AI in Mineral Crushing and Grinding: Pros and Cons

AI can trim the energy use and downtime of a comminution circuit, but the published proof is thinner than most brochures suggest. Crushing and grinding, together called comminution (breaking ore into smaller pieces so the valuable minerals can be separated), is estimated to take between a quarter and just over half of a mine’s energy, depending on the study you read. One university study that used plant history projected a power cut of up to 2% for a rod mill, while equipment changes such as high-pressure grinding rolls have reported cuts of 21% to 34%. This page lists the jobs AI does in a circuit, the pros and cons with sources, how it compares with other upgrades, and a calculator for what a small saving is worth.

It is written for mining business owners, plant managers and consultants who are weighing a proposal, not for engineers who already run model-based control. We do not rate or endorse any supplier. Where a figure comes from a company, we say so.

AI in crushing and grinding: the short answer

Yes, AI works in some circuits, and the gains reported so far are modest, site-specific and mostly written up by the companies that sell the systems. The honest summary from the sources we opened is below.

• Where it helps most: keeping a mill or crusher near its best operating point as ore hardness, feed size and liner wear change, and warning staff before a part fails.
• What the evidence is: a 2024 Queen’s University thesis forecast mill power with R² scores of 0.50 to 0.86, and a Newcrest model predictive control project at Cadia reported results well above its target but published no percentage.
• What it costs you: sensors, clean historical data, integration with older control systems and people who can maintain the models.
• What else to compare: mature advanced process control, laser wear scanning, high-pressure grinding rolls, better blasting and sensor-based sorting often target the same energy.

The sections below show the numbers behind each point. Mining Technology’s feature on AI in comminution quotes both ABB, which sees AI as an important piece of the puzzle, and Rockwell Automation, whose manager says AI is not necessarily the best way to improve efficiency. That split is a fair picture of the field.

What AI infrasolutions means in a crushing plant

“AI infrasolutions” is not a standard engineering term, and this page does not review any company that uses the name. We use it as shorthand for the data and control layer wrapped around the physical plant: sensors on crushers, mills, conveyors and screens, a historian that stores the readings, models that predict what happens next, and a control or advisory screen that tells operators what to change. The crusher, screen and mill are the crushing solutions; the software layer is the AI part.

In practice that layer has four parts you can price and test separately.

• Sensing: motor power, feed rate, mill speed, water flow, vibration, liner thickness and sometimes camera images of the feed.
• Data storage: a plant historian with months or years of readings in one place.
• Models: statistical or machine-learning models that forecast power draw, particle size or wear.
• Action: either advice shown to an operator, or set-points sent back to the control system within limits.

A 2025 scoping review in Frontiers in Energy Research covered more than 75 publications on machine learning in mining energy systems. It lists predictive maintenance, demand forecasting and real-time process control as the main uses, which matches the four parts above. Our own video explains the three jobs an AI mining tool can do, from finding places to look to running a mine.

▶ Plain explanation of the three jobs an AI mining tool can do and who builds one. · 3:52

How much energy do crushing and grinding use?

Studies disagree, because they count different things. Within a mine, comminution is usually the largest single energy user, but the share runs from about 25% to 53% depending on the source and the scope.

Chart

Comminution’s share of a mine’s energy, by source

Percent; different studies and scopes, so read as a range

• A report commissioned by the Weir Group, which analysed data from more than 40 published studies, found comminution is typically 25% of mining’s final energy consumption, or about 1% of energy use worldwide.
• The University of Queensland’s Sustainable Minerals Institute says comminution takes approximately 40% of a mining operation’s energy (2021 page) and, for base and precious metals alone, 0.4% of global electricity.
• ABB experts quoted by Mining Technology put it at up to 50% of a mine’s total energy consumption.
• A 2019 Metso article cites the Coalition for Eco-Efficient Comminution (CEEC), an industry body, for 53% of mine energy and 3% of electricity generated worldwide, while the CEEC website says 1.8% of all electric power generated in the world.

For a plant owner the useful reading is a range. If your circuit takes a third to a half of site energy, even a low single-digit percentage saving can pay for a project, which is why the calculator near the end lets you enter your own figures.

Where AI can help: five jobs in a mineral circuit

Most AI proposals for comminution fall into five jobs. They differ a lot in maturity.

Chart

Mine energy by process step

Percent of mine energy, CEEC figures as quoted by Metso (2019)

1. Supervisory control of set-points

A model adjusts mill speed, feed rate and water flow within safe limits. ABB experts told Mining Technology that machine-learning models can analyse real-time sensor data to predict equipment wear, adjust mill loading and change operating parameters on the fly. The best documented example we found uses model predictive control, a mature form of advanced process control, rather than machine learning (see the evidence section).

2. Forecasting and soft sensors

A soft sensor estimates something you cannot easily measure, such as particle size or the power draw 20 minutes ahead, from things you can. The Queen’s University thesis did exactly that for rod and ball mills.

3. Wear and condition monitoring

Liner wear changes how a crusher or mill breaks rock. Laser scanning of liners is established: Maptek’s case study reports a scan took less than 30 minutes and staff did not need to enter the crusher. Metso’s CrusherMapper case at an iron ore mine in South Africa reports thickness readings accurate to ±3 mm and automatic flagging of high-wear zones. Both are supplier documents. Adding machine learning on top, to predict the remaining liner life, is the newer step.

4. Digital twins

A digital twin is a computer model of the circuit. ABB says AI-powered twins let a mine test different operating scenarios before changing the real plant.

5. Energy and load forecasting

Models forecast hourly power demand so that big loads can be shifted away from expensive periods. This is the least specific to crushing, and the Frontiers review notes that mining sites often have sensor gaps, weak connections and legacy control systems that slow it down.

One sign of where this is heading: SBM Mineral Processing’s Remax 600 mobile impact crusher uses AI and sensors to adapt itself, and a UK contractor bought the first one for aggregate and recycling work. The maker quotes up to 600 tonnes per hour at a maximum feed size of 900 mm. That is a quarry and recycling machine and the capacity is the maker’s claim, so do not read it as proof for a hard-rock mine.

▶ Follows one gold nugget from rock to ring, including the crushing, dissolving and refining steps. · 5:27

What the evidence shows so far

Three pieces of evidence are worth knowing, because they show both the promise and the limits.

A university forecast of mill power

A 2024 Queen’s University (Canada) master’s thesis built models to forecast power use in a rod mill and a ball mill 20 minutes ahead, using historical data from one plant. For the rod mill, the R² scores (1.0 would be a perfect fit) were 0.82, 0.73, 0.84 and 0.86 for XGBoost, random forest, support vector regression and deep neural network models. For the whole circuit, the same models scored 0.74, 0.73, 0.70 and 0.50. The author then built a reinforcement-learning recommendation engine that suggested changes to rod-mill water flow, with a projected power reduction of up to 2%, equal to 299.24 megawatt hours a year. That is a projection from historical data, not a measured result from running the plant.

A model predictive control project at Cadia

Newcrest, the Australian gold miner that Newmont acquired in November 2023, studied 12 months of history from the CON2 grinding circuit at Cadia in New South Wales, then ran a Rockwell Automation model predictive controller that adjusted SAG mill speed, ore feed and water flow together. Rockwell’s case study says the application was designed, implemented and validated in nine months, and that the circuit is delivering well above its throughput target. The study publishes the target and the result as words, not percentages. It is a supplier-written case study, and model predictive control is an older technique than machine learning.

Research on wear-aware control

The University of Queensland’s JKMRC notes that grinding circuits take nearly 50% of a mine’s energy, yet most control systems assume the plant stays the same while liners wear. A 2026 student project there is testing deep reinforcement learning that adapts set-points as liners wear in a simulated SAG, ball mill and crusher circuit. It is promising research, not a plant result.

The chart at the end of this section shows how well the four models forecast power in the rod mill compared with the whole circuit. Forecasts get worse as the system gets bigger and more connected.

Chart

Power forecast accuracy, four models

R-squared multiplied by 100; 100 is a perfect fit; one plant, 20 minutes ahead

Pros of AI in crushing and grinding

• Steadier operation. Mining Technology notes that these processes have traditionally relied on human expertise and manual operation, which carries a risk of human error.
• Possible energy savings. The Queen’s thesis projected up to 2% less rod-mill power from one change; the calculator shows how a 2% saving scales with tonnage and power price.
• Throughput gains. The Cadia project reported production above target, and the Cadia process-control superintendent, quoted by Rockwell, says the control solution stabilised the grinding circuit and lifted production tonnes.
• Fewer surprises from wear. Wear data lets a plant plan relines instead of reacting to failures. Metso’s case describes forecasting to reduce unnecessary liner changeout shutdowns.
• Safer inspection. Maptek’s scanning approach keeps people out of the crusher cavity.
• Testing without risk. A digital twin lets staff try a new strategy on a computer first.
• Handles harder ore. ABB points out that declining ore grades mean larger throughputs, and harder ores need more energy to break. Control that adapts to variation matters more as ore gets harder.

Cons and risks of AI in crushing and grinding

• Data is the bottleneck. The Frontiers review lists sensor sparsity, unreliable connectivity in off-grid locations and reliance on legacy control systems as barriers specific to mining. It also names high implementation cost and limited interpretability.
• Forecast accuracy drops at circuit level. In the Queen’s study the weakest circuit-wide R² was 0.50, against 0.86 for the best rod-mill model.
• Models age. JKMRC points out that liner wear changes power draw, throughput and efficiency over time, so a model trained on new liners drifts.
• Most reported results come from suppliers. The Cadia study and the scanning cases are written by the companies that sold the system. Ask for site-named figures with a baseline and a time period.
• AI is not always the first fix. Rockwell’s manager told Mining Technology that the best approach is to understand the miner’s problem first and then pick the technology; AI is in the mix but the newest tool may not solve the problem.
• Integration and security. The Frontiers review flags integration with SCADA platforms, outdated software, incompatible data formats and data privacy as recurring problems.
• Skills. Someone has to maintain the models. Plan training and a named owner on site.

AI or other upgrades? Six options compared

The same energy bill can be attacked in several ways. This table lists what each option targets, what the sources we opened report, and who it suits. The ranges are not like-for-like: they come from different plants, scopes and years.

The savings chart at the end of this section shows the upper end of each reported range side by side. The HPGR figures come from a Weir article in Global Mining Review, and Weir makes HPGRs, so treat them as a supplier’s summary of University of British Columbia research. The CEEC roadmap was written by 37 industry experts and describes reductions of 15% to 30% in the short term, and 50% within 10 years, from a full energy strategy, not from AI alone. The HPGR article says rolls reduced power at Huckleberry gold mine by 21% at a grind of 160 microns and 34% at 75 microns.

The lesson is that software tuning often delivers single digits, while changing the equipment or the energy strategy can deliver tens of percent, at a much larger capital cost. For a related look at equipment choices, see our guide to industrial gold mining equipment, and for the cost side see reducing costs and improving efficiency in copper mining.

Chart

Upper end of reported comminution power savings

Percent; ‘up to’ values; different plants and bases

Solution mining pros and cons: when there is nothing to crush

Solution mining, also called in-situ leaching or in-situ recovery, skips crushing altogether. Instead of removing rock, a fluid is pumped into the orebody to dissolve the metal and brought to the surface for recovery. The World Nuclear Association says this leaves little surface disturbance and no tailings or waste rock. It is mainly used for uranium: in 2022 in-situ leaching produced 27,773 tonnes of uranium, 56% of the world total, up from 16% in 2000. The Association adds that it can also be applied to copper and gold.

The pros and cons are different from those of a crushing plant.

• Pros: no crushing or grinding energy, little surface disturbance, no tailings or waste rock, and wells are sealed or capped at closure.
• Cons: it only suits deposits in permeable sand or sandstone that sit below the water table and are confined above and below by impermeable layers; fluids must be kept inside the ore zone; and groundwater must be restored afterwards. In the United States the law requires the aquifer to be restored for its pre-mining use, which often needs reverse osmosis and leaves a brine stream to dispose of.

AI is not the main question for solution mining; geology and groundwater are. The sources we opened give no measured results for AI in leaching wellfields, so we do not claim any. If your ore must be crushed, the rest of this page applies. If your deposit is a sandstone-hosted uranium type, solution mining may avoid the crusher entirely.

How to pilot AI on a crushing circuit in six steps

These steps follow what the sources describe: start from the problem, use history, test in advice mode and keep operators in charge. They are a practical outline, not a vendor method.

1. Name the goal. Pick one: throughput, kilowatt hours per tonne, or liner life. Rockwell’s manager asks the same question: is it throughput, yield or energy savings?
2. Audit the data. Check sensor coverage, gaps and clock errors. The Cadia team started with 12 months of history.
3. Set a baseline. Record current kilowatt hours per tonne and tonnes per hour over a normal period, including ore type.
4. Run in advice mode. Let the model suggest changes for operators to accept or reject, as the Queen’s recommendation engine does for rod-mill water.
5. Close the loop with limits. Allow set-point changes only inside agreed safe ranges, with an operator override.
6. Review and decide. Compare against the baseline after a full ore-type cycle. Cadia’s project took nine months from design to validation, so budget time before judging.

What is a small energy saving worth?

A saving of 1% to 2% sounds small. On a plant that grinds millions of tonnes a year it is not. Use the calculator below with your own numbers; all inputs are examples, and the power price must be today’s price at your site. The Queen’s thesis projected 2% for one rod mill, and Rockwell’s Newcrest case study says a one percent improvement can translate into substantial cost savings, so we use 2% only as a starting example.

Remember what the calculator leaves out: throughput gains, liner savings, licence fees for the software and the cost of extra sensors. Add those to the project cost if you know them.

What to ask a vendor before you sign

Because most evidence comes from suppliers, use questions that force specifics.

• Which operating mine ran this system, for how long, and where can I speak to the metallurgist?
• What was the baseline kilowatt hours per tonne and tonnes per hour, and how was ore variation handled?
• Is the model advising or controlling, and what happens when a sensor fails?
• How does the model cope with liner wear and a change of ore type?
• Who owns the data and the trained model if we stop paying?
• How does it connect to our control system and cyber-security rules?
• What is the total cost, including sensors, integration, training and yearly licence fees?

Treat any round efficiency number that has no named study behind it as unverified, and ask what it measured.

Where satellite data fits before the crusher

Ore variability, the thing that makes crushers and mills hard to control, starts in the ground. Our satellite-based mineral detection work screens a licence area from space to flag target zones, alteration and structures before field crews go in, and delivers in 5 to 20 business days depending on size and mineral. It does not model a crushing circuit. If you need a screen of your own ground, you can request a quote. To see how operators cut emissions across a site, read our piece on mining decarbonisation.

Frequently asked questions

What is AI in crushing and grinding?

It means using sensors and machine-learning or model-based control to forecast and adjust crushers and mills, since comminution is estimated at about 25% to 53% of a mine’s energy depending on the source. In practice it covers forecasting, supervisory control, wear monitoring and digital twins.

How much energy can AI save in grinding?

The only measured-style figure we found in an academic source is a projected reduction of up to 2% for a rod mill, from a 2024 Queen’s University thesis. Larger figures come from equipment changes and energy programmes, such as 15% to 30% in the CEEC roadmap, or from vendor material.

What is comminution?

Comminution is the process of reducing solid ore to smaller sizes by crushing, grinding and other methods, according to the Coalition for Eco-Efficient Comminution. It is the step before the minerals are separated.

Is AI better than advanced process control?

Not necessarily. Rockwell’s manager told Mining Technology that the right technology depends on the problem, and ABB says AI is not a stand-alone solution and works best alongside advanced automation and analytics. The Cadia project used model predictive control, not machine learning.

What data do I need to start?

You need reliable sensor readings for power, feed rate, mill speed, water flow and throughput over a long period. The Cadia team studied 12 months of history before designing the controller.

Will AI replace operators in a crushing plant?

The cases we found keep operators in the loop, either accepting advice or supervising limits. The Frontiers review highlights explainable AI for decision traceability, which points to people staying responsible.

What does AI infrasolutions mean?

It is not a standard engineering term. On this page it means the sensors, data storage, models and control screens placed around crushers, screens and mills.

What are the pros and cons of solution mining?

Pros: no crushing, little surface disturbance and no tailings or waste rock. Cons: it needs permeable, confined ore below the water table, and the aquifer must be restored afterwards.

Does AI work on small crushers?

Some machines now ship with AI built in, such as SBM’s Remax 600 impact crusher, which a UK contractor bought for aggregate work. Performance numbers are the maker’s, so ask for a site trial.

How long does a pilot take?

The Cadia grinding project took nine months from design to validation, and the Lihir flotation project in the same Rockwell case study took 12 months. Allow time for data checks before that.

Can satellite data help a crushing plant?

Not directly. It helps earlier, by mapping where ore types and alteration zones are on a licence, which is one source of the feed variation a plant has to handle.