The Geotechnical Digital Twin — A Decision System, Not a Model

The term digital twin is used for quite different things in mining, including geological models, monitoring dashboards and systems that connect observations with numerical analysis. Before assessing their value, I would want to understand what is being updated and how that update affects the engineering work.

Consider a pit slope where measured movement begins to depart from prediction. Displaying the discrepancy is useful, but the engineer still has to establish what caused it. Groundwater conditions may have changed, an exposed structure may alter the interpretation, or the model may be missing a relevant mechanism. Adjusting parameters until the curves agree may improve the fit without resolving that question.

For me, a useful geotechnical digital twin should help the team investigate such discrepancies. It should preserve the assumptions behind the current interpretation, bring relevant observations together, and make it easier to assess whether the design or operating response needs review.

What is being updated?

I find the distinction between a model, a digital shadow and a twin useful, provided we are clear about what we mean. A geological or numerical model records an interpretation of the mine at a particular stage. Engineers can revise it as new information becomes available, but that revision may be a separate, periodic task.

A digital shadow connects the representation to incoming observations. It may show the latest displacement, pore pressure or excavation geometry alongside the model. That gives the team a more current view of conditions, although the assumptions behind the model may remain unchanged.

I would expect a geotechnical digital twin to go further by supporting regular reconciliation between observations and predictions, with a record of the resulting changes to the interpretation. The frequency and extent of updating should suit the engineering problem. A live data feed alone would not meet that expectation.

Working through a discrepancy

In the slope example, the team would need to examine the movement alongside groundwater records, excavation progress and structural mapping. If two explanations remain plausible, the system should retain both long enough to establish what evidence could distinguish them. It should also record why an interpretation was accepted, revised or rejected.

This is where the connection between datasets becomes useful. An engineer reviewing the slope should be able to establish which geometry, groundwater conditions and material assumptions supported the prediction, and whether those conditions still apply. Without that history, a succession of updated models can obscure why the design was considered acceptable in the first place.

The result of the review may be a revised model, additional investigation, a change in monitoring or a design review. It may also be a reasoned decision to retain the existing interpretation. Any prescribed operational response still needs to follow the site’s trigger-action arrangements; investigating the discrepancy should not delay a required action.

The limits of updating

Bayesian updating is one possible approach to combining prior estimates with new observations and revising the uncertainty attached to them. Its usefulness depends on the assumptions in the update process and the quality of the observations. It cannot be relied on to identify a mechanism that the framework does not allow for.

The same limitation applies to automated calibration. If an unrecognised structure controls movement, adjusting rock-mass properties may reproduce the measurements while leaving the underlying explanation wrong. The team needs to be able to question the conceptual model, rather than continually compensate for it through parameter changes.

That is the practical test I would apply to a proposed digital twin: take a discrepancy from the mine and follow it through the system. Can the team trace the original assumptions, compare plausible explanations, identify what remains unresolved and record the basis for its response? That demonstration would tell me more about its engineering value than the sophistication of its visualisation.