Uncertainty Quantification Fundamentals

Technical provenance

Applicable standards / specifications

• ASME V&V 10 (2019) — Standard for Verification and Validation in Computational Solid Mechanics
• ASME VVUQ 10.2 (2021) — The Role of Uncertainty Quantification in Verification and Validation of Computational Solid Mechanics Models

References

• ASME Verification, Validation and Uncertainty Quantification (VVUQ) — Authoritative standards family for computational-model verification and validation.
• Bathe, K.-J. — Finite Element Procedures — Background reference for finite-element formulation, discretisation, solution and verification.

What Is It?

Uncertainty quantification (UQ) is the systematic representation and propagation of uncertainty through engineering models to produce a probabilistic or bounded characterisation of the result. Instead of producing a single deterministic answer, UQ produces a distribution, a range or a set of bounds that reflects the combined effect of all recognised uncertainties in the inputs. It distinguishes between aleatory uncertainty — inherent variability that cannot be reduced — and epistemic uncertainty — incomplete knowledge that could be reduced with more information. Understanding this distinction is essential for choosing the right UQ method and for interpreting the results correctly.

Why It Matters

A deterministic analysis produces a single number — a stress, a displacement, a margin. This number implies a precision that the inputs do not possess. Material properties vary from batch to batch; loads vary from event to event; geometric dimensions vary within tolerance; boundary conditions are approximations. A single deterministic result does not capture the range of possible outcomes. UQ provides the range — and, with probabilistic methods, the likelihood of different outcomes within that range. This allows the engineer to assess not just whether the nominal case is acceptable, but whether the worst credible case is acceptable, and how sensitive the margin is to the recognised uncertainties.

A deterministic result implies a precision the inputs do not possess. UQ characterises the range of possible outcomes and the likelihood within that range — enabling risk-informed decisions, not just nominal-case decisions.

Aleatory vs Epistemic Uncertainty

The distinction between aleatory and epistemic uncertainty is fundamental to UQ. Aleatory uncertainty is the inherent randomness or variability in a quantity — the scatter in material yield strength from batch to batch, the variation in wind load from gust to gust, the variability in manufacturing dimensions within tolerance. It cannot be reduced by more measurement; more data characterises the distribution better but does not remove the variability. Epistemic uncertainty is incomplete knowledge — the lack of a material test, an unmeasured boundary stiffness, a model form whose accuracy is unknown. It can be reduced by more information — a test, a measurement, a validation study. The distinction matters because the two types of uncertainty are treated differently in UQ.

Representing Uncertainty in Inputs

The first step in UQ is representing the uncertainty in each input. For aleatory uncertainty, the input is represented as a probability distribution — normal, lognormal, uniform, Weibull — chosen based on the physical nature of the variability and any available data. For epistemic uncertainty, the input may be represented as an interval (a lower and upper bound) if no distributional information is available, or as a distribution if some knowledge of the likely shape exists. The choice of representation should be guided by what is actually known — not by assuming a distribution when only bounds are available, and not by using bounds when distributional data exists.

Input uncertainty representations:

Aleatory:    x  ~  N(μ, σ²)     (normal distribution)
             x  ~  U(a, b)       (uniform distribution)
             x  ~  f(x)           (any PDF)

Epistemic:   x  ∈  [x_low, x_high]   (interval)
             x  ~  f(x)  with  f imprecisely known   (imprecise probability)

Propagation Methods

Once the input uncertainties are represented, they must be propagated through the model to characterise the uncertainty in the result. Several methods exist, each with strengths and limitations. The choice depends on the number of uncertain inputs, the cost of each model run, the smoothness of the model response and the type of result required (a full distribution, just bounds, or specific percentiles).

Monte Carlo Sampling

Monte Carlo is the most general and intuitive UQ method. Each uncertain input is sampled from its distribution, the model is run with that set of inputs, and the result is recorded. This is repeated many times — typically hundreds to thousands — building up a sample of the result distribution. The distribution of results reflects the combined effect of all input uncertainties. Monte Carlo does not require the model to be smooth, linear or simple — it works for any model that can be run. Its limitation is convergence: the error in the estimated statistics decreases only as the square root of the number of samples. Halving the error requires quadrupling the runs. For expensive models, this can be prohibitive.

Monte Carlo convergence:

Error  ∝  1 / √N

where:
N  =  number of samples

To halve the error:  N → 4N
To reduce error by factor of 10:  N → 100N

For percentile estimates (e.g. 99% value), more samples
are needed than for mean or standard deviation.

CONSIDERATION: Monte Carlo convergence is slow for tail statistics. Estimating a 99th percentile reliably requires far more samples than estimating a mean. Use Latin hypercube or response-surface methods for expensive models.

Characterising the Result

The output of a UQ analysis is not a single number but a characterisation of the result distribution or range. The most common characterisations are the mean and standard deviation, specific percentiles (e.g. 95th, 99th), the full probability density function, or bounds (for interval analysis). The characterisation should match the engineering decision. If the decision is about the nominal case, the mean may suffice. If the decision is about the worst credible case, a high percentile (95th, 99th) or the upper bound is needed. If the decision involves probability of failure, the full distribution and its tail are needed.

• Mean — the expected or average result; suitable for nominal-case assessment
• Standard deviation — the spread of the result; indicates how much the result varies
• Percentiles — the value below which the result falls with a given probability (e.g. 95th, 99th)
• Probability density function — the full distribution; needed for probability-of-failure assessment
• Bounds — the range of possible results; suitable for epistemic uncertainty and worst-case assessment

Probability of Failure and Reliability

When the result distribution and a limit (allowable stress, yield strength, displacement limit) are known, the probability that the result exceeds the limit can be computed. This is the probability of failure — or, conversely, the reliability is one minus the probability of failure. This is the most quantitative outcome of UQ: it translates input uncertainty into a risk metric. However, the accuracy of the probability of failure depends critically on the accuracy of the input distributions and on the tail of the result distribution — the region near the limit. Small errors in the tail can produce large errors in the probability of failure. UQ-based reliability estimates should be treated as indicative, not precise, especially for very low probabilities.

Probability of failure:

P_fail  =  P( R  >  R_limit )
         =  ∫_{R_limit}^{∞}  f_R(r)  dr

where:
R        =  result (random variable)
R_limit  =  limiting value (allowable, yield, displacement limit)
f_R(r)   =  probability density function of the result

Reliability  =  1  −  P_fail

Separating Aleatory and Epistemic

In a rigorous UQ treatment, aleatory and epistemic uncertainties are separated and propagated differently. Aleatory uncertainty, being irreducible variability, is represented by probability distributions and propagated to produce a distribution of results. Epistemic uncertainty, being lack of knowledge, is represented by intervals or imprecise probabilities and propagated to produce bounds on the result distribution. The combined result is a family of distributions — each member of the family representing a possible "true" distribution consistent with current knowledge. This separation prevents the common error of treating lack of knowledge as if it were variability, which can produce an artificially precise — and misleadingly narrow — result distribution.

Separating aleatory and epistemic uncertainty:

  Aleatory (variability)     →  Result distribution
  Epistemic (ignorance)      →  Bounds on that distribution

  Combined result:
  ┌─────────────────────────┐
  │    ╱╲                   │
  │   ╱  ╲   ← distribution │  ← family of distributions
  │  ╱    ╲    shape varies  │     bounded by epistemic
  │ ╱      ╲   with epistemic│     uncertainty
  │╱        ╲  uncertainty   │
  └─────────────────────────┘
   ↑ lower bound    ↑ upper bound
   (epistemic)       (epistemic)

When UQ Is and Is Not Needed

Not every analysis requires formal UQ. For a well-characterised structure with generous margins, a deterministic analysis with conservative inputs may be sufficient — the margin absorbs the uncertainty. UQ adds value when margins are tight, when the controlling uncertainty is not well-characterised, when the cost of over-design is significant or when a probabilistic assessment is required by a code or specification. UQ is also valuable when the dominant uncertainty is epistemic — because it identifies which uncertainties, if reduced, would most improve the confidence in the result. A sensitivity analysis (the precursor to UQ) identifies the controlling inputs; UQ characterises the result given the uncertainty in those inputs.

MISTAKE: Treating epistemic uncertainty — lack of knowledge — as if it were aleatory variability by assigning a distribution without basis. This produces a result distribution that appears precise but is actually founded on an assumed shape for an unknown quantity.

Key Takeaways

• UQ characterises the range and likelihood of results given input uncertainty — not a single deterministic number
• Aleatory uncertainty is irreducible variability; epistemic uncertainty is reducible ignorance
• The two types are represented and propagated differently — do not conflate them
• Monte Carlo is general but slow to converge; response surfaces enable UQ for expensive models
• Probability of failure translates input uncertainty into a risk metric — but tail accuracy is critical
• UQ is most valuable when margins are tight, uncertainty is decision-relevant or probabilistic assessment is required