Home / Courses / Computational Fluid Dynamics / Verification

Computational Fluid Dynamics · Chapter 10 of 10 · Advanced

Errors, Verification, and Validation

A CFD result without an uncertainty is an opinion. This chapter turns the verify-before-trust habit into numbers: the observed order, the extrapolated answer, and a percentage band on the result.

Readiness check

This closing chapter quantifies trust. Tick only what you can do closed-notes.

Recall that a finer mesh is more accurate.
Take a logarithm and a ratio.
Recall the residual as a convergence measure.
Compute a percentage relative difference.
Distinguish a numerical error from a modeling error.

0 or 1 weak itemsContinue with this chapter.

2 weak itemsRevisit residuals and convergence in Chapter 6.

3 or more weak itemsReview the verification habit in Chapter 1.

The core idea

Solutions on successively finer grids converge toward the exact answer at a known rate. From three grids you can find that rate, extrapolate the exact value, and put a percentage uncertainty on the result.

p = ln[(f₃ − f₂)/(f₂ − f₁)]/ln rf_exact ≈ f₁ + (f₁ − f₂)/(r^p − 1)GCI = F_s|ε|/(r^p − 1)

Every CFD result carries error from three sources: discretization (finite mesh and time step), iterative (incomplete convergence), and modeling (turbulence and other physics approximations). Discretization error shrinks as the grid is refined, at a rate set by the scheme's order of accuracy. Running three systematically refined grids reveals the observed order p from the ratio of successive differences, and Richardson extrapolation estimates the value the grid would reach if it were infinitely fine. The grid convergence index then converts the remaining error into a reported percentage band. Verification asks whether the equations are solved correctly (a math question, settled by these tools); validation asks whether they are the right equations (a physics question, settled by comparison with experiment).

The skill works when: you confirm the observed order, extrapolate, and report a GCI band.

The skill breaks down when: a single mesh is trusted, or good agreement with data hides a compensating numerical and modeling error.

The concept. As the mesh is refined (h to the left), the result converges toward the extrapolated exact value. The rate of approach is the observed order, and the gap that remains is the numerical uncertainty.

The skills, taught in order

Five skills name the error sources, define mesh independence, find the order, compute the GCI, and separate verification from validation.

10.1 Sources of error

A CFD result deviates from reality through discretization error (finite grid and time step), iterative error (stopping before full convergence), and modeling error (turbulence and physics approximations). Round-off is usually negligible. Each is reduced differently, so naming the dominant one matters.

Error	Source	Reduced by
Discretization	finite mesh and time step	refining the grid
Iterative	incomplete convergence	tighter residual tolerance
Modeling	turbulence and physics models	better models and validation

10.2 Mesh independence

A solution is mesh-independent when refining the grid no longer changes the result meaningfully. Demonstrating this with a refinement study is the minimum standard before any CFD number is reported; a result that still moves on refinement is not yet converged in space.

10.3 Order of accuracy and Richardson extrapolation

With three grids refined by a ratio r, the observed order is p = ln[(f₃ − f₂)/(f₂ − f₁)]/ln r, and Richardson extrapolation estimates the zero-mesh value f_exact ≈ f₁ + (f₁ − f₂)/(r^p − 1). A p near the scheme's formal order confirms the solution is in the asymptotic range.

10.4 The grid convergence index

The GCI turns the extrapolation into a reported uncertainty: GCI = F_s|ε|/(r^p − 1), with the relative difference ε = (f₂ − f₁)/f₁ and a safety factor F_s (1.25 for three grids). It is the standard way to attach an error bar to a CFD result.

10.5 Verification versus validation

Verification asks whether the equations are solved right, a mathematics question answered by mesh and iterative convergence. Validation asks whether they are the right equations, a physics question answered by comparison with experiment. A model can be verified yet invalid, so both are needed before trusting a result.

Engineering connection: this chapter closes the verify-before-trust loop opened in Chapter 1, turning every earlier method into a defensible, quantified result.

Worked example 1: observed order and Richardson extrapolation

A drag coefficient is computed on three grids refined by a ratio r = 2: fine f₁ = 1.00, medium f₂ = 1.01, coarse f₃ = 1.05. Find the observed order of accuracy and the Richardson-extrapolated value.

Figure 1. The successive differences fall by a factor of four as the grid halves, so the observed order is two. Richardson extrapolation projects the fine and medium values to the zero-mesh limit.

ProblemFind the observed order p and the extrapolated value for the grids in Figure 1.
Given / findr = 2, f₁ = 1.00 (fine), f₂ = 1.01, f₃ = 1.05 (coarse). Find p and f_exact.
AssumptionsSystematic refinement by a constant ratio; the solutions are in the asymptotic range.
ModelObserved order from the ratio of successive differences; Richardson extrapolation for the zero-mesh value.
Equationsp = ln[(f₃ − f₂)/(f₂ − f₁)]/ln rf_exact = f₁ + (f₁ − f₂)/(r^p − 1)
Solve(f₃ − f₂)/(f₂ − f₁) = 0.04/0.01 = 4, so p = ln 4/ln 2 = 2. Then f_exact = 1.00 + (1.00 − 1.01)/(2² − 1) = 1.00 − 0.01/3 = 0.997.
CheckAn observed order of 2 matches a second-order scheme, confirming the grids are in the asymptotic range, so the extrapolation is trustworthy. The exact value sits just below the fine-grid result.
ConclusionThree grids give both the convergence rate and an estimate of the answer the mesh is approaching, the foundation for a quantified uncertainty.

Result. Observed order p = 2; extrapolated value f_exact ≈ 0.997.

Worked example 2: the grid convergence index

Using the same three grids (p = 2, r = 2, fine f₁ = 1.00, medium f₂ = 1.01), compute the grid convergence index for the fine-grid result with a safety factor F_s = 1.25.

Figure 2. The grid convergence index converts the fine-to-medium difference into a reported uncertainty band on the fine-grid result, here about half a percent.

ProblemFind the GCI for the fine-grid result in Figure 2.
Given / findp = 2, r = 2, f₁ = 1.00, f₂ = 1.01, F_s = 1.25. Find the GCI.
AssumptionsSolutions in the asymptotic range; three-grid safety factor 1.25.
ModelRelative difference ε = (f₂ − f₁)/f₁, then GCI = F_s|ε|/(r^p − 1).
Equationsε = (f₂ − f₁)/f₁GCI = F_s|ε|/(r^p − 1)
Solveε = (1.01 − 1.00)/1.00 = 0.01 = 1 percent. GCI = 1.25 × 0.01/(2² − 1) = 0.0125/3 = 0.00417 = 0.42 percent.
CheckThe GCI is smaller than the raw 1 percent fine-to-medium difference because the second-order convergence (r^p − 1 = 3) credits the refinement; the safety factor then guards against over-optimism. The fine-grid drag is 1.00 ± 0.42 percent from discretization.
ConclusionThe GCI is the number that should accompany every reported CFD result. With this band and a validation comparison, the simulation becomes evidence rather than a picture.

Result. GCI ≈ 0.42 percent on the fine-grid result.

Misconceptions and diagnostics

Mistake	Symptom	Diagnostic question	Correction
Trusting one mesh	A result with no uncertainty	"Did the answer change on refinement?"	Run at least three grids and report a GCI.
Confusing verification and validation	Calling a converged run validated	"Right equations, or right solution of them?"	Verification is numerical; validation needs experimental data.
Out of the asymptotic range	Observed order far from formal order	"Is p near the scheme's order?"	Refine further until the order is recovered before extrapolating.
Compensating errors	Good data match on a coarse mesh	"Could numerical and modeling errors be cancelling?"	Verify numerically first, then validate against data.

Practice ladder

Level 1 · Direct skill

Three grids (r = 2) give f₁ = 2.00, f₂ = 2.02, f₃ = 2.10. Find the observed order.

Show answer

(f₃ − f₂)/(f₂ − f₁) = 0.08/0.02 = 4, so p = ln 4/ln 2 = 2.

Level 2 · Mixed concept

For the Worked Example 2 result, what is the GCI if the fine-to-medium difference is instead 2 percent?

Show answer

GCI = 1.25 × 0.02/3 = 0.0083 = 0.83 percent. The band scales linearly with the relative difference.

Level 3 · Independent problem

For f₁ = 0.50, f₂ = 0.52, r = 2, p = 2, F_s = 1.25, find the GCI and the Richardson-extrapolated value.

Show answer

ε = (0.52 − 0.50)/0.50 = 0.04. GCI = 1.25 × 0.04/3 = 0.0167 = 1.67 percent. f_exact = 0.50 + (0.50 − 0.52)/3 = 0.50 − 0.0067 = 0.493.

Level 4 · Transfer to real engineering

For a real CFD study, describe the verification and validation plan you would run before reporting a result to a decision-maker.

What good work looks like

A three-grid refinement study with an observed order and GCI, residual convergence, and a comparison with experimental or benchmark data, with both numerical and modeling uncertainty stated.

Working with AI, and proving it yourself

Use AI as an examiner, not a solver

"Check that my observed order is near the scheme's formal order before I extrapolate."

"Give me three-grid results; I will compute the order and GCI."

"Is my result accurate?" Computing the GCI yourself is the skill.

"Is it validated?" Separating verification from validation is the point.

Portfolio task

Run a three-grid study on a real case: report the observed order, the Richardson-extrapolated value, and the GCI, and compare the result to experimental or benchmark data.

Must include: an observed order, an extrapolated value, a GCI, and a validation comparison.

Retrieval and spaced review

Closed notes. Answer out loud, then reveal.

1. Name the three main error sources.

Discretization, iterative, and modeling error.

2. What is mesh independence?

The result no longer changes meaningfully as the grid is refined.

3. Write the observed order.

p = ln[(f₃ − f₂)/(f₂ − f₁)]/ln r.

4. What does the GCI report?

A percentage uncertainty band on the result: GCI = F_s|ε|/(r^p − 1).

5. Verification versus validation?

Verification: solving the equations right (numerical). Validation: solving the right equations (against experiment).

TodayFinish this quiz and Levels 1 and 2 of the ladder.

+1 dayRe-derive the observed order and GCI from a blank page.

+3 daysRun a three-grid analysis on new data.

+7 daysCombine meshing, modeling, and verification into one study.

+30 daysRevisit the whole course through the CFD hub.

Textbook mapping

Item	Mapping
Primary source	Versteeg and Malalasekera, An Introduction to Computational Fluid Dynamics, Chapter 10 and Appendix A (Errors and Accuracy)
Cross-reference	Roache, Verification and Validation in Computational Science · Experimentation and Measurements
Core topics	10.1 Error sources · 10.2 Mesh independence · 10.3 Order and Richardson · 10.4 GCI · 10.5 Verification vs validation
Engineering connection	Closes the verify-before-trust loop opened in Chapter 1.
Read next	Return to the CFD hub and integrate the workflow.