The Grid Convergence Index (GCI) is a standardized method to estimate CFD simulation error. It is also used to determine whether the solution is mesh independent. This method is formalized in ASME V&V 20.

Assumption

For a continuous and infinitely differentiable function $\phi(x)$, a Taylor series expansion is,

\[\phi(x+h) = \phi(x) + \frac{\partial \phi}{\partial x} h + \frac{\partial^2 \phi}{\partial x^2} \frac{h^2}{2!}+...+\frac{\partial^p \phi}{\partial x^p} \frac{h^p}{p!}+0(h^{p+1})\]

It is assumed that the CFD solution $\phi$ (a scalar solution such as pressure drop) can be approximated by Taylor series expansion.

\[\phi(h) = \phi + A_1 h + A_2 h^2+...+A_p h^p+0(h^{p+1})\]

Or, in general,

\[\phi(h) = \phi + C \cdot h^p+0(h^{p+1})\]

$\phi$ is the theoritical solution. It is an approximation when $h \rightarrow 0$.

Richardson Extrapolation

Using two solutions $\phi_1$, $\phi_2$ with two different grid size $h_1$, $h_2$,

\[\phi_1 \approx \phi + A_p {h_1}^p\] \[\phi_2 \approx \phi + A_p {h_2}^p\]

By adjusting the terms,

\[\frac{\phi_2 - \phi}{\phi_1 - \phi}=(\frac{h_2}{h_1})^p=r^p\]

where, \(r=\frac{h_2}{h_1}\)

We then got the Richardson extrapolation.

\[\phi = \phi_1 + \frac{\phi_1 - \phi_2}{r^p-1}\]

Estimated Error

From Richardson extrapolation, the estimated discretization error of the fine grid solution $\phi_1$, because $\phi_1 \approx \phi$,

\[E = \frac{\phi-\phi_1}{\phi} \approx \frac{\phi_1-\phi_2}{\phi_1 (r^p-1)}\]

Roache Approach

The Roache approach uses solutions $(\phi)$ of three meshes (fine, medium and coarse). Again, the solution can be pressure drop, drag coefficient $C_D$, lift coefficient $C_L$ etc.

1. Representative Cell Size $(h)$

The average representative grid size $h$ is estimated from total volume $V$ and cell count $N$.

\[h \approx (\frac{V}{N})^{1/3}\]

2. Refinement Ratio $(r)$

The grid refinement ratio,

\[r_{21}=\frac{h_2}{h_1}, \quad r_{32}=\frac{h_3}{h_2}\]

ASME V&V 20 requires the refinement ratio to be larger than 1.3. For simplicit, it can be set constant: $r_{21}=r_{32}=r$

3. Solution Differences $(\epsilon)$

The solution differences are,

\[\epsilon_{21}=\phi_2 - \phi_1, \quad \epsilon_{32}=\phi_3 - \phi_2\]

4. Order of Accuracy $(p)$

From the Taylor series approximation,

\[\phi_1 = \phi + C h_1^p\] \[\phi_2 = \phi + C h_2^p\] \[\phi_3 = \phi + C h_3^p\]

The errors,

\[\epsilon_{21} = \phi_2 - \phi_1 = C(h_2^p-h_1^p)=Ch_1^p (r_{21}^p-1)\] \[\epsilon_{32} = \phi_3 - \phi_2 = C(h_3^p-h_2^p)=Ch_2^p (r_{32}^p-1)\]

This leads to,

\[\frac{\epsilon_{32}}{\epsilon_{21}} = r_{21}^p \frac{r_{32}^p-1}{r_{21}^p-1}=r^p\]

The apparent order of accuracy is estimated,

\[p = \frac{\ln |\epsilon_{32} / \epsilon_{21}|}{\ln(r)}\]

5. Estimated Error

Calculate the estimated error using solutions of the fine grid and the medium grid.

\[E_{21} = \frac{\phi_1-\phi_2}{\phi_1 (r^p-1)}\]

6. Grid Convergence Index (GCI)

The grid convergence index (GCI) is defined,

\[GCI = F_s \cdot E_{21} \times 100 \%\]

Where $F_s$ is the safety factor. $F_s=1.25$ when using 3 or more grid. $F_s=3.0$ when using only 2 grids. For an engineering simulation, a criterion of $GCI < 1 \%$ is usually used.

7. Metrics

Convergence Ratio $(R)$

\[R=\frac{\epsilon_{21}}{\epsilon_{32}}\]
  • $0 < R < 1$: monotonic convergence
  • $-1 < R < 0$: Oscillatory convergence
  • $R > 1$ or $R<-1$: Divergence

Asymptotic Ratio $(AR)$

\[AR = \frac{GCI_{32}}{r^p\cdot GCI_{21}}\]

When asymptotic ratio is close to 1.0, the grid refinement is within the asymtotic range.

CFD Result Error

The CFD result error can be estimated with the formula below,

\[\phi = \phi_1 \pm (\phi_1 \cdot GCI_{21})\]

8. Report

It might not be accurate to say the solution is mesh independent. It is about the CFD solution error. It would be appropriate to use the wording β€œthe solution prediction is mesh independent within $1\%$”.

It is also important to state this for a particular quantity of interest. For example, it might be accurate for pressure drop but might not be accurate for the total force.