Course 25 | Module 7 of 12

Simulation Credibility and VVUQ

Separate numerical correctness, physical adequacy, uncertainty, and decision-specific credibility.

MAP

Module map

Learning outcomes

  • Distinguish code verification, solution verification, validation, calibration, and uncertainty quantification.
  • Estimate discretization and input uncertainty for simple models.
  • Interpret validation evidence without equating calibration with prediction.
  • Build a context-of-use and risk-informed credibility argument.

Evidence standard

Complete all four lessons, reproduce the worked checks, run the lab, and correct the weekly quiz. Treat AI output as candidate evidence until independently verified.

7.1

The VVUQ taxonomy and error-source map

Why this lesson matters

A single percent agreement number cannot tell whether discrepancy came from code, discretization, inputs, experiment, or model form.

Learning objectives

  • Define and distinguish Code verification and Solution verification.
  • Apply the lesson method to the worked the vvuq taxonomy and error-source map case.
  • Evaluate evidence, uncertainty, and AI-assisted output before making a claim.

Readiness check

Before continuing, explain what decision this topic supports and name one upstream source that must be controlled.

Check your response

A sound answer names a specific engineering decision, its configuration, and a controlled requirement, model, dataset, interface, or standard that constrains the work.

Core idea

Verification asks whether the mathematical model is solved correctly. Validation asks whether the model adequately represents reality for its intended use. UQ characterizes how uncertainty in inputs, models, numerics, and measurements affects quantities of interest and decisions.

Key concepts

Code verificationEvidence that software correctly implements the intended mathematical algorithms.
Solution verificationEstimation of numerical error for a specific simulation solution.
ValidationAssessment of model adequacy through comparison with relevant experimental reality.
Uncertainty quantificationCharacterization and propagation of uncertainty to model outputs and decisions.

Step-by-step explanation

  1. Define the quantity of interest and intended decision.
  2. Map implementation, iteration, discretization, round-off, input, parameter, measurement, and model-form sources.
  3. Verify code with exact, manufactured, or benchmark solutions where appropriate.
  4. Verify the specific solution using convergence and numerical-error evidence.
  5. Compare with relevant experiments and propagate uncertainty before making a credibility claim.

Worked example

A CFD pressure drop is 4.8 kPa and a test reports 5.1 kPa. The mesh study estimates 0.15 kPa numerical uncertainty and the experiment reports 0.20 kPa standard uncertainty.

  1. 1

    Comparison error is 4.8 - 5.1 = -0.3 kPa, or -5.88% relative to the test.

  2. 2

    Keep numerical and experimental uncertainty separate before combining under declared assumptions.

  3. 3

    Do not call the remaining gap model-form error exactly; uncertain inputs and boundary conditions also contribute.

  4. 4

    Check code and solution verification before interpreting physical-model adequacy.

Result. The -0.3 kPa difference is an observed comparison, not a complete diagnosis. Credibility requires an uncertainty-aware decomposition and relevance to use.

Independent check. Each discrepancy source is named and no component is assigned more certainty than evidence supports.

Common misconceptions

MisconceptionCorrection
Validation means the model is trueValidation provides evidence of adequacy for defined conditions and quantities; it does not prove universal truth.
A tool output closes the questionA result remains a candidate until its inputs, method, configuration, uncertainty, and relevance have been checked.

Diagnostic questions

Where does model-form uncertainty enter?

Through imperfect governing assumptions, closures, constitutive relations, geometry idealization, and omitted physics.

What would make this work reproducible?

Controlled inputs, method or code, versions, assumptions, outputs, and a stated interpretation tied to the decision.

Practice ladder

Basic

Classify ten VVUQ activities into code verification, solution verification, validation, calibration, or UQ.

Intermediate

Build an error-source map for a finite-element stress result.

Advanced

Explain why agreement can be accidental when two errors cancel.

AI-assisted engineering task

Ask AI to classify an evidence list using definitions, then identify statements that overclaim certainty.

How to prove the AI output yourself

  1. Check each label against the VVUQ taxonomy.
  2. Inspect actual procedures and quantities of interest.
  3. Require a technical reviewer to resolve ambiguous activities.

Retrieval and spaced review

Answer closed-notes today, then again after 1, 3, 7, and 30 days.

Define Code verification.

Evidence that software correctly implements the intended mathematical algorithms.

What role does Solution verification play here?

Estimation of numerical error for a specific simulation solution.

What must a reviewer be able to reconstruct?

Each discrepancy source is named and no component is assigned more certainty than evidence supports.

End-of-lesson summary

Verification asks whether the mathematical model is solved correctly. Validation asks whether the model adequately represents reality for its intended use. UQ characterizes how uncertainty in inputs, models, numerics, and measurements affects quantities of interest and decisions.

Student notes

Draw one error-source tree per simulation before reporting an agreement percentage.

Recommended readings

Instructor notes

Insist on nouns: code, equations, solution, experiment, quantity, decision. This prevents slogan-level V&V definitions.

7.2

Code and solution verification through beam deflection

Why this lesson matters

A plausible result can hide coding defects, inadequate discretization, or an equation applied outside its assumptions.

Learning objectives

  • Define and distinguish Manufactured solution and Discretization error.
  • Apply the lesson method to the worked code and solution verification through beam deflection case.
  • Evaluate evidence, uncertainty, and AI-assisted output before making a claim.

Readiness check

Before continuing, explain what decision this topic supports and name one upstream source that must be controlled.

Check your response

A sound answer names a specific engineering decision, its configuration, and a controlled requirement, model, dataset, interface, or standard that constrains the work.

Core idea

Code verification tests implementation against known mathematical truth. Solution verification estimates remaining numerical error in a particular calculation. Analytical checks and convergence studies are complementary, not substitutes for one another.

Key concepts

Manufactured solutionA constructed exact solution used to test code implementation and observed order.
Discretization errorDifference introduced by representing continuous equations on finite spatial or temporal resolution.
Observed orderThe convergence rate inferred from a systematic refinement sequence.
Asymptotic rangeRefinement regime where the theoretical error model approximately governs behavior.

Step-by-step explanation

  1. Reproduce a known analytical or manufactured solution to challenge implementation.
  2. Refine mesh or time step systematically with fixed physics and solver tolerances.
  3. Check monotonicity, observed order, and whether results enter an asymptotic range.
  4. Estimate numerical uncertainty using a stated method and assumptions.
  5. Keep model applicability separate from numerical accuracy.

Worked example

For the beam in Module 4, analytical deflection is 2.061 mm. Three finite-element meshes predict 1.93, 2.02, and 2.05 mm.

  1. 1

    Compute absolute errors relative to the analytical solution: 0.131, 0.041, and 0.011 mm.

  2. 2

    The sequence approaches the reference and error decreases, supporting convergence.

  3. 3

    Do not infer exact order without refinement ratios and element formulation.

  4. 4

    Check that the analytical and FEA models share simply supported boundaries, center load, linear elasticity, and Euler-Bernoulli applicability.

Result. The finest result differs by -0.011 mm, about -0.53%. This is useful verification evidence under aligned mathematical assumptions, not validation of beam reality.

Independent check. Reference solution, refinement definition, solver tolerance, quantity extraction, and assumptions are documented and reproducible.

Common misconceptions

MisconceptionCorrection
Mesh convergence validates the modelIt addresses numerical solution behavior, not physical adequacy against reality.
A tool output closes the questionA result remains a candidate until its inputs, method, configuration, uncertainty, and relevance have been checked.

Diagnostic questions

What must remain fixed in a mesh study?

Physics, geometry, boundary conditions, solver settings, extraction definition, and convergence tolerances except the intended refinement.

What would make this work reproducible?

Controlled inputs, method or code, versions, assumptions, outputs, and a stated interpretation tied to the decision.

Practice ladder

Basic

Compute the three absolute and relative errors.

Intermediate

Design a mesh-refinement table with element size, degrees of freedom, result, iteration error, and estimated discretization error.

Advanced

Explain how shear deformation could make both a converged FEA solution and Euler-Bernoulli reference inadequate for a deep beam.

AI-assisted engineering task

Ask AI to check refinement-table arithmetic and identify missing columns, without inferring an observed order from incomplete data.

How to prove the AI output yourself

  1. Recompute errors.
  2. Plot result versus resolution.
  3. Check solver and iteration convergence.
  4. Compare assumptions with the reference solution.

Retrieval and spaced review

Answer closed-notes today, then again after 1, 3, 7, and 30 days.

Define Manufactured solution.

A constructed exact solution used to test code implementation and observed order.

What role does Discretization error play here?

Difference introduced by representing continuous equations on finite spatial or temporal resolution.

What must a reviewer be able to reconstruct?

Reference solution, refinement definition, solver tolerance, quantity extraction, and assumptions are documented and reproducible.

End-of-lesson summary

Code verification tests implementation against known mathematical truth. Solution verification estimates remaining numerical error in a particular calculation. Analytical checks and convergence studies are complementary, not substitutes for one another.

Student notes

Keep code verification evidence, solution verification evidence, and model-validation evidence in separate notebook sections.

Recommended readings

Instructor notes

The arithmetic is intentionally simple. Spend time on what the comparison does and does not establish.

7.3

Validation, calibration, and uncertainty sources

Why this lesson matters

A model tuned to one dataset can reproduce it while making poor predictions elsewhere. Calibration and validation must remain distinguishable.

Learning objectives

  • Define and distinguish Calibration and Validation experiment.
  • Apply the lesson method to the worked validation, calibration, and uncertainty sources case.
  • Evaluate evidence, uncertainty, and AI-assisted output before making a claim.

Readiness check

Before continuing, explain what decision this topic supports and name one upstream source that must be controlled.

Check your response

A sound answer names a specific engineering decision, its configuration, and a controlled requirement, model, dataset, interface, or standard that constrains the work.

Core idea

Calibration estimates model parameters using data under stated assumptions. Validation evaluates predictive adequacy using relevant, preferably independent data. Aleatory variability and epistemic uncertainty require different interpretations even when both are represented probabilistically.

Key concepts

CalibrationInference of parameter values or distributions from observations and a model.
Validation experimentAn experiment designed to measure quantities and conditions relevant to assessing model adequacy.
Aleatory uncertaintyVariability represented as irreducible within the chosen model and information state.
Epistemic uncertaintyLack of knowledge that may be reduced or better characterized with information.

Step-by-step explanation

  1. Define the prediction question and validation hierarchy.
  2. Design experiments with controlled inputs, measured boundary conditions, and uncertainty estimates.
  3. Separate data used for parameter calibration from data used for predictive assessment.
  4. Compare quantities using metrics that preserve sign, scale, uncertainty, and physical meaning.
  5. Report the validated domain and prediction limitations rather than a universal validity label.

Worked example

A convection coefficient is calibrated to make a thermal model match a 100 W steady test. The same coefficient is then used to predict a 500 W transient.

  1. 1

    The 100 W agreement is calibration evidence because h was chosen using that dataset.

  2. 2

    It cannot independently validate the model or the 500 W transient prediction.

  3. 3

    Collect separate tests spanning flow, temperature, and transient regimes with measured boundary conditions.

  4. 4

    Assess whether a constant h model remains physically adequate or only compensates for omitted contact and radiation effects.

Result. Calibration can improve fit while hiding model discrepancy. Independent, relevant validation and uncertainty analysis are needed for the 500 W decision.

Independent check. Calibration and validation datasets, parameters, objectives, uncertainty, and applicability domains are explicitly separated.

Common misconceptions

MisconceptionCorrection
A calibrated model is validatedCalibration shows a fit under its data and assumptions; predictive validation requires relevant independent evidence.
A tool output closes the questionA result remains a candidate until its inputs, method, configuration, uncertainty, and relevance have been checked.

Diagnostic questions

Can aleatory and epistemic categories change?

Yes. The classification depends on the model, information, and decision context; what is unknown today may be measured tomorrow.

What would make this work reproducible?

Controlled inputs, method or code, versions, assumptions, outputs, and a stated interpretation tied to the decision.

Practice ladder

Basic

Classify four datasets as calibration, validation, or operational monitoring.

Intermediate

Design a validation matrix across heat load, flow, and ambient temperature.

Advanced

Explain parameter compensation and identifiability when contact resistance and convection are both tuned.

AI-assisted engineering task

Ask AI to identify whether each dataset influenced parameter fitting and to flag leakage between calibration and validation.

How to prove the AI output yourself

  1. Inspect code and data splits.
  2. Re-run calibration with held-out cases.
  3. Check sensitivity and parameter identifiability.
  4. Review experimental uncertainty and boundary measurements.

Retrieval and spaced review

Answer closed-notes today, then again after 1, 3, 7, and 30 days.

Define Calibration.

Inference of parameter values or distributions from observations and a model.

What role does Validation experiment play here?

An experiment designed to measure quantities and conditions relevant to assessing model adequacy.

What must a reviewer be able to reconstruct?

Calibration and validation datasets, parameters, objectives, uncertainty, and applicability domains are explicitly separated.

End-of-lesson summary

Calibration estimates model parameters using data under stated assumptions. Validation evaluates predictive adequacy using relevant, preferably independent data. Aleatory variability and epistemic uncertainty require different interpretations even when both are represented probabilistically.

Student notes

Label every dataset by role before fitting: calibration, validation, monitoring, or final decision evidence.

Recommended readings

Instructor notes

Use a visibly excellent calibrated fit and a poor held-out prediction. Students should distrust fit quality without data provenance.

7.4

Context of use, model risk, and credibility claims

Why this lesson matters

Credibility is not an intrinsic badge on a model. Required evidence depends on how the model influences a decision and the consequence of a wrong decision.

Learning objectives

  • Define and distinguish Context of use and Model influence.
  • Apply the lesson method to the worked context of use, model risk, and credibility claims case.
  • Evaluate evidence, uncertainty, and AI-assisted output before making a claim.

Readiness check

Before continuing, explain what decision this topic supports and name one upstream source that must be controlled.

Check your response

A sound answer names a specific engineering decision, its configuration, and a controlled requirement, model, dataset, interface, or standard that constrains the work.

Core idea

A context of use states the specific role and scope of a model in a decision. Risk-informed credibility scales verification, validation, and UQ rigor with model influence and decision consequence. The result is a bounded credibility claim with limitations, not 'the model is validated'.

Key concepts

Context of useA precise statement of how and for what decision a model will be used.
Model influenceThe degree to which the model determines or supports the decision.
Decision consequenceThe severity of adverse outcome if the model contributes to a wrong decision.
Credibility claimA bounded statement of confidence supported by specified evidence for a context of use.

Step-by-step explanation

  1. State the question of interest and decision owner.
  2. Write the model's context of use, including outputs, conditions, and role.
  3. Assess model influence and consequence to determine model risk.
  4. Set credibility goals and plan verification, validation, UQ, review, and reporting.
  5. Compare achieved evidence with goals and issue a decision-specific claim and limitations.

Worked example

The beam model may be used in two ways: A, rank three concepts before prototype; B, certify that a flight-critical support meets its final deflection limit without physical testing.

  1. 1

    Use A has moderate influence and limited consequence because later test can catch error; analytical checks and bounded comparisons may suffice.

  2. 2

    Use B gives the model high influence and high consequence; substantially stronger configuration control, verification, validation, uncertainty, independent review, and justification are required.

  3. 3

    Do not transfer evidence from A automatically to B because context, quantity, design maturity, and consequence differ.

  4. 4

    Write separate credibility claims and unresolved limitations.

Result. The same code can be adequate for concept ranking and inadequate for certification. Credibility belongs to the model-use-evidence combination.

Independent check. The evidence plan is commensurate with influence and consequence, and the final claim never exceeds the validated conditions or quantities.

Common misconceptions

MisconceptionCorrection
A model is validated onceValidation evidence is conditional on quantity, domain, configuration, and context of use.
A tool output closes the questionA result remains a candidate until its inputs, method, configuration, uncertainty, and relevance have been checked.

Diagnostic questions

Can a low-fidelity model be credible?

Yes for a bounded low-risk use if evidence shows it is adequate for that decision.

What would make this work reproducible?

Controlled inputs, method or code, versions, assumptions, outputs, and a stated interpretation tied to the decision.

Practice ladder

Basic

Write a one-sentence context of use for concept ranking.

Intermediate

Compare credibility goals for ranking, sizing, and certification uses.

Advanced

Build a model-credibility assessment table with goals, achieved evidence, gaps, and residual decision risk.

AI-assisted engineering task

Ask AI to critique whether a credibility claim exceeds its context of use and cited evidence.

How to prove the AI output yourself

  1. Read the full evidence, not the summary.
  2. Check validated domain and configuration.
  3. Compare planned and achieved credibility activities.
  4. Require accountable decision-owner acceptance.

Retrieval and spaced review

Answer closed-notes today, then again after 1, 3, 7, and 30 days.

Define Context of use.

A precise statement of how and for what decision a model will be used.

What role does Model influence play here?

The degree to which the model determines or supports the decision.

What must a reviewer be able to reconstruct?

The evidence plan is commensurate with influence and consequence, and the final claim never exceeds the validated conditions or quantities.

End-of-lesson summary

A context of use states the specific role and scope of a model in a decision. Risk-informed credibility scales verification, validation, and UQ rigor with model influence and decision consequence. The result is a bounded credibility claim with limitations, not 'the model is validated'.

Student notes

Never write 'validated model' alone. Complete the sentence: validated for which quantity, conditions, configuration, and decision role?

Recommended readings

Instructor notes

Note that NASA-STD-7009B is the current edition and supersedes the A edition named in the original course brief. Use the public historical A text for comparison, but teach current status accurately.

LAB 7

Lab 7: Propagate uncertainty through a beam-deflection model

Lab objective

Use Monte Carlo sampling to propagate load, modulus, width, and height uncertainty and estimate the probability of exceeding a deflection limit.

Engineering context

The simply supported aluminum beam has a 2.50 mm deflection requirement. Height uncertainty is especially important because deflection scales with h^-3.

Input data

  • P = 1200 +/- 30 N
  • E = 69 +/- 1.5 GPa
  • b = 40.0 +/- 0.2 mm
  • h = 30.0 +/- 0.15 mm
  • L = 0.8 m
  • limit = 2.50 mm

Step-by-step task

  1. Sample independent normal inputs as a teaching assumption
  2. Compute I and deflection
  3. Summarize mean and 95% interval
  4. Estimate exceedance probability and rank correlations

Python code

import numpy as np

rng = np.random.default_rng(17)
n = 100_000
P = rng.normal(1200.0, 30.0, n)
E = rng.normal(69e9, 1.5e9, n)
b = rng.normal(0.0400, 0.0002, n)
h = rng.normal(0.0300, 0.00015, n)
L = 0.8
I = b * h**3 / 12.0
delta_mm = P * L**3 / (48.0 * E * I) * 1000.0
low, high = np.quantile(delta_mm, [0.025, 0.975])
print(f"mean={delta_mm.mean():.3f} mm")
print(f"95% interval=[{low:.3f}, {high:.3f}] mm")
print(f"P(delta > 2.50 mm)={(delta_mm > 2.50).mean():.5f}")
for name, values in {"P": P, "E": E, "b": b, "h": h}.items():
    print(name, np.corrcoef(values, delta_mm)[0, 1])

Explanation of code

Step 1 sample independent normal inputs as a teaching assumption Step 2 compute I and deflection Step 3 summarize mean and 95% interval Step 4 estimate exceedance probability and rank correlations

Expected output

A mean near the nominal 2.06 mm, a finite 95% interval, and a small exceedance estimate that is reproducible with seed 17.

Interpretation

The calculation illustrates propagation, not a validated probability model. Distribution choice, correlation, tolerances, model form, and manufacturing data require justification.

Common errors

  • Treating +/- values as normal standard deviations without justification
  • Ignoring correlation
  • Reporting Monte Carlo precision as physical certainty

Extension tasks

  • Use bounded or tolerance-derived distributions
  • Add model discrepancy
  • Converge the exceedance estimate and compute sampling confidence

Reflection questions

  • Why does height dominate?
  • What uncertainty is missing?
  • Would a nominal pass be sufficient for a high-consequence decision?
WEEK 7

Weekly quiz and concept check

Closed notes. Answer each item, then use the key to correct in a different color.

  1. What does code verification ask?
  2. What does solution verification ask?
  3. What does validation ask?
  4. Why is calibration not validation?
  5. What determines credibility rigor?
  6. What is a defensible credibility claim?
Answer key
  1. 1. Whether software correctly implements the intended mathematical method.
  2. 2. How much numerical error remains in a particular solution.
  3. 3. Whether the model is adequately representative of reality for a defined use.
  4. 4. The same data used to tune parameters cannot independently test prediction.
  5. 5. Context of use, model influence, decision consequence, and residual uncertainty.
  6. 6. A bounded statement tied to quantities, conditions, configuration, evidence, use, and limitations.
SOURCES

Module source map

SourceHow it is used
Oberkampf and Roy, Verification and Validation in Scientific ComputingVerification, validation, numerical error, uncertainty, prediction, and simulation credibility.
NASA-STD-7009, Standard for Models and SimulationsModel and simulation lifecycle, credibility products, acceptance criteria, and reporting. NASA-STD-7009B supersedes 7009A.
ASME V&V 10, Computational Solid MechanicsCommon language and general VVUQ guidance for computational solid mechanics.
AIAA G-077, CFD Verification and Validation GuideCFD credibility, verification, validation, and numerical uncertainty.
ASME V&V 40, Risk-informed Model CredibilityContext of use, model risk, and evidence rigor commensurate with decision consequence.
FDA Guidance on Computational Modeling and Simulation CredibilityRisk-informed credibility assessment and transparent reporting of computational evidence.

Access labels and full-course source notes are on the course home page. Paywalled standards are not paraphrased as if their full text were accessed.