Two bridges cross the same river, and both design reports give a calculated scour depth of 2.4 metres. On the first, that figure comes from a narrow band of plausible values and the foundation at 3 metres is never approached. On the second, the same 2.4 is the average of a range running from 1.1 to 4.6 metres, and roughly one flood season in twenty puts the scour below the foundation level.
The two reports are identical in the number everyone reads first, and they describe completely different structures. That gap is the entire argument for probabilistic assessment: the decision does not depend on the expected value but on how often the tolerance is exceeded and by how much.
Monte Carlo simulation is the standard method for producing that answer, and it is conceptually simple enough to explain in a paragraph — which is precisely why its results are so often misread.
What the method actually does
Take a model that converts inputs into an output: soil stiffness, layer thickness and load into settlement, for instance. In a deterministic calculation each input receives one value and the model returns one answer.
In a Monte Carlo run, each input receives a probability distribution instead. Soil stiffness is not 30 MPa, it is a lognormal distribution with a mean of 30 and a coefficient of variation of 0.25. The model is then run thousands of times, each run drawing a random value from each input distribution, and the outputs are collected.
The result is not an answer but a histogram: the frequency with which each outcome occurred across all runs. From that histogram come the numbers that actually matter — the probability of exceeding a limit state, the value at the 95th percentile, the shape of the tail.
Why the average is the least useful output
The two bridges above make the point, and the same pattern appears in every quantity an engineer estimates. A settlement calculation returning a mean of 45 millimetres against a tolerance of 70 is reassuring if the distribution is narrow and meaningless if it carries a long right tail in which three percent of runs cross the limit.
The mean says nothing about that difference. The 95th percentile and the exceedance probability say everything. Reports that lead with the mean and bury the percentiles are answering a question nobody asked.
This is also why sensitivity to the input distributions matters more than sensitivity to the input means. Widening the coefficient of variation on a single dominant parameter can double the exceedance probability while leaving the mean untouched.
How many runs are enough
The most common practical question, and it has a usable answer.
Convergence is not uniform across the output. A mean stabilises quickly — often within a few thousand runs. A 95th percentile takes considerably longer. A 99.9th percentile, or a probability of failure in the range of one in ten thousand, requires either enormous sample sizes or a variance reduction technique such as importance sampling.
The working rule: the further into the tail the answer lies, the more runs are needed to pin it down, and the relationship is not linear. Estimating a probability of 0.01 to within a useful margin needs on the order of tens of thousands of runs; estimating 0.0001 by brute force needs on the order of millions.
The practical test is convergence itself. Plot the estimate against the number of runs. If the line is still moving at the end of the simulation, the simulation is too short — regardless of how many runs it contained.
The inputs decide everything
A Monte Carlo run inherits the quality of its input distributions, and this is where most results go wrong.
Distributions chosen for convenience rather than evidence are the first problem. A normal distribution assigned to a strictly positive quantity — stiffness, thickness, strength — will generate negative values in the tail, and those runs either crash or, worse, quietly return nonsense. Lognormal is usually the correct default for such parameters.
Correlation is the second. Soil parameters are rarely independent: density and stiffness move together, and sampling them independently produces combinations that do not occur in reality. Ignoring correlation typically widens the output distribution and overstates the tail — which sounds conservative and is simply wrong.
The third is spatial variability. A single distribution for a soil parameter across an entire site treats every point as independently random, when in practice values close together are similar. Random field modelling exists for this reason, and the difference between it and a naive approach can be an order of magnitude in the calculated failure probability.
Reading the output honestly
Four habits separate a useful simulation report from a decorative one.
State the percentiles, not just the mean and standard deviation. Two distributions with identical first two moments can behave completely differently at the limit state.
Show the convergence plot. It is the only evidence that the number of runs was sufficient, and it takes one figure.
Report the assumed distributions and correlations explicitly. A result that cannot be reproduced because its inputs were not documented is a number, not a finding.
And keep the deterministic calculation alongside. If the Monte Carlo mean and the deterministic result diverge substantially, one of them contains an error, and finding out which is cheaper before construction than after.
The same structure outside engineering
The reasoning behind Monte Carlo methods is not specific to geotechnics. It applies wherever a single trial is unpredictable and the aggregate is not, and one of the clearest non-technical examples is the family of instant games built on a rising multiplier.
The mechanic is a single draw. A round starts, a multiplier climbs from 1.00, and it stops at a value taken from a fixed distribution decided before the round begins. The player's only decision is when to exit. One round is a single sample; a thousand rounds are a histogram.
What makes the parallel exact is the vocabulary. These games publish a return-to-player figure, which is nothing other than an expected value expressed as a percentage — the same mean that, on its own, tells an engineer almost nothing. Alongside it sits volatility, which is the spread of the same distribution, and it is what actually determines the experience: two games with an identical RTP of 97% behave completely differently if one pays small multipliers constantly and the other concentrates its return in rare high ones. That is the bridge and the tail, in a different setting.
The practice modes follow from this directly. A distribution is not observable in ten rounds, and it is reasonably visible in five hundred, which is why free versions exist and why they are worth running before anything else — Spanish-language guides to chicken road juego demo describe the mechanic on exactly that basis, as a distribution to be observed in aggregate rather than a sequence to be predicted.
The independence assumption is where both fields lose people in the same way. In a correctly implemented generator each round is drawn independently, so a run of low results carries no information about the next one — exactly as a sequence of favourable soil samples says nothing about the next borehole unless a spatial correlation structure has been established and documented. Reading a pattern into independent draws is the same error in a laboratory and at a screen; the engineering version simply costs more.
The transferable point runs against the usual intuition. A result that looks surprising in one trial is not evidence of anything. A result that stays stable across ten thousand trials is. Confusing the two is the origin of most misread simulations, in either field.
Where the method is worth using
Monte Carlo is not the right tool everywhere, and applying it to problems that do not need it is a common way to waste a week.
It earns its place where the output is sensitive to input uncertainty and the consequence of exceedance is significant: slope stability with variable soil properties, settlement over heterogeneous ground, scour depth under uncertain hydraulic loading, remaining life of a structure under stochastic traffic loading.
It is unnecessary where a parameter dominates and its range is narrow, or where the design is governed by a code check with prescribed partial factors that already embed the statistical treatment. In those cases a sensitivity study on two or three parameters gives the same insight in an afternoon.
In summary
A Monte Carlo simulation converts uncertainty in the inputs into a distribution of outputs, and the value of the exercise lies in that distribution rather than in any single number drawn from it.
Choose input distributions from evidence, respect correlation and spatial structure, run until the quantity of interest converges rather than until a round number of runs is reached, and report percentiles alongside the mean. Done that way, the method answers the question a design decision actually poses: not what will happen, but how often the tolerance will be exceeded, and by how much.
