Download as pdf or txt
Download as pdf or txt
You are on page 1of 3

Fast and Accurate

Approximations for the


Dag Biberg
Colebrook Equation
Schlumberger Norway Technology Center,
Gamle Borgenvei 3,
The standard expression for pipe friction calculations, the Colebrook equation, is in an
Asker 1383, Norway
implicit form. Here, we present two accurate approximate solutions, given by replacing
e-mail: DBiberg@slb.com
the numerically unstable term in Keady’s exact Lambert function solution with a trun-
cated series expansion. The resulting expressions have a higher accuracy than most
advanced approximations and a lower computational cost than basic engineering formu-
las. The simplest expression, given by retaining only three terms in the series expansion,
has a maximum error of less than 0.153% for Re  4000. The slightly more involved
expression, based on five terms, has a maximum error of 0.0061%.
[DOI: 10.1115/1.4034950]

1 Introduction
The standard expression for pipe friction calculations, the relates the frictional pressure loss DPF over a distance L to the
Colebrook [1] equation, is in an implicit form. This has led to the fluid density q and bulk velocity U. The force balance over the
development of a large number of explicit approximations over corresponding pipe section ADPF ¼ SLsw (where A and S are
the years, starting with Moody’s formula [2]. Extensive reviews the pipe inner area and circumference) relates the frictional pres-
of these approximations and their performance are given by, for sure loss to the wall shear stress
example, Brkić [3], Genić et al. [4], and Lipovka and Lipovka [5].
In approximations to the Colebrook equation, there is a tradeoff k qU 2
sw ¼ (2)
between accuracy and simplicity. Many of the basic expressions, 4 2
such as those given by Churchill [6], Swamee and Jain [7], and
Håland [8], have sufficient accuracy for most engineering tasks. The Colebrook [1] equation
However, more demanding applications may require higher  
accuracy. 1 2:51 ks
pffiffiffi ¼ 2 log10 pffiffiffi þ (3)
Chen [9], Barr [10], Zigrang and Sylvester [11], Serghides [12], k Re k 3:7D
Buzzelli [13], and Romeo et al. [14] developed more advanced
approximations, reported by Brkić [3] to be one order of magni- is the standard expression for pipe friction calculations. It deter-
tude more accurate than the basic engineering expressions. Keady mines the friction factor for a given Reynolds number Re ¼ qUD/
[15] obtained an exact, but numerically unstable solution for the l (where l is the fluid viscosity) and equivalent sand wall rough-
Colebrook equation, in terms of the Lambert or Wright omega ness ks. The wall roughness is determined indirectly from pressure
functions. Sonnad and Goudar [16] developed an extremely accu- drop and flow rate measurements for fully rough flow, using
rate approximation, based on an efficient reformulation of the Eqs. (1) and (3). The Colebrook equation (3) is implicit. Some
Colebrook equation and a continued fraction expansion. Clamond effort is thus required to obtain the friction factor. In practice, this
[17] developed a highly efficient iterative scheme for the Wright is solved either by numerical iteration or by use of an explicit
omega function in Keady’s solution, which solves the Colebrook approximation, such as Håland’s [8] approximation
equation to machine precision in one or two (quartic) iteration "
steps. Mikata and Walczak [18] introduced a recursive formula  1:11 #
1 6:9 ks
for the Wright omega function, which solves the Colebrook pffiffiffi ¼ 1:8 log10 þ (4)
k Re 3:7D
equation to any degree of accuracy.
The more advanced approximations to the Colebrook equation
tend to be more complex and have a higher computational cost which gives the friction factor to within 1.5% of the Colebrook
than the basic engineering formulas. This can become an issue in equation (3) for 4000 6 Re 6 108 and 0 6 ks 6 0:05. However, the
simulations of pipelines or pipeline networks, which require a Colebrook equation (3) also has an exact explicit solution.
large number of function evaluations for every time step of the
simulation. 3 Keady’s Explicit Solution
By chance, using the symbolic solver MAPLE, Keady [15]
2 The Colebrook Equation discovered that the Colebrook equation (3) has an exact explicit
In a fully developed steady flow in a straight pipe with constant solution in terms of the Lambert function W(z). Reformulating
inner diameter D, the Darcy friction factor, k, defined by Keady’s solution slightly we have

1 Re ks
L qU2 pffiffiffi ¼ aW ðexpð xÞÞ  (5)
DPF ¼ k (1) k bc D
D 2
where
Contributed by the Fluids Engineering Division of ASME for publication in the  
JOURNAL OF FLUIDS ENGINEERING. Manuscript received February 19, 2016; final Re Re ks
manuscript received October 4, 2016; published online December 7, 2016. Assoc. x ¼ ln þ (6)
Editor: Praveen Ramaprabhu. ab abc D

Journal of Fluids Engineering Copyright V


C 2017 by ASME MARCH 2017, Vol. 139 / 031401-1

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 12/18/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


with constants The accuracy of the solution depends on the number of terms
retained in the series.
2 The truncation error in Eq. (11) after n terms is on the order of
a¼ b ¼ 2:51 c ¼ 3:7: (7)
lnð10Þ ðlnðxÞ=xÞn . The maximum truncation error using n terms is thus of
the order of ðlnð7:51Þ=7:51Þn  0:269n for x  7.51, which corre-
Noting that lnð10Þ ¼ lnðxÞ= log10 ðxÞ, we see that the constants sponds to fully turbulent flow and ReP 4000 in Eq. (6).
(Eq. (7)) are the same as those occurring in the Colebrook equa- The accuracy of the Colebrook equation (3) versus pipe flow
tion (3) after switching to natural logarithms. data is of the order of a few percent, provided that the wall rough-
Although Eq. (5) is exact, it cannot, in general, be applied in ness is of the “commercial pipe” quality, for which the Colebrook
numerical computations. The problem is given by the exponential equation (3) was developed. It thus seems appropriate to aim at an
argument of the Lambert function expðxÞ, which can become very approximation that has an accuracy, which is one or two orders of
large for large Reynolds numbers and lead to overflow in a com- magnitude higher, that is 0.01–0.1%, because this will not reduce
puter code, see for example, Sonnad and Goudar [19]. accuracy further. Higher accuracy does not seem to be necessary,
Mikata et al. [18] noted that xðxÞ ¼ Wðexp ðxÞÞ is the Wright given the inherent uncertainty of the Colebrook equation (3).
omega function, which may be represented by the recursion Applying only the first three terms in the series (Eq. (11)), we
obtain the approximation
xðxÞ ¼ lim xn ðxÞ  
n!1
1
x0 ðxÞ ¼ x (8) W ðexpð xÞÞ  x þ  1 lnð xÞ (12)
x
xn ðxÞ ¼ x  lnðxn1 ðxÞÞ
which has a maximum error of 0.053% for xP7:51. This approxi-
see also Lawrence et al. [20]. Applying the truncation mation involves only one logarithm and is therefore more compu-
xðxÞ ¼ limn!N xn ðxÞ, where N is an integer of finite value, gives tationally efficient than Eq. (9). It is also more accurate.
a numerically stable algorithm for computing the Lambert Combining Eqs. (12) and (5), the left-hand term in Eq. (5) can-
function term in Keady’s solution (Eq. (5)) and an approximate cels against a term in the combination of Eqs. (6) and (12). The
solution for the Colebrook equation (3). Taking, for example, result is the approximate solution for the Colebrook equation (3),
N ¼ 3 gives given by
   
Wðexp ðxÞÞ  x  lnðx  lnðx  lnðxÞÞ (9) 1 Re
pffiffiffi ¼ a ln þ Gð x Þ (13)
k ab
Noting that the argument x as given in Eq. (6) is limited to
x P 7:51 for fully turbulent flow, for which Re P 4000, we find where
that this expression has a maximum error of 0.142% in the range
 
of interest here. Applying the approximation (Eq. (9)) in Keady’s 1
solution (Eq. (5)) gives an approximation to the Colebrook equa- Gð x Þ ¼  1 lnð xÞ (14)
x
tion (3), with a maximum error of 0.283% for Re P 4000. In prin-
cipal, any degree of accuracy can be achieved by taking N in which x and a, b, and c are given by Eqs. (6) and (7). The maxi-
sufficiently large, that is by retaining a sufficient number of terms mum error in the approximation (Eqs. (13) and (14)) is less
in the recursion equation (8). However, the increasing number of than 0.153% for ReP4000 and drops to 0.138% in the parameter
logarithms involved leads to increasingly inefficient expressions range considered by Brkić [3], that is, 4000 6 Re 6 108 and
unsuited for large-scale pipeline simulations. ks =D < 0:01. The approximation (Eqs. (13) and (14)) thus has an
accuracy, which is on a par with the most accurate expressions
in Brkić’s [3] review of approximations to the Colebrook
4 Present Approximation equation (3).
The approximation (Eqs. (13) and (14)) is sufficiently accurate
Here, we consider a numerically stable and efficient approxima- for most applications. However, including two more terms in the
tion for the Colebrook equation (3), based on Keady’s exact solu- series (Eq. (11)) gives
tion (Eq. (5)) and the absolutely convergent series expansion
 
W ðzÞ ¼ lnðzÞ  lnðlnðzÞÞ z2
Gð xÞ ¼ lnð xÞ zð1 þ zðz  1ÞÞ  1 þ lnð xÞð3 þ zð2lnð xÞ  9ÞÞ
6
1 X
X 1
ð1Þk lnðlnðzÞÞm
þ jsðk þ m; k þ 1Þj (10) (15)
k¼0 m¼1
m! lnðzÞkþm
where z ¼ 1=x. Applying this expression in Eq. (13) reduces the
where sðk þ m; k þ 1Þ are the Stirling numbers of the first kind, maximum error to around 0.0061% for ReP 4000.
see Corless et al. [21]. Applying the series (Eq. (10)) to represent The approximation (Eqs. (13) and (15)) is thus considerably
WðexpðxÞÞ in Eq. (5) gives more accurate than all the approximations considered in Brkić’s
[3] review of approximations to the Colebrook equation (13), but
lnð xÞ lnð xÞðlnð xÞ  2Þ
W ðexpð xÞÞ ¼ x  lnð xÞ þ þ not as accurate as the extremely accurate solutions given by Son-
x 2x2 nad and Goudar [16], Clamond [17], and Mikata & Walczak [18].
 2 
lnð xÞ 2ln ð xÞ  9lnð xÞ þ 6 However, in Sec. 5, we will see that the present approximations
þ þ  (11) (Eq. (13)) with either Eq. (14) or (15) have a much lower compu-
6x3
tational cost.
in which the lnðzÞ terms in Eq. (10) have removed all the occur- Figure 1 shows the percent error in the approximation
rences of the exponential function in the argument of the Lambert (Eq. (13)) using both Eqs. (14) and (15) for 4000 6 Re 6 108 and
function, since lnðzÞ ! lnðexpðxÞÞ ¼ x. The associated tendency ks =D ¼ 0:00005. The errors are seen to be largest for the lowest
to overflow is thus also removed. Applying the series to represent Reynolds numbers and somewhat smaller for the higher Reynolds
the Lambert function term in Keady’s solution (Eq. (5)) thus gives numbers. Comparisons using other values for the wall roughness
a numerically stable solution of the Colebrook equation (3). show similar trends.

031401-2 / Vol. 139, MARCH 2017 Transactions of the ASME

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 12/18/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


6 Conclusion
The approximations developed here are among the most accu-
rate approximate solutions to the Colebrook equation currently
available, but still have a lower computational cost than basic
engineering formulas. The residual errors in the approximations
are one or two orders of magnitude lower than the inherent error
in the Colebrook equation and will thus not add significantly to
the uncertainty of friction predictions. The approximations are
applied to speed up the friction calculations in Schlumberger’s
multiphase pipe flow simulator OLGA from 2016 and onward.

Acknowledgment
This work was supported by the Schlumberger Norway Tech-
nology Center. A special thanks goes to Dr. G. Staff for running
the Fortran speed test.
Fig. 1 Percentage error in Eq. (13) using Eq. (14) (full drawn
line) and Eq. (15) (dashed line) for 4000 < Re < 108 and ks =D
5 0:00005 References
[1] Colebrook, C. F., 1939, “Turbulent Flow in Pipes, With Particular Reference to
the Transition Region Between the Smooth and Rough Pipe Laws,” J. Inst. Civ.
5 Computational Cost Eng., 11(4), pp. 133–156.
[2] Moody, L. F., 1947, “An Approximate Formula for Pipe Friction Factors,”
Assessing the computational cost, we distinguish between Trans. ASME, 69, pp. 1005–1006.
expensive functions, such as logarithms, noninteger powers and [3] Brkić, D., 2011, “Review of Explicit Approximations to the Colebrook Relation
for Flow Friction,” J. Pet. Sci. Eng., 77(1), pp. 34–48.
divisions, and the low-cost operators—addition, subtraction, and [4] Genić, S., Arandjelović, I., Kolendić, P., Jarić, M., Budimir, N., and Genić, V.,
multiplication. Håland’s approximation (Eq. (4)) seems to be the 2011, “A Review of Explicit Approximations of Colebrook’s Equation,” FME
most accurate of the basic Colebrook approximations and will Trans., 39(2), pp. 67–71.
therefore be used as a reference here, see Genić et al. [4]. One [5] Lipovka, A. Y., and Lipovka, Y. L., 2014, “Determining Hydraulic Friction
Factor for Pipeline Systems,” J. Sib. Fed. Univ. Eng. Technol., 1(7), pp. 62–82.
evaluation of Håland’s approximation (Eq. (4)) involves one (base [6] Churchill, S. W., 1973, “Empirical Expressions for the Shear Stress in
10) logarithm, one noninteger power, and three divisions, in addi- Turbulent Flow in Commercial Pipe,” AIChE J., 19(2), pp. 375–376.
tion to three low-cost operations. This is typical for many of the [7] Swamee, P. K., and Jain, A. K., 1976, “Explicit Equations for Pipe-Flow
basic approximations, such as Churchill [6] and Swamee and Problems,” J. Hydraul. Div., 102(5), pp. 657–664.
[8] Håland, S. E., 1983, “Simple Explicit Formulas for the Friction Factor in
Jain [7]. Turbulent Pipe Flow,” ASME J. Fluids Eng., 105(1), pp. 89–90.
It is clear from the number of logarithms, noninteger powers, [9] Chen, N. H., 1979, “An Explicit Equation for Friction Factor in Pipes,” Ind.
and divisions involved in the more advanced approximations con- Eng. Chem. Fundam., 18(3), pp. 296–297.
sidered by Brkić [3], as well as in the expressions developed by [10] Barr, D. I. H., 1981, “Solutions of the Colebrook-White Function for Resistance
to Uniform Turbulent Flow,” Proc. Inst. Civil. Eng., 71(2), pp. 529–535.
Sonnad and Goudar [16], Clamond [17], and Mikata and Walczak [11] Zigrang, D. J., and Sylvester, N. D., 1982, “Explicit Approximations to the
[18], that they all have a higher computational cost than Håland’s Solution of Colebrook’s Friction Factor Equation,” AIChE J., 28(3), pp.
approximation (Eq. (4)). 514–515.
This is not the case, however, for the present approximations [12] Serghides, T. K., 1984, “Estimate Friction Factor Accurately,” Chem. Eng. J.,
91(5), pp. 63–64.
(Eq. (13)) with either Eq. (14) or (15), which are also very accu- [13] Buzzelli, D., 2008, “Calculating Friction in One Step,” Mach. Des., 80(12),
rate, but still have a lower computational cost than Håland’s p. 54.
approximation (Eq. (4)). These expressions can be evaluated using [14] Romeo, E., Royo, C., and Monz on, A., 2002, “Improved Explicit Equation for
only two logarithms and two divisions each in addition to 9 or 22 Estimation of the Friction Factor in Rough and Smooth Pipes,” Chem. Eng. J.,
86(3), pp. 369–374.
low-cost operations, respectively, which makes them very effi- [15] Keady, G., 1998, “Colebrook–White Formula for Pipe Flows,” J. Hydraul.
cient. A speed test, based on timing a very large number of ran- Eng., 124(1), pp. 96–97.
dom evaluations in the parameter range 4000 6 Re 6 108 and [16] Sonnad, J. R., and Goudar, C. T., 2007, “Explicit Reformulation of the
0 6 ks 6 0:05, using Fortran and the Fortran function CPU_TIME, Colebrook–White Equation for Turbulent Flow,” Ind. Eng. Chem. Res., 46(8),
pp. 2593–2600.
revealed that Eq. (13) with Eq. (14) or (15) evaluates on average [17] Clamond, D., 2009, “Efficient Resolution of the Colebrook Equation,” Ind.
about 32% or 20% faster than Håland’s approximation (Eq. (4)), Eng. Chem. Res., 48(7), pp. 3665–3671.
respectively. [18] Mikata, Y., and Walczak, W., 2015, “Exact Analytical Solutions of the
For comparison, Clamond’s [17] iterative scheme, which Colebrook–White Equation,” J. Hydraul. Eng., 142(2), p. 1061.
[19] Sonnad, J. R., and Goudar, C. T., 2004, “Constraints for Using Lambert W
depending on the input involves two or three logarithms, three or Function-Based Explicit Colebrook–White Equation,” J. Hydraul. Eng., 130(9),
five divisions, and 24 or 43 low-cost operations, was found to pp. 929–931.
evaluate on average about 59% slower than Håland’s approxima- [20] Lawrence, P. W., Corless, R. M., and Jeffrey, D. J., 2012, “Algorithm 917:
tion (Eq. (4)). This number was confirmed by Lipovka and Lip- Complex Double-Precision Evaluation of the Wright x Function,” ACM Trans.
Math. Software, 38(3), pp. 1–17.
ovka [5], who also found the Sonnad and Goudar [16] solution to [21] Corless, R. M., Gonnet, G. H., Hare, D. E. G., Jeffrey, D. J., and Knuth, D. E.,
be 46% slower than the Håland’s approximation. 1996, “On the Lambert W Function,” Adv. Comput. Math., 5(1), pp. 329–359.

Journal of Fluids Engineering MARCH 2017, Vol. 139 / 031401-3

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 12/18/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

You might also like