Production targets and the safety case define what success looks like.

Plant data management brings in the historian record, cleans it, reconciles the mass balance and adds the operating regimes recorded by operators. It creates two datasets: one for training and one of hold-out campaigns reserved for a fair test.

Controller learning chooses a model, learns from the operating record, transfers useful knowledge from similar circuits and tunes its hyperparameters. The result is a trained controller.

Verification tests that controller in the digital twin against the hold-out campaigns. It checks performance against the safety case and interlocks, and confirms that the flowsheet constraints are respected.

If the results fall short, the performance gap report sends the work back to data preparation or learning. Once it passes, the controller is ready for deployment.

Cameras, instruments and acoustic sensors watch the circuit.

Analysis combines ML soft sensors and computer vision with mass-balance calculations and familiar KPIs to build a picture of the current state.

Planning uses neural APC and reinforcement learning alongside established MPC and PID strategies to choose a response.

Execution applies approved changes to setpoints, feeders or the rock breaker. The plant responds, and the cycle starts again.