Key Takeaways

  • Grinding (SAG/ball mills) is one of the single largest energy consumers on a mine site — comminution overall is roughly half of total plant energy use, so small efficiency gains are worth real money.
  • Closed-loop AI controllers now write mill setpoints directly, rather than just advising an operator, using deep neural networks trained on historical plant data instead of a first-principles model.
  • Vendors report 2-5% throughput gains, lower specific energy, fewer liner strikes/shutdowns, and better downstream recovery from AI-driven grinding control.
  • This is a mature, commercially available category — not an emerging research idea — with multiple named vendors actively selling into mining today.

TL;DR

You do AI-assisted grinding mostly by adopting a closed-loop advanced process control (APC) platform — ABB Ability Expert Optimizer, Imubit’s Closed Loop AI Optimization, or Metso’s control stack built on the MillSense charge sensor — that learns your mill’s behavior from historical data and continuously adjusts feed rate, water addition, and mill speed in real time.

How Do I Optimize Grinding With AI?

The practical path starts with instrumentation, not modeling: you need reliable sensor data on mill charge, throughput, power draw, and particle size before any AI system has something to learn from. Metso’s MillSense system, a wireless sensor bolted to the mill liner, measures the toe and shoulder angle of the charge directly — a much better real-time proxy for what’s happening inside the mill than power draw alone — and is often bundled with acoustic/vibration sensing (SmartEar) for a fuller picture.

Once that data is flowing, the control layer is where AI actually earns its keep. ABB Ability Expert Optimizer applies model predictive control (MPC) across the grinding and flotation circuit, stabilizing feed rate, separator speed, and reject rate against a process model. Imubit takes a different, more explicitly AI-native approach: instead of building a first-principles model, it trains a deep neural network directly on years of historical operating data to predict how the mill will respond to setpoint changes, then closes the loop — writing optimal setpoints to the DCS in real time while keeping the previous control system as a safety fallback. Imubit has 90+ deployments in refining and chemicals and is actively expanding into mining, reporting throughput gains in the low single digits alongside energy and downtime reductions.

For a mine site adopting this, the realistic rollout looks like: instrument the mill (charge/vibration sensors), historize the data for 6-12 months if you don’t already have it, run the AI system in advisory/shadow mode to validate its setpoint recommendations against operator judgment, then progressively hand over closed-loop control with operator override always available. This is not experimental — it’s a standard APC modernization project with an AI-based control core instead of (or alongside) a classical model-based one.

Try It With Geocluster

Deciding whether closed-loop grinding AI is worth the instrumentation investment at your specific mill requires digging through vendor case studies, published throughput/energy results, and your own plant’s data maturity — exactly the kind of multi-source research Geocluster is built to accelerate. Point it at your comminution circuit question and let it pull together the evidence before you write the business case.