340
11. Efficiency and Accuracy Improvement
error estimation, and optimization should be related to the desired level of
accuracy of the solutions. If only qualitative features of the flow are sought,
a quick job may be acceptable, but for a quantitatively accurate results at
reasonable cost, high grid quality is required.
Comparison with experimental data requires that the experimental uncertainty be known. It is best to compare only fully converged results (ones
from which iteration and discretization error have been removed) with experimental data, because this is the only way that the effect of a model can be
assessed. Experimental uncertainty bars usually extend on both sides of the
reported value. If discretization errors are not small compared to the experimental uncertainty, nothing can be learned about the value of the models
used. Estimation of modeling errors is the most difficult task in CFD.
In many cases the exact boundary conditions are not known and one
has to make assumptions. Examples are far-field conditions for flows around
bodies and inlet turbulence properties. In such a case, it is essential to vary
the critical parameter (location of far-field boundary, turbulence quantities)
over a substantial range to estimate the sensitivity of the solution to this
factor. Often it is possible to obtain good agreement for reasonable values of
the parameters but, unless the experimental data provide them, this amounts
to little more than sophisticated curve fitting. That is why it is essential to
choose experimental data that provide all of the necessary quantities and
to discuss the importance of taking such data with the people who make
the measurements. Some turbulence models are very sensitive to inlet and
free-stream turbulence levels, leading to substantial changes in results for
relatively minor variation in the parameters.
If a number of variants of the same geometry are to be studied, one can
often rely on a validation performed for a typical representative case. It is
reasonable to assume that the same grid resolution and the same model will
produce discretization and modeling errors of the same order as those in the
test case. While this is true in many cases, it may not always be so and
care is needed. Changes in geometry may lead to the appearance of new flow
phenomena (separation, secondary flow, instability etc.), which the model
used may not capture. Thus, the modeling error may increase dramatically
from one case to another although the change in geometry may be minor (e.g.
the computation of flow around an engine valve may be accurate within 3 %
for one valve opening and qualitatively wrong for a slightly smaller opening;
see Lilek a t al., 1991, for a more detailed description).
General Suggestions. The definition of rigid rules for validation of CFD
codes and results is both difficult and impractical. While use of Richardson
extrapolation is recommended wherever possible for estimating discretization
errors, it may be difficult to obtain conclusive answers on all questions (e.g.
the order may turn out to be different for different quantities). When several
types of models are employed (for turbulence, two-phase flow, free-surface
effects etc.) it may be difficult to separate different effects from one another.
11. Efficiency and Accuracy Improvement
error estimation, and optimization should be related to the desired level of
accuracy of the solutions. If only qualitative features of the flow are sought,
a quick job may be acceptable, but for a quantitatively accurate results at
reasonable cost, high grid quality is required.
Comparison with experimental data requires that the experimental uncertainty be known. It is best to compare only fully converged results (ones
from which iteration and discretization error have been removed) with experimental data, because this is the only way that the effect of a model can be
assessed. Experimental uncertainty bars usually extend on both sides of the
reported value. If discretization errors are not small compared to the experimental uncertainty, nothing can be learned about the value of the models
used. Estimation of modeling errors is the most difficult task in CFD.
In many cases the exact boundary conditions are not known and one
has to make assumptions. Examples are far-field conditions for flows around
bodies and inlet turbulence properties. In such a case, it is essential to vary
the critical parameter (location of far-field boundary, turbulence quantities)
over a substantial range to estimate the sensitivity of the solution to this
factor. Often it is possible to obtain good agreement for reasonable values of
the parameters but, unless the experimental data provide them, this amounts
to little more than sophisticated curve fitting. That is why it is essential to
choose experimental data that provide all of the necessary quantities and
to discuss the importance of taking such data with the people who make
the measurements. Some turbulence models are very sensitive to inlet and
free-stream turbulence levels, leading to substantial changes in results for
relatively minor variation in the parameters.
If a number of variants of the same geometry are to be studied, one can
often rely on a validation performed for a typical representative case. It is
reasonable to assume that the same grid resolution and the same model will
produce discretization and modeling errors of the same order as those in the
test case. While this is true in many cases, it may not always be so and
care is needed. Changes in geometry may lead to the appearance of new flow
phenomena (separation, secondary flow, instability etc.), which the model
used may not capture. Thus, the modeling error may increase dramatically
from one case to another although the change in geometry may be minor (e.g.
the computation of flow around an engine valve may be accurate within 3 %
for one valve opening and qualitatively wrong for a slightly smaller opening;
see Lilek a t al., 1991, for a more detailed description).
General Suggestions. The definition of rigid rules for validation of CFD
codes and results is both difficult and impractical. While use of Richardson
extrapolation is recommended wherever possible for estimating discretization
errors, it may be difficult to obtain conclusive answers on all questions (e.g.
the order may turn out to be different for different quantities). When several
types of models are employed (for turbulence, two-phase flow, free-surface
effects etc.) it may be difficult to separate different effects from one another.