Verification and Validation in CFD: How to Know Whether Your Simulation Can Be Trusted
Learn how to assess the credibility of CFD simulations through numerical verification, mesh and time-step independence, residual convergence, conservation checks, Richardson extrapolation, Grid Convergence Index (GCI), experimental validation, uncertainty quantification, and practical OpenFOAM workflows.
1. Introduction: A Converged Simulation Is Not Necessarily a Correct Simulation
Computational Fluid Dynamics (CFD) has become an essential engineering tool for analysing fluid flow, heat transfer, turbulence, combustion, aerodynamics, and many other physical processes.
Modern CFD software can generate impressive visualisations of velocity fields, pressure distributions, temperature contours, and turbulent structures. With a few clicks, an engineer can create a computational domain, generate a mesh, select a turbulence model, assign boundary conditions, and run a simulation.
However, an important question remains:
How do we know whether the results produced by our CFD simulation are actually reliable?
Imagine that you are simulating airflow through a ventilation duct. Your simulation converges, the residuals decrease, and the velocity contours look physically reasonable.
But what if:
- The mesh is too coarse to resolve the velocity gradients near the walls?
- The time step is too large to capture the transient flow behaviour?
- The turbulence model does not represent the actual flow physics?
- The inlet velocity profile is unrealistic?
- The numerical solution does not conserve mass?
- The predicted pressure drop differs significantly from experimental measurements?
A simulation can converge mathematically and still produce physically inaccurate results.
This is why Verification and Validation (V&V) are fundamental parts of reliable CFD engineering.
Verification asks:
Are we solving the mathematical equations correctly?
Validation asks:
Do the mathematical equations and models represent the real physical problem sufficiently well?
These two activities help establish confidence in numerical predictions before they are used for engineering decisions.
This article explains the fundamental principles of CFD verification and validation using simple physical examples, mathematical formulations, practical engineering interpretations, and OpenFOAM workflows.
What you will learn
By the end of this article, you should understand:
- The difference between verification and validation.
- Why solver convergence does not guarantee accuracy.
- How to perform mesh independence studies.
- How to assess time-step independence.
- How to interpret residuals and monitor physical quantities.
- How to check mass and energy conservation.
- What Richardson extrapolation and Grid Convergence Index (GCI) mean.
- How to compare CFD results with analytical and experimental data.
- How uncertainty affects numerical predictions.
- How to organise a systematic CFD V&V workflow in OpenFOAM.
2. Understanding Verification and Validation
Verification and validation are closely related, but they answer different questions.
A simple way to understand the distinction is to consider a vehicle speedometer.
Suppose a speedometer displays 100 km/h.
Verification: Does the speedometer’s internal calculation correctly convert the wheel rotation signal into a speed value?
Validation: Does the displayed speed agree sufficiently well with the vehicle’s actual speed measured using a reliable reference instrument?
The first question concerns whether the calculation is implemented correctly. The second concerns whether the result represents reality.
The same distinction applies to CFD.
2.1 Verification: Are We Solving the Equations Correctly?
Verification examines whether the numerical solution correctly represents the mathematical equations being solved.
In CFD, governing equations such as the Navier–Stokes equations are converted into algebraic equations that can be solved numerically.
This process introduces several possible sources of error:
- Spatial discretisation.
- Temporal discretisation.
- Iterative solver tolerances.
- Mesh resolution.
- Numerical implementation.
- Programming and software errors.
Verification examines whether these numerical procedures produce a sufficiently accurate solution to the intended mathematical model.
For example, consider laminar flow through a straight circular pipe.
An analytical solution exists for fully developed laminar pipe flow. We can compare our CFD velocity profile against this analytical solution.
If the CFD solution approaches the analytical result as the mesh is refined, this provides evidence that the numerical implementation is behaving correctly.
This is an example of numerical verification using a known mathematical solution.
2.2 Validation: Are We Solving the Right Physics?
Validation examines whether the mathematical model adequately represents the real physical process.
Even if a CFD solver accurately solves the selected equations, the selected equations or physical models may not represent reality.
For example, consider turbulent airflow around a vehicle.
A CFD model may solve the Reynolds-averaged Navier–Stokes equations accurately, but the predicted drag coefficient can still differ from experimental measurements because of:
- Limitations of the turbulence model.
- Inaccurate representation of the vehicle geometry.
- Incorrect inlet turbulence conditions.
- Surface roughness assumptions.
- Wind tunnel blockage effects.
- Experimental measurement uncertainty.
Validation investigates whether the model predictions agree with relevant experimental observations within an acceptable level of uncertainty.
2.3 Verification vs Validation
| Aspect | Verification | Validation |
|---|---|---|
| Main question | Are we solving the equations correctly? | Are the equations representative of reality? |
| Focus | Numerical accuracy | Physical accuracy |
| Typical comparison | Analytical solutions, manufactured solutions, numerical benchmarks | Experimental measurements |
| Main concerns | Mesh, time step, discretisation, solver errors | Physical models, assumptions, experimental uncertainty |
| Example | Checking a pipe-flow velocity profile against Hagen–Poiseuille theory | Comparing predicted pipe pressure drop with measured pressure drop |
A useful distinction is:
Verification concerns the numerical solution of the mathematical model. Validation concerns the physical model’s ability to reproduce reality.
2.4 The relationship between verification and validation
The two activities complement one another.
A numerical solution should first be demonstrated to be sufficiently accurate for the mathematical model being solved. The model can then be compared with experimental evidence.
- Physical Engineering Problem
- Mathematical Model
- Numerical Implementation
- Verification
- Verified Numerical Solution
- Validation Against Experiments
- Assessment of Model Accuracy
- Engineering Application
Verification and validation are not one-time activities. They should be repeated whenever important aspects of the simulation change.
For example, changing the mesh, turbulence model, boundary conditions, or physical assumptions may require additional verification or validation work.
3. Understanding Different Sources of CFD Error
Before discussing verification techniques, it is useful to understand where errors originate.
A CFD prediction is affected by several different sources of uncertainty and error.
3.1 Discretisation Error
The governing equations describe continuous physical quantities.
However, a computer cannot directly solve continuous differential equations over an infinitely detailed domain.
Instead, the computational domain is divided into a finite number of cells.
This process is called discretisation.
For example, consider a one-dimensional temperature distribution along a heated rod.
The exact temperature field is continuous:
A numerical simulation calculates temperature at a finite number of computational locations.
The approximation introduces discretisation error.
The error generally depends on:
- Mesh spacing.
- Numerical interpolation.
- Spatial discretisation schemes.
- Time-step size.
- Numerical order of accuracy.
3.2 Iterative Error
Many CFD algorithms solve nonlinear systems through repeated iterations.
The solution is updated until the difference between successive iterations becomes sufficiently small.
However, stopping the calculation too early leaves an iterative error.
For example, a pressure-based solver may stop before pressure and velocity are sufficiently consistent with the continuity equation.
3.3 Modelling Error
The physical equations and models used in CFD are approximations of reality.
Examples include:
- Reynolds-averaged turbulence models.
- Wall functions.
- Combustion models.
- Radiation models.
- Porous-media models.
- Simplified material-property models.
A turbulence model such as the standard model does not resolve every turbulent fluctuation. It approximates their average effects.
Even a numerically accurate solution can therefore contain modelling error.
3.4 Input and Boundary-Condition Uncertainty
Simulation results depend on the values supplied by the engineer.
Examples include:
- Inlet velocity.
- Fluid density.
- Dynamic viscosity.
- Thermal conductivity.
- Wall temperature.
- Surface roughness.
- Turbulence intensity.
If these values are inaccurate, the resulting CFD prediction may also be inaccurate.
3.5 Experimental Uncertainty
Experimental data are not perfectly accurate either.
Measurements can be affected by:
- Instrument calibration.
- Sensor accuracy.
- Measurement resolution.
- Environmental conditions.
- Experimental repeatability.
- Data-processing procedures.
Consequently, CFD results should not be expected to match experimental measurements exactly.
A meaningful comparison must consider the uncertainty associated with both numerical and experimental results.
Summary of major CFD error sources
| Error source | Physical or numerical meaning | Example |
|---|---|---|
| Discretisation | Approximation of continuous equations | Coarse mesh |
| Iterative | Incomplete convergence | Residuals remain high |
| Modelling | Simplification of physical behaviour | Turbulence-model limitations |
| Input | Uncertainty in supplied parameters | Incorrect viscosity |
| Boundary conditions | Imperfect representation of surroundings | Unrealistic inlet profile |
| Experimental | Measurement uncertainty | Pressure sensor accuracy |
One of the most important principles in CFD is that reducing one source of error does not automatically eliminate the others.
A very fine mesh cannot compensate for an inappropriate physical model.
4. Numerical Convergence: What Does It Actually Mean?
Numerical convergence is often one of the first things engineers examine when evaluating a CFD simulation.
Most CFD solvers report residuals during the solution process.
But what do residuals actually represent?
4.1 Understanding residuals
Consider a general algebraic equation:
where:
- is the coefficient matrix.
- is the unknown solution vector.
- is the source vector.
The exact solution satisfies:
During an iterative calculation, the current approximation may not satisfy the equation exactly.
The residual is defined as:
Here, represents the remaining imbalance in the algebraic equation.
If the residual is large, the current solution does not satisfy the discretised equation particularly well.
As the iterative solver progresses, the residual should generally decrease.
4.2 Absolute and relative residuals
An absolute residual represents the magnitude of the algebraic imbalance.
A relative residual compares the current residual with a reference residual.
A simplified expression is:
where:
- is the current residual.
- is the initial or reference residual.
- represents an appropriate norm.
For example, if the initial residual is 1 and the final residual is , the residual has decreased by five orders of magnitude.
This is generally a positive indication of iterative convergence.
However, it does not prove that the physical solution is accurate.
4.3 Why residual reduction is not sufficient
Imagine simulating airflow through a duct.
The solver reports that the residuals have decreased by several orders of magnitude.
However, the pressure drop across the duct is still changing with each iteration.
This means the solution may not have reached a sufficiently stable state for the quantity of interest.
For this reason, CFD engineers should monitor both:
- Numerical residuals.
- Important physical quantities.
Examples of physical quantities to monitor include:
- Pressure drop.
- Mass flow rate.
- Drag coefficient.
- Lift coefficient.
- Average temperature.
- Heat transfer rate.
- Velocity at a selected location.
- Turbulent kinetic energy.
4.4 Residual convergence and physical convergence
Consider a steady-state simulation.
The residuals may decrease rapidly during the initial iterations and then approach a plateau.
At the same time, the pressure drop may become nearly constant.
This combination provides stronger evidence of convergence than residual reduction alone.
For transient simulations, the situation is different.
The physical solution is expected to change with time. Therefore, the objective is not to make the physical quantities constant across time steps.
Instead, each time step should be sufficiently converged, and the resulting time-dependent behaviour should be physically meaningful.
4.5 Practical convergence checklist
Before accepting a steady-state CFD solution, examine:
- Have the residuals decreased sufficiently?
- Have the monitored physical quantities stabilised?
- Is the mass imbalance acceptably small?
- Are the pressure and velocity fields physically reasonable?
- Are there unexpected oscillations or divergence?
- Is the final result insensitive to tighter solver tolerances?
There is no universal residual threshold that guarantees accuracy for every CFD problem.
A residual of may be sufficient for one engineering application, while another may require tighter convergence.
The required convergence level depends on the problem, the solver, the quantity of interest, and the accuracy requirements.
5. Conservation Checks: Mass, Momentum and Energy
Conservation laws are the foundation of fluid mechanics and heat transfer.
A physically meaningful CFD solution should satisfy the relevant conservation equations.
Three important conservation principles are:
- Conservation of mass.
- Conservation of momentum.
- Conservation of energy.
5.1 Conservation of mass
For an incompressible fluid, the continuity equation is:
Here:
- is the spatial differential operator.
- is the velocity vector.
- The divergence operator measures the net outward flow from a small volume.
Physically, this means that an incompressible fluid cannot continuously accumulate inside a control volume.
For a simple duct:
where is the mass flow rate.
The mass flow rate is:
where:
- is density.
- is cross-sectional area.
- is the average normal velocity.
Suppose water enters a pipe at a mass flow rate of 2 kg/s.
For a steady incompressible flow, approximately 2 kg/s should leave the pipe.
If the outlet mass flow rate is 1.8 kg/s, the simulation has a mass imbalance of 0.2 kg/s.
This should be investigated before accepting the results.
5.2 Mass imbalance calculation
A useful relative mass imbalance is:
For example:
Therefore:
The relative mass imbalance is 1%.
Whether this is acceptable depends on the application and the accuracy requirements.
5.3 Conservation of momentum
The momentum equation describes how fluid motion changes under the influence of forces.
For an incompressible Newtonian fluid with constant viscosity:
The terms represent:
- Local acceleration.
- Convective acceleration.
- Pressure forces.
- Viscous forces.
- Body forces such as gravity.
A numerical solution should represent the balance between these contributions.
For example, in fully developed laminar pipe flow, the pressure gradient drives the fluid while viscous stresses resist its motion.
The pressure force and viscous resistance balance each other.
5.4 Conservation of energy
For a simplified heat-transfer problem with constant properties and no internal heat generation:
where:
- is specific heat capacity.
- is temperature.
- is thermal conductivity.
For a steady system, the energy entering a control volume should balance the energy leaving it, together with any heat added or removed.
For example, consider a heated pipe.
The energy balance should account for:
- Energy transported by the incoming fluid.
- Energy transported by the outgoing fluid.
- Heat supplied through the pipe wall.
- Heat losses to the surroundings, if modelled.
An energy imbalance may indicate an issue with boundary conditions, numerical convergence, or the implementation of the energy equation.
5.5 Why conservation checks matter
A simulation may show visually smooth contours while still containing a significant conservation error.
For engineering applications, this can lead to incorrect predictions of:
- Pressure drop.
- Heat transfer.
- Pumping power.
- Flow distribution.
- Thermal performance.
Conservation checks should therefore be included in every serious CFD verification workflow.
6. Mesh Independence Study: Does the Grid Affect the Answer?
One of the most important parts of CFD verification is assessing the influence of mesh resolution.
A mesh divides the computational domain into smaller control volumes.
The governing equations are then solved over these discrete cells.
A coarse mesh contains fewer cells and generally requires less computational time.
A fine mesh contains more cells and can represent smaller spatial variations.
However, a finer mesh does not automatically guarantee a better physical model or a more accurate result.
6.1 Why does mesh resolution matter?
Consider fluid flowing through a circular pipe.
The velocity is zero at the wall because of the no-slip condition.
As we move away from the wall, the velocity increases until it reaches a maximum near the centreline.
This creates a velocity gradient near the wall.
A coarse mesh may not capture this gradient accurately.
A finer mesh provides more spatial resolution and can represent the velocity profile more accurately.
The same principle applies to:
- Boundary layers.
- Separation regions.
- Vortices.
- Shock waves.
- Thermal gradients.
- Mixing layers.
- Shear layers.
6.2 Coarse, medium and fine meshes
A common approach is to create at least three systematically refined meshes.
For example:
| Mesh | Number of cells | Relative resolution |
|---|---|---|
| Coarse | 20,000 | Low |
| Medium | 80,000 | Medium |
| Fine | 320,000 | High |
The mesh should be refined in a controlled way.
Ideally, the geometry, boundary conditions, physical models, and numerical schemes should remain unchanged.
Only the mesh resolution should change.
6.3 Mesh refinement ratio
For a three-dimensional mesh, an approximate characteristic cell size can be defined as:
where:
- is the total computational volume.
- is the number of cells.
- is a characteristic grid spacing.
For a uniform refinement study, the refinement ratio is:
For example, if the characteristic mesh spacing is reduced by a factor of two:
In three dimensions, reducing the characteristic cell size by a factor of two can increase the number of cells by approximately a factor of eight.
This is why fine-mesh simulations can become computationally expensive very quickly.
For non-uniform meshes, cell count alone is not sufficient to describe mesh quality or refinement. Important local regions should also be examined.
6.4 What should we compare?
Do not compare only the total number of cells.
Instead, compare engineering quantities that are important to the problem.
Examples include:
| CFD application | Quantity to monitor |
|---|---|
| External aerodynamics | Drag and lift coefficients |
| Internal flow | Pressure drop |
| Heat transfer | Average heat transfer coefficient |
| Electronics cooling | Maximum component temperature |
| Turbomachinery | Pressure ratio and efficiency |
| Pipe flow | Centreline velocity and friction factor |
| Combustion | Flame temperature and species concentration |
For example, suppose you are simulating airflow through a duct.
You may obtain the following pressure-drop values:
| Mesh | Cells | Pressure drop |
|---|---|---|
| Coarse | 20,000 | 145 Pa |
| Medium | 80,000 | 153 Pa |
| Fine | 320,000 | 155 Pa |
The pressure drop changes significantly between the coarse and medium meshes, but only slightly between the medium and fine meshes.
This suggests that the result is becoming less sensitive to mesh resolution.
The percentage difference between two solutions can be calculated as:
Using the medium and fine mesh values:
The pressure drop differs by approximately 1.29%.
This may be acceptable for some engineering applications, but not for others.
The acceptable difference must be defined according to the engineering requirements.
6.5 Visualising mesh independence
- Coarse Mesh
- Medium Mesh
- Fine Mesh
- Compare Quantity of Interest
6.6 Important considerations
A mesh independence study should not simply increase the cell count everywhere.
Refinement should be physically meaningful.
For example:
- Refine near walls to resolve boundary layers.
- Refine around sharp edges where gradients are large.
- Refine near separation and recirculation regions.
- Refine in regions with strong thermal gradients.
- Refine around jets, wakes, and shear layers.
Mesh quality should also be monitored.
Important mesh-quality indicators include:
- Non-orthogonality.
- Skewness.
- Aspect ratio.
- Cell-volume variation.
- Face alignment.
- Boundary-layer resolution.
A mesh with many cells but poor cell quality can still produce inaccurate or unstable results.
6.7 Mesh independence does not mean exact accuracy
Suppose three meshes produce almost identical pressure drops.
This suggests that the numerical result is becoming insensitive to further mesh refinement.
However, all three meshes may still use an inappropriate turbulence model or unrealistic boundary conditions.
Mesh independence is evidence of spatial numerical convergence, not proof that the simulation represents reality.
7. Richardson Extrapolation and Grid Convergence Index (GCI)
Comparing results from three meshes provides useful information, but a more systematic method can estimate the remaining discretisation error.
Two important concepts are:
- Richardson extrapolation.
- Grid Convergence Index (GCI).
These methods are commonly associated with systematic grid-refinement studies.
7.1 Richardson extrapolation
Richardson extrapolation estimates the solution that would theoretically be obtained as the grid spacing approaches zero.
This is called the grid-independent or zero-grid-spacing solution.
For a numerical quantity , the error can often be approximated as:
where:
- is the numerical result on a mesh of characteristic spacing .
- is the extrapolated solution as .
- is an error coefficient.
- is the observed order of convergence.
As the grid spacing becomes smaller, the discretisation error is expected to decrease when the solution is in the appropriate convergence regime.
7.2 Observed order of convergence
For three systematically refined meshes with a constant refinement ratio , define:
- : Fine-grid solution.
- : Medium-grid solution.
- : Coarse-grid solution.
For a monotonic, asymptotic convergence pattern:
This expression assumes that the mesh refinement ratio is the same between successive grids and that the solutions exhibit the expected convergence behaviour.
7.3 A practical numerical example
Suppose a CFD simulation produces the following pressure-drop results:
| Grid | Pressure drop |
|---|---|
| Fine | 155.0 Pa |
| Medium | 153.0 Pa |
| Coarse | 145.0 Pa |
Assume the characteristic grid spacing is refined by a factor of two:
The observed order is:
The observed order is approximately 2.
This is consistent with second-order convergence for this illustrative dataset.
However, this conclusion should only be accepted when the grids are systematically refined and the solutions are within the appropriate asymptotic convergence regime.
7.4 Extrapolated solution
For three systematically refined grids with equal refinement ratio, the extrapolated solution can be estimated using:
Substituting the values:
The estimated zero-grid-spacing pressure drop is approximately 155.67 Pa.
This is an extrapolation, not an exact physical truth.
7.5 Grid Convergence Index
The Grid Convergence Index provides a systematic estimate of grid-related numerical uncertainty.
For the fine grid, a commonly used expression is:
where:
- is a safety factor.
- is the fine-grid result.
- is the medium-grid result.
- is the refinement ratio.
- is the observed order of convergence.
A commonly used safety factor for a three-grid study is:
Using the illustrative pressure-drop values:
The estimated fine-grid discretisation uncertainty is approximately 0.54% under the assumptions of the method.
7.6 Important limitations of GCI
GCI should not be treated as a universal accuracy indicator.
Its application requires careful attention to:
- Systematic grid refinement.
- Appropriate characteristic mesh sizes.
- Consistent numerical methods.
- A meaningful quantity of interest.
- Monotonic or otherwise properly analysed convergence behaviour.
- The asymptotic range of convergence.
If the solutions oscillate between meshes, the simple monotonic formula may not be appropriate.
For strongly non-uniform meshes, complex geometries, or different local refinement strategies, defining a representative mesh spacing requires additional care.
GCI estimates discretisation uncertainty. It does not include all modelling, input, or experimental uncertainties.
8. Time-Step Independence: Is the Simulation Resolving Time Correctly?
Mesh independence addresses spatial resolution.
However, transient CFD simulations must also resolve temporal behaviour.
A simulation can use a very fine mesh and still produce inaccurate results if the time step is too large.
8.1 Why does time-step size matter?
Consider fluid flowing past a circular cylinder.
The flow separates from the cylinder surface and forms vortices downstream.
These vortices are shed periodically.
This phenomenon is known as vortex shedding.
The flow is inherently time-dependent.
If the time step is too large, the simulation may fail to capture the correct vortex-shedding frequency and amplitude.
The resulting velocity and force histories may therefore be inaccurate.
8.2 Temporal discretisation
A time-dependent quantity can be approximated using discrete time levels.
For example:
where:
- is a physical quantity.
- represents the current time level.
- represents the next time level.
- is the time-step size.
The smaller the time step, the finer the temporal resolution.
However, reducing the time step increases computational cost.
8.3 Courant number
The Courant number is an important parameter in transient CFD.
It describes how far information is transported across a computational cell during one time step.
A simplified expression is:
where:
- is the local velocity.
- is the time step.
- is the characteristic cell size.
For example, consider:
Then:
The Courant number is 0.2.
This indicates that the fluid travels approximately 20% of the characteristic cell length during one time step.
For multidimensional and non-uniform meshes, the local Courant number depends on the fluxes through the cell faces rather than only one velocity and one cell dimension.
8.4 Does a low Courant number guarantee accuracy?
No.
A low Courant number can be helpful for numerical stability and temporal resolution, but it does not automatically guarantee an accurate solution.
Accuracy also depends on:
- Temporal discretisation scheme.
- Spatial discretisation.
- Flow physics.
- Frequency of physical phenomena.
- Mesh resolution.
- Solver tolerances.
The required time step should be established through a time-step sensitivity study.
8.5 Performing a time-step independence study
Consider a transient simulation of vortex shedding behind a cylinder.
Run the same simulation using three different time steps:
| Simulation | Time step |
|---|---|
| Case A | 0.004 s |
| Case B | 0.002 s |
| Case C | 0.001 s |
Keep the mesh, physical models, and other simulation settings unchanged.
Compare important transient quantities:
- Vortex-shedding frequency.
- Drag coefficient.
- Lift coefficient.
- Maximum velocity.
- Pressure fluctuations.
Suppose the results are:
| Time step | Strouhal number |
|---|---|
| 0.004 s | 0.190 |
| 0.002 s | 0.198 |
| 0.001 s | 0.200 |
The Strouhal number approaches a stable value as the time step is reduced.
This suggests that temporal discretisation error is decreasing.
The relative difference between the two finest time steps is:
The difference is 1%.
Whether this is acceptable depends on the required accuracy.
8.6 Important considerations for transient simulations
A good transient CFD study should consider:
- The initial transient period.
- The time required to reach statistically stationary behaviour, if applicable.
- The time step used to resolve the physical process.
- The total simulation duration.
- The number of samples used for statistical averaging.
- The sensitivity of important quantities to time-step size.
For turbulent flows, a sufficiently long sampling period is important.
A short simulation may produce misleading averages because the statistical quantities have not been sampled adequately.
9. Validation Against Analytical Solutions
Analytical solutions provide an important reference for numerical verification.
Before comparing CFD with experiments, it is often useful to test the numerical setup against a physical problem with a known mathematical solution.
One excellent example is fully developed laminar flow through a circular pipe.
9.1 Hagen–Poiseuille flow
For steady, incompressible, fully developed laminar flow through a straight circular pipe, the velocity profile is parabolic.
The analytical velocity distribution is:
where:
- is the axial velocity at radial position .
- is the pressure drop.
- is dynamic viscosity.
- is pipe length.
- is pipe radius.
The maximum velocity occurs at the centreline:
The velocity is zero at the wall:
This is a direct consequence of the no-slip boundary condition.
The average velocity is:
Therefore:
This relationship provides a useful check for a CFD simulation.
9.2 Pressure-drop prediction
The Hagen–Poiseuille equation can also be expressed as:
where is the pipe diameter.
Suppose:
- Pipe diameter:
- Pipe length:
- Dynamic viscosity:
- Average velocity:
The pressure drop is:
A CFD model of the same fully developed laminar flow can be compared against this analytical prediction.
9.3 Reynolds number
The Reynolds number is:
For water with:
the Reynolds number is:
This is not a suitable laminar-flow example.
To obtain a laminar case, reduce the average velocity.
For example:
Then:
This is within the laminar regime for a conventional pipe-flow setup.
The corresponding pressure drop becomes:
9.4 What should be compared?
For the laminar pipe-flow case, compare:
- Centreline velocity.
- Radial velocity profile.
- Average velocity.
- Pressure drop.
- Wall shear stress.
A useful comparison is to plot the CFD velocity profile against the analytical parabolic profile.
If the numerical solution approaches the analytical curve with mesh refinement, this provides evidence of spatial verification.
9.5 Important assumptions
The Hagen–Poiseuille solution assumes:
- Steady flow.
- Incompressible Newtonian fluid.
- Constant viscosity.
- Fully developed flow.
- Laminar conditions.
- Straight circular pipe.
- No-slip walls.
A CFD simulation must satisfy these assumptions sufficiently well before a direct comparison is meaningful.
For example, if the simulation includes a short entrance region, the velocity profile may still be developing.
Comparing that profile directly with the fully developed analytical solution would not be appropriate.
10. Experimental Validation: Comparing CFD with Reality
Analytical solutions are valuable, but most industrial CFD problems are too complex to solve exactly.
For example:
- Airflow around a complete vehicle.
- Cooling inside a data centre.
- Combustion inside an engine.
- Flow through a turbine.
- Heat transfer in an electronic enclosure.
In these situations, experimental data provide an important reference for assessing the physical model.
10.1 What is experimental validation?
Experimental validation compares CFD predictions with measurements obtained from a physical experiment.
Examples include:
| CFD quantity | Experimental measurement |
|---|---|
| Velocity | Particle Image Velocimetry (PIV) |
| Pressure | Pressure transducer |
| Temperature | Thermocouple or infrared camera |
| Flow rate | Flow meter |
| Drag force | Force balance |
| Heat transfer | Heat-flux sensor or calorimetry |
The comparison should be performed at equivalent physical locations and under equivalent operating conditions.
10.2 Example: CFD validation of pipe pressure drop
Suppose an experiment measures the pressure drop across a pipe.
The measured pressure drop is:
The CFD simulation predicts:
The absolute difference is:
The relative difference is:
The CFD prediction differs from the measured pressure drop by 3.2%.
This is useful information, but it does not automatically establish whether the model is acceptable.
We must also consider experimental uncertainty, numerical uncertainty, and the engineering accuracy requirement.
10.3 Comparing complete profiles instead of one value
A single engineering quantity may not reveal all the differences between CFD and experiment.
For example, two simulations may predict nearly identical pressure drops but produce different velocity distributions.
For a more complete validation exercise, compare velocity profiles at multiple locations.
For each measurement location:
where:
- is the predicted velocity.
- is the measured velocity.
The error distribution can reveal whether the model consistently overpredicts or underpredicts the velocity.
10.4 Root Mean Square Error
A commonly used comparison metric is Root Mean Square Error (RMSE):
where is the number of measurement points.
RMSE gives greater weight to larger differences because the errors are squared.
It is useful when comparing predicted and measured profiles.
However, RMSE should not be used alone.
It does not automatically account for experimental uncertainty, and it does not indicate whether the differences are physically important.
10.5 Normalised error
For quantities with different scales, normalised error can be useful.
One possible definition is:
where is an appropriate reference velocity.
The choice of reference value should be stated clearly.
10.6 Validation is not curve fitting
An important mistake is adjusting CFD parameters until the results match one experimental dataset.
For example, an engineer may repeatedly change:
- Turbulence intensity.
- Wall roughness.
- Turbulence model coefficients.
- Inlet velocity profile.
- Numerical settings.
until the predicted pressure drop agrees with an experiment.
This may produce a good match for that particular operating condition, but it does not necessarily demonstrate that the model is generally valid.
A stronger validation process uses relevant experimental data, documents all assumptions, and evaluates whether the model can reproduce additional conditions without arbitrary tuning.
11. Uncertainty Quantification: How Certain Are Our Results?
No numerical simulation or physical experiment is perfectly accurate.
Therefore, a credible CFD assessment should consider uncertainty.
Uncertainty quantification (UQ) examines the uncertainty associated with model inputs, numerical methods, physical models, and measurements.
11.1 Sources of uncertainty
Important sources include:
Numerical uncertainty
- Mesh resolution.
- Time-step size.
- Iterative convergence.
- Numerical schemes.
Input uncertainty
- Fluid density.
- Viscosity.
- Thermal conductivity.
- Boundary conditions.
- Operating conditions.
Model uncertainty
- Turbulence model.
- Wall treatment.
- Combustion model.
- Radiation model.
Experimental uncertainty
- Sensor accuracy.
- Calibration.
- Measurement repeatability.
- Data acquisition.
11.2 Example: Uncertainty in pressure-drop prediction
Suppose the CFD pressure-drop prediction is:
The estimated numerical uncertainty is:
The estimated input uncertainty contribution is:
Assuming the contributions are independent and expressed using compatible uncertainty measures, a simplified combined uncertainty can be estimated by:
This is a simplified uncertainty combination.
In a rigorous uncertainty analysis, the uncertainty measures, probability distributions, sensitivity coefficients, correlations, and coverage factors must be considered.
11.3 Uncertainty does not mean the simulation is useless
Uncertainty is a normal part of engineering analysis.
The objective is not to eliminate all uncertainty.
The objective is to understand, quantify, and communicate it sufficiently well to support engineering decisions.
For example, if two cooling designs differ in predicted maximum temperature by only 0.5 K, but the uncertainty in the prediction is approximately 2 K, the simulation may not provide sufficient evidence to claim that one design is thermally superior.
This is why uncertainty matters when making engineering decisions.
12. OpenFOAM Practical Workflow for CFD Verification and Validation
The principles discussed above can be implemented systematically using OpenFOAM.
Consider a simple laminar pipe-flow simulation.
The objective is to verify the numerical solution against the analytical Hagen–Poiseuille solution.
12.1 Define the physical problem
Consider the following case:
| Parameter | Value |
|---|---|
| Pipe diameter | 0.05 m |
| Pipe length | 2 m |
| Average velocity | 0.01 m/s |
| Density | 1000 kg/m³ |
| Dynamic viscosity | 0.001 Pa·s |
| Reynolds number | 500 |
| Flow regime | Laminar |
The analytical pressure drop is:
The analytical centreline velocity is:
These values provide reference quantities for the CFD simulation.
12.2 Create three meshes
Create three systematically refined meshes.
For example:
| Case | Mesh resolution |
|---|---|
| pipe_coarse | 20 × 100 |
| pipe_medium | 40 × 200 |
| pipe_fine | 80 × 400 |
These are illustrative mesh dimensions for a simplified two-dimensional axisymmetric representation.
The exact mesh design should be appropriate for the geometry and solver.
Keep the following consistent:
- Geometry.
- Boundary conditions.
- Fluid properties.
- Solver.
- Numerical schemes.
- Convergence criteria.
Only change the mesh resolution.
12.3 Select the solver
For an incompressible steady laminar flow, a pressure-based steady solver such as simpleFoam can be used.
The turbulence model should be disabled or configured for laminar flow, depending on the OpenFOAM version and case setup.
The choice of solver must match the physical problem.
12.4 Monitor convergence
Monitor:
- Velocity residual.
- Pressure residual.
- Mass flow rate.
- Pressure drop.
- Centreline velocity.
A converged solution should show stable monitored quantities and sufficiently reduced residuals.
12.5 Extract the velocity profile
Use OpenFOAM’s sampling functionality or post-processing tools to extract the axial velocity across a pipe cross-section.
Compare the numerical profile against:
This form uses the average velocity directly.
For the specified case:
Plot both curves.
The numerical velocity profile should approach the analytical solution as the mesh is refined, provided the setup and numerical treatment are appropriate.
12.6 Compare pressure drop
Calculate the pressure difference between two locations in the fully developed region.
Compare the numerical result against:
Make sure the numerical pressure difference is measured over the same physical pipe length used in the analytical calculation.
12.7 Perform a mesh refinement analysis
Record the pressure drop and centreline velocity for each mesh.
| Mesh | Pressure drop | Centreline velocity |
|---|---|---|
| Coarse | CFD result | CFD result |
| Medium | CFD result | CFD result |
| Fine | CFD result | CFD result |
| Analytical | 0.256 Pa | 0.020 m/s |
Calculate relative differences against the analytical solution.
The results should approach the analytical values as the mesh is refined.
12.8 Document the verification results
A useful project report should include:
- Geometry and physical assumptions.
- Governing equations.
- Analytical reference solution.
- Mesh descriptions.
- Mesh-quality information.
- Solver and numerical schemes.
- Residual histories.
- Mass conservation results.
- Velocity-profile comparison.
- Pressure-drop comparison.
- Mesh sensitivity assessment.
- Conclusions and limitations.
This creates a reproducible verification exercise that can be shared through a GitHub repository.
13. Common Mistakes in CFD Verification and Validation
Even experienced CFD engineers can make mistakes when assessing simulation credibility.
The following are particularly important.
Mistake 1: Assuming low residuals mean accurate results
Low residuals indicate that the discretised equations have been solved to a certain iterative tolerance.
They do not prove that the mesh, physical models, or boundary conditions are appropriate.
Better approach: Monitor residuals, physical quantities, conservation errors, and numerical sensitivity.
Mistake 2: Using only one mesh
A single mesh cannot demonstrate mesh independence.
Better approach: Perform systematic refinement studies using multiple meshes.
Mistake 3: Comparing simulations with different physical assumptions
Changing the mesh, turbulence model, boundary conditions, and numerical schemes simultaneously makes it difficult to identify why the results changed.
Better approach: Change one major factor at a time when investigating numerical sensitivity.
Mistake 4: Refining the mesh without considering mesh quality
More cells do not automatically mean better numerical accuracy.
Poorly shaped cells can introduce numerical errors and convergence difficulties.
Better approach: Evaluate mesh quality and refine regions according to physical gradients.
Mistake 5: Ignoring time-step sensitivity
A transient simulation may produce incorrect frequencies, amplitudes, or phase relationships if the time step is too large.
Better approach: Perform a time-step independence study for important transient quantities.
Mistake 6: Comparing CFD results with experiments without considering uncertainty
Experimental measurements contain uncertainty.
A small difference between CFD and experiment may be insignificant, while a seemingly small percentage difference may be important for a critical engineering application.
Better approach: Consider experimental uncertainty, numerical uncertainty, and engineering requirements.
Mistake 7: Treating GCI as a universal accuracy percentage
GCI estimates discretisation uncertainty under specific assumptions.
It does not quantify every source of uncertainty in the simulation.
Better approach: State the assumptions and limitations of the grid-convergence analysis.
Mistake 8: Validating against only one quantity
Matching a single pressure-drop value does not guarantee that the complete velocity and pressure fields are correct.
Better approach: Compare multiple relevant quantities and, where possible, spatial profiles.
14. A Practical Checklist Before Trusting a CFD Simulation
Before using CFD results to support an engineering decision, consider the following checklist.
Physical model
- Is the selected governing equation appropriate?
- Are the fluid properties realistic?
- Is the flow regime correctly identified?
- Is the turbulence model appropriate, if required?
- Are the boundary conditions physically justified?
Numerical setup
- Is the mesh suitable for the physical gradients?
- Are mesh-quality indicators acceptable?
- Are numerical schemes appropriate?
- Are solver tolerances sufficiently tight?
- Is the time step appropriate for transient simulations?
Verification
- Have residuals decreased sufficiently?
- Are important physical quantities stable?
- Is mass conservation acceptable?
- Is energy conservation acceptable where relevant?
- Has mesh sensitivity been assessed?
- Has time-step sensitivity been assessed for transient cases?
- Has the solution been compared with an analytical or numerical benchmark where possible?
Validation
- Are suitable experimental data available?
- Are the experimental conditions comparable to the CFD setup?
- Are relevant quantities compared at equivalent locations?
- Has experimental uncertainty been considered?
- Are model limitations documented?
Reporting
- Are assumptions clearly stated?
- Are mesh and time-step studies documented?
- Are numerical and experimental comparisons presented?
- Are uncertainties discussed?
- Are conclusions supported by evidence?
15. Final Thoughts: Trust in CFD Must Be Earned
CFD is a powerful engineering tool, but a simulation result should never be accepted simply because the solver finishes successfully or produces visually convincing contours.
A reliable CFD analysis requires a systematic understanding of both numerical accuracy and physical accuracy.
Verification helps establish that the numerical methods are solving the selected mathematical equations correctly.
Validation helps determine whether the mathematical models represent the real physical system sufficiently well.
Mesh independence, time-step sensitivity, residual convergence, conservation checks, analytical comparisons, experimental validation, and uncertainty quantification all contribute to building confidence in CFD predictions.
For engineers, this is particularly important because simulation results often influence decisions involving:
- Equipment design.
- Thermal management.
- Energy efficiency.
- Aerodynamic performance.
- Structural and thermal safety.
- Industrial process optimisation.
The ultimate objective is not simply to obtain a converged solution.
It is to produce a numerical prediction whose accuracy, limitations, and uncertainty are understood.
A good CFD engineer does not just ask whether the simulation converged. A good CFD engineer asks whether the result can be trusted.
References and Further Learning
The following resources provide additional technical information on CFD verification, validation, numerical uncertainty, and experimental comparison.
1. NASA – Verification and Validation
NASA’s CFD resources provide useful explanations of numerical verification, validation, and best practices for computational simulations.
-
NASA – CFD Verification and Validation:
https://www.grc.nasa.gov/www/wind/valid/tutorial/tutorial.html -
NASA Turbulence Modeling Resource:
https://tmbwg.github.io/turbmodels/
2. Grid Convergence Index
A foundational reference for systematic grid-refinement studies is:
-
Roache, P. J. (1994). Perspective: A Method for Uniform Reporting of Grid Refinement Studies. Journal of Fluids Engineering, 116(3), 405–413.
https://doi.org/10.1115/1.2910291 -
Celik, I. B., Ghia, U., Roache, P. J., Freitas, C. J., Coleman, H., & Raad, P. E. (2008). Procedure for Estimation and Reporting of Uncertainty Due to Discretization in CFD Applications. Journal of Fluids Engineering, 130(7).
https://doi.org/10.1115/1.2960953
3. OpenFOAM Documentation
Official documentation for understanding solver settings, numerical schemes, mesh generation, and post-processing.
-
OpenFOAM Documentation:
https://doc.openfoam.com/ -
OpenFOAM User Guide:
https://www.openfoam.com/documentation/user-guide/ -
OpenFOAM Tutorials:
https://www.openfoam.com/documentation/tutorial-guide/
This article is intended as an educational introduction to CFD verification and validation. The numerical examples are illustrative. For engineering applications, select verification procedures, uncertainty methods, and acceptance criteria appropriate to the governing physics, numerical method, and intended use of the simulation.