Conducting &

Reporting CFD
Analyses

1
 Objective

Objective is to achieve a high level of credibility and confidence in the

results from CFD performed as part of the design and
analysis of a propulsion system.

Credibility and confidence are obtained by demonstrating acceptable

levels of error and uncertainty as assessed through verification and
validation.

This discussion is focused on the question…

How good are your CFD results?

Determining this precisely is often difficult, but we can try to understand

sources of error and ways to demonstrate quality CFD results.
2
 Flow Characteristics

The characteristics of flows of interest may include:

• Mach number range (Static,Mach 0 to hypersonic, Mach 25)
• Shock waves.
• High Reynolds numbers (laminar transitioning to turbulent).
• Turbulent, viscous flow.
• Shock / boundary layer interactions.
• Boundary layer separation.
• High-temperature effects at hypersonic Mach numbers.
• Compressor face effects.
• Unsteady flows (hammershocks, unstart)
3
 Objectives of a CFD Analysis

The CFD analysis may provide:

• Steady-state flow at cruise conditions.
• Features of the shock system (positions, cowl stand-off).
• Features of the boundary layers (thickness, separation).
• Pressure recovery at throat and compressor face.
• Spillage (side walls, cowl).
• Bleed requirements (slots, bleed holes).
• Distortion at compressor face.
• Drag.
• Off-design flow features.
• Unsteady flow features (unstart, stability).
• Aerodynamic and thermal loads.

4
 Levels of the Use of CFD Results

1. Provide Qualitative Information.

Provide features of flow field (conceptual, rough estimates
of quantities). Accuracy requirements are low.

2. Provide Incremental Quantities, P.

• P = Pbaseline + P
• Errors partially cancel for P
 P = (P2-P1)actual + (E2-E1) = Pactual + E
• Moderate accuracy requirement.

3. Provide Absolute Quantities, P.

Requires high level of accuracy.
5
 Verification and Validation

Verification
“The process of determining that a model implementation
accurately represents the developer’s conceptual description
of the model and the solution to the model.” (AIAA-G-077-
1998).
Verification checks that there are no programming errors and that
the coding correctly implements the equations and models. It also
examines discretization errors in the CFD calculations.

Validation
“The process of determining the degree to which a model is an
accurate representation of the real world from the perspective
of the intended uses of the model.” (AIAA-G-077-1998).
Validation checks that the CFD results agree with reality as
6
observed in experiments, flight tests, or applications.
 Uncertainty and Error

Uncertainty
“A potential deficiency in any phase or activity of the modeling process
that is due to the lack of knowledge.” (AIAA-G-077-1998).
The use of potential indicates that deficiencies may or may not exist.

Error
“A recognizable deficiency in any phase or activity and simulation that is
not due to lack of knowledge.” (AIAA-G-077-1998).
The deficiencies are knowable up examination. Acknowledged errors have
procedures for identifying and removing them. Unacknowledged
errors are undiscovered and no set procedures exist to find them.

7
 Errors in CFD Results

Total Error = Physical Modeling Error

+ Discretization Error
+ Programming Error
+ Computer Round-off Error
+ Usage Error
+ Post-Processing Error

Physical modeling errors are usually acknowledged and may be quantified.

Discretization errors are grid dependent but a priori knowledge of the

proper grid for a desired level of accuracy is generally lacking.
Programming errors should have been minimized by programmers.
Computer round-off errors are usually negligible or controllable.
Usage errors are minimized by proper training.
Post-processing errors are usually controllable. 8
 Physical Models in CFD

Many physical models are used in CFD and we must understand the
limits and possible errors with each model:
• Spatial Dimensions. (2D, axisymmetric, 3D)
• Temporal Dimensions. (steady-state, unsteady)
• Equations. (Navier-Stokes, approximations)
• Turbulence Model.
• Thermodynamic and Transport Property model.
• Chemistry Model.
• Flow Boundary Conditions (inflow, outflow).
• Compressor Face Model.
• Bleed / Blowing Model.
• Vortex Generator Model.

9
 Discretization Error

Errors that exist because continuum flow equations and

physical models are represented in a discrete domain of
space (grid) and time.

Level of error is function of interactions between the solution

and the grid. Since solution is unknown before-hand, the grid
generation may not produce the optimum grid.

Discretization error is quantified through verification

methods that examine grid convergence.

Discretization error is controllable.

10
 Iterative and Grid Convergence

Convergence used in two ways:

Iterative Convergence
Simulations should demonstrate iterative convergence. As the algorithm
iterates the solution, the simulation results approach a fixed value
(residuals drop and level off). This applies to both steady and unsteady
flow solutions. In the case of unsteady flows, the iterative convergence
applies to iterations over a time step.

Grid Convergence
As the grid spacing is reduced, the simulation results become insensitive
to the grid and approach the continuum results.

11
 Iterative Convergence

Iterative Convergence can be demonstrated in several ways:

1. Value of the largest change in the solution over an iteration, Qbig.
2. Iteration history of the L2 norm of the equation residual.
3. Examine conservative variables.
4. Iteration history of a local or global quantity of the flow.

Usually some measure of all of these are used to determine iterative

convergence.

The largest change in the solution over an iteration is simply


 Qˆ Big max  Qˆ i ,m 
12
 L2 Residual

The L2 norm of Q is commonly called the residual of the equation solution

and is computed as

 
N M 2
 Qˆ i ,m
L2  i 1 m 1
M *N

where N, total number of grid points

M, number of elements in Q (5 for Navier-Stokes Equations)

• It represents the change in the solution over an iteration averaged over all
the grid points and equations.
• The L2 residual is usually displayed and plotted as its log values, [log(L2)].

• This shows the order-of-magnitude of the change in the solution.

• Generally desire the L2 residual to approach zero with iterations.
13
 Residual Plots
The behavior of the
residual is not always Navier-Stokes Equations
so smooth.

Unsteadiness (real or
Order-of-magnitude drop (about 5)
pseudo) in the
solution (typical of
turbulent and
separated flows) may
psuedo-unsteadiness
limit how far the
residual drops.
Leveling off of residual
Often a couple of
orders-of-magnitude
drop is all one
achieves.

14
Residual Plots (continued)

Spalart-Allmaras turbulence model equations

15
 Examine Conservative Variables
The solution of the conservation equations must satisfy those conservative
principles. Iterative convergence can be demonstrated through examination of
the conservative variables:

For example, one can examine the conservation of mass through a duct:
Statistical analysis of the variation. This
Variation of the mass flow indicates the level of iterative
through the duct. convergence error in determining the
mass flow through the duct.

16
 Change in Local or Global Quantity
The change in a local or global property is perhaps the best measure of iterative
convergence since it typically directly monitors the quantity that is important for
the engineering analysis.
Local quantity is point
value. Global quanity is
integrated over a portion of
the flow field.
Examples include:
• Lift on a wing
• Drag on a wing
• Inlet recovery Lift on the ONERA M6 wing
with number of cycles.
• Heat flux

Global quantities usually

show a smooth
convergence. 17
 Verification Assessment

Verification assessment has two aspects:

Verification of a Code
Verifies that code has no programming errors and that the coding
correctly implements the equations and models. The methods
involve examining modules of code, checking basic assumptions
(mass conservation), and comparing results to analytic results.

Verification of a Calculation
Verifies that a calculation (simulation) demonstrates a certain level
of accuracy. The primary method for the verification of a
calculation
is the grid convergence study.
18
 Grid Convergence Study

Quantifies:
1. “ordered” discretization error band (related to grid size by order p)
2. p, order of grid convergence (order-of-accuracy)
3. continuum or “zero grid spacing’’ value of observed quantity
Approach:
• Assumed or demonstrated that all other error terms are negligible,
minimized, accepted, or under control.
• Perform CFD solution on two or more grids of increased refinement.
• Solutions must be in the “asymptotic range of convergence’’.
• If three solutions, can then compute order of convergence, p.
• Use Richardson extrapolation to compute continuum value.
• Compute Grid Convergence Indices (GCI) as discretization error band.

19
 Grid Generation Considerations

• Viscous wall spacing set for y+ < 1.0 for fine grid.
• Flow field resolution set to resolve shock system.
• Grid density to keep grid spacing ratio below 15%.
• Use ICEM CFD to construct and GRIDGEN to improve.
• Subset grid to remove some sub-layer grid points for
wall function.
• Best if grid refinement ratio, r = h / h = 2.0
2 1

• h is grid spacing.
• Minimum, r > 1.1
• WIND allows for grid sequencing (r = 2).

20
 Example: Supersonic Diffuser

• Example grid convergence

study with three grids.
• Mach 2.35.
• r = h2 / h1 = 2.
• Asymptotic range observed
in plot.
• Richardson extrapolation.

21
 Order of Convergence

• Order of grid convergence, p.

• Errors reduce as grid is refined.

E  f h  f exact C h p  H .O.T .
• h is the grid spacing.
• C is a constant.
• Use three values to solve for order of convergence.

 f3  f 2 
p ln  lnr 
 f 2  f1 
• For this example, p = 1.786

22
 Richardson Extrapolation

• Extrapolation of two quantities to continuum value at

“zero grid spacing”.
• Provides estimate of error.
• Extrapolate is of higher order.
f1  f 2
f h 0  f1  p
r 1

• For this example, (p2/p0) = 0.97130

h=0

23
 Grid Convergence Index (CGI)

• Standardized method for reporting discretization error.

• Based on Richardson extrapolation.
• Error band on fine grid solution.
• Expressed as percentage.
Fs  f 2  f1
GCI fine  
r p
 1 f1

• FS is factor of safety (FS = 1.25 if three or more grids)

• For this example,
– GCI12 = 0.1031% (0.0010)
– GCI23 = 0.3562% (0.0035)
• “Pressure recovery is 0.971 with grid error of 0.001.”
24
 Asymptotic Range of Convergence

• Solutions are in range in which errors decrease at rate

denoted by order of convergence, p.
• Check that constant C is indeed constant over solutions,
E
C
hp

• Check with GCI values over three solutions,

GCI 23 r p GCI12

• For this example, ratio = 1.002

25
 Validation Assessment

• Validation assessment addresses if CFD code agrees with

reality through comparison to experiments.
• Experiment must also address its own errors.
• Unit  Benchmark  Subsystem  Complete System.
• Set of validation cases are being collected and studied that
cover the range of flow features present in inlets
(turbulent boundary layers, shocks, shock / boundary layer
interactions, subsonic ducts, supersonic diffusers).
• NPARC Alliance Validation Archive:
www.grc.nasa.gov/www/wind/valid
26
 Reporting CFD Results

In summary, for a CFD analysis, report the following:

• Objective of CFD analysis
• Geometry simplifications
• Grid resolution
• Boundary and initial conditions
• Equations and physical models
• Algorithm settings
• Iterative convergence criteria
• Grid convergence criteria
• Results (values, plots, visualizations)
• Errors that can be quantified
• Sensitivity to models and parameters (turbulence, chemistry)
27
 Further Information on CFD

A sample of books and resources for learning more on CFD:

Anderson, D.A., Tannehill, J.C., and Pletcher, R.H., Computational Fluid

Mechanics and Heat Transfer, McGraw-Hill, 1984.
Anderson, J.D., Computational Fluid Dynamics, McGraw-Hill, 1995.
Hirsch, C., Numerical Computation of Internal and External Flows (2
Volumes), John Wiley & Sons, 1988 & 1990.
Hoffmann, K.A. and S.T. Chiang, Computational Fluid Dynamics (3
Volumes), EES Books, 2000. (also AIAA coarse).
Wilcox, D.C., Turbulence Modeling for CFD, DCW Industries, 1998.
AIAA Courses (CFD, Turbulence)
Grid Generation training (ICEM, Pointwise, VGRID)
28