Even well-constructed geotechnical models are often trusted beyond the limits of the knowledge they represent.

​

In modern practice, we have partly exchanged the disorder of the field for the orderliness of the screen. We produce factors of safety to three decimal places, stress–strain plots that look authoritative, and simulations that suggest a far greater degree of control than the ground itself is willing to offer. Yet a model is not reality. It is a simplified hypothesis about reality, built from limited evidence and constrained by assumptions that are often much stronger than they first appear.

​

That distinction matters because geotechnical engineering remains, at its core, a data-limited discipline. Unlike structural or mechanical engineering, we do not work with materials manufactured to specification. We work with natural materials: heterogeneous, partly concealed, scale-dependent, and resistant to neat representation. Starfield and Cundall made this point clearly many years ago. The art of modelling is not to make a model as complicated as the world, but to decide which features of geology and mechanism are essential to the decision and which can be neglected without distorting it. Once a model becomes as complicated as the real system, it stops clarifying and starts reproducing confusion in digital form.

​

The difficulty is not modelling itself. The difficulty is what happens around modelling. A numerical result can look clean long before the underlying understanding becomes strong. This is where the process becomes dangerous. In practice, calibration is often mistaken for validation. A model is adjusted until it reproduces observed behaviour, and that agreement is quietly taken as proof that the model is now trustworthy. But calibration only shows that the model can be tuned to fit what has already happened. Validation asks a harder question: whether the model is actually adequate for predictive use beyond the data on which it has been tuned.

​

In a geological system that is only partly observed, that distinction is critical. Different geological interpretations can often produce similar calibrated outputs. Conflicting assumptions about structure continuity, groundwater behaviour, or deformability may all generate acceptable back-analysis fits. That is why inverse analysis and back-analysis are inherently vulnerable to non-uniqueness. A model that matches yesterday’s displacement record is not necessarily true. It is one plausible explanation among several.

​

Rock engineering makes this problem even more acute because the ground is not a tidy engineering material. It is discontinuous, inhomogeneous, anisotropic, and very often not elastic in any meaningful engineering sense. The DIANE shorthand remains useful precisely because it reminds us how far the ground sits from the assumptions that make many models behave so neatly. Numerical methods do not automatically account for scaling effects, hidden structural persistence, progressive rock-bridge degradation, localized pore-pressure build-up, or other processes that only become important when they begin to control behaviour. Those mechanisms have to be identified, interpreted, and represented deliberately. If they are not, the model may still run smoothly, but the understanding remains conditional.

​

This is one of the peculiar difficulties of geotechnical systems. The ground is not simply the medium being analysed. It also participates in the loading, restraint, deformation, and failure of the system itself. In other words, the engineer is not modelling a passive material alone, but a natural system that helps generate its own response. If the main physical drivers are not understood clearly enough, then uncertainty is not really being reduced. It is being absorbed into a computational framework and made less visible.

​

That is why the most important question is rarely, Does the model show a stable slope? The more important question is whether the decision remains robust to what the model does not see, does not represent, or does not yet understand.

​

This is where models are most useful when they are used with discipline. Their role is not to provide absolute answers. Their role is to act as structured test benches for engineering judgement. A good modelling process explores sensitivity, compares mechanisms, examines alternative scenarios, and exposes the drivers that matter most to failure or instability. It expands understanding of uncertainty. It does not compress uncertainty into a single comfortable number.

​

Used in that way, modelling becomes more valuable, not less. It helps reveal which assumptions are decisive, where additional information may actually change the decision, and where confidence is running ahead of the evidence. Used poorly, however, it can do the opposite. It can create the impression that uncertainty has been mastered when it has merely been reorganized.

​

That is why the real issue is not whether models are right or wrong in some absolute sense. The real issue is whether they are being trusted appropriately. In high-consequence decisions, the greatest danger is often not that the model is crude, but that it is believed too easily because it is precise, well presented, and analytically sophisticated.

​

This is also where independent judgement matters most. The value is often not in producing another simulation, but in asking harder questions of the simulation already on the table. What geological assumptions are carrying the result? Which uncertainties have been represented, and which have been left outside the frame? What mechanisms would change the decision if they proved more important than expected? And how much of the apparent confidence comes from the physics of the problem, rather than from the discipline of the software?

​

That, to me, is the proper role of modelling in geotechnical engineering. Not to replace judgement, and not to conceal uncertainty behind precision, but to sharpen the decision by exposing where the real uncertainties still lie.

​

This is also the level at which VSKY.GEO is intended to operate: not by treating models as authority in themselves, but by using them to test assumptions, expose uncertainty, and support decisions that remain defensible under incomplete knowledge.

​

When models are trusted too easily, uncertainty does not disappear from the ground. It only disappears from view.

​

​