Verification and Validation in Computational Science

Verification and Validation in Computational Science
💡No image available
Overview
Scope	Verification and validation (V&V) for computational models and simulations
Primary Goal	Ensure correctness of implementation and credibility relative to observations
Common Methods	Analytic checks, method-of-manufactured-solutions, convergence studies, uncertainty quantification, and comparison to experiments

Verification and validation (V&V) are core processes in computational science used to establish that numerical models are implemented correctly and that their results are credible for their intended applications. Verification addresses whether a simulation solves the governing equations as intended, while validation assesses whether the model outputs agree with experimental data or other trusted references. Together, V&V support scientific reliability in fields ranging from computational fluid dynamics to climate modeling.

Overview

In computational science, models typically represent physical or engineering systems using mathematical equations and numerical discretizations. A simulation may be “correct” relative to its code and equations yet still be wrong for the real-world phenomenon being studied. This distinction motivates the separation between verification, which focuses on numerical and implementation fidelity, and validation, which focuses on empirical adequacy.

A widely used framework distinguishes these two objectives as part of broader model quality practices that often include uncertainty quantification. In high-stakes contexts such as aerospace design and nuclear engineering, V&V is closely connected with computational modeling, numerical analysis, and uncertainty quantification. Standards and guidance documents in these areas frequently reference related topics such as model calibration and statistical model checking.

Verification

Verification aims to determine whether the computational model is solving the intended set of equations and whether the implemented algorithm behaves consistently with mathematical expectations. Typical verification activities include code review and documentation, unit testing, and systematic checks for numerical accuracy.

Numerical verification commonly uses techniques such as convergence testing by refining the mesh or time step to determine whether the solution approaches a stable limit. Another method is the method of manufactured solutions, which creates a known exact solution and then derives forcing terms so that numerical error can be quantified even when analytic solutions to the original equations are unavailable. For time-dependent problems, verification may also involve verifying stability properties consistent with the chosen discretization scheme described in references such as finite difference method or finite element method.

In addition to these accuracy checks, verification often considers software quality attributes. Practices such as reproducibility and regression testing help ensure that changes do not degrade correctness over successive versions of a code base. When simulations are parallelized, verification may also include tests of deterministic results across processor counts to detect race conditions or inconsistent reductions.

Validation

Validation examines whether a computational model is an adequate representation of the real system for the intended use. This process typically involves comparing simulation outputs with experimental observations, benchmark problems, or other high-confidence references.

Validation is frequently tied to the selection of measurable quantities and acceptance criteria. For example, in computational fluid dynamics, one may compare predicted drag force or velocity profiles against wind-tunnel measurements across a range of operating conditions. In climate and geoscience applications, validation may involve assessing simulated fields such as temperature or precipitation patterns against observations from datasets produced by established observational programs.

A major challenge in validation is that discrepancies can arise from multiple sources: incomplete physics, modeling assumptions, parameter uncertainty, or unmodeled processes. Accordingly, validation is often conducted in conjunction with model calibration to adjust parameters within defensible ranges. The relationship between validation outcomes and uncertainty is commonly analyzed using tools from Bayesian inference and related statistical frameworks, especially when reporting credibility intervals for predictions.

Quantification of uncertainty and error

Even when verification and validation are both performed, residual errors and uncertainties remain. Computational scientists therefore frequently augment V&V with uncertainty quantification to characterize how input uncertainties propagate to model outputs and to how confident users should be in simulation results.

Uncertainty sources may include uncertain boundary conditions, stochasticity in initial states, variability in material properties, and numerical discretization error. Methods for disentangling these contributions often leverage statistical sampling (e.g., Monte Carlo method) and sensitivity analysis. Error budgets can also be constructed to separate discretization effects from modeling-form uncertainty, helping analysts avoid over-attributing discrepancies to a single cause.

In some domains, V&V is extended to include model management practices such as tracking versions, documenting assumptions, and ensuring that numerical experiments are repeatable. This is especially relevant for large-scale models developed over long time periods, where changes to solvers, meshes, or parameter sets can affect outcomes even if the governing equations remain nominally unchanged.

Standards, practice, and relevance

V&V is often formalized through organizational processes and standards that guide documentation, testing, and reporting. These practices reflect the need for traceability from model equations and code to reported predictions, and they support decision-making in safety-critical and regulatory contexts.

In applied research and engineering, V&V is commonly integrated into a simulation workflow: specify intended use, select or derive the governing equations, implement and verify the numerical methods, validate against appropriate datasets, then quantify uncertainties relevant to the final application. Discussions of these workflows intersect with broader quality approaches found in software testing and scientific software practices, including reproducibility principles that are increasingly emphasized in computational disciplines.

Because computational models can be used for forecasting, design optimization, and risk assessment, V&V helps clarify the limits of model applicability. When validation is incomplete or performed only for narrow regimes, results should be interpreted as conditional and not assumed to generalize beyond the tested domain.