Mining operations are collecting more data, but the useful outcome is not the data itself. It is an interpretable decision that changes what happens on a physical asset or mining system.
The Digital Twins in Mechanised Mining position paper responds to a practical problem: the term “digital twin” is used for everything from a geometric model to a dashboard, simulator or fully connected control system. Without a shared definition, organisations can purchase or develop very different systems under the same label.
A five-part framework
RAMMS describes a digital twin through five connected dimensions.
The physical entity is the asset or system being twinned. It may be a component, machine, production process or mine-wide system. Sensors, communications, edge or cloud computing and cybersecurity sit around this entity, but should be chosen for the decision rather than treated as ends in themselves.
The digital model represents relevant structure and behaviour. It can be geometric, physics-based, data-driven or hybrid. The most detailed model is not automatically the best: fidelity should be sufficient for the decision and compatible with available computing time and measurements.
The physical-to-digital connection updates the model from measured reality. This can involve direct sensor values, probabilistic state estimation, Kalman filtering or machine-learning-based updating.
The digital-to-physical connection carries a recommendation or command back to the real operation. This is where a model becomes operational: an operator receives a warning, a maintenance action is scheduled or a controller changes the machine state.
Finally, optimisation and uncertainty quantification identify the best intervention and show how much confidence should be placed in it. A useful twin should not present a precise recommendation without exposing the uncertainty that surrounds the estimate.
Eight properties—not a maturity ladder
The report also evaluates twins through properties such as fidelity, scalability, interoperability, update rate, interpretability and the ability to support optimisation. These are design choices, not a simple ladder where more is always better.
A real-time high-fidelity model may be essential for braking control, while a slower, lower-fidelity model may be entirely adequate for weekly maintenance planning. The framework therefore begins with the required intervention and works backward to the model, data and infrastructure.
Why hybrid modelling suits mining
Purely data-driven models can struggle when faults are rare, labels are expensive and equipment operates beyond the conditions represented in training data. Purely physical models can require unavailable parameters and may omit difficult real-world behaviour.
Hybrid modelling combines their strengths. Physical equations preserve engineering structure and support generalisation; learned components capture residual behaviour, map accessible measurements to unmeasured states or reduce the computational cost of a high-fidelity simulation.
This approach appears across RAMMS: Bismarck Louw’s engine model maps sparse ECU measurements into physical KPI calculations; Hanno Moes uses reduced-order estimates inside an LHD model; and Luke van Eyk’s traction work updates interpretable terrain parameters from onboard data.
Mining case studies and programme opportunities
The full report evaluates case studies involving roadheaders, hydraulic support, mine ventilation, operations and maintenance, and conveyor motors. Each example is decomposed into the five framework dimensions so that the real digital-twin contribution becomes visible.
The same framework is then applied to RAMMS opportunities in hard-rock cutting and drilling, productivity and maintenance optimisation, and equipment utilisation and condition monitoring. That analysis helped shape work on engines, LHD braking, traction, conveyors, machine vision, core drilling and shaft conveyances.
The actionable test
A detailed model without an operational connection remains a model. A one-way data dashboard remains a digital representation. The defining test is whether synchronised knowledge of the physical system supports an actionable intervention.
That intervention can still involve a person. “Actionable” does not require full autonomy; it requires a clear and defensible path from the model’s estimate to a decision that can improve safety, productivity or asset integrity.