11.1 Error Analysis and Estimation
337
of the experimental data is essential if they are to be used for validation
purposes.
One should also note that modeling errors differ for different quantities;
for example, computed pressure drag may agree well with the measured value,
but the computed friction drag may be substantially in error. Mean velocity
profiles are sometimes well predicted, while the turbulence quantities may
be under- or over-predicted by a factor of two. It is important to compare
results with a variety of quantities in order to assure that the model really is
accurate.
Detection of Programming and User Errors. A kind of errors that is
difficult to quantify is the programming error. These may be simple "bugs"
(typing errors that do not prevent the code from compiling) or serious algorithmic errors. The analysis of iteration and discretization errors usually
helps the developer find them, but some may be so consistent that they
remain undiscovered for years (if ever), especially when there are no exact
reference solutions to compare.
A critical analysis of results is essential for the discovery of potential
user errors; it is therefore crucial that the user have solid knowledge of fluid
dynamics in general and of the problem to be solved in particular. Even
if the CFD code that is being used has been validated on other flows, the
user can make errors in setting-up the simulation so that the results may
be significantly in error (e.g. due to errors in geometry representation, in
boundary conditions, in flow parameters etc.). User errors may be difficult to
spot (e.g. when an error in scaling is made and the computed flow corresponds
to a different Reynolds number than anticipated); the results should therefore
be critically evaluated, if possible also by someone other than the person who
performed the computation.
11.1.3 Recommended Practice for CFD Uncertainty Analysis
One should distinguish between validation of a newly developed CFD code
(or new features added to an existed code) and validation of an established
code for a particular problem.
Validation of a CFD Code. Any new code or added feature should undergo systematic analysis with the aim of assessing the discretization errors
(both spatial and temporal), of defining convergence criteria in order to assure small iteration errors, and of eliminating as many 'bugs' as possible.
For this purpose one has to select a set of test cases representative of the
range of problems solvable by the code, and for which sufficiently accurate
solutions (analytical or numerical) are available. Since one wants to assure
that the equations are correctly solved for the specified boundary conditions,
experimental data are not the best way to measure the quality of numerical
solutions. Reference solutions are needed to locate errors in the algorithm or
337
of the experimental data is essential if they are to be used for validation
purposes.
One should also note that modeling errors differ for different quantities;
for example, computed pressure drag may agree well with the measured value,
but the computed friction drag may be substantially in error. Mean velocity
profiles are sometimes well predicted, while the turbulence quantities may
be under- or over-predicted by a factor of two. It is important to compare
results with a variety of quantities in order to assure that the model really is
accurate.
Detection of Programming and User Errors. A kind of errors that is
difficult to quantify is the programming error. These may be simple "bugs"
(typing errors that do not prevent the code from compiling) or serious algorithmic errors. The analysis of iteration and discretization errors usually
helps the developer find them, but some may be so consistent that they
remain undiscovered for years (if ever), especially when there are no exact
reference solutions to compare.
A critical analysis of results is essential for the discovery of potential
user errors; it is therefore crucial that the user have solid knowledge of fluid
dynamics in general and of the problem to be solved in particular. Even
if the CFD code that is being used has been validated on other flows, the
user can make errors in setting-up the simulation so that the results may
be significantly in error (e.g. due to errors in geometry representation, in
boundary conditions, in flow parameters etc.). User errors may be difficult to
spot (e.g. when an error in scaling is made and the computed flow corresponds
to a different Reynolds number than anticipated); the results should therefore
be critically evaluated, if possible also by someone other than the person who
performed the computation.
11.1.3 Recommended Practice for CFD Uncertainty Analysis
One should distinguish between validation of a newly developed CFD code
(or new features added to an existed code) and validation of an established
code for a particular problem.
Validation of a CFD Code. Any new code or added feature should undergo systematic analysis with the aim of assessing the discretization errors
(both spatial and temporal), of defining convergence criteria in order to assure small iteration errors, and of eliminating as many 'bugs' as possible.
For this purpose one has to select a set of test cases representative of the
range of problems solvable by the code, and for which sufficiently accurate
solutions (analytical or numerical) are available. Since one wants to assure
that the equations are correctly solved for the specified boundary conditions,
experimental data are not the best way to measure the quality of numerical
solutions. Reference solutions are needed to locate errors in the algorithm or