--- title: 'VVUQ: Verification, Validation & Uncertainty Quantification' url: https://www.emergentmind.com/topics/verification-validation-and-uncertainty-quantification-vvuq type: topic --- # VVUQ: Verification, Validation & Uncertainty Quantification Verification, validation, and uncertainty quantification (VVUQ) denotes a fit-for-purpose, risk-informed framework for establishing the credibility of computational models and simulations. Verification asks whether the model equations are being solved right; validation asks whether the equations are right for the intended use; uncertainty quantification characterizes and propagates uncertainty in inputs, measurements, model structure, and surrogate approximations to decision-relevant predictions. Across multiscale simulation, Bayesian calibration, formal probabilistic verification, and data-driven surrogates, VVUQ links numerical correctness, empirical adequacy, and predictive uncertainty into a single workflow oriented toward predictive distributions rather than single-valued best estimates [2105.00553][2509.21697]. ## 1. Core concepts and scope VVUQ is commonly organized around three distinct but coupled tasks. Verification assesses numerical correctness and includes code verification and solution verification. Validation assesses model adequacy by comparing predictions with experimental or observational data in the relevant domain of use. Uncertainty quantification propagates quantified uncertainties to predictions and risk metrics, so that safety margins, acceptance criteria, or other performance thresholds can be evaluated probabilistically rather than pointwise [2105.00553]. | Pillar | Core question | Typical focus | |---|---|---| | Verification | “are we solving the model equations right?” | code verification, solution verification | | Validation | “are the equations the right ones for the intended use?” | comparison with data, adequacy for context of use | | Uncertainty quantification | predictive distributions, not single-valued best estimates | uncertainty propagation to predictions and risk metrics | In application-specific credibility frameworks, VVUQ is explicitly tied to context of use and model risk. In radiotheranostics, for example, PBPK-based theranostic digital twins are framed through a fit-for-purpose, risk-informed foundation in which credibility goals and activities are aligned with conceptual and code verification, internal and external validation, global sensitivity analysis, uncertainty propagation, and decision-focused prediction evaluation. In nuclear systems modeling, the Best Estimate plus Uncertainty paradigm similarly requires predictive distributions for safety-relevant quantities so that acceptance criteria are demonstrated with quantified uncertainty rather than nominal predictions alone [2509.21697][2105.00553]. A recurring implication is that VVUQ is not reducible to a single metric or a single stage in model development. Verification without validation can establish numerical consistency while leaving physical adequacy unresolved. Validation without uncertainty quantification can show nominal agreement while obscuring whether that agreement is robust to parameter, measurement, or model-form uncertainty. UQ without verification and validation can propagate errors whose numerical or physical meaning has not been established. ## 2. Verification: numerical correctness and quantified numerical uncertainty In PDE-based simulation, verification is primarily concerned with discretization, solver convergence, and implementation correctness. Standard solution-verification analyses adopt an asymptotic error model \[ f(h) - f_0 = A h^p + o(h^p), \] where \(f(h)\) is the quantity of interest at mesh size \(h\), \(f_0\) is the exact-but-unknown value, and \(p\) is the observed order. With three successively refined meshes and refinement ratio \(r\), Richardson extrapolation estimates \[ \hat f_0 = f(h_1) + \frac{f(h_1)-f(h_2)}{r^{\hat p}-1}, \] and the Grid Convergence Index uses a safety factor \(F_s\) to form an interval around the finest-grid result. A complementary Bayesian paradigm models discretization uncertainty with Gaussian processes and produces a posterior mean and credible interval for \(f_0\), rather than a deterministic extrapolation interval [2103.14559]. In multiphysics and ODE-based applications, solution verification also includes time-step refinement and solver comparisons. For PBPK models, one recommended convergence study evaluates \[ \mathcal{R} = \frac{\|y_{\Delta t} - y_{\Delta t/2}\|}{\|y_{\Delta t/2} - y_{\Delta t/4}\|} \to 2^p, \] alongside unit tests, dimensional checks, mass-balance tests, regression tests, monitoring of step rejection, and Jacobian conditioning. Manufactured or analytical solutions for simplified subsystems provide an additional route for code and solution verification when exact reference behavior is available [2509.21697]. Verification also appears in formal probabilistic settings. For discrete-time Markov chains, confidence-interval quantitative verification propagates parameter confidence intervals to property-level confidence intervals for PCTL requirements. Decision rules are then stated directly in terms of whether the current property interval lies entirely on the satisfying side of a threshold, entirely on the violating side, or still straddles the decision boundary. When the boundary is unresolved, additional unit-testing data are acquired adaptively to reduce epistemic uncertainty in the most decision-critical components [2109.02984]. A common misconception is that verification is synonymous with validation. The distinction remains fundamental: convergence, conservation, solver residuals, or confidence-interval model checking establish correctness of the computational procedure for a specified mathematical object, not empirical adequacy of that object relative to reality. ## 3. Validation: data comparison, adequacy, and calibration of uncertainty Validation compares predictions with measurements, benchmarks, or observational data in the intended domain of use. In mechanistic models this comparison may be internal, using the data employed during fitting, or external, using held-out or independent data. Internal validation commonly relies on visual fit of time-activity curves, residual-versus-time or residual-versus-predicted diagnostics, histograms centered at zero, SSE, MSE, \(R^2\), parameter coefficients of variation, correlation matrices, and simulation-based checks such as Visual Predictive Checks, Normalized Prediction Distribution Errors, and posterior predictive checks. External validation extends this with RMSE, MAE, likelihood-based criteria, coverage, PIT histograms, concordance correlation coefficients, and Bland–Altman plots [2509.21697]. In multiscale and ensemble simulation, validation may be distributional rather than pointwise. VECMAtk implements a ValidationSimilarity pattern that compares probability distributions of quantities of interest from simulations and experiments using Hellinger distance, Jensen–Shannon distance, and Wasserstein metrics. It also implements EnsembleValidation, which validates each run in an ensemble and aggregates results into a combined validation metric [2010.03923]. Bayesian validation approaches explicitly compare hypotheses. In a BEPU-oriented framework, the calibrated and uncalibrated models are treated as competing hypotheses, \[ B_{01} = \frac{p(D^{\mathrm{VAL}}\mid H_0)}{p(D^{\mathrm{VAL}}\mid H_1)}, \] where \(H_0\) corresponds to the calibrated model and \(H_1\) to the uncalibrated one. The Bayes factor then quantifies support in validation data for posterior-informed versus prior-informed parameter uncertainty and can be propagated into prediction through Bayesian model averaging [2105.00553]. Validation of uncertainty estimates is itself a separate problem. For variance-based UQ, consistency asks whether \[ \mathbb{E}\!\left[(Y-\mu(X))^2 \mid \sigma(X)=\sigma\right] \approx \sigma^2, \] while adaptivity asks whether calibration remains valid across the input domain rather than only on average. For interval-based UQ, the corresponding condition is \[ \mathbb{P}\!\left(Y\in I_{\alpha}(X)\mid X\in S\right)\approx 1-\alpha \] for subsets \(S\) of feature space. Reliability diagrams, calibration curves, local coverage analyses, and confidence curves therefore test different aspects of UQ validity; none of them, by itself, covers all desirable properties [2303.07170]. This suggests that validation is not exhausted by low RMSE or favorable residual plots. Distributional similarity, hypothesis support, coverage, and feature-conditional reliability can disagree, and VVUQ practice increasingly treats those disagreements as diagnostic rather than exceptional. ## 4. Uncertainty quantification and sensitivity analysis UQ in contemporary VVUQ explicitly distinguishes multiple uncertainty sources. A BEPU-oriented taxonomy includes parameter uncertainty, experimental or measurement uncertainty, model uncertainty caused by missing or incomplete physics and numerical approximation errors, and code uncertainty when surrogate models are used [2105.00553]. Application-specific taxonomies often recast these sources into aleatoric and epistemic terms. In PBPK dosimetry, aleatoric uncertainty includes imaging noise, ROI segmentation, TAC sampling schedules, counting statistics, partial volume effects, motion, and biological variability, whereas epistemic uncertainty includes parameter uncertainty, model structure, S-values, and decay-product biodistribution [2509.21697]. In ML surrogates, a closely related category is code/interpolation uncertainty, arising from data coverage, extrapolation, imperfect architecture, and stochastic training [2503.17385][2206.14615]. Non-intrusive propagation methods are widely used. EasyVVUQ supports stochastic collocation, polynomial chaos expansion, variance-based sensitivity analysis, Sobol analysis, bootstrapping, and basic statistics; it can also integrate surrogates and emulators for model-form uncertainty [2010.03923]. In CFD-oriented workflows, standard and probabilistic polynomial chaos expansions are used to propagate uncertainty from design variables and observation-dependent noise, while heteroscedastic Gaussian process regression supplies surrogates with input-dependent predictive variance [2007.07071]. In fusion virtual qualification, nested sampling over aleatoric and epistemic inputs yields an ensemble of CDFs whose envelope defines a probability box, \[ F_{\mathrm{lower}}(x)=\min_i F_i(x), \qquad F_{\mathrm{upper}}(x)=\max_i F_i(x), \] so that both types of uncertainty are retained without collapsing to a single precise distribution [2605.11886]. Global sensitivity analysis provides the main bridge between UQ and model reduction. Variance-based Sobol decomposition writes \[ \operatorname{Var}(Y)=\sum_i V_i+\sum_{i