Roache 1997

November 28, 1996 12:21 Annual Reviews ROACHE AR23-05
Annu. Rev. Fluid. Mech. 1997. 29:12360

Copyright c 1997 by Annual Reviews Inc. All rights reserved
QUANTIFICATION OF UNCERTAINTY
IN COMPUTATIONAL FLUID
DYNAMICS
P. J. Roache
Ecodynamics Research Associates, Inc., P.O. Box 9229, Albuquerque,
New Mexico 87119
KEY WORDS: numerical uncertainty, verification, validation, error estimation, computational

fluid dynamics
ABSTRACT
This review covers Verification, Validation, Confirmation and related subjects for
computational fluid dynamics (CFD), including error taxonomies, error estima-
tion and banding, convergence rates, surrogate estimators, nonlinear dynamics,
and error estimation for grid adaptation vs Quantification of Uncertainty.
Introduction
The focus of this review is the quantification of uncertainty, the estimation
or banding of the numerical error of a final calculation in computational fluid
dynamics (CFD) and related fields in computational physics, mathematics, and
engineering. By final calculation I mean one considered to be used as is. The
motivation for this calculation is different than that for an error estimate used for
the process of solution-adaptive grid generation. Although the present methods
may be applicable to grid adaptation, and some of the methods described herein
were motivated by that problem, the grid adaptation problem has vastly different
(though usually unacknowledged) requirements.
The key word is quantification of uncertainty, as opposed to vague and all
too common qualitative assessments. Quantification of uncertainty may also
involve more than just obtaining a good error estimate; in fact, the more conser-
vative procedure based on the grid-convergence index (described below) reports
an error band equal to three times an error estimate. Furthermore, I consider
123
0066-4189/97/0115-0123$08.00
124 ROACHE
here only a posteriori error estimation, being of the opinion that useful a priori
estimation is not possible for nontrivial fluid mechanics problems.
Policy Statements on Numerical Uncertainty
In 1986, the editors of the ASME Journal of Fluids Engineering (JFE) saw fit to
publish a brief policy statement (Roache et al 1986) requiring at least minimal
attention to the quantification1 of numerical accuracy. Although the statement
seemed innocuous and seemed to address an obvious need, it still met significant
resistance. (Experience with the implementation of the policy is given in Roache
1990.) Since then, other journals have adopted similar explicit policies (ASME
Editorial Board 1994, AIAA 1994, Gresho & Taylor 1994). The JFE has
expanded its original policy, including a prohibition of methods with first-order
spatial accuracy (Freitas 1993a, 1995b; see discussion in Shyy & Shindir 1994,
Vanka 1994, Roache 1994b, Freitas 1994; Leonard 1995). An associate editor
of the AIAA Journal (W Oberkampf 1995, personal communication) estimates
that he needs to cite the journal policy in approximately three quarters of his
communications with manuscript authors.
The topic has also been the subject of several American Society of Mechanical
Engineers (ASME) symposia and resulting series of special publications (Celik
& Freitas 1990, Celik et al 1993, Johnson & Hughes 1995), from which the
present article takes its title. Also see the related ASME publications on the
CFD Triathlons (Freitas 1993b, 1995a) and other benchmark comparison
exercises, for example, the international HYDROCOIN and PSACOIN projects
(OECD 1988, 1990). These exercises give a fair cross-section of present CFD
practices in regard to numerical accuracy and generally are far from edifying.
Oldenburg & Pruess (1995) give an example of qualitatively different flow
revealed by even a cursory grid resolution study that was not discovered by any
participants in the HYDROCOIN code intercomparison exercise.
Verification and Validation: Numerical versus Conceptual
Modeling
It is useful at the onset to make the important semantic distinction between
verification and validation. Following Boehm (1981) and Blottner (1990),
I adopt the succinct description of verification as solving the equations right,
and of validation as solving the right equations. The code author defines pre-
cisely what partial differential equations (PDEs) are being solved and demon-
strates convincingly that they are solved correctlythat is, usually with some
1 Other aspects of confidence building in CFD, including broader issues of code quality-
assurance, confirmation, calibration, tuning, and certification, are found in Mehta (1989, 1991,
1995), Roache et al (1990), Aeschliman et al (1995), Oberkampf (1994), Oberkampf et al (1995),
Cosner (1995), Melnik et al (1995), and Roache (1995).
QUANTIFICATION OF UNCERTAINTY IN CFD 125
order of accuracy and always consistently, so that as some measure of dis-

cretization 1 (e.g. the mesh increments) approaches zero, the code produces
a solution to the continuum PDEs; this is verification. Whether or not those
equations and that solution bear any relation to a physical problem of interest to
the code user is the subject of validation. In a meaningful sense, a code cannot
be validated; only a calculation (or range of calculations with a code) can be
validated. In my experience, dealing with other than algorithm developers, this
is a difficult concept and requires frequent reiteration.
Another way to make the distinction (i.e. to get to the idea behind the
words, beyond mere semantics) is to speak of numerical errors versus con-
ceptual modeling errors. An example of conceptual modeling versus numerical
modeling is the assumption of incompressibility. This is clearly a conceptual
modeling assumption. Is it the code builders fault, or any criticism of the code
itself, if the user incorrectly applies it? For example, dynamic stall involves
compressibility at a surprisingly low free-stream Mach number. Results from
an incompressible code may not agree with experiment very well, but we can-
not say that the code fails verification because it was applied to a compressible
flow problem, although we may have some sympathy for the user who is fooled
by dynamic stall. But no one would have sympathy for a user who applied
an incompressible flow code to a reentry vehicle at Mach 20. In such a case,
the lack of agreement with experiment is not a code problem but a modeling
problem. The same is true of many practical aspects of applying a CFD code.
The model includes more than the code. Model includes conceptual model-
ing assumptions (e.g. incompressibility, symmetry, etc). It also includes data
input to the codee.g. geometry data, which often are not so easy to determine
accurately as one might expect, and boundary conditions and initial conditions.
Incorrect determination of any of these items can lead to failure of validation
of a model, with possibly no criticism of the code.
The question of solution uniqueness always arises with nonlinear equations,
and its position in the verification-validation distinction is important. If, as
stated above, we take verification to mean simply that a solution to the con-
tinuum PDEs is obtained, then the problem of PDE nonuniqueness is avoided.
This is probably being too easy on the code developer in some cases. Physi-
cally inadmissable solutions should be eliminated by the codefor example, a
shock-tube simulator should be required to eliminate expansion shock solutions;
but nature abounds with physically nonunique solutions in fluid dynamicsfor
example, hysteresis of airfoil stall and recovery, and bi-stable fluid amplifiers.
If nature cannot decide which solution to produce, we cannot expect more of
a code. Note that obtaining a solution to the continuum PDEs may involve
sorting through multiple numerical solutions for nonlinear problems; for exam-
ple, see later discussion of Yee et al (1991) and Lafon & Yee (1992). See also
126 ROACHE
especially Stephens & Shubin (1981), whose study of Euler solutions indicates
that the multiple numerical solutions converge to the same solution as the grid
is refined.
A related question is the stability of a solution. Although people usu-
ally think of artificial (i.e. numerical artifact) instability as a code difficulty,
the equally common situation is artificial stability, for example, a first-order
time-dependent code whose artificial viscosity damps out disturbances so that
accurate calculations of physical fluid instability would require impractically
high resolution. In my opinion, it is asking too much of a code to exactly mimic
stability boundaries except in the limit of 1 0. For example, it is not a fail-
ure of verification for a full Navier-Stokes code to produce steady shear-layer
solutions at Reynolds numbers known to be unstable; an unstable solution is
still a valid solution.
The choice of using verification or validation was originally arbitrary and is
now recommended solely because of common developing use. (I have pub-
lished articles using the opposite definition.) In a common English thesaurus,
the terms verify, validate, and confirm are all synonyms, but the words are
used herein, and generally in code quality-assurance (QA), as technical terms
with more context-specific meaning. Such technical terms are preferably re-
lated to common use, but each terms technical meaning is defined independent
of common use and in a specific technical context. This is not a universally
accepted attitude toward semantics. In a widely quoted paper that has been re-
cently described as brilliant in an otherwise excellent Scientific American article
(Horgan 1995), Oreskes et al (1994) think that we can find the real meaning of
a technical term by inquiring about its common meaning. They make much of
supposed intrinsic meaning in the words verify and validate and, as in a Greek
morality play, agonize over truth. They come to the remarkable conclusion
that it is impossible to verify or validate a numerical model of a natural sys-
tem. Now most of their concern is with groundwater flow codes, and indeed,
in geophysics problems, validation is very difficult. But they extend this to
all physical sciences. They clearly have no intuitive concept of error toler-
ance, or of range of applicability, or of common sense. My impression is that
they, like most lay readers, actually think Newtons law of gravity was proven
wrong by Einstein, rather than that Einstein defined the limits of applicability
of Newton. But Oreskes et al (1994) go much further, quoting with approval
(in their footnote 36) various modern philosophers who question not only
whether we can prove any hypothesis true, but also whether we can in fact
prove a hypothesis false. They are talking about physical lawsnot just
codes but any physical law. Specifically, we can neither validate nor inval-
idate Newtons Law of Gravity. (What shall we do? No hazardous waste
disposals, no bridges, no airplanes, no . . . .) See also Konikow & Bredehoeft

(1992) and a rebuttal discussion by Leijnse & Hassanizadeh (1994). Clearly,
we are not interested in such worthless semantics and effete philosophizing,
but in practical definitions, applied in the context of engineering and science
accuracy.
Inadequate Error Taxonomies

Several taxonomies of errors given in the literature are inadequate and mislead-
ing, in my opinion. Not all lists are taxonomies. For example, the list flora,
fauna, mammals, dogs is not an adequate biological taxonomy. Mammals are
not separate from fauna but are part of it, as are dogs. But we see this kind of
false taxonomy often in CFD.
For example, so-called grid-generation errors (Ferziger 1993) are not sepa-
rate from discretization errors. For the verification of a code or a calculation,
there are no such things as grid-generation errors [nor are there errors associ-
ated with coordinate transformations (Ferziger 1993)]. Indeed, bad grids add
to discretization error size, but they do not add new terms. This does not mean
that one grid is as good as another, or that a really bad grid cannot magnify
errors, but only that these so-called grid-generation errors do not need to be
considered separately from other discretization errors in a grid-convergence
test. If the grid-convergence test is performed, and the errors are shown to
reduce as O(12 ), for example, then all discretization errors are verified. One
does not need to separately estimate or band the grid-generation errors.
Likewise for Karniadakiss (1995) proposed numerical error bar (see also
Vanka 1995), which consists of separately estimated numerical errors from
boundary conditions, computational domain size, temporal errors, and spatial
errors. This is clearly a false taxonomy. Numerical errors at boundaries can
be ordered in 1 [e.g. the various approaches for wall vorticity (Roache 1972)]
or possibly nonordered in 1 [e.g. the boundary-layer-like p/n = 0, or
the downstream (outflow) boundary location (Roache 1972)]. Ordered errors
will tend to zero as the discretization improves, so that a boundary error from
wall vorticity evaluation need not be considered separately from spatial errors.
That taxonomy (Karniadakis 1995) already includes temporal errors and spatial
errors and computational domain size errors, so that both ordered and non-
ordered boundary errors are already counted elsewhere in the taxonomy. Since
the intention in (Karniadakis 1995) is to provide a quantitative breakdown in
the sources of numerical error in an error band, the proposed taxonomy is not
only confused but misleading.
Note, however, that outflow boundary errors may prove to be ordered not in
1 but in 1/l, where l is the distance from the region of interest (e.g. an airfoil)
128 ROACHE
to the outflow boundary. [See Zinggs (1993) data, shown in Roache (1994a)
to be first-order in 1/l.]
The subject of outflow boundary conditions does produce some fuzziness
in categorization of Verification vs. Validation, in my opinion. The error can
be ordered, as above, and therefore can be part of verification. That is, it is
up to the user (who is doing the conceptual modeling) to estimate or band
the error caused by the position of the outflow boundary. But if the code has
some sophisticated outflow condition (e.g. a simple vortex condition for Euler
equations), then the distinction is not so clear. Certainly the equations used are
clear, and the code may solve the equations right (i.e. verification), yet there
exists another benchmark solely from the mathematics (the case with infinite
boundary distance) which could be used to justify the outflow condition without
recourse to physical experiment (which would be associated with validation).
Another example of the same semantic failure arises when we consider bench-
marking a turbulent boundary-layer code or parabolized Navier-Stokes (PNS)
code against a Reynolds-stress-averaged full Navier-Stokes (NS) code. Pre-
sume that both codes are convincingly verified, that is, they correctly solve their
respective equations. Suppose that the PNS code results agree well with the
NS code results for some range of parameters (e.g. including angle of attack).
This agreement is not included in the term verification, since the verification
of the PNS code has already been completed prior to the NS benchmarking.
Then we could say that the agreement has demonstrated that the PNS code is
solving the right equations in one sense, that is, it justifies the use of parabolic
marching equations. Yet to claim validation would be over-reaching, since we
have not demonstrated the adequacy of the turbulence model by comparison
with experiment. We have solved the right equations only in an intermediate
sense of demonstrating that the PNS equations adequately represent the full
NS equations, but not in the ultimate sense of validation, of solving the right
physical equations.
Unfortunately, such mere semantics may become of vital interest to code
QA when dealing with regulatory agencies such as the EPA or with legal defi-
nitions in a NASA contract.
Taxonomy for Additional Information for Error Estimates

Once we have produced a CFD solution of the governing partial differential
equations, it is clear that we require some additional information in order to
quantitatively estimate the uncertainty or numerical accuracy. The following
taxonomy of sources of this additional information provides a framework for
the discussion. By the word grid, I refer to any measure of discretization
(e.g. Cartesian grid, nonorthogonal grid, number of Fourier modes in a spectral
solution, number of discrete vortices).
Sources of Additional Information for Error Estimation, Given

a CFD Solution of the Governing PDEs on a Grid
A. Additional Solution(s) of the Governing Equations on Other Grids
A.1 Grid Refinement
A.2 Grid Coarsening
A.3 Other Unrelated Grid(s)
B. Additional Solution(s) of the Governing Equations on the Same Grid
B.1 Higher-order Accuracy Solution(s)
B.2 Lower-order Accuracy Solution(s)
C. Auxiliary PDE Solutions on the Same Grid
D. Auxiliary Algebraic Evaluations on the Same Grid; Surrogate Estimators
D.1 Nonconservation of Conservation Variables
D.2 Nonconservation of Higher Moments
D.3 Zhu-Zienkiewicz-type Estimators
D.4 Convergence of Higher-order Quadratures
The following are brief remarks on this taxonomy, some of which will be
justified and amplified in later discussion:
Categories A and B involve the direct, unambiguous evaluation of any error
measure of engineering or scientific interest. For Category A, no additional
CFD code development or modifications are required. For Category B, no
additional grid generation is required.
The approach of Category C (e.g. Van Straalen et al 1995) does not simply
involve a local evaluation of something. The key aspect here is that errors are
transported (i.e. advected, diffused). However, a simple local evaluation of
something, without advection, is just what you want to guide solution adap-
tation, hence, the different needs of error estimation for solution adaptation
versus quantification of uncertainty for a final calculation.
Category D, involving algebraic evaluations on the same grid, is cheap, needs
no additional grid generation, and uses no significant dynamic memory. How-
ever, the error evaluated usually has no direct relation to any error measure of
engineering or scientific interest.
Grid Refinement and Coarsening
Systematic grid-convergence studies are the most common, most straightfor-
ward and arguably constitute the most reliable technique for the quantification
130 ROACHE
of numerical uncertainty. Unlike the other methods available, this approach

can be used to dependably consider the convergence of any quantity of interest,
as well as the usual L 2 and L norms.
By grid-convergence studies people usually mean Category A.1 (grid refine-
ment), but Category A.2 (grid coarsening) usually would make more sense. If
completely solved solutions are obtained on two grids, presumably the finer
grid solution would be used, so the coarse grid solution could be used to es-
timate the error. Whether one refines or coarsens just depends on which grid
was calculated first. So for completely solved solutions, grid refinement and
coarsening are identical.
A disadvantage of Category A methods is that multiple-grid generation is
required. Cartesian grids obviously pose no problem. For boundary-fitted
structured grids, the simplest method for grid doubling (halving) is to generate
the finest grid first, using whatever method is preferred (e.g. see Knupp &
Steinberg 1994) and then obtain the coarser grids by removing every other
point (e.g. see Zingg 1993). For noninteger grid refinement (coarsening),
the same generating equations and parameters should be used. See Roache
(1993a, 1994) for further discussion, including nonstructured grid refinement
(see also Pelletier & Ignat 1995) and structured refinement of nonstructured
grids.
Error estimation using other unrelated grids (Category A.3) poses an interest-
ing challenge. By unrelated grids I mean two or more grids that are overlapping
but not simply obtained one from the other. For example, say grid A is finer than
grid B in some regions but coarser in others, such as might be obtained in two
steps of an r -type (redistribution) solution-adaptive grid. It would seem that
the two solutions on unrelated grids would provide the additional information
necessary to estimate the uncertainty in either, but to my knowledge a method
for doing so has not been invented.
In order to quantify the uncertainty with systematically refined (coarsened)
grids, we need the convergence rate p to estimate the error. For now, we are
assuming that p is known, that is, that we are using a rigorously verified code
and are now concerned with quantification of the uncertainty of a particular
calculation using two grid solutions. The same methods can be converted to
verify a code, that is, to verify (or determine) p. Code verification is discussed
in a later section.
A Method for Uniform Reporting of Grid-Convergence Studies:
The Grid-Convergence Index
This section presents a summary of the main results of Roache (1993a,b; 1994a).
The grid-convergence index (GCI) presents a simple method for uniform re-
porting of grid-convergence studies without any restriction to integer refinement
(e.g. grid doubling). The GCI is based on generalized Richardson Extrapolation

involving comparison of discrete solutions at two different grid spacings.
Richardson Extrapolation
Richardson Extrapolation, also known as h 2 extrapolation, the deferred ap-
proach to the limit, and iterated extrapolation, was first used by Richardson in
1910 and later embellished in 1927. The discrete solutions f are assumed to
have a series representation, in the grid spacing h, of
f = f [exact] + g1 h + g2 h 2 + g3 h 3 + (1)
The functions g1 , g2 , and so on are defined in the continuum and do not depend
on any discretization. For infinitely differentiable solutions, these functions are
related to all orders to the solution derivatives through the elementary Taylor
series expansions, but this is not a necessary assumption for Richardson Extrap-
olation, nor is the infinite series indicated in Equation 1. It is only necessary
that Equation 1 be a valid definition for the order of the discretization. Thus,
the extrapolation may be valid for finite element solutions, and so on.
For a second-order method, g1 = 0. Then the idea is to combine two
separate discrete solutions, f 1 and f 2 , on two different grids with (uniform)
discrete spacings of h 1 (fine grid) and h 2 (coarse grid), so as to eliminate the
leading-order error terms in the assumed error expansion, that is, to solve for g2
at the grid points in Equation 1, substitute this value into Equation 1, and obtain
a more accurate estimate of f [exact]. The result is the original statement
(Richardson 1927) for h 2 extrapolation:
2
f = f [exact] + h 22 f 1 h 21 f 2 h 2 h 21 + H.O.T. (2)
where H.O.T. are higher-order terms. Using the grid refinement ratio r =
h 2 / h 1 , this result can be conveniently expressed in terms of a correction to the
fine-grid solution f 1 , dropping H.O.T.:
f [exact]
= f 1 + ( f 1 f 2 )/(r 2 1). (3)
The most common use of this method is with a grid doubling, or halving.
(As noted earlier, these are identical, that is, we just have a coarse grid and a
fine grid.) With r = 2, Equation 3 becomes
f [exact]
= 4/3 f 1 1/3 f 2 . (4)
It is often stated that Equation 4 is fourth-order accurate if f 1 and f 2 are
second-order accurate. Actually, as known by Richardson, this statement is
true only if odd powers are absent in the expansion (Equation 1), which he
achieved by assuming the exclusive use of second-order centered differences.
If uncentered differences are used (e.g. upstream weighting of advection terms),
132 ROACHE
even if these are second-order accurate (3-point upstream), the h 2 extrapolation

is third-order accurate, not fourth. As a practical limitation, even extrapola-
tions based on centered differences do not display the anticipated fourth-order
accuracy until the cell Reynolds number Rc is reduced; for the 1-D advection-
diffusion equation with Dirichlet boundary conditions, Rc < 3 is required
(Roache & Knupp 1993).
Although Richardson Extrapolation is most commonly applied to grid doubl-
ing and is often stated to be applicable only to integer grid refinement (e.g.
Conte & DeBoor 1965), this is not required. In order to use Equation 3, it is
necessary to have values of f 1 and f 2 at the same points, which would seem
to require commonality of the discrete solutions and, therefore, integer grid
refinement ratios r (grid doubling, tripling, and so on). However, even in his
1910 paper, Richardson looked forward to defining a continuum f 2 by higher-
order interpolation, and in his 1927 paper had a specific approach worked out.
Ferziger (1993) alludes to this approach with less detail but more generality.
Similarly, Richardson Extrapolation is commonly applied only to obtaining a
higher-order estimate on the coarse grid with h 2 = 2h 1 , but Roache & Knupp
(1993) show how to obtain fourth-order accuracy on all fine-grid points by
simple second-order interpolation, not of the solution values f 2 , but of the ex-
trapolated correction from Equation 4, that is, by second-order interpolation of
1/3( f 1 f 2 ). The use of simple second-order interpolation avoids complexities
with nonuniform grids and near-boundary points.
Richardson (1910, 1927) also considered sixth-order extrapolation (using
three grid solutions to eliminate g2 and g4 ), parabolic and elliptic equations,
staggered grids (then called interpenetrating lattices), rapid oscillations and the
2h wavelength limit, a priori error estimates, singularities, integral equations,
statistical problems, Fourier coefficients, and other noncalculus problems. For
example, Richardson (1927) showed the power of the method in an elegant
example of extrapolating two very crude approximations to a circle, namely
an inscribed square and an inscribed hexagon, to get an estimate of with
three-figure accuracy.
The usual assumptions of smoothness apply, as well as the assumption
(often verified) that the local error order is indicative of the global error order.
The extrapolation must be used with considerable caution, since it involves the
additional assumption of monotone truncation-error convergence in the mesh
spacing h. This assumption may not be valid for coarse grids. Also, the ex-
trapolation magnifies machine round-off errors and incomplete iteration errors
(Roache 1972). In spite of these caveats, the method is extremely convenient to
use compared to forming and solving direct fourth-order discretizations, which
involve more complicated stencils, wider bandwidth matrices, special con-
siderations for near-boundary points and non-Dirichlet boundary conditions,
additional stability analyses, and so on, especially in nonorthogonal coordi-

nates that generate cross-derivative terms and generally complicated equations.
Such an application was given by the present author in Roache (1982).
The method is in fact oblivious to the equations being discretized and to
the dimensionality of the problem and can easily be applied as a postproces-
sor (Roache 1982) to solutions on two grids with no reference to the codes,
algorithms, or governing equations that produced the solutions, as long as the
original solutions are indeed second-order accurate. (By this somewhat abu-
sive but common terminology, I mean a solution obtained with a second-order
accurate method applied in its asymptotic range.) The difference between the
second-order solution and the extrapolated fourth-order solution is itself a use-
ful diagnostic tool, obviously being an error estimator (although it does not
provide a true bound on the error except possibly for certain trivial problems).
It was used very carefully, with an experimental determination rather than an
assumption of the local order of convergence, by de Vahl Davis (1983) in his
classic benchmark study of a model free-convection problem. See Nguyen &
Maclaine-Cross (1988) for application to heat exchanger pressure drop coeffi-
cients. Zingg (1993) applied the Richardson error estimator to airfoil lift and
drag calculations in body-fitted grids. Zinggs work demonstrates the necessity
of grid-convergence testing even when experimental data are available. In four
of seven cases, experimental agreement was better with coarse-grid calcula-
tions than with fine! Also, his data indicate that Richardson Extrapolation can
be applied to the estimation of far-field boundary errors, with the error being
first order in the inverse of distance to the boundary (Roache 1994a). Blottner
(1990) has used the same procedure to estimate effects of artificial dissipation
terms in hypersonic flow calculations.
An important aspect of Richardson Extrapolation is that it applies not only to
point-by-point solution values but also to solution functionals (e.g. lift coeffi-
cient C L for an aerodynamics problem or integrated discharge for a groundwa-
ter flow problem), provided that consistent or higher-order methods are used in
the evaluations (e.g. second- or higher-order quadratures for lift) as well as the
basic assumption that the order of the method applies globally as well as locally.
If Richardson Extrapolation is applied to produce (say) fourth-order accurate
grid values, one could in principle calculate a fourth-order accurate functional
like C L from the grid values, but it would require careful implementation of
fourth-order accurate quadratures. It is much simpler to apply the extrapo-
lation directly to the C L s obtained in each grid, requiring only second-order
quadratures. Indeed, this is a major attraction of Richardson Extrapolation
compared to using fourth-order accurate stencils solved either directly or by
deferred corrections. (Note, however, that the two approaches yield different
answers, although both are fourth-order accurate if done properly.)
134 ROACHE
A significant yet overlooked disadvantage of Richardson Extrapolation is that

the extrapolated solution generally is not conservative in the sense of maintain-
ing conservation properties (e.g. Roache 1972). This shortcoming could well
dictate that Richardson Extrapolation not be used. For example, if it were used
on the groundwater flow simulations for the Waste Isolation Pilot Plant (WIPP
PA Dept. 1992), it would be more accurate in some norm but would introduce
additional nonconservative (i.e. lack of conservation property) source terms
into the radionuclide transport equation. de Vahl Davis (1983, p. 256) pointed
out the more fundamental problem that the extrapolated solution is no longer
internally consistent; the values of all the variables do not satisfy a system of
finite difference approximations. It is also noteworthy that Richardson (1927)
pointed out that the accuracy of the extrapolation does not apply to arbitrarily
high derivatives of the solution. The extrapolation can introduce noise to the
solution, which, although low level, may decrease the accuracy of the solution
higher derivatives.
Thus, it is not advocated (Roache 1993a,b, 1994a) that Richardson Extrapo-
lation necessarily be used to improve the reported solution, since that decision
involves these considerations and possibly others. What is advocated is that,
regardless of whether Richardson Extrapolation is used to improve the solution,
the GCI based on the generalized theory of Richardson Extrapolation should
be used to uniformly report grid-convergence tests.
A fine-grid Richardson error estimator approximates the error in a fine-grid
solution, f 1 , by comparing this solution to that of a coarse grid, f 2 , and is
defined as
f ine
E1 = , (5)
1rp
while a coarse-grid Richardson error estimator approximates the error in a
coarse-grid solution, f 2 , by comparing the solution to that of a fine grid, f 1 ,
and is defined as
r p
E 2coar se = , (6)
1rp
where
= f2 f1, (7)
f 2 = a coarse-grid numerical solution obtained with grid spacing h 2 ,
f 1 = a fine-grid numerical solution obtained with grid spacing h 1 ,
r = refinement factor between the coarse and fine grid
(r = h 2 / h 1 > 1), and
p = formal order of accuracy of the algorithm.
The actual error A1 of the fine-grid solution may be expressed (Roache 1993a,
1994a) as

A1 = E 1 + O h p+l , E 12 , (8)
where l = 1 generally or l = 2 if centered differences have been used. Thus,
E 1 is an ordered estimator and is a good approximation when E 1 1.
As noted above, it is neither necessary nor often desirable to use r = 2, or
grid doubling (halving). Accurate application of these generalized Richardson-
based grid-error estimators requires that the observed convergence rate equals
the formal convergence rate. This requirement implies that the leading-order
truncation error term in the error series truly dominates the error.
To account for the uncertainty in these generalized Richardson-based error
estimates due to various factors and to put all grid-convergence studies on the
same basis as grid doubling with a second-order method, a safety factor is
incorporated into these estimators and the GCI is defined for fine and coarse
grids as
f ine
GCI1 = Fs |E 1 | (9)
GCI coar se
2 = Fs |E 2 | (10)
Fs > 1 can be interpreted as a safety factor, since Fs = 1 gives GCI = |E|.
That is, the error band reduces to the best estimate of the error, analogous to
a 50% error band of experimental data. I recommended (Roache 1993a, b;
1994a) a more conservative value of Fs = 3. This value also has the advantage
of relating any grid-convergence study (any r and p) to one with a grid doubling
and a second-order method (r = 2, p = 2). I emphasize that the GCIs are not
error estimators but are three (or Fs ) times the error estimators, representing
error bands in a loose statistical sense. Ostensibly, if we have a fine-grid and
a coarse-grid solution, we would be expected to use the fine-grid solution, so
reporting of the above fine-grid evaluation of GCI would apply. However, a
practical scenario occurs for which the contrary situation applies, i.e. we use
the coarse-grid solution. Consider a parametric study in which hundreds of
variations are to be run. (For example, consider a 3-D time-dependent study of
dynamic stall, with perhaps 3 Mach numbers, 6 Reynolds numbers, 6 airfoil-
thickness ratios, 3 rotor-tip designs, and 2 turbulence models, for a total of 648
combinations.) A scrupulous approach would require a grid-refinement study
for each case, but most engineers would be satisfied with one or a few good
grid-refinement tests, expecting, for example, that a grid adequate for a NACA
0012 airfoil could be assumed to be adequate for a NACA 0015 airfoil. (In fact,
this is often not justified by experience. For example, stall characteristics can
be quite sensitive to thickness ratio.) So for the bulk of the stack of calculations,
136 ROACHE
we would be using the coarse-grid solution, and we want a GCI for this solution.
That is, we derive a GCI from Equation 5, not as the correction to the fine-grid
solution f 1 , but as the correction to the coarse-grid solution f 2 . In this case,
the error estimate changes and must be less optimistic.
A recent application of the GCI to airfoil calculations is given by Lotz et al
(1995), which demonstrates the power of the method without the need for integer
grid refinement and its application with solution-adaptive grids. Pelletier &
Ignat (1995) have shown that the GCI is applicable, at least in a rough sense,
to unstructured grid refinement.
The motivation for using Fs > 1 is that Fs = 1 is analogous to a 50% error
band on experimental data, which is not adequate. My originally recommended
value (Roache 1993a,b; 1994a) of Fs = 3 is conservative and relates the grid
convergence study to one with a grid doubling with a second-order method. For
many reasons (see below) this is not unduly conservative when only two grids
are used in the study. However, it is now clear that Fs = 3 is overly conservative
for scrupulously performed grid convergence studies using three or more grid
solutions to experimentally determine the observed order of convergence p
(e.g. see the papers in Johnson & Hughes 1995). For such high-quality studies,
a modest and more palatable value of Fs = 1.25 appears to be adequately
conservative. However, for the more common two-grid study (often performed
reluctantly, at the insistence of journal editors) I still recommend the value
Fs = 3 for the sake of uniform reporting and adequate conservatism.
Code Verification
So far, this article has been concerned with error estimation, or verification of a
calculation. It has been assumed throughout that the code itself has been verified
independently. If a code has an error (not an ordered error, but a mistake), then
performance of grid-convergence studies or other error-estimation techniques
will not faithfully produce quantification of the uncertainty. However, the
same techniques can be used to verify the code separately, including its order of
convergence, by monitoring grid convergence toward a nontrivial exact solution.
It is always possible to obtain a nontrivial exact solution for this procedure, if
necessary, by specifying the exact solution and adding the appropriate forcing
function to the governing PDEs (Steinberg & Roache 1985, Shih et al 1989,
Ethier & Steinmen 1994, Westerink & Roache 1995). The necessary condition
is that the solution be nontrivial, that is, that it have significant solution structure
to exercise higher-derivative calculations. Significantly, it is not at all necessary
that the solution be realistic in any sense. Physically realistic solutions will
be more convincing to the mathematically naive (and therefore have value in
the real world of regulatory agencies, contract management, and so on), but
the code verification is just as well accomplished by nonphysical solutions.
(For the alternative philosophy, see Oberkampf 1994, Oberkampf et al 1995,

and Aeschliman et al 1995.) Additional realistic solutions (such as Sudicky &
Frind 1982, corrected) can be used in additional confirmation exercises (Roache
1995).
I claim that this method of code verification via systematic grid-convergence
testing (whether or not the GCI is used) is both general and rigorous, and has
proven remarkably sensitive in revealing small and subtle coding errors and
usage inconsistencies, as indicated by the following examples (Westerink &
Roache 1995; Salari et al 1995):
A. In verification tests of a commercial groundwater flow code, a first-order

error in a single corner cell in a strongly elliptic problem caused the observed
convergence to be first-order accurate (Roache et al 1990).
B. In verification tests of our SECO FLOW 2D variable density groundwater
flow code, first-order extrapolation for ghost cell values of only one quantity
(aquifer thickness) along one boundary caused the observed convergence
to be first-order accurate (Roache 1993c,d).
C. In contaminant transport calculations (advection-diffusion + decay, retar-
dation, and matrix diffusion), use of a plausible single-grid-block represen-
tation for a point source as the grid is refined introduces error in a finite
volume (or block-centered finite difference) formulation. In this cell con-
figuration, the cell faces align with the boundaries of the computational
domain, and doubling the number of cells requires the location of the single
cell representing the source to shift by 1/2. It is to be expected that the
solution accuracy in the neighborhood of the source would be affected. But
surprisingly, the accuracy of time-integrated discharge across boundaries
far from the source was also degraded to first-order accuracy (Salari et al
1995).
D. The observed convergence rate of ostensibly second-order-accurate turbu-
lent-boundary-layer codes (Wilcox 1993) can be degraded, apparently by
conditional statements limiting eddy viscosity and defining the boundary-
layer edge (DC Wilcox 1995, personal communication,).
E. Airfoil codes can exhibit the expected second-order convergence rates for
lift and drag but less for moment, possibly because of approximations in-
volved in applying quasi-periodicity across cut-planes of a C-grid
(K Salari 1995, personal communication).
F. When a 2-D compressible driven cavity problem is solved with an even
number of vertical grid lines, one must average velocity profiles from grid
138 ROACHE
lines on each side of the centerline rather than use only one; otherwise, the
convergence will be first-order (Roache & Salari 1990).
Note that in Examples C and F apparently first-order behavior was obtained
with second-order verified codes; the problem in Example C was a subtle con-
ceptual modeling error and in Example F was a not-so-subtle postprocessing
error. In both cases, the codes had no coding errors and were applied in the
asymptotic range.
Esoteric Errors in Code Verification
General CFD or computational physics codes [more general than the simple
Poisson equation in nonorthogonal coordinates of Steinberg & Roache (1985)]
would be difficult to include in a theorem because of esoteric errors. The
difficult aspects of the codes are not algebraic complexity [in Steinberg &
Roache (1985) we convincingly verified 1800 lines of dense Fortran]; the more
difficult and vexing problems come from option combinations and conditional
differencing. Esoteric errors can arise because of nonlinear flux limiters like
FCT, TVD, hybrid or type-dependent differencing, and so on (Steinberg &
Roache 1985).
Code Verification versus Verification of Calculations
In this review, I distinguish between verification of codes and verification of
individual calculations. Clearly, code verification must be done before the
verification of a (real) calculation can proceed. However, this distinction works
both ways; it is not enough to perform a calculation of a new problem with
a verified code. A code may be rigorously verified to be (say) second-order
accurate, but when applied to a new problem, this fact provides no estimate of
accuracy or confidence interval, that is, of the size of the error (as opposed to the
order of the error). Even though the real problem solution may be converging
at the same rate (say second-order) as the code verification problem, it is still
necessary to perform grid-convergence tests in order to band the numerical
error. (It would be preferable to have different words for these two verification
activities, but I am at a loss for a clarifying term.) The very important point,
independent of the semantics, is that use of a verified code is not enough. This
point is probably well recognized by present readers, but it is not universally
so. Especially in the commercial CFD arena, user expectations are often that
the purchase and use of a really good code will remove from the user the
obligation of doing his homework, that is, the straightforward but tedious
work of verification of calculations via systematic grid-convergence studies.
This unrealistic hope is sometimes encouraged by advertising.
Unfortunately, the situation may be even more difficult than this, in that a
new calculation may not even exhibit the verified order of convergence. If the
code verification problem(s) are in some sense close to the real problem of in-
terest, the order of convergence (e.g. second-order) from the code verification
exercise may be assumed to hold for the real problem. However, there is a
distinction to be made between the formal order of convergence (as indicated
by analyses of the discretization), the actual asymptotic order of convergence,
and the observed (or apparent) order of convergence. Note that there may be
more than one formal analysis, especially for nonlinear problems, so that formal
order of convergence is not necessarily unique. Then the actual asymptotic rate
of convergence may differ from a formal rate, due to competition of truncation
error terms for a particular problem (Westerink & Roache 1995). This situa-
tion typically arises when higher grid resolution reveals higher modes of the
solution, which often occurs in atmospheric and ocean modeling (Westerink
& Roache 1995, Dietrich et al 1990) and could be expected in high Reynolds
number DNS (Direct Numerical Simulation) calculations. It is also possible,
and not altogether unusual, to observe superconvergence, that is, an actual
asymptotic convergence rate higher than a formal rate, due to cancellation (or
partial cancellation) of space and time truncation errors. This observation is
evident in the classical situation of (what are generally) first- or second-order
methods for constant velocity advection problems producing the exact answers
for unity Courant number (Roache 1972). Finally, the observed convergence
rate p may differ from the actual asymptotic convergence rate due to failure to
achieve the asymptotic range of grid resolution. This is especially obvious if
solution singularities are present, in which case the method of grid refinement
still can be used to estimate uncertainty, but local values of observed p near the
singularity are required (de Vahl Davis 1983).
The cause of the discrepancy between an observed convergence rate and the
actual asymptotic convergence rate may or may not be revealed by formal anal-
ysis. For example, in Roache & Knupp (1993), a method was presented for
completing classical Richardson Extrapolation. The classical method combines
two second-order solutions on a coarse grid and a doubled fine grid to produce
a fourth-order accurate solution, but only on the coarse grid. The completed
algorithm was devised to produce a fourth-order solution on the entire fine grid.
However, initial grid-convergence studies indicated only a third-order conver-
gence rate for the steady-state Burgers equation, even though the formal rate
was fourth order. Further analysis by PM Knupp then showed that, because
of truncation error competition particular to the steady Burgers solution, the
higher-order terms in the error expansion include the Reynolds number param-
eter Re, and that the asymptotic range for fourth-order accuracy occurs only for
Re 1x 3.
For these reasons, it appears to be preferable to verify a code with a non-
realistic problem that has nontrivial but regular solution behavior (e.g. a tanh
140 ROACHE
function) rather than to use a realistic problem that displays the confusing trun-
cation error competition.
Extraction of Convergence Rate from Grid-Convergence Tests
When using grid-convergence studies to estimate the size of the discretization
error, it is necessary to have at least two grid solutions and an a priori knowledge
of the convergence rate p, for example, to have verified that the code is second-
order accurate ( p = 2). If an exact solution is known or constructed (see
above), it is straightforward to extract the order of convergence p from results
of a systematic grid-convergence test using at least two grid solutions; this
serves to verify a code. However, it is also desirable to verify the order p for an
actual problem, since the observed order of convergence depends on achieving
the asymptotic range, which is problem dependent, and the observed order
may differ from the formal (theoretical) order, or from the order verified for a
test case, for a variety of reasons (see discussion above and a more complete
discussion in Westerink & Roache 1995). Without an exact solution for the
actual problem, it is necessary to have at least three grid solutions to extract p.
Thus, if there is any suspicion that the grid resolution is not in the asymptotic
range, three grid solutions are necessary to verify (or determine) the observed
rate of convergence and thereby allow estimation of the error.
Blottner (1990) and others use graphical means, plotting the error on log paper
and extracting the order from the slope. This procedure requires evaluation of
the error itself, which is of course generally not known. If the finest grid solution
is taken to be the reference value (unfortunately, often called the exact value,
which it obviously is not), then the observed order will be accurate only for
those grids far from the finest, and the calculated order approaching the finest
grid will be indeterminate. Blottner (1990) improves on this by estimating the
exact value by Richardson Extrapolation (see also Shirazi & Truman 1989),
but this procedure is somewhat ambiguous since the order is already needed in
order to perform the Richardson Extrapolation.
If the grid refinement is performed with constant r (not necessarily r = 2,
but constant), then the order can be extracted directly from three grid solutions,
without a need for estimating the exact solution, following de Vahl Davis (1983,
p. 254). With subscript 1 indicating the finest grid in the present notation,

f3 f2
p = ln / ln(r ). (11)
f2 f1
A generalization of this procedure, not restricted to constant r , is possible using
the generalized theory of Richardson Extrapolation (Roache 1993a, b; 1994a).
For constant r , Equation 19 (typo corrected) of Roache (1994a) may be used
to verify an assumed order p. (It is not necessary to use the GCI itself.) One
calculates
f ine f ine
= GC I12 GC I23 . (12)
If r , then p is the observed order. However, Equation 12 also requires
p
r to be constant over the three grid set, and it cannot be used to calculate p
directly since p is implicitly present in the GCIs. The more general procedure
(Roache 1995a) is to solve the equation

p 23 = r12 p
p 12 (13)
r23 1 r12 1
for p. This is simple for r constant (not necessarily 2 or integer), giving
p = ln(23 /12 )/ ln r. (14)
(Note that Equation 14 differs slightly from Equation 11 but is equivalent to the
order of the approximation involved.)
If r is not constant during the grid refinement, Equation 13 is transcendental
in p. Usual solution techniques (e.g. Newton-Raphson) can apply, but one
should allow for observed p < 1. This result can happen even for simple
problems at least locally (de Vahl Davis 1983), and in some cases the observed
p < 0 (unfortunately, behavior far away from asymptotic convergence can be
nonmonotone). Also, r 2 will be easier to solve than r 1 + , and r 2
is probably not of much interest. For well-behaved synthetic cases, simple
substitution iteration with a relaxation factor 0.5 works well. With =
previous iterate for p, the iteration equation is
ln()
p = + (1 ) (15a)
ln(r12 )

r12 1 23
= . (15b)
r23 1 12
Note this form of the iteration gives the exact answer in one step for the case
of r = 2 and = 0.
Predicting the Required Grid Resolution
Once p is known with some confidence, one may predict the next level of grid
refinement r necessary to achieve a target accuracy, expressed as a target error
estimate E 1 or a target GC I1 called GC I . With GC I23 being the value from
Equation 2 for the previous two grids,
p
r = p GCI /GCI23 . (16)
This result, of course, depends only on the assumed definition of order of the
discretization error, that is, only on c = error/1 p and not on the GCI theory
itself.
142 ROACHE
Special Considerations for Turbulence Modeling

Special considerations are required for turbulence modeling and for other fields
with multiple scales. Here, the code theoretical performance can be verified
(within a tolerance) for a range of parameters but could fail in another range.
It is necessary to get the grid resolution into the asymptotic range in order to
do grid-convergence testing. Virtually any grid is in the asymptotic range for a
simple Laplace equation. For any boundary-layer calculation, it is clear that the
initial (coarse) grid must get some points into the boundary layer. For turbulence
modeling without wall functions (Wilcox 1993, Shirazi & Truman 1989) the
grid must get some points into the wall layer. For turbulence modeling with wall
functions, the grid should not get into the wall layer (Celik & Zhang 1995). In
my interpretation, the wall functions should be viewed as an elaborate nonlinear
boundary condition, and the grid-convergence exercise should be done from the
edge of the wall layer out. Similarly, for large eddy simulations (LES), as used
in aerodynamic turbulence research and in atmospheric and ocean modeling
with sub-grid turbulence modeling, the grid convergence must not go to zero,
or else the Reynolds stresses will be counted twice, once from the full Navier-
Stokes terms and once modeled from the LES terms. Also, the presence of
any switching functions, such as length determinations for the Baldwin-Lomax
turbulence model (Wilcox 1993) can easily corrupt second-order convergence
rates.
Finally, the grid-resolution requirements are much more demanding for tur-
bulent boundary layers, just as laminar boundary layers are much more de-
manding than inviscid flows. For example, Claus & Vanka (1992) found that
2.4 million nodes (256 96 96) did not demonstrate grid independence of
the computed velocity and turbulence fields of crossflow jets.
Special Considerations for Artificial Dissipation
Special considerations are required for any numerical treatment that effectively
changes the governing PDEs as the grid is refined. Sub-grid turbulence mod-
eling was mentioned earlier. The other important but troublesome numerical
methods are those that treat shock-like problems with any kind of artificial
dissipation that depends on the grid size. (Shock-like problems means either
true shocks or simply regions of unresolvable large solution structure involving
solution gradients, curvatures, and various combinations.) This includes both
explicit artificial dissipation (in which an identifiable additional term is added
to the governing PDEs, whether treated explicitly or implicitly in the solution
algorithm) and implicit dissipation introduced locally via nonlinear flux lim-
iters (FCT, TVD, ENO, PPM, and so on). If not addressed, these terms corrupt
the convergence behavior. See Kuruvila & Anderson (1985) for discussion of
difficulties and pitfalls of doing convergence studies with artificial dissipation
terms in the equations. The explicit dissipation approach can be treated cor-
rectly and rigorously, as in Blottner (1990) and Shirazi & Truman (1989).
The artificial viscosity of first-order upstream differencing (Roache 1972)
requires no special consideration other than its slow convergence. However,
hybrid methods in their various forms cause more problems. By adaptively
changing the weighting given to first-order and second-order stencils, depend-
ing on the local cell Reynolds number Rc, these methods by definition are
ultimately second-order accurate, but only when the resolution is so fine that
the hybrid algorithm itself is inoperative, typically when Rc < 2. Over the
practical range of resolution of interest, the order is ill-defined, but the empiri-
cally observed order is, as expected, somewhere vaguely between 1 and 2 (e.g.
see Celik & Zhang 1995). Unfortunately, hybrid methods are commonly used
in commercial codes, wherein the emphasis is often on code robustness at the
sacrifice of accuracy. The same ill-defined behavior is observed for the class
of Allen-Southwell methods (Roache 1972).
Error Estimation from Higher-Order Solutions
on the Same Grid
Category B.1 methods estimate the accuracy of a base solution by comparison
with higher-order accuracy solution(s). (Again, note the somewhat abusive
but common terminology.) Richardson (1908) again scooped modern error-
estimation papers by inventing Category B.1, error estimation from higher-order
solutions on the same grid, noting that the difference between a second-order
accurate solution and a fourth-order accurate solution is itself an ordered error
estimator. The higher-order accuracy solution might be obtained via a new
solution of higher-order discretizations using FDM (finite difference methods),
FVM (finite volume methods), FEM (finite element methods), and so on, or by
deferred corrections, compact differences (again direct or deferred corrections,
and so on).
For the technique of error estimation from higher-order solutions on the
same grid, much the same advantages and limitations apply as with the grid-
convergence technique. It applies to all point values and functionals (e.g. lift,
drag); x and t errors may be estimated independently or coupled; and the error
estimate includes nonlinear coupling. These methods are not commonly used
because they require additional code capability, unlike grid-convergence tests
(Category A). On the other hand, these Category B methods do not require
additional grid generation.
The development costs of the additional code capability may be minimized
by noting several points. If we were intending to use the higher-order solution
itself, many restrictions and requirements could apply, such as full iterative
convergence, strict conservation, and so on. However, if our only use of the
144 ROACHE
higher-order solution is to estimate the error of the base solution, these consid-
erations are not so important. The point is that error estimation of, for example,
a second-order solution using fourth order methods is less demanding numer-
ically than obtaining a fourth-order solution for direct use. Also, though not
generally recognized, directional splitting works. (The following development
has not been completed for cross-derivative terms, which in any case may re-
quire some careful formulation for higher-order stencils and should be verified
rigorously.)
We use a notation similar to that above for multiple grid solutions: f denotes
the exact solution, f 2 the second-order accurate solution, and f 4 the fourth-order
accurate solution; C2x are the coefficients of the Taylors theorem expansion
for the second-order solution in the x-direction, and so on; and R2x are the
remaining terms in the complete second-order expansion in x, and so on. Then,
in two dimensions,
f = f 2 + C2x 1x 2 + R2x + C2y 1y 2 + R2y (17)
and
f = f 4 + C4x 1x 4 + R4x + C4y 1y 4 + R4y . (18)
Defining the error of the second-order solution E 2 to be
E2 = f f2 (19)
and substituting for f from Equation 18, we obtain
E 2 = f 4 f 2 + 0(14 ). (20)
This is the basic (Richardson 1908) result that the difference between the
second- and fourth-order solutions on the same grid is itself a fourth-order
error estimate for the second-order solution (a result so obvious that it hardly
deserves to be named a theorem). Likewise obvious, if we define the error of
the fourth-order solution E 4 to be
E4 = f f4 (21)
and substituting for f from Equation 18, we obtain
E 4 = f 2 f 4 + 0(12 ) (22)
for a second-order error estimator for the fourth-order solution.

The somewhat less obvious result is for the directional splitting of the higher-
order solutions, which can be much easier to implement than a fully directional
higher-order solution. Let f 4x denote the solution obtained with fourth-order
discretization in x and second-order in y, and f 4y denote the solution obtained

with fourth-order discretization in y and second-order in x. The C2x coefficients
are unchanged from the previous definitions. Then
f = f 4x + C4x 1x 4 + R4x + C2y 1y 2 + R2y (23)
and
f = f 4y + C4y 1y 4 + R4y + C2x 1x 2 + R2x . (24)
We now estimate the difference between f 4x and f 2 , using Equations 23 and 17,
respectively (if we dropped the higher-order remainder terms at this point, we
would prove the resulting estimator to be second-order accurate, but by retaining
these terms presently, we show that the estimator is fourth-order accurate)
f 4x f 2 = f C4x 1x 4 R4x C2y 1y 2 R2y
{ f C2x 1x 2 R2x C2y 1y 2 R2y } (25)
= C2x 1x 2 + R2x C4x 1x 4 R4x .
Similarly, we estimate the difference between f 4y and f 2 using Equations 24
and 17, respectively:
f 4y f 2 = f C4y 1y 4 R4y C2x 1x 2 R2x
{ f C2x 1x 2 R2x C2y 1y 2 R2y } (26)
= C2y 1y 2 + R2y C4y 1y 4 R4y .
Adding Equations 25 and 26 gives
f 4x + f 4y 2 f 2 = C2x 1x 2 + R2x + C2y 1y 2 + R2y
(27)
C4x 1x 4 R4x C4y 1y 4 R4y .
In comparison with Equation 17, the first four terms on the right-hand side of
Equation 27 are identically equal to f f 2 , without approximation. (That
is, the remainder terms R2x and R2y are included.) The last four terms on the
right-hand side of Equation 27 are O(14 ). With the definition given in Equation
19, we have the error of the second-order solution E 2 to be estimated as
E 2 = f 4x + f 4y 2 f 2 + O(14 ). (28)
Thus the estimate for the second-order solution can be obtained to fourth order
by directionally split fourth-order solutions.
Error Estimation from Lower-Order Solutions
on the Same Grid
Certainly a lower-order solution on the same grid could be used to estimate the
uncertainty of the higher-order solution (Category B.2), in the same way that
grid coarsening can be used rather than grid refinement, but I know of no such
applications.
146 ROACHE
Residual Evaluation and Transport

In one type of Category C method (auxiliary PDE solutions on the same grid)
and in Category D methods as well, residuals of the solution are evaluated.
In the Category C methods, these residuals are used as a source term to a
linearized error equation that mimics the governing equations. That is, the
errors are advected, diffused, and so on. In the various flavors of this approach,
much is made of the distinction between exact error residual equations and
approximate error residual equations. There is less than meets the eye in this
distinction, because somewhere in the development, additional approximations
are required. Indeed, there can be no exact evaluation of the error of the discrete
point-wise solution by passing it through the continuum operator(s) of the
governing PDEs, since the domain of the PDEs requires more than point-wise
values. The philosophy of FEM, or the underlying conceptual model of FEM
practitioners, is usually distinct from FVM and FDM in this regard, often leading
to disconnects in communications. To most FDM and FVM practitioners, the
numerical solution consists of point-wise, discrete values of the flow variables
(as well as related solution functionals, such as drag coefficients and so on,
obtained by numerical quadratures of the discrete values.) To FDM and FVM
people, the term residual ordinarily refers to the algebraic residual, the leftovers
from the numerical solutions of the discrete algebraic equations, that will result
from incomplete iterative solution or from accumulated round-off errors in a
direct matrix solver. Therefore, it makes no sense to FDM or FVM people
to speak about estimating the discretization error of a PDE solution, which
involves real-number system arithmetic, by evaluating the (algebraic) residual,
which may be made arbitrarily small no matter how crude and inaccurate the
discretization may be. (In fact, the cruder the discretization, the easier it is to
reduce this algebraic residual.)
FEM practitioners have a different viewpoint. They invariably conceptualize
the numerical solution as more than point-wise, since the selection and con-
struction of the inter-element basis functions is an essential part of the FEM
methodology. Thus, they conceptualize their numerical solution as a continuum
solution and, thence, are led to consider the (differential) residual resulting from
passing this solution function through the PDE operator. When an FEM author
talks about residuals, he or she refers to what I distinguish here as differential
residuals. (Incidentally, I believe that this difference in attitude and semantics
can be traced to the historical accident that FDM practice grew up with iter-
ative solvers, whereas FEM practice grew up with direct solvers.) Note the
major distinction between these approaches: The second (differential) residual
would be nonzero even if the computer used true real-number arithmetic, but
the first (algebraic) residual actually can be driven to machine zero in many
(not necessarily all) algorithms/codes.
Once this useful semantic distinction is made, we can examine the differ-
ences in the approach. The practices are indeed different, but the difference is
not as essential as one might be led to think. First of all, we note that most
(moderate-order) FEM solutions, although they produce functions, cannot be
passed through the PDE operator everywhere because they are not smooth at
element boundaries. This limitation can be sidestepped by using the weak or
integral form of the governing equations. Note, however, that conceptualizing
the solution as a continuum function engenders some problems. For example,
the solution may not be truly (locally and globally) conservative, that is, con-
serving of mass, momentum, and/or energy, even though the conservation form
(Roache 1972) of the equations was used. Also, when these FEM solutions
are graphed for presentation of results, and often when they are postprocessed
for evaluation of solution functionals such as drag coefficients, the basis func-
tions are not used. The quadratures can use the same algorithms used for FDM
solutions or for any point-wise data. Conversely, on the FDM/FVM side, it
is not necessary to conceptualize the solution as consisting of only the dis-
crete (node or volume) points. Indeed, a simple methodology for deriving a
finite difference method involves fitting polynomials through node point values
(e.g. a parabolic fit leads to standard second-order centered difference sten-
cils). Therefore, there is no reason why FEM approaches to error estimation
(either for solution adaptive grids or for quantification of uncertainty) cannot
be applied to FDM solutions.
All of this discussion is preamble to the most interesting aspect of residual
evaluation. Regardless of the approach (FEM, FDM, FVM), if a (differential)
residual is to be evaluated for error evaluation, it must be based upon a functional
form distinct from that used in the discretization that produced the solution.
If not, the differential residual evaluation will be practically zero, except for
some algebraic garbage caused by incomplete iterative convergence or machine
round-off. Although arguments may be made about which functional form to
use for the (differential) residual evaluation, they all essentially work for error
estimation and, therefore, the process (a) is somewhat arbitrary, and (b) applies
to FDM and FVM as well as FEM.
Note that each conceptualization of the underlying (sub-cell) functional form
of the solution leads to a discretization, but the discretization is not necessarily
unique to the conceptualization. For example, the variable coefficient diffusion
problem can be conceptualized as having smooth property variation or piece-
wise constant property variation. (In groundwater flow and transport problems,
148 ROACHE
the first conceptualization typically is held by geophysicists and engineers, and

the second by geologists.) Nevertheless, for a single grid, both conceptualiza-
tions can lead to the identical discretization and therefore identical discrete solu-
tion. This suggests some tolerance in the construction of sub-cell solutions. To
put it another way, just because your point-wise solution agrees with experiment
does not necessarily prove the literalness of your underlying conceptualization.
Auxiliary PDE Solutions on the Same Grid
The procedure for error estimation in Category C (auxiliary PDE solutions
on the same grid) then involves selection of an alternate functional form for
the intercell (interelement) solution and evaluation of the (differential) residual.
Again, the derivatives may not exist at element boundaries. Whether one selects
a C2 smooth function to begin with, or blends (interpolates) the differentials
obtained at element (or cell) centers to produce smooth functions over the
entire region, or puts off such questions to the quadrature rules, etc, is not very
significant. The residuals can then be used as local error estimators (Category D)
or transported in an error equation (Category C), which is indeed a significant
distinction. The major point is that the residual evaluation must involve a
discretization rule different from that used in obtaining the solution. In this
way, these methods are after all not so distinct from Category B methods, which
use higher- (or lower-) order stencils to evaluate the error. In Category C, a
higher-order stencil is used to obtain a higher-order solution, but if the stencil
were used only locally to obtain a local residual, it would be like Category D.
The approach in which a rigorously iterated solution is replaced with a less
completely converged solution is intermediate, and the globalness of the error
estimate becomes dependent on the amount and type of iteration.
For examples of the error transport equation approach, see Ferziger (1993)
and Van Straalen et al (1995). In all cases, the error transport equation will be
some linearized version of the governing equations. It will be somewhat cheaper
to solve than the original fluid dynamics equations (e.g. full Navier-Stokes) but,
as expected, will be a less accurate error estimator near separations and so on.
This approach is truly global but not as reliable as Category A or B methods.
Auxiliary Algebraic Evaluations on the Same Grid;
Surrogate Estimators
Category D methods, involving algebraic evaluations on the same grid, are
cheap, need no additional grid generation, and use no significant dynamic
memory. However, the error evaluated usually has no direct relation to any
error measure of engineering or scientific interest. Rather, the error measure
is defined for mathematical convenience (as are the L 2 and L norms) or by
physical intuition (as is kinetic energy conservation). These measures can be
used as a surrogate for measures of engineering and scientific interest if and
only if a correlation is established via numerical experimentation over a suite

of problems and range of problem parameters.
Nonconservation of Conservation Variables

Category D.1 (nonconservation of conservation variables) applies only to codes
that do not use fully conservative algorithms for conservation variables. For
example, mass is not identically conserved in most pre-1980 boundary-layer
codes, in many FEM codes, as well as other CFD codes. The error estimate
then involves numerical evaluation by quadrature for the erroneous loss or gain
of mass in the computational domain. If the coded algorithm is consistent,
this error will approach zero but only in the limit of 1 0 for a nontrivial
problem. (Even for a nominally mass-conserving full Navier-Stokes code, the
satisfaction of mass conservation will usually depend on the degree of strict
iterative convergence achieved, but this gives no indication of the discretization
error of the solution, which is our interest here.) Momentum, vorticity, and
internal energy are other possibilities. Note that the evaluation of this error
depends on the accuracy of the quadrature, which probably should be consistent
with the algorithm for solving the PDEs (see also discussion of Category D.4
below), but this does not appear to be a strict requirement. Conservative codes
and algorithms are generally preferred, but nonconservative codes are seen to
offer a readily evaluated and understood (although still surrogate) error measure.
Nonconservation of Higher Moments

Category D.2 involves evaluation of conservation errors for higher-order mo-
ments. In typical turbulent (Reynolds-Averaged) Navier-Stokes codes, fully
conservative discretization may be used for the primitive variables of mass,
momentum, and internal energy. However, turbulent kinetic energy is typically
not identically conserved. An evaluation of its global and local conservation by
quadrature, including carefully evaluated boundary inflow terms and dissipation
(Haworth et al 1993; Chang & Haworth 1995, 1996), then gives a surrogate
indication of general discretization errors. In Chang & Haworth (1995), it
is also used to reliably guide local grid refinement for solution-adaptive grid
generation.
In an earlier section, I noted that, if there is any suspicion that the grid reso-
lution is not in the asymptotic range, three grid solutions are necessary to verify
(or determine) the rate of convergence and thereby estimate the error, when the
exact solution is not known. The qualifier used, when the exact solution is
not known, may seem redundant when the context is a realistic problem, but
there is an important and useful distinction to be made in regard to Category
D.1 and D.2 methods that are based on conservation errors. Although the con-
servation imbalance may not be of direct engineering or scientific relevance, it
150 ROACHE
has the advantage that the exact solution of this quantity is known, namely zero.
Thus, in this surrogate error measure, one requires only a single grid solution
to calculate the error if p is known, and only two grid solutions to extract the
observed convergence rate p (see Chang & Haworth 1996).
Zhu-Zienkiewicz-Type Estimators
Category D.3 methods (the Zhu-Zienkiewicz-type estimators) were developed
and intended primarily to guide grid adaptation. See Zhu & Zienkiewicz (1990),
Zienkiewicz & Zhu (1987, 1992), and other FEM methods that have a similar
flavor, such as those of Strouboulis & Oden (1990), Oden et al (1993), Babuska
et al (1994), and Ewing et al (1990). These methods allow the global energy
norm to be well estimated and give good evaluation of local errors (and provide
local estimate of stress accuracy, certainly important for structures problems).
More relevant in the present context of the quantification of uncertainty, these es-
timators can be used as surrogate error estimators, and they are relatively cheap.
Although developed for FEM, the Zhu-Zienkiewicz approach is adaptable
to FDM and FVM (D Pelletier, personal communication), using the concepts
discussed above in the section Residual Evaluation and Transport.
The Zhu-Zienkiewicz-type methods share the shortcoming of all surrogate
estimators for fluid dynamicsother than guidance for grid adaptation, there is
little inherent engineering or scientific interest in the error measure as defined.
Therefore, unless the only interest is mathematics for its own sake, it is neces-
sary to establish a correlation of the Zhu-Zienkiewicz estimator with an error
measure of interest. This can be accomplished only by expensive numerical
experimentation for a given class of problems. For example, a correlation based
on numerical experiments for internal combustion engine modeling would be
unlikely to provide any guidance for external aerodynamics, nor even for a large
range of flow parameters for geometrically similar problems. However, when
one is involved in extensive suites of calculations, one may build up such correla-
tions by experience and arrive at a practical and very inexpensive error estimator.
This need for establishing a correlation in error measures is also true if one
direct error measure (e.g. obtained by a grid-convergence study) is to be used
as a surrogate for another. For example, in an airfoil calculation, does a 1%
error estimate on C L ensure a 1% (or 5%, and so on) error estimate for C M ?
The correlation providing the (fuzzy) answer will be valid only for a restricted
range of angle of attack, Reynolds number, Mach number, and airfoil class.
Convergence of Higher-Order Quadratures

Category D.4 methods are the simplest, involving convergence of higher-order
quadratures (numerical integration of a known function). For example, quadra-
ture for pressure and shear force on the surface leads to evaluation of a drag
coefficient C D . For nontrivial problems (i.e. problems with significant solution

structure), second-order quadrature will give a different answer than fourth-
order accurate quadrature. As the discretization refines, these two quadratures
will converge to each other. For coarse grids, the difference may be used as a
surrogate error estimator. The same philosophy can be used point-wise. The
simplest approach ultimately reduces to comparing the calculated point-wise
value of some variable of interest with the value interpolated between the points
neighbors. Indeed, these values will converge as the solution converges and,
therefore, can be (and have been) used as surrogate error estimates. However,
the poverty of the concept is shown by the fact that the error estimate is obtained
from a single solution without recourse to the governing PDEs. It is difficult to
take seriously an error estimate for point-wise pressures from an incompressible
Navier-Stokes code when the same error estimate would result from the same
point-wise values in an M = 20 solution or in Darcy flow in a porous medium.
In fact, since the governing PDEs are not invoked, these error estimators would
be better named simply resolution indicators.
Time Accuracy Estimation

Estimation of the numerical error of the time discretization can be done in
the same manner as the spatial errors, but simpler methods may also be used.
In fact, it is relatively straightforward to estimate the temporal error as the
calculation evolves, and to build a code with a solution-adaptive time step to
control the temporal error to a predetermined level, even when only first-order
time differencing is used.
Consider the following inexpensive temporal error estimator (Oden et al
1993; Roache 1993c, d) for a fully implicit (backward) time differencing
method. We write the backward time method for a general equation system in
terms of an operator L (not necessarily linear) as
n+1
f i, j f i,n j 1t = L n+1 . (29)
Generally, the solution will involve an expensive matrix solution for all f n+1 .
The time-error estimator uses the difference between backward and forward
time integration, implemented as an extrapolation. The method is very cheap to
implement because it does not require another implicit matrix solution, nor even
another explicit stencil evaluation. It includes the effects of time-dependent
boundary conditions and source terms.
In the current time-step, Equation 29 is advancing the solution for f from time
level n to time level (n + 1) with increment 1t using fully implicit (backward)
time differencing, so that the operator L is being evaluated at (n + 1). In
the previous time-step, the solution was advanced from (n 1) relative to the
152 ROACHE
current indexing to n with increment 1tOLD and the L evaluated at n. We could

explicitly evaluate L n and make a separate, parallel estimate of the values f n+1
with an explicit step, as in
n+1
Fi, j f i,n j 1t = L n (30)
where F n+1 signifies the new value of f predicted by the explicit algorithm.
(Since the left-hand side involves values of f only at the location (i, j) but not
at the neighboring locations (i 1, j) and (i, j 1) the equation is explicit,
that is, it does not involve a matrix solution of f at all values of i and j
simultaneously.) The difference between the new value of f n+1 predicted by
the implicit algorithm and F n+1 predicted by the explicit algorithm, both of
which predictions are O(1t) accurate, is itself an error estimator of accuracy
O(1t 2 ) for the time discretization error for that time step. Explicit evaluation
of L n would not be expensive compared with the computer time necessary for
the implicit matrix solution, but it does involve coding storage penalties and
complexities (storage of old values of boundary conditions, source terms, and
so on). An economical and elegant approach is to recognize that the L from the
previous implicit step is identical to the L for the present explicit step. Thus,
the L for the present explicit step can be evaluated from the knowledge of the
previous change in f , requiring only the temporary storage of previous solution
arrays. The explicit solution for f at (n + 1), F n+1 , is then obtained by simple
linear extrapolation of previous solutions. For constant 1t,
j = 2 f i, j f i, j .
Fi,n+1 n n1
(31)
For the more general case of variable 1t,
1t n
j = f i, j +
Fi,n+1 f i, j f i,n1 .
n
(32)
1tOLD j
The point-wise difference F n+1 f n+1 is then a point-wise temporal error

estimator. Usually, a user would be interested in the maximum over the spatial
domain of the percentage error, so that one would evaluate the error estimator
E ET for the single time-step as

E E T = 100 max abs Fi,n+1
j f i, j
n+1
f RANGE . (33)
i, j
E E T is thus calculated as the percent maximum deviation (or L norm) of the

absolute value of the difference between the new f values f n+1 , predicted by
the fully implicit algorithm, and F n+1 , predicted by the explicit algorithm, nor-
malized by f RANGE , which is the total range of f n+1 . (In practice, ghost-point
evaluations with Dirichlet boundary conditions tend to exaggerate the error, so
error estimates and f RANGE should be calculated only over interior points.)
Note that the explicit calculation is used only as an error estimator within
a time step of an implicit method, not as the solution algorithm. That is, its
effects do not accumulate. Thus, stability limitations and/or conservation issues
of explicit time-stepping as a solution algorithm are irrelevant.
Although the extrapolation procedure is equivalent to explicit time-stepping,
the extrapolation cannot be started until there are two time levels. Also, in the
event that the initial conditions are set arbitrarily by the analyst (without setting
initial conditions as a steady-state solution), the initial conditions are likely
to be incompatible with the boundary conditions applied at the first time-step.
This means that the change in boundary values during the first time-step is
approximately fixed, that is, it does not depend much on the time resolution.
Consequently, the first time-step would not provide a meaningful estimate of
f /t, and the error estimator would be invalid.
E E T can be used readily (internal to the code) as the basis for a solution-
adaptive time-stepping algorithm, adjusting 1t so that the error estimate E E T
is acceptable, either by recalculating the previous time-step (preferred) or by
simply adjusting the next time-step.
Nonlinear Dynamics Solutions

Yee et al (1991) and Lafon & Yee (1992) have given a large number of examples
of nonlinear (chaotic) dynamics solutions in which spurious (grossly erroneous)
solutions for steady-state problems with strongly nonlinear source terms are
obtained for stable implicit algorithms applied beyond the linear stability limit.
Their point is the danger of being misled by such calculations, and the need
for further research to guarantee a priori accuracy. I believe the real lesson
of the examples is simply the need for a posteriori error estimation. In all
but one of the examples, the most cursory time-step convergence study would
quickly reveal the inadequacy of the temporal resolution, with no reasonable
chance of being misled into a nonphysical solution. The one remaining example
(involving a fourth-order Runge-Kutta integration) could be misleading with a
cursory study but would be revealed in a thorough systematic study. However, it
is worth noting and is perhaps nonintuitive that even steady-state solutions may
require time-step resolution (or equivalently, relaxation parameter resolution)
when significant nonlinearities are present.
Cafe Curves
Different fields of endeavor have different requirements for the variables and
functionals of interest. To generalize too much, engineering (mechanical,
aerospace, chemical) often requires functionals of the solution such as lift coef-
ficient, heat transfer rates, mixing rate, and so on. It is much less common that
practical interest is in accuracy of the entire field calculation; one example that
154 ROACHE
comes to mind is aero-optical propagation (Truman & Lee 1990). However,

meteorological and ocean modelers are usually interested in accuracy of all the
field variables. In these cases, it is meaningful to emphasize global accuracy of
the entire computational field, and difficulty arises in attempting to characterize
global accuracy with just one or a few numbers.
Often L 2 or L error norms are used as global error indicators. However,
questions arise about whether these error norms are always reliable, and more
significantly how they can be interpreted in any general sense. For example, in a
tidal convergence study (Luettich & Westerink 1995, Westerink et al 1994), L
norms are poor indicators of error because shifting amphidromes, associated
with very high gradient in response amplitude and phase, cause extremely high
localized errors that do not converge smoothly or at the same rate as an L 2 norm.
Furthermore, there is the general problem of interpreting how representative an
L 2 or L norm is to the response as a whole, or to aspects of the solution of
particular interest.
Luettich & Westerink (1995) introduced the concept of cumulative area frac-
tion error (CAFE) curves to present domain errors in a more complete and
meaningful way. These curves plot the fraction of the total domain that ex-
ceeds a particular error level (y-axis) against that error level (x-axis). Under-
and over-prediction are indicated separately. Assuming that under- and over-
prediction are distributed approximately evenly throughout the domain, median
under- and over-prediction errors correspond approximately to a cumulative
area fraction of 0.25. For an example of the cumulative-area-fraction error
curves, see the grid intercomparisons for tidal flow computations reported by
Westerink et al (1994); see also Westerink & Roache (1995).
The Myth of the Converged Solution

CFD practitioners, especially industrial engineers using general-purpose com-
mercial CFD codes, would like to be able to obtain, at least in principle, a con-
verged solution (converged both iteratively and in time-space discretization)
and then be able to confidently postprocess that solution for any engineering
measure of interest. Unfortunately, this is not possible in any general sense.
Again, the problem is the a priori unknown correlation between possible error
measures of valid engineering and scientific interest. Of course, expert judg-
ment (based on experience with a nearby class of problems) can be valid, but
the problem is not soluble in any general sense.
For example, in our experience with dynamic stall and oscillating airfoil
and wing calculations (Salari & Roache 1990, Salari et al 1994), lift was well
converged at the finest discretization achieved (141 55 55 in 3-D, and
461 71 in 2-D), drag was more problematical, and moment definitely was
not converged. If one had performed the grid-convergence study (or other
error-estimation exercise) examining only the lift, one could only infer from
engineering judgment (experience on related problems) whether or not reso-
lution was adequate for drag and moment. If a new error measure became of
interest, say position of transition, or location of trailing-edge separation, or
second harmonic component of the unsteady pressure coefficient (of interest
for helicopter dynamic stall analysis; McCroskey 1981), the convergence study
would have to be repeated for this error measure, strictly speaking. Calculations
of the branching between symmetric and nonsymmetric flow patterns require
special attention to grid resolution (see Sengupta et al 1995). Rosenfeld (1994)
used grid resolution studies of bluff-body wakes with resolution up to 513 513
points, and examined convergence in both the physical and Fourier domains; he
showed that phase velocities of vortices converge much slower than amplitudes,
so that coarser grids may be acceptable for prediction of force coefficients.
Certainly, it is always possible to devise some error measure that is exquisitely
sensitive to discretization error, for example, some high-order statistical corre-
lations in turbulent flow calculations, so that this measure is far from converged
even when other, more benign measures are well converged. A practical exam-
ple is a boundary-layer stability calculation. For most engineering applications,
skin friction and wall-heat transfer are of principal interest, and these are sen-
sitive to the first normal derivative of velocity at the wall. However, it is known
from stability theory (Lin 1967) that laminar-boundary layer stability is de-
pendent on the diffusion of vorticity across the critical layer (the y-position at
which the mean flow speed equals the disturbance wave speed). The principal
component of boundary-layer vorticity magnitude is u/ y, which means that
the appropriate measure of accuracy is of the term 3 u/ y 3 , which may be
expected to be more difficult to converge than simply the velocity component
u. Conversely, accurate convergence of pressure may be unnecessary.
Thus, the concept of a converged solution, ascertained to be so independent
of the intended error measure, is a myth.
Error Estimation for Grid Adaptation versus Quantification
of Uncertainty
Local error estimators are used for solution-adaptive grid-generation algo-
rithms, and they are usually successful for this purpose (e.g. see Schonauer
et al 1981, Oden et al 1993). However, almost anything intuitive is successful
for adaptation purposes, e.g. minimizing solution curvature or adapting to so-
lution gradients (even though solution gradients per se cause no error in most
discretization schemes). In 1-D problems (e.g. Salari & Steinberg 1994) or
quasi-1-D problems (Dwyer et al 1980), the gains in computational efficiency
from solution adaptivity of the r -type (redistribution) in structured grids are
very impressive. In real multidimensional problems, gains are usually modest
156 ROACHE
(Roache et al 1984, Hall & Zingg 1995) but are somewhat more significant for
unstructured grid adaptation (e.g. see Morgan et al 1991, Lohner 1989, Hetu
& Pelletier 1992, Pelletier & Ignat 1995).
But this task of grid adaptation has little connection to the quantification of
uncertainty for a final calculation with a useful error measure. So the success
of these local error estimators in guiding grid adaptation must not be taken as
demonstration of their efficacy for the quantification of uncertainty. Estimates
of local errors usually are not what we want for verification of calculations;
we need global errors. By global I do not mean just a global summing up
of local values (as is often used in the FEM literature) but an evaluation that
includes nonlocal effects, that is, one taking into account the fact that errors
are advected, diffused, and so on. For error estimation of useful scientific or
engineering measures, local estimates are suspect, and their validity as surrogate
estimators for measures of interest must be established anew for each family of
nearby problems.
Internet Archive
An Internet archive of a bibliography for CFD Verification and Validation
is maintained by R Barron of the University of Windsor, Ontario, Canada,
and can be accessed at the following address: http://www.lpac.ac.uk/SEL-
HPC/Articles/ValCFD.html
Final Remarks
The various approaches to error estimation and quantification of uncertainty in
CFD have their relative merits, involving algebraic simplicity of local estimators
vs the complexity of additional PDE formulations or higher-order accurate so-
lutions, single-grid vs multiple-grid generations, postprocessing vs CFD code
modifications, and so on. Systematic grid-convergence studies are the most
common, most straightforward and arguably constitute the most reliable tech-
nique for the quantification of numerical uncertainty. This approach requires no
code modifications nor algorithm developments but does require multiple grid
generations. Significantly, it is not necessary to double the grid; noninteger
grid refinement and grid coarsening are economical alternatives.
However, regardless of each approachs relative merits and of the possibility
for future improvements, it is clear that methods are available now to convinc-
ingly assess the numerical uncertainty of CFD calculations. [It is certainly
not necessary, as claimed in a recent SIAM article by Johnson et al (1995), to
analytically and a priori solve the hydrodynamics stability problem in order
to estimate CFD numerical errors!] Journal editors and reviewers can insist
on reasonably thorough assessment without placing impractical demands on
authors, to the improvement of journal quality.
ACKNOWLEDGMENTS
This work was partially supported by Sandia National Laboratories and the
United States Department of Energy under Contract DE-AC04-76DP00789,
Dr. M Fewell, Technical Monitor.
Visit the Annual Reviews home page at

http://www.annurev.org.
Literature Cited
Aeschliman DP, Oberkampf WL, Blottner FG. ances. Int. J. Num. Meth. Fluids. In press
1995. A proposed methodology for compu- Claus RW, Vanka SP. 1992. Multigrid calcula-
tational fluid dynamics code verification cal- tions of a jet in crossflow. J. Propulsion Power
ibration and validation. Proc. Int. Congr. In- 8:18593
strumentation Aerospace Simulation Facil- Cosner RR. 1995. CFD Validation requirements
ities. Dayton, Ohio: Wright-Patterson Air for technology. AIAA 26th Fluid Dynamics
Force Base Conf. San Diego, CA
AIAA. 1994. Policy statement on numerical ac- de Vahl Davis G. 1983. Natural convection of air
curacy and experimental uncertainty. AIAA J. in a square cavity: a bench mark numerical
32(1):3 (editorial) solution. Int. J. Num. Meth. Fluids 3(3):249
ASME Editorial Board. 1994. Journal of Heat 64
Transfer editorial policy statement on nu- Dietrich DE, Roache PJ, Marietta MG. 1990.
merical accuracy. ASME J. Heat Transfer Convergence studies with the Sandia Ocean
116(11):79798 Modeling System. Int. J. Num. Meth. Fluids
Babuska I, Strouboulis T, Upadhyay CS. 1994. 11:12750
A model study of the quality of a posteriori Dwyer HA, Kee R, Sanders B. 1980. Adaptive
error estimators for linear elliptic problems. grid method for problems in fluid mechanics
Error estimation in the interior of patchwise and heat transfer. AIAA J. 18:120512
uniform grid of triangles. Comp. Meth. Appl. Ethier CR, Steinmen DA. 1994. Exact fully 3D
Mech. Engineer. 114:30778 Navier-Stokes solutions for benchmarking.
Blottner FG. 1990. Accurate Navier-Stokes re- Int. J. Num. Meth. Fluids 19:36975
sults for the hypersonic flow over a spher- Ewing RE. 1990. A posteriori error estimation.
ical nosetip. AIAA J. Spacecraft Rockets Comp. Meth. Appl. Mech. Engineer. 82:59
27(2):11322 72
Boehm BW. 1981. Software Engineering Eco- Ferziger JH. 1993. Estimation and reduction of
nomics. New York: Prentice-Hall numerical error. See Celik et al 1993, p. 18
Celik I, Chen CJ, Roache PJ, Scheurer G, eds. Freitas CJ. 1993a. Editorial policy statement on
1993. Symposium on Quantification of Un- the control of numerical accuracy. ASME J.
certainty in Computational Fluid Dynamics. Fluids Engineer. 115(3):33940
FED-ASME, Vol. 158. New York: Am. Soc. Freitas CJ, ed. 1993b. The CFD Triathlon:
Mech. Eng. Three Laminar Flow Simulations by Com-
Celik I, Freitas CJ, eds. 1990. Benchmark Test mercial CFD Codes. FED-ASME, Vol. 160.
Cases for Computational Fluid Dynamics. New York: Am. Soc. Mech. Eng.
FED-ASME, Vol. 93. New York: Am. Soc. Freitas CJ. 1994. Response: To the Comments
Mech. Eng. by Drs. W. Shyy and M. Sindir. ASME J. Flu-
Celik I, Zhang WM. 1995. Calculation of nu- ids Engineer. 116(2):198
merical uncertainty using Richardson extrap- Freitas CJ. 1995a. Perspective: Selected bench-
olation: application to some simple turbulent marks from commercial CFD codes. ASME
flow calculations. ASME J. Fluids Engineer. J. Fluids Engineer. 117(2):20818
117:43945 Freitas CJ. 1995b. The end of justification, the
Chang S, Haworth DC. 1995. Kinetic-energy- beginning of implementation. ASME J. Flu-
balance based solution-adaptive mesh refine- ids Engineer. 117(1):9
ment. See Johnson & Hughes 1995, pp. 712 Gresho PM, Taylor C. 1994. Int. J. Num. Meth.
Chang S, Haworth DC. 1996. Adaptive grid re- Fluids 19:iii (editorial)
finement using cell-level and global imbal- Hall DJ, Zingg DW. 1995. Viscous Airfoil Com-
158 ROACHE
putations Using Adaptive Grid Redistribu- AIAA Paper 89-1670

tion. AIAA Journal 33 (7):1205-10 Mehta U. 1991. Some aspects of uncertainty in
Haworth DC, El Tahry SH, Huebler MS. computational fluid dynamics results. ASME
1993. A global approach to error estimation J. Fluids Engineer. 113:53843
and physical diagnostics in multidimensional Mehta U. 1995. Guide to credible computa-
computational fluid dynamics. Int. J. Num. tional fluid dynamics simulations. AIAA Pa-
Meth. Fluids 17:7597 per 95-2225
Hetu JF, Pelletier DH. 1992. Adaptive remesh- Melnik RE, Siclari MJ, Marconi F, Barger T,
ing for viscous incompressible flows. AIAA Verhoff A. 1995. An overview of a recent in-
J. 30:198692 dustry effort at CFD code certification. AIAA
Horgan J. 1995. From complexity to perplexity. Paper 95-2229
Sci. Am. 272(6):1049 Morgan K, Peraire J, Peiro J, Hassan O, Probert
Johnson C, Rannacher R, Boman M. 1995. Nu- EJ. 1991. Unstructured Grid Methods for
merics and hydrodynamic stability: toward Compressible Flows. In Computational Me-
error control in computational fluid dynam- chanics, ed. YK Cheung, JHW Lee, AYY
ics. SIAM J. Num. Anal. 32:105879 Leung, 2:148997. Rotterdam: Balkema
Johnson RW, Hughes ED, eds. 1995. Quantifi- Nguyen TV, Maclaine-Cross IL. 1988. Incre-
cation of Uncertainty in Computational Flu- mental pressure drop number in parallel-
ids Dynamics1995. FED-ASME, Vol. 213. plate heat exchangers. ASME J. Fluids Eng.
New York: Am. Soc. Mech. Eng. 110:9396
Karniadakis GE. 1995. Toward a numerical er- Oberkampf WL. 1994. A proposed framework
ror bar in CFD. ASME J. Fluids Engineer. for computational fluid dynamics code cali-
117(1):79 bration/validation. AIAA Paper 94-2540
Knupp P, Steinburg S. 1994. Fundamentals of Oberkampf WL, Blottner FG, Aeschliman
Grid Generation. Boca Raton: CRC DP. 1995. Methodology for Computa-
Konikow LF, Bredehoeft JD. 1992. Groundwa- tional Fluid Dynamics Code Verifica-
ter models cannot be validated. Adv. Water tion/Validation. AIAA Paper 95-2226
Res. 15:7583 Oden JT, Kennon SR, Tworzydlo WW, Bass
Kuruvila G, Anderson J. 1985. A study of the JM, Berry C. 1993. Progress on adaptive h-
effects of numerical dissipation on the calcu- p finite element methods for the incompress-
lations of supersonic separated flows. AIAA ible Navier-Stokes equations. Comput. Mech.
Paper 85-0301 11:42132
Lafon A, Yee HC. 1992. On the numerical treat- Oldenburg CM, Pruess K. 1995. Dispersive
ment of nonlinear source terms in reaction- transport dynamics in a strongly coupled
convection equations. AIAA Paper 92-0419 groundwater-brine flow system. Water Re-
Leijnse A, Hassanizadeh SM. 1994. Model def- sources Res. 31(2):289302
inition and model validation. Adv. Water Res. Oreskes N, Shrader-Frechette K, Belitz K. 1994.
17:197200 Verification validation and confirmation of
Leonard BP. 1995. Comments on the policy numerical models in the earth sciences. Sci-
statement on numerical accuracy. ASME J. ence 263:64146
Fluids Engineer. 117(1):56 Organization for Economic Cooperation and
Lin CC. 1967. The Theory of Hydrodynamic Development. 1988. The International HY-
Stability. London: Cambridge Univ. Press DROCOIN Project Groundwater Hydrology
Lohner R. 1989. Adaptive remeshing for tran- Modelling Strategies for Performance As-
sient problems. Comp. Meth. Appl. Mech. En- sessment of Nuclear Waste Disposal Level
gineer. 61:32338 1: Code Verification. OECD Report 71617.
Lotz RD, Thompson BE, Konings CA, Paris: OEDC
Davoudzadeh F. 1995. Numerical uncertain- Organization for Economic Cooperation and
ties in transonic flow calculations for airfoils Development. 1990. PSACOIN Level 1A In-
with blunt trailing edges. Int. J. Num. Meth. tercomparison. Paris: OECD
Fluids. In press Patankar SV. 1980. Numerical Heat Transfer
Luettich RA, Westerink JJ. 1995. Continen- and Fluid Flow. New York: McGraw-Hill
tal shelf scale convergence studies with a Pelletier D, Ignat L. 1995. On the accuracy
barotropic tidal model. In Quantitative Skill of the grid convergence index and the Zhu-
Assessment for Coastal Ocean Models, ed. Zienkiewicz error estimator. See Johnson &
DR Lynch, AM Davies. Washington: Am. Hughes 1995, pp. 3136
Geophys. Union. In press Polya G. 1957. How to Solve It; A New As-
McCroskey WJ. 1981. The Phenomenon of Dy- pect of Mathematical Method. Princeton, NJ:
namic Stall. NASA TM-81264 Princeton Univ. Press
Mehta U. 1989. Computational requirements Putre M. 1994. Verifying analytical models.
for hypersonic flight performance estimates. ASME Mech. Engineer. 116(11):7074
Richardson LF. 1908. The approximate arith- Roache PJ, Steinberg S, Moeny WM. 1984. In-
metical solution by finite differences of phys- teractive electric field calculations for lasers.
ical problems involving differential equations AIAA Paper 84-1655
with an application to the stresses in a ma- Rosenfeld M. 1994. Grid refinement test of
sonry dam. Trans. R. Soc. London Ser. A, time-periodic flows over bluff bodies. Comp.
210:30757 Fluids 23(5):693709
Richardson LF. 1927. The deferred approach Salari K, Roache PJ. 1990. The influence of
to the limit. Trans. R. Soc. London Ser. A sweep on dynamic stall produced by a rapidly
226:229361 pitching wing. AIAA Paper 90-0581
Roache PJ. 1972. Computational Fluid Dynam- Salari K, Steinberg S. 1994. Flux-corrected
ics. Albuquerque, NM: Hermosa. Rev. 1976 transport in a moving grid. J. Comput.
Roache PJ. 1982. Semidirect/Marching Meth- Physics 111(1):2432
ods and Elliptic Grid Generation. In Proceed- Salari K, Blaine RL, Economy K, Roache PJ.
ings Symposium on the Num. Generation of 1995. Grid resolution studies of radionu-
Curvilinear Coordinate Systems and Use in clide transport in fractured porous media. See
the Numerical Solution of Partial Differen- Johnson & Hughes 1995, pp. 2530
tial Equations, ed. JF Thompson, pp. 72737. Salari K, Roache PJ, Wilcox DC. 1994. Numer-
Amsterdam: North-Holland ical Simulation of Dynamic Stall Using the
Roache PJ. 1990. Need for control of numerical Two-Equation K- Turbulence Model, ed. K
accuracy. J. Spacecraft Rockets 27(2):98102 Ghia, M Hafez. ASME-FED, 196:17. New
Roache PJ. 1993a. A method for uniform re- York: Am. Soc. Mech. Eng.
porting of grid refinement studies. See Celik Schonauer W, Raith K, Glotz G. 1981. The prin-
et al 1993, pp. 10920 ciple of difference of difference quotients as a
Roache PJ. 1993b. A method for uniform re- key to the self adaptive solution of nonlinear
porting of grid refinement studies. Proc. 11th partial differential equations. Comp. Meth.
AIAA Comput. Fluid Dynam. Conf., Orlando, Appl. Mech. Engineer. 28(3):32759
FL, 2:105758 Sengupta R, Orkwis PD, Davis SM. 1995. Un-
Roache PJ. 1993c. The SECO code algorithms steady solutions transients and local grid res-
for groundwater flow and transport. In Fi- olution requirements for asymmetric vortex
nite elements in fluids: new trends and ap- formation. AIAA Paper 95-0364
plications. Proc. VIII Int. Conf. Finite Ele- Shih TM, Tan CH, Hwang BC. 1989. Effects of
ments Fluids, ed. K Morgan, E Onate, J Pe- grid staggering on numerical schemes. Third
riaux, J Peraire, OC Zienkiewicz, 2:93948. Int. J. Num. Meth. Fluids 9(2):193212
Barcelona, Spain: Univ. Politecnica Cataluna Shirazi SA, Truman CR. 1989. Evaluation of
Roache PJ. 1993d. The SECO suite of codes algebraic turbulence models for PNS predic-
for site performance assessment. In Proceed- tions of supersonic flow past a sphere-cone.
ings 1993 Intl. High-Level Radioactive Waste AIAA J. 27(5):56068
Management Conf. Las Vegas, NV Shyy W, Sindir M. 1994. Comments on Policy
Roache PJ. 1994a. Perspective: a method for statement on the control of numerical accu-
uniform reporting of grid refinement studies. racy. ASME J. Fluids Engineer. 116(2):196
ASME J. Fluids Engineer. 116(3):40513 97
Roache PJ. 1994b. Response: To the Comments Steinberg S, Roache PJ. 1985. Symbolic manip-
by Drs. W. Shyy and M. Sindir. ASME J. Flu- ulation and computational fluid dynamics. J.
ids Engineer. 116(2):19899 Comput. Physics 57(2):25184
Roache PJ. 1995. Verification of codes and cal- Stephens AB, Shubin GR. 1981. Multiple so-
culations. AIAA Paper 95-2224;AIAI J. In lutions and bifurcation of finite difference
press approximations to some steady problems in
Roache PJ, Knupp PM. 1993. Completed fluid dynamics. SIAM J. Sci. Stat. Comput.
Richardson extrapolation. Comm. Appl. Num. 2(4):40415
Methods 9:36574 Strouboulis T, Oden JT. 1990. A posteriori error
Roache PJ, Salari K. 1990. Weakly compress- estimation of finite element approximations
ible Navier-Stokes Solutions with an implicit in fluid mechanics. Comp. Meth. Appl. Mech.
approximate factorization code. AIAA Paper Engineer. 78:20142
90-0235 Sudicky EA, Frind EO. 1982. Contaminant
Roache PJ, Ghia K, White F. 1986. Editorial transport in fractured porous media: analyti-
policy statement on the control of numerical cal solutions for a system of parallel fractures.
accuracy. ASME J. Fluids Engineer. 108(1):2 Water Resources Res. 18(6):163442
Roache PJ, Knupp PM, Steinberg S, Blaine RL. Truman CR, Lee MJ. 1990. Effects of orga-
1990. Experience with benchmark test cases nized turbulence structures on the phase dis-
for groundwater flow. See Celik & Freitas tortion in a coherent optical beam propagat-
1990, pp. 4956 ing through a turbulent shear flow. Physics
160 ROACHE
Fluids A 2(5):85157 with 40 CFR 191 Subpart B Vol. 2: Tech-

Vanka P. 1994. Response: To the Comments by nical Basis. Albuquerque, NM: Performance
Drs. W. Shyy and M. Sindir. ASME J. Fluids Assess. Dept., Sandia Natl. Lab.
Engineer. 116(2):19798 Yee HC, Sweby PK, Griffiths DF. 1991. Dynam-
Vanka P. 1995. Comment on Toward a numerical approach study of spurious steady-state
ical error bar in CFD. ASME J. Fluids Engi- numerical solutions of nonlinear differential
neer. 117(1):910 equations. I. The dynamics of time discretiza-
Van Straalen BP, Simpson RB, Stubley GD. tion and its implications for algorithm devel-
1995. A posteriori error estimation for finite opment in computational fluid dynamics. J.
volume simulations of fluid flow transport. In Comput. Physics 97:249310
Proc. Third Annu. Conf. CFD Soc. Canada, Zhu JZ, Zienkiewicz OC. 1990. Superconver-
Vol. 1, ed. PA Thibault, DM Bergeron. Banff, gence recovery technique and a posteriori er-
Alberta ror estimates. Int. J. Num. Meth. Engineer.
Westerink JJ, Luettich RA, Muccino JC. 1994. 30:132139
Modeling tides in the western North At- Zienkiewicz OC, Zhu JZ. 1987. A simple error
lantic using unstructured graded grids. Tellus estimator and adaptive procedure for practi-
46A:18799 cal engineering analysis. Int. J. Num. Meth.
Westerink JJ, Roache PJ. 1995. Issues in con- Engineer. 24:33757
vergence studies in geophysical flow compu- Zienkiewicz OC, Zhu JZ. 1992. The supercon-
tations. Soc. Industr. Appl. Math. In press vergent patch recovery and a posteriori error
Wilcox DC. 1993. Turbulence Modeling for estimates. I: The recovery technique. Int. J.
CFD. La Canada, CA: DCW Industries Num. Meth. Engineer. 33:133164
WIPP PA Dept. 1992. Annual Performance As- Zingg DW. 1993. Grid studies for thin-layer
sessment for the Waste Isolation Pilot Plant Navier-Stokes computations of airfoil flow-
Dec. 1992 Vol. 1 Preliminary Comparison field. AIAA J. 30(10):256164

Roache 1997

Uploaded by

Copyright:

Available Formats

Roache 1997

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Roache 1997

Uploaded by

Copyright:

Available Formats

November 28, 1996 12:21 Annual Reviews ROACHE AR23-05

Annu. Rev. Fluid. Mech. 1997. 29:12360

KEY WORDS: numerical uncertainty, verification, validation, error estimation, computational

QUANTIFICATION OF UNCERTAINTY IN CFD 125

order of accuracy and always consistently, so that as some measure of dis-

QUANTIFICATION OF UNCERTAINTY IN CFD 127

disposals, no bridges, no airplanes, no . . . .) See also Konikow & Bredehoeft

Inadequate Error Taxonomies

Taxonomy for Additional Information for Error Estimates

QUANTIFICATION OF UNCERTAINTY IN CFD 129

Sources of Additional Information for Error Estimation, Given

of numerical uncertainty. Unlike the other methods available, this approach

QUANTIFICATION OF UNCERTAINTY IN CFD 131

(e.g. grid doubling). The GCI is based on generalized Richardson Extrapolation

even if these are second-order accurate (3-point upstream), the h 2 extrapolation

QUANTIFICATION OF UNCERTAINTY IN CFD 133

additional stability analyses, and so on, especially in nonorthogonal coordi-

A significant yet overlooked disadvantage of Richardson Extrapolation is that

QUANTIFICATION OF UNCERTAINTY IN CFD 135

QUANTIFICATION OF UNCERTAINTY IN CFD 137

(For the alternative philosophy, see Oberkampf 1994, Oberkampf et al 1995,

A. In verification tests of a commercial groundwater flow code, a first-order

QUANTIFICATION OF UNCERTAINTY IN CFD 139

QUANTIFICATION OF UNCERTAINTY IN CFD 141

Special Considerations for Turbulence Modeling

QUANTIFICATION OF UNCERTAINTY IN CFD 143

and substituting for f from Equation 18, we obtain

for a second-order error estimator for the fourth-order solution.

QUANTIFICATION OF UNCERTAINTY IN CFD 145

discretization in x and second-order in y, and f 4y denote the solution obtained

Residual Evaluation and Transport

QUANTIFICATION OF UNCERTAINTY IN CFD 147

the first conceptualization typically is held by geophysicists and engineers, and

QUANTIFICATION OF UNCERTAINTY IN CFD 149

only if a correlation is established via numerical experimentation over a suite

Nonconservation of Conservation Variables

Nonconservation of Higher Moments

Convergence of Higher-Order Quadratures

QUANTIFICATION OF UNCERTAINTY IN CFD 151

coefficient C D . For nontrivial problems (i.e. problems with significant solution

Time Accuracy Estimation

current indexing to n with increment 1tOLD and the L evaluated at n. We could

The point-wise difference F n+1 f n+1 is then a point-wise temporal error

E E T is thus calculated as the percent maximum deviation (or L norm) of the

QUANTIFICATION OF UNCERTAINTY IN CFD 153

Nonlinear Dynamics Solutions

comes to mind is aero-optical propagation (Truman & Lee 1990). However,

The Myth of the Converged Solution

QUANTIFICATION OF UNCERTAINTY IN CFD 155

QUANTIFICATION OF UNCERTAINTY IN CFD 157

Visit the Annual Reviews home page at

putations Using Adaptive Grid Redistribu- AIAA Paper 89-1670

QUANTIFICATION OF UNCERTAINTY IN CFD 159

Fluids A 2(5):85157 with 40 CFR 191 Subpart B Vol. 2: Tech-

You might also like