcome Public reports of Mineral Resources and Ore Reserves in Australia must be based on documentation prepared by a Competent Person: a member of the AusIMM, the Australian Institute of Geoscientists or a Recognised Professional Organisation, with at least five years of relevant experience, who gives prior written consent to how their work is reported. (Source 2) Daily dig calls are not public reports, but in our reading the reconciliation they feed becomes part of the evidence behind those estimates. Sensor calibrations, model versions and every automated diversion should be logged so a Competent Person can trace them. JORC said in March 2026 that a provisional 2026 Code was to go to ASX, ASIC and peer bodies in early Q2 2026, with ASX consultation and a transition period to follow; the requirements above are those of the 2012 Edition. (Source 2) ## How to think about adoption Start with the orebody, not the sensor. Test whether grade varies at the scale a sensor can act on, using bore core and production data rather than smoothed block models. Measure the current loop first: hours from drilling to assay, from assay to dig line, and how often reconciliation disagrees. Run sensors alongside the laboratory before letting them divert ore. Treat developer throughput claims and vendor recovery figures as hypotheses to test on your own benches, and agree with your Competent Person how sensor and model outputs will be documented. ## How we read this evidence BFCS.ai does not fabricate figures. Throughput, timing and cost figures for bulk sorting come from the technology's developers, and the truck and copper figures from the vendor MineSense; we label both. The peer-reviewed work we found measures grade variability and model accuracy, not end-to-end time from drilling to dig decision, and we found no independent study of that. ## Sources 1. Pitard, F. F., "Blasthole Sampling for Grade Control: The Many Problems and Solutions," Sampling 2008 Conference, The Australasian Institute of Mining and Metallurgy (AusIMM), 2008 (abstract read on the AusIMM and OneMine listings). [ausimm.com](https://www.ausimm.com/publications/conference-proceedings/sampling-2008-conference/blasthole-sampling-for-grade-control---the-many-problems-and-solutions/) **What we took from it:** Conventional blasthole sampling has "an extremely bad reputation"; biases arise in delimitation, extraction, weighting and preparation; "the miner wants the ore grade control geologist to make up his mind within two or three days at most," leaving not enough time for samplers, preparation, laboratory and resources staff; the resulting ore grade misclassification is described as responsible for severe reconciliation problems and "massive financial losses every year". 2. Joint Ore Reserves Committee (JORC), "Competent Persons" and "Code Update" pages, read October 2026. [jorc.org](https://jorc.org/competent/) Code update: [jorc.org](https://www.jorc.org/code-update/) **What we took from it:** Public reports must be based on and fairly reflect documentation prepared by a Competent Person, who must be a Member or Fellow of the AusIMM, the Australian Institute of Geoscientists or a Recognised Professional Organisation, have at least five years' experience relevant to the style of mineralisation and the activity, and give prior written consent to how their work is reported (Clause 9 of the 2012 Edition; ASX Listing Rules 5.6, 5.22 and 5.24). an exposure draft was released on 1 August 2024; in March 2026 JORC said a provisional JORC Code 2026 was being finalised and was to be shared with CRIRSCO, ASX, ASIC and JORC Parent Body representatives in early Q2 2026; the latest update was a webinar on 20 August 2026; once final there will be a mandatory ASX consultation and the new Code will operate after an agreed transition period. 3. CSIRO bulk sensing research. [doi.org](https://doi.org/10.1016/j.gsf.2014.09.003) Coghill, P. J., Miljak, D. & McGann, A., "Modelling Grade Variation in Crushed Ore at Cadia Ridgeway and Its Potential Consequences for Bulk Ore Sorting," Third AusIMM International Geometallurgy Conference, 2016, pp. 77-88 (conference paper; abstract only, read on the AusIMM listing). [ausimm.com](https://www.ausimm.com/publications/conference-proceedings/third-ausimm-international-geometallurgy-conference-2016/modelling-grade-variation-in-crushed-ore-at-cadia-ridgeway-and-its-potential-consequences-for-bulk-ore-sorting/) **What we took from it:** Coghill, P., Miljak, D. & Williams, E., "Consequences of fractal grade distribution for bulk sorting of a copper porphyry deposit," Geoscience Frontiers 6(4), 2015, pp. 477-480 (peer-reviewed; abstract only, read via the Semantic Scholar and ANU repository records; the Elsevier page refused automated access). Fractal ordering of copper grade found in bore core from the Cadia Ridgeway porphyry; the authors "measure its persistence after mining" by monitoring mine output every 20 seconds for a month with a large-scale, zero-field magnetic resonance sensor (the abstract does not state the result); fractal distributions have many low-grade pods and large-scale clustering, both favourable for segregating ore. CSIRO trialled a magnetic resonance sensor measuring chalcopyrite, and by inference total copper grade, on the 1,300 t/h Ridgeway portal conveyor, producing data "down to a 1 t scale for months at a time"; the authors measure the persistence of the fractal ordering in crushed rock from the block cave over one month (again, the abstract does not state the result). 4. Developer-reported data (CSIRO technology commercialised by NextOre, a joint venture of CSIRO, RFC Ambrian and Worley; not peer-reviewed). [csiro.au](https://www.csiro.au/en/news/all/articles/2017/december/sorting-out-waste) NextOre partner release hosted by CSIRO, "NextOre wins major mining technology solution award," 6 December 2022: [csiro.au](https://www.csiro.au/en/news/All/News/2022/December/NextOre-wins-major-mining-technology-solution-award) **What we took from it:** CSIRO, "Sorting out waste," 11 December 2017: the magnetic resonance system reads ore in real time and informs sorters "in seconds"; the article (not a quoted spokesperson) states it can pass material at 5,000 tonnes per hour; in 2017 the partners forming NextOre were named as CSIRO, Advisian and RFC Ambrian; CSIRO's David Miljak estimates processing costs could fall by "as much as 20 per cent in some mines" per pound of copper or ounce of gold, based on the team's own numbers. grade readings in seconds; typical rates exceeding 5,000 t/h per conveyor; commercially deployed at several copper operations; the Kansanshi unit (First Quantum) operates at 2,800 t/h, which NextOre claims is the highest-capacity ore sorting plant in the world. 5. CRC ORE (Cooperative Research Centre for Optimising Resource Extraction). [im-mining.com](https://im-mining.com/?p=73334) Rutter, J. & Dunham, S., "Heterogeneity and bulk ore sorting: methods to estimate ore sorting performance," AusIMM, 2020 (abstract only, read on OneMine): [onemine.org](https://onemine.org/documents/heterogeneity-and-bulk-ore-sorting-methods-to-estimate-ore-sorting-performance) **What we took from it:** International Mining, "CRC ORE simplifies complexity for value," 18 February 2021, reporting CRC ORE Principal Geologist Jon Rutter's assertion that "bulk ore sorting implementation is not a plug and play opportunity" and quoting him that "without a heterogeneity signature, we cannot implement bulk ore sorting"; the article says a selective mining unit may be a dig block of about 15,000 t, a truck 100 to 300 t and a shovel typically 50 to 100 t; in its own worked example (not attributed to a speaker), at 2,000 t/h a PGNAA or PFTNA sensor with a 30-second integration time gives a sensor diversion unit of 16.7 t, and an XRF sensor at 10 seconds 5.6 t. estimating performance requires the grade-tonnage curve at the scale of the sensor diversion unit; "an estimator that smooths the SDU distribution will not correctly predict ore/waste separation or economic performance"; ordinary kriging and inverse distance weighting smooth the grade profile and are described as less suited to modelling local heterogeneity. 6. Vendor case study (self-reported, not peer-reviewed). [cim.org](https://www.cim.org/industry-partners/spotlights-2024/minesense/) **What we took from it:** CIM Magazine, MineSense industry partner spotlight, December 2023 to January 2024 issue, content provided by the vendor. ShovelSense uses X-ray fluorescence sensors on shovels and machine learning to characterise each bucket, aggregated to trucks routed through fleet management. Across nine open-pit copper mines for part of 2023, on average 5.9% of measured trucks were sent to waste that had been designated ore and 7.2% to ore that had been designated waste. At one unnamed site, shifts with 80% to 100% coverage were observed to have a "four per cent improvement in recovery" and "eight per cent more copper in the concentrate" than shifts under 20% coverage (whether relative or in percentage points is not stated; the vendor calls this a direct result of ShovelSense, but the comparison is observational, with no period or sample size); extrapolated to a full year at 90% utilisation, digging units with ShovelSense recovered 500 to 9,000 tonnes more copper per year. Mines, sample sizes and price assumptions are not given. 7. Leung, R., Balamurali, M. & Lowe, A., "Surface Warping Incorporating Machine Learning Assisted Domain Likelihood Estimation: A New Paradigm in Mine Geology Modelling and Automation," Mathematical Geosciences, 2021 (peer-reviewed; read as the arXiv author version, v3, 14 September 2021), supported by the Australian Centre for Field Robotics and the Rio Tinto Centre for Mine Automation. [doi.org](https://doi.org/10.1007/s11004-021-09967-5) (author version: [arxiv.org](https://arxiv.org/abs/2103.03923) **What we took from it:** Geological surfaces built from sparse exploration data are updated with production blasthole assays; machine learning replaces a manually calibrated domain likelihood. Trained on 35,733 exploration assays from a Pilbara iron ore deposit; evaluated on two deposits. The authors conclude that machine learning warped models achieved a performance gain similar to the expert-calibrated (y-chart) model, and that high-capacity estimators (random forest, multilayer perceptron, gradient boosting) have a good prospect of matching or exceeding it. The published Springer version was not opened; figures are from the author version. 8. Peer-reviewed reviews and laboratory work. [doi.org](https://doi.org/10.17159/2411-9717/1250/2022) Goldstein, D. M., Aldrich, C. & O'Connor, L., "A Review of Orebody Knowledge Enhancement Using Machine Learning on Open-Pit Mine Measure-While-Drilling Data," Machine Learning and Knowledge Extraction 6(2), 2024, article 63 (abstract only, read via Crossref and Semantic Scholar; the MDPI page refused automated access): [doi.org](https://doi.org/10.3390/make6020063) Okada, N. et al., "Optimizing multi-spectral ore sorting incorporating wavelength selection utilizing neighborhood component analysis for effective arsenic mineral detection," Scientific Reports 14, 11544, 2024: [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC11109168/) **What we took from it:** Mahboob, M. A., Celik, T. & Genc, B., "Review of machine learning-based Mineral Resource estimation," Journal of the Southern African Institute of Mining and Metallurgy 122(11), 2022 (abstract only, read on SciELO): systematic review of 31 studies published 1990 to 2019; machine learning methods outperformed conventional grade estimation methods. MWD data from blast holes are cheap and plentiful compared with sparse exploration data; machine learning allows real-time or near real-time analysis; applications cover rock mechanics, lithology and geochemical species; the most important MWD metrics are rate of penetration, torque, weight on bit, bit air pressure and rotation speed; multilayer perceptrons are most used, mainly to identify rock types. 14 samples with simulated grades made by diluting powdered Chuquicamata copper ore with quartz; samples classified by arsenic level with identification accuracy of 91.9% or more using eight or more selected bands (abstract figure); laboratory imaging under idealised lighting, with field trials named as future work. Putting this to work ## Working through this in your own organisation? BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use. [Talk to Inhouse CX →](https://inhouse.cx/engage/) Continue the thread - [Can AI Optimise Haulage Costs?](https://bfcs.ai/articles/can-ai-optimise-haulage-costs/): More in Mining · MINING · CHEAPER · 6 min · Read → - [Can AI Predict Equipment Failure?](https://bfcs.ai/articles/can-ai-predict-equipment-failure/): More in Mining · MINING · BETTER · 6 min · Read → - [Can AI Increase Line Throughput?](https://bfcs.ai/articles/can-ai-increase-line-throughput/): Also Faster · MANUFACTURING · FASTER · 7 min · Read → --- Canonical page: https://bfcs.ai/articles/can-ai-speed-up-ore-grading/ This is the Markdown version of the page, generated from its HTML at build time. Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml