What AI Changes in Geotechnical Modelling

In 1988, Starfield and Cundall described rock mechanics problems as data-limited. That remains a useful way to think about numerical modelling today.

​

The models have become dramatically more sophisticated. We can represent complex excavation sequences, structures, groundwater, stress redistribution and progressive rock-mass behaviour. Monitoring can now provide continuous evidence of how a slope or excavation is actually responding.

​

However, the underlying condition has not changed nearly as much. The ground is still only partly known.

​

Structures are inferred between observations. Properties are estimated from limited samples. Groundwater is imperfectly constrained. Failure mechanisms may remain uncertain until excavation reveals more of the system.

​

The important question is not only how sophisticated our models can become. It is how we use them while our interpretation of the ground remains provisional.

​

Calibration was an important step forward

​

One of the most useful developments in geotechnical modelling has been the move from purely forward prediction toward calibration against observed behaviour.

​

We build a model.

We make a prediction.

The mine advances.

We compare the prediction with what actually happened.

Parameters are adjusted and the model is rerun.

​

This is clearly better than modelling without reconciliation.

​

Back-analysis can reveal information that drilling and laboratory testing may never provide directly at mine scale.

​

Here there is an important distinction - a model that reproduces observed behaviour has demonstrated that it can reproduce observed behaviour.  It has not necessarily demonstrated why that behaviour occurred.

​

Li made this point clearly in 2022: calibration is not the same as validation or prediction.

 

Predictive capability still needs to be tested against observations that were not used in calibration. This matters particularly in rock engineering because back-analysis is often non-unique.

​

The same displacement history may be reproduced by changing stiffness, strength, pore pressure, stress, structure persistence, boundary conditions, or some combination of them.

​

A different constitutive model may fit.

A different structural interpretation may fit.

Sometimes a different mechanism may fit.

​

An excellent calibration can still leave a difficult question:

Have we identified the mechanism — or simply found one explanation that reproduces the observations?

 

Calibration remains extremely useful. But we should be careful about what we claim it has achieved.

​

Three ways of using a model

​

I find it useful to think about three progressively richer ways in which numerical models are used. They are not strict historical generations. All three exist in current practice. The difference is in the question being asked.

​

Forward modelling

 

What does our current interpretation predict?

 

This remains fundamental.

​

A good forward model can help us understand mechanisms, compare design options and test sensitivities. The result remains conditional on the interpretation we started with.

​

Calibrated modelling

 

Can our model reproduce what the ground has already done?

 

This is a major improvement.

​

Monitoring becomes part of the modelling process. Parameters can be informed by actual mine-scale behaviour.

​

Still calibration introduces a trap - as the fit improves, confidence in the model often improves with it.

​

Yet a better numerical fit does not necessarily mean that the conceptual model has become more correct.

​

Imagine that movement in a slope is actually controlled by an unrecognized persistent structure.

​

We may compensate for that missing structure by reducing rock-mass stiffness, adjusting strength or modifying pore pressure until the calculated deformation resembles what the radar is seeing.

​

The final fit may be excellent. The mechanism may still be wrong.

​

Adaptive / observational modelling

​

The next step is not simply to calibrate the model once.

​

It is to keep confronting it with reality.

​

Mining advances.

New structures appear.

Groundwater changes.

Deformation develops.

The interpretation is updated.

​

The question becomes:

Does our current interpretation remain credible as new evidence arrives?

 

This is where the distinction between calibration and reinterpretation becomes important.

​

If every discrepancy is resolved by adjusting stiffness, strength, pore pressure or another parameter while the underlying conceptual explanation is preserved indefinitely, we may still be doing little more than continuous calibration.

​

Sometimes the right response is not:

     How do we recalibrate this model?

It is:  

    What if this is no longer the best explanation?

​

That is the more important step beyond calibration.

​

Adaptive modelling should allow not only parameters to change, but the interpretation itself.

 

Geometry may change.

Structural continuity may be reconsidered.

Boundary conditions may change.

The assumed mechanism may change.

And where necessary, another conceptual model may replace the one we have been using.

 

The point is not to preserve multiple models. It is to preserve the ability to change the model when the evidence demands it.

​

Alternative models have a place

​

There will be situations where two or more interpretations remain genuinely plausible.

​

Perhaps deformation could be explained by groundwater response or by a persistent structure.

Perhaps two mechanisms reproduce the available monitoring equally well.

​

In those cases, analysing alternative models can be useful — particularly where the difference could materially affect a future decision.

​

The purpose is not to maintain several models indefinitely.

​

It is to understand where they differ, what each would predict next, and what additional evidence could help resolve the uncertainty.

​

That is a selective tool within adaptive modelling, not an objective in itself.

​

This may be where AI fits better

​

Most discussion around AI in numerical modelling focuses naturally on automation.

​

Model setup.

Parameter estimation.

Back-analysis.

Sensitivity studies.

Optimization.

Scenario generation.

​

All of this will be useful. But much of it simply makes existing modelling workflows faster.

 

There is also a risk.

​

If the conceptual model is wrong, a faster optimizer may simply produce a more convincing calibration of the wrong explanation.

​

AI may therefore make calibrated modelling much better without addressing its fundamental limitation.

​

The more interesting role may sit around the model rather than inside the calibration loop.

 

An AI-assisted system could continuously keep track of:

• the current accepted interpretation;

• the assumptions on which it depends;

• observations that support it;

• observations that remain poorly explained;

• successive calibration changes;

• whether parameters are being pushed away from independently supported ranges;

• alternative mechanisms that may deserve reconsideration;

• and when unresolved uncertainty begins to matter to a future decision.

 

The numerical solver still does the mechanics. The engineer still decides what is geologically and geotechnically plausible. AI helps maintain the reasoning around both.

​

AI as a challenger

​

Perhaps the most useful role for AI is not:

 

     Find the parameter set that best matches the observations.

 

It may be:

    Show me what the current model still cannot explain.

Or:

    Are successive calibrations becoming increasingly artificial?

Or:

    Has new evidence appeared that should make us reconsider the conceptual model?

​

That changes AI from an optimizer into something closer to a technical challenger. And that may be a better fit for geotechnical engineering.

​

One of the persistent weaknesses of modelling is not our inability to build coherent explanations. It is our tendency, once we have one, to keep improving it.

​

The model acquires history.

It becomes familiar.

It begins to carry authority.

Eventually it becomes the model.

​

AI may help us resist that tendency.

​

Beyond calibration

​

Numerical modelling has progressed from prediction toward calibration and increasingly toward continuous reconciliation with observed behaviour.

​

AI will accelerate all three. But its more important contribution may be different.

​

It may help us recognize when better calibration is no longer the answer — and when the interpretation itself needs to change.

​

That may be the real step beyond calibration. And it connects directly with a broader question.

​

A mine still has to commit to a slope angle, a stope geometry and a sequence. So the practical question becomes:

​

How do we allow our technical interpretation to evolve while keeping the mine plan stable enough to execute?

​

That is the question for the next essay.