2. Uncertainty taxonomy in structural engineering

Structural reliability analysis begins with a deceptively simple question: what exactly is uncertain? Before any limit-state function can be written or any probability of failure computed, the engineer must classify the sources of uncertainty that enter the model. This classification is not academic bookkeeping — it determines which variables can be treated as random, which must be bounded, which require Bayesian updating, and which can only be reduced by collecting more data. The standard taxonomy separates aleatory uncertainty (inherent randomness that cannot be reduced by further study) from epistemic uncertainty (lack of knowledge that can, in principle, be reduced). A third category — model uncertainty — captures the error introduced by the mathematical idealization itself.

Aleatory vs. Epistemic: The Practical Distinction

Aleatory uncertainty arises from natural variability. Concrete compressive strength from a single ready-mix plant varies from truck to truck even under identical mix design; wind gust speeds at a site vary from storm to storm; the yield strength of nominally identical steel bars varies within a heat. These quantities are physically random — repeating the "experiment" produces different outcomes, and no amount of additional measurement eliminates the scatter. They are modeled as random variables with a probability density function (PDF), typically characterized by a mean, coefficient of variation (CoV), and distribution type.

Epistemic uncertainty, by contrast, stems from incomplete knowledge. The mean concrete strength of a specific structure may be unknown because only three cores were tested; the soil profile at a bridge pier may be inferred from a single borehole; the future traffic load model may be based on outdated weigh-in-motion data. These uncertainties are reducible: more cores, more boreholes, or updated traffic surveys shrink the uncertainty. In a Bayesian framework, epistemic uncertainty is represented as a distribution over distribution parameters (a hierarchical or "distribution of distributions" model), and it narrows as data accumulate.

Random Variables: Formal Representation

A random variable X is defined by its cumulative distribution function (CDF) F_X(x) = P(X ≤ x). In structural reliability, the most common distributions are:

• Normal (Gaussian): used for geometric dimensions, elastic modulus, and self-weight when the CoV is small (< 0.15).
• Lognormal: used for material strengths (concrete, steel) and resistances, since these are strictly positive and right-skewed.
• Gumbel (Type I extreme value): used for maximum wind speeds, snow loads, and flood levels — the maximum of many independent events.
• Weibull: used for fatigue life and brittle material strength.

For a lognormal variable with median m and CoV V, the parameters are λ = ln(m) and ζ = √(ln(1+V²)). For example, concrete cylinder strength with m = 30 MPa and V = 0.15 gives ζ = 0.149 and λ = 3.401. The 5th percentile — often used as the characteristic value f_ck — is exp(λ − 1.645ζ) = 23.4 MPa, which is consistent with the Eurocode relationship f_ck = f_cm(1 − 1.645V) for small V.

Stochastic Fields: When Location Matters

Many structural properties vary not only in magnitude but in space. The concrete strength in a 30 m span beam is not a single random variable — it is a random field X(t, θ) where t is the spatial coordinate and θ indexes the random realization. The key descriptor is the autocorrelation function ρ(τ), which measures how strongly values at two points separated by distance τ are correlated. A common model is the exponential (Markov) form:

ρ(τ) = exp(−|τ| / ℓ)

where ℓ is the correlation length. For concrete, ℓ is typically 1–3 m; for wind pressure on a tall building, ℓ may be 10–50 m; for soil properties, ℓ can range from 0.5 m (friction angle) to tens of meters (undrained shear strength). The correlation length controls how much spatial averaging occurs: a large ℓ means the whole member behaves almost as one random variable, while a small ℓ means local weak zones can trigger failure before the global average is reached.

In finite-element reliability analysis, a random field is discretized into a finite set of correlated random variables using methods such as the Karhunen–Loève expansion or the midpoint method. The number of terms retained depends on ℓ relative to the domain size: a domain of length L with correlation length ℓ requires roughly L/ℓ terms to capture the dominant variability.

Model Uncertainty

Even with perfect knowledge of all input variables, the model itself may be wrong. Model uncertainty is represented by a multiplicative or additive correction factor θ_model applied to the resistance or load prediction. For example, the ACI 318 flexural strength model has a bias factor (mean predicted/actual) of approximately 1.05 and a CoV of 0.10–0.15. This factor is treated as a random variable in the reliability calculation, and it is epistemic in nature — a better model would reduce it.

Worked Example: Classifying Uncertainties for a Reinforced Concrete Beam

Consider a simply supported RC beam with span 6 m, cross-section 300 × 600 mm, and 3Ø25 tension reinforcement. The limit-state function for flexure is g = M_R − M_S, where M_R is the resistance and M_S is the applied moment from dead and live loads. The uncertainty classification is:

Note that the live load is modeled as Gumbel because it represents the maximum occupancy over the design life, while the dead load is Normal because it is a sum of many small contributions. The model factor is epistemic — a refined finite-element model or a full-scale test would reduce its CoV.

From Taxonomy to Computation

The classification directly affects the reliability method chosen. If all uncertainties are aleatory and the limit-state function is smooth, first-order reliability methods (FORM) or second-order reliability methods (SORM) are efficient. If epistemic uncertainty dominates, a Bayesian approach with posterior predictive distributions is more appropriate. If random fields are present, the discretization scheme and the number of random variables determine whether Monte Carlo simulation, importance sampling, or subset simulation is feasible.

The following diagram summarizes the taxonomy and its mapping to computational treatment:

The diagram shows how each uncertainty class maps to a specific computational strategy. Aleatory variables enter directly as random variables in the limit-state function. Epistemic variables require Bayesian treatment or sensitivity studies. Model uncertainty is applied as a correction factor. Stochastic fields require discretization into correlated random variables before any reliability algorithm can be applied.

In practice, the boundary between aleatory and epistemic is not always sharp, and the classification may shift as knowledge improves. A variable initially treated as aleatory (e.g., concrete strength) may become partially epistemic when the engineer realizes that the mean value for a specific structure is unknown. The key is to document the classification explicitly, justify it with available data, and propagate the resulting uncertainty through the reliability calculation. This discipline prevents the common error of treating epistemic uncertainty as aleatory — which overestimates the irreducible randomness — or treating aleatory uncertainty as epistemic — which falsely suggests that more testing can eliminate natural variability.