Comminution — the process of crushing and grinding ore down to a size where you can extract the valuable minerals — is the single biggest energy consumer at most mining operations. It typically accounts for 30-50% of total site energy use, and at large operations, that translates to tens of millions of dollars annually. Even small percentage improvements in crusher and grinding circuit efficiency can shift the economics of a project significantly.

That’s why AI-driven optimisation of crusher circuits has attracted so much attention over the past couple of years. And unlike some of the more speculative AI applications in mining, this one is producing measurable results at operating mines right now.

The Problem with Traditional Control

Most crusher circuits are controlled by a combination of PLC logic and operator experience. The PLC handles basic parameters — closed-side setting, feed rate, motor current limits — while operators make adjustments based on what they’re seeing and hearing. An experienced crusher operator can do a surprisingly good job. They’ll notice when the feed material gets harder, when there’s a surge in fines, or when the crusher is starting to labour under too much load.

The issue is consistency. Operators work shifts. They have good days and bad days. They can’t process the 50+ sensor readings from across the circuit simultaneously and continuously. And they tend to run conservative settings to avoid blockages and overloads, which means the circuit rarely operates at its true optimum.

AI optimisation systems take in all of those sensor feeds — vibration, power draw, feed rate, particle size distribution from image analysis, oil temperatures, bearing conditions — and make adjustments in real time. They’re constantly finding the sweet spot between maximum throughput and equipment protection, something that’s essentially impossible for a human operator to do consistently across a 12-hour shift.

What the Numbers Show

A copper operation in South Australia implemented an AI optimisation system on their primary and secondary crushing circuit in mid-2025. After a three-month tuning period, they reported a 7% increase in average hourly throughput with a simultaneous 4% reduction in specific energy consumption (kWh per tonne crushed). That doesn’t sound dramatic until you do the maths: at 8 million tonnes per year, a 7% throughput increase represents an additional 560,000 tonnes through the circuit annually.

A Team400 project with a Queensland gold operation focused on optimising the SAG mill and secondary crusher interaction. The challenge there was managing the recirculating load — when too much oversize material cycles back through the crusher, it creates a bottleneck that cascades through the entire processing plant. The AI system learned to anticipate oversize surges based on upstream ore characteristics and preemptively adjust the crusher’s closed-side setting.

BHP’s Olympic Dam operation has been particularly open about their work in this area. Their published results from a broader processing plant optimisation program, which includes crusher circuit AI, showed a 5-8% improvement in overall plant recovery.

The Data Challenge

Getting these systems to work isn’t just about plugging in an AI model. The biggest challenge I’ve seen across multiple implementations is data quality and sensor reliability. Crushers are brutal environments. Dust, vibration, and impact loading destroy sensors at an alarming rate. If your particle size camera lens is coated in dust or your load cell is drifting because of temperature effects, the AI is making decisions on bad data.

The operations that succeed with AI crusher optimisation invest heavily in sensor maintenance. That means daily cleaning schedules for cameras, regular calibration of load cells, and redundancy on critical measurements. It’s unglamorous work, but it’s the foundation everything else sits on.

You also need a proper data infrastructure. Many older processing plants have sensor data trapped in isolated SCADA systems that don’t talk to each other. Before any AI work can begin, there’s usually a significant integration effort required to get all the data flowing into a single platform where the models can access it.

Edge Cases and Limitations

AI crusher optimisation isn’t a set-and-forget solution. The models need retraining when ore characteristics change significantly — moving into a different geological domain, seasonal moisture variations, or a shift from oxide to sulphide ore. A model trained on dry, competent ore will make poor decisions when suddenly confronted with wet, clay-rich material.

Most current systems handle this through continuous learning, where the model updates itself as conditions change. But there’s usually a lag period where performance dips before the system adapts. Operators need to understand this and be ready to intervene during transition periods.

The other limitation is that crusher optimisation in isolation only captures part of the opportunity. The real gains come from optimising the entire comminution circuit as a system — from the primary crusher through to the grinding mills and classification circuits. An AI consultancy working with a nickel operation in Western Australia found that optimising the crusher alone delivered a 5% throughput gain, but extending the optimisation across the full grinding circuit pushed that to 12%.

Where Things Are Heading

The integration of ore characterisation data from the mine face is the next frontier. If the processing plant AI knows what’s coming in the next few truck loads — hardness, moisture content, grade distribution — it can pre-adjust the entire circuit before the ore even hits the primary crusher. A few operations are piloting this approach using hyperspectral imaging on haul trucks, and early results suggest it could be worth another 3-5% on top of existing optimisation gains.

For operations still running crushers on PLC logic and operator intuition, the business case for AI optimisation is strong and getting stronger. The technology has matured past the pilot phase. The question isn’t whether it works, but whether your data infrastructure is ready to support it.