CFD verification vs validation

The core difference

These questions are connected but independent. A numerical solution can be verified yet disagree with reality because the physical model, boundary conditions or properties are inadequate. A result may agree with one experiment by compensation of errors while remaining numerically unresolved.

The complete CFD credibility chain

A reviewer usually evaluates credibility in a sequence. Evidence later in the chain cannot repair missing evidence earlier in the chain.

• Problem definition.

  Geometry, operating conditions, required outputs and intended use are stated clearly.

• Physical model.

  Governing models, turbulence, heat transfer, phases, materials and assumptions are justified.

• Iterative/statistical convergence.

  Each numerical case is solved sufficiently for the reported outputs.

• Solution verification.

  Spatial, temporal and other numerical sensitivities are quantified.

• Validation.

  Predictions are compared with appropriate physical or benchmark evidence.

• Application prediction.

  The verified and validated model is applied within a stated domain and its uncertainty and limitations are respected.

Convergence is necessary—but not sufficient

Convergence describes whether the numerical solution process or sampled statistics have settled adequately. It does not establish that the converged solution is accurate.

Low residuals can coexist with a wrong answer. Conversely, a transient or statistically stationary simulation may show oscillatory instantaneous residuals while producing stable, repeatable statistics. Use convergence evidence appropriate to the physics.

Code verification: are the equations implemented correctly?

Code verification focuses on finding programming, algorithmic or discretization-implementation errors. It is especially important when developing a solver, adding custom source terms, implementing a new boundary condition or modifying OpenFOAM code.

Typical code-verification evidence

• Comparison with exact analytical solutions
• Manufactured solutions with known forcing terms
• Canonical problems with known convergence behaviour
• Unit tests and regression tests for custom implementations
• Comparison against an independently implemented method
• Demonstration that formal discretization order is recovered where expected

Solution verification: quantify numerical error in your case

Solution verification evaluates the numerical uncertainty of the actual simulation. It does not ask whether the model is physically correct; it asks whether the numerical approximation is sufficiently resolved.

Spatial verification

Use systematically refined meshes, quantities of interest, observed order, Richardson extrapolation and Grid Convergence Index where assumptions are satisfied. Report local resolution such as y+, prism layers, wakes, gaps and interfaces when local outputs matter.

Temporal verification

Repeat the simulation using at least one smaller timestep while keeping spatial resolution and physical duration appropriate. Compare mean values, peak values, phase, dominant frequencies and transient events—not only Courant number.

Other numerical sensitivities

• Domain extent and inlet/outlet placement
• Numerical schemes and limiter choices
• Interface interpolation and overset resolution
• Particle counts or parcel statistics
• Angular resolution for rotating machinery
• Sampling duration for unsteady statistics

Validation: compare with physical evidence

Validation evaluates the combined mathematical model, assumptions, properties, boundary conditions and numerical prediction against evidence representing the physical system.

A credible validation comparison requires alignment

• Geometry and characteristic dimensions
• Operating conditions and boundary conditions
• Fluid and solid properties
• Measurement location and definition
• Averaging method and reference quantities
• Flow regime and relevant dimensionless parameters
• Experimental uncertainty and repeatability where available

A close curve is not enough if the experiment and CFD do not represent the same conditions. Document differences and explain how they limit the validation claim.

Separate the main uncertainty sources

The difference between CFD and experiment contains contributions from several sources. Reporting a single percentage error without discussing these sources can make the comparison look more certain than it is.

Use validation metrics that match the evidence

Choose metrics appropriate to the quantity, scale and research claim.

Possible metrics

• Absolute and relative error for scalar outputs
• Root-mean-square error for profiles or time series
• Normalized mean bias for systematic over/underprediction
• Correlation or phase error when pattern and timing matter
• Integrated differences for forces, heat rates or performance maps
• Confidence or uncertainty intervals when data support them

A metric must not hide the engineering meaning. A small average error may conceal an unacceptable peak temperature or incorrect separation location.

What if no new experimental data are available?

Lack of a dedicated experiment does not justify skipping validation. Build a validation strategy from the strongest available evidence and limit claims accordingly.

• Analytical and limiting solutions.

  Test conduction, laminar flow, hydrostatic, fully developed or energy-balance limits where they apply.

• Canonical benchmarks.

  Validate submodels using recognized cavities, channels, jets, waves, rotating passages or other standard cases.

• Published experiments.

  Reproduce a sufficiently similar configuration with traceable geometry and operating conditions.

• Component-level validation.

  Validate heat transfer, pressure loss, turbulence or interface behaviour separately before combining them.

• Cross-code comparison.

  Use as supporting evidence, not physical validation; two codes can share the same modelling deficiency.

• Trend and physical-consistency checks.

  Confirm conservation, asymptotic behaviour, dimensionless trends and known physical limits.

Build a validation hierarchy

Complex systems are easier to defend when evidence is organized from simple physics to full application.

• Level 1 — Unit physics.

  Test individual mechanisms such as conduction, pressure loss, free convection or a turbulence benchmark.

• Level 2 — Components.

  Validate a channel, heat exchanger element, blade passage, jet, valve or free-surface subsystem.

• Level 3 — Coupled assembly.

  Test interactions between components and physics with controlled data.

• Level 4 — Application.

  Compare system-level outputs under representative operating conditions.

This hierarchy identifies which parts of the prediction are strongly supported and where uncertainty remains. It is often more credible than a single system-level comparison with poorly known inputs.

How to write verification and validation sections

Verification subsection

Report convergence criteria, conservation, mesh family, quantities of interest, refinement ratios, observed convergence, spatial uncertainty, timestep sensitivity and production-mesh selection.

Validation subsection

Describe the reference evidence, matching geometry and conditions, measurement definitions, experimental uncertainty, comparison metrics, deviations, limitations and the domain over which the validation applies.

Replace placeholders with actual evidence. Avoid universal claims such as “the CFD model is validated” when only one output or operating point was compared.

Common journal-review objections

• “Residuals alone do not demonstrate convergence.”

  Add engineering monitors, conservation and statistical or periodic stability.

• “Mesh independence is incorrectly presented as validation.”

  Separate solution verification from physical-reference comparison.

• “The validation case does not match the application.”

  Explain similarity in geometry, regime and dimensionless parameters—or reduce the claim.

• “Experimental uncertainty is missing.”

  Report available uncertainty and avoid overinterpreting differences below the evidence resolution.

• “Boundary conditions were tuned to the validation data.”

  Disclose calibration and add an independent validation condition.

• “Only one operating point was validated.”

  Add conditions spanning the intended prediction range or explicitly limit applicability.

• “The turbulence/model choice is unsupported.”

  Use physics-based justification, sensitivity comparison and relevant benchmark evidence.

Match every claim to the available evidence

The strongest CFD papers do not claim certainty. They show a transparent chain of numerical and physical evidence, then state conclusions at exactly the level that evidence supports.