Uncertainty Quantification (UQ)

Uncertainty quantification (UQ) is the study of how uncertainty in inputs, models, and data propagates through mathematical and computational systems to affect outputs. It combines probability, statistics, and numerical methods to produce calibrated predictions and quantify confidence in results. UQ is widely used in scientific computing, engineering design, and machine learning, including settings such as Bayesian inference and stochastic modeling.

Overview

In many applications, quantities of interest are computed with models that depend on uncertain parameters, boundary conditions, initial states, or observational noise. UQ provides a framework to describe these uncertainties using probability distributions and then to determine how they influence model outputs. The resulting uncertainty can be represented through predictive intervals, distributions, or risk metrics.

A typical workflow begins by identifying sources of uncertainty, such as measurement error or incomplete knowledge. Uncertainty may be reduced through additional data, but the process of updating beliefs is usually formalized with methods like Bayesian inference. When a system includes random inputs, uncertainty propagation can be analyzed using stochastic processes, with Monte Carlo method as a common baseline approach.

Uncertainty modeling and sources

Uncertainty in a model can be classified in several ways, depending on origin and characterization. Aleatoric uncertainty refers to irreducible randomness inherent in the system (for example, noise in measurements), while epistemic uncertainty reflects lack of knowledge about parameters or structure. In practice, distinguishing these categories guides whether more data will reduce uncertainty and how it should be modeled.

Parametric uncertainty often arises from uncertain model coefficients, which may be estimated with regression or parameter fitting. When calibration data are available, UQ frequently uses likelihood-based approaches and posterior distributions. For models with uncertain inputs, probability distributions can be specified directly or inferred from historical data using tools such as maximum likelihood estimation and Markov chain Monte Carlo in Bayesian settings.

Uncertainty propagation

Uncertainty propagation describes how input uncertainty transforms into output uncertainty. In its simplest form, a quantity of interest (Y) is treated as a function of random inputs (X), and the goal is to characterize the distribution of (Y=f(X)). Direct propagation using Monte Carlo method is widely used because it is broadly applicable, but it may be computationally expensive when (f) represents a costly simulation.

For high-dimensional problems, more efficient techniques are often required. Variance-based sensitivity analysis quantifies which inputs most influence output variability and can be used to guide experimental design. Methods related to [Sobol indices](/wiki/Sobol_sequence and Sobol indices) are commonly used to apportion output variance across factors and interactions. When a model is described by a set of differential equations, discretization error also contributes uncertainty and may be analyzed using numerical methods such as finite element method.

Statistical inference and model calibration

Model calibration combines observations with a modeling framework to estimate uncertain parameters and assess predictive performance. In Bayesian UQ, the posterior distribution over parameters updates prior beliefs using observed data. This approach is closely associated with probabilistic modeling and can incorporate hierarchical structures, model discrepancies, and measurement processes.

When sampling from the posterior is required, Markov chain methods such as Markov chain Monte Carlo may be used, while approximate approaches rely on techniques like variational methods. For computational speed, uncertainty-aware approximations can also be built using surrogate models, allowing repeated evaluation of expensive simulators. Surrogates often integrate with Gaussian process regression to provide both mean predictions and uncertainty estimates.

Surrogate models and computational strategies

Many UQ tasks require numerous evaluations of a forward model, which can be prohibitive for complex simulations. Surrogate modeling replaces the expensive model with an approximation that is cheaper to evaluate while retaining uncertainty information. In Bayesian surrogate frameworks, Gaussian process models provide uncertainty estimates that can be used for active learning and to adaptively refine the surrogate.

Other computational strategies include polynomial approximations and reduced-order models. Polynomial chaos expansion represents the model output as a series of orthogonal polynomials of random inputs, enabling efficient estimation of moments and distributions under specific assumptions. For gradient-free uncertainty analysis and expensive black-box models, alternatives include design of experiments and emulation based on metamodels.

See also

• Bayesian inference
• Monte Carlo method
• Markov chain Monte Carlo
• Gaussian process regression
• Polynomial chaos expansion

Categories: Uncertainty, Statistical modeling, Computational science, Numerical analysis, Probabilistic modeling