You are on page 1of 26



Comparative Analysis of Four Second-Moment Turbulence Closure Models for the

Oceanic Mixed Layer
Institut für Meereskunde, Universität Hamburg, Hamburg, Germany

Space Applications Institute, CEC—Joint Research Centre, Ispra, Italy

(Manuscript received 1 March 2000, in final form 21 August 2000)

In this comparative study, four different algebraic second-moment turbulence closure models are investigated
in detail. These closure schemes differ in the number of terms considered for the closure of the pressure–strain
correlations. These four turbulence closures result in the eddy-diffusivity principle such that the closure as-
sumptions are contained in dimensionless so-called stability functions. Their performance in terms of Prandtl
number, Monin–Obukhov similarity theory, and length scale ratios are first tested against data for simple flows.
The turbulence closure is then completed by means of a k–e two-equation model, but other models such as the
two-equation model by Mellor and Yamada could also be used. The concept of the steady-state Richardson
number for homogeneous shear layers is exploited for calibrating the sensitivity of the four models to shear and
stable stratification. Idealized simulations of mixed layer entrainment into stably stratified flow due to surface
stress and due to free convection are carried out. For the latter experiment, comparison to recent large eddy
simulation data is made. Finally, the well-known temperature profile data at OWS Papa are simulated for an
annual cycle. The main result of this paper is that the overall performance of the new second-moment closure
model by Canuto et al.—expressed as nondimensional stability functions—is superior compared to the others
in terms of physical soundness, predictability, computational economy, and numerical robustness.

1. Introduction ating (due to turbulence) fields. A system of infinitely

many differential equations for higher statistical cor-
The intention of this paper is to construct a turbulence relations that is equivalent to the Reynolds equations
closure model that can be recommended for implemen-
and the heat and salt equations can be derived. It has
tation into a wide range of three-dimensional numerical
been the major work of turbulence modelers in the last
ocean models. The following criteria will be applied for
the evaluation of such a model: (i) derivation from sec- decades to suggest closures of this system of equations
ond-moment transport equations, (ii) physical sound- on various levels of sophistication. Milestones in this
ness, (iii) high predictability, (iv) computational econ- process of finding practical solutions for marine and
omy, and (v) numerical robustness. The physics relevant atmospheric modeling were the work of Launder and
for ocean dynamics is included in the Navier–Stokes colleagues (Launder and Spalding 1972; Launder 1975;
equations and the molecular transport equations for heat Launder et al. 1975) and the work of Mellor and Yamada
and salt. Due to the importance of numerically unre- (1974, 1982). To date, numerous modifications to these
solvable small-scale turbulence to large-scale processes models have been suggested (see, e.g., Galperin et al.
in the ocean, model assumptions are necessary in order 1988; Mellor 1989; Kantha and Clayson 1994; Burchard
to achieve an applicable ocean model. Two main schools and Baumert 1995; D’Alessio et al. 1998; Canuto et al.
can be differentiated in the literature: statistical closure 2001). This shows that entirely satisfying solutions have
models and empirical approaches. Statistical models are not been found yet and that they might never be found.
based on the Reynolds decomposition of momentum and Due to this fact, empirical approaches have been popular
scalar fields into mean (ensemble average) and fluctu- throughout recent decades. Even the most simple pa-
rameterizations such as constant eddy viscosity and dif-
fusivity are still applied with some success (see, e.g.,
Corresponding author address: Dr. Hans Burchard, Institut für
Meereskunde, Universität Hamburg, Troplowitzstrasse 7, 22529
Roussenov et al. 1995). In oceanography, the present
Hamburg, Germany. most widely spread empirical turbulence closure is the
E-mail: K-profile parameterization (KPP) model introduced by

q 2001 American Meteorological Society 1943


Large et al. (1994) (see also Large and Gent 1999) in notation is explained. To our experience, differences
which profiles for eddy viscosity and eddy diffusivity in notation have long been a major obstacle for ex-
are constructed based on the Monin–Obukhov similarity change between various schools of statistical model-
theory and Deardorff’s countergradient fluxes without ing. In the appendix sections d–f, exact forms of the
any consideration of statistical moments. The success stability functions suggested by Kantha and Clayson
of this model can only be explained by the fact that (1994), Rodi (1980), and Hossain (1980) (version of
statistical models still do not fulfil some basic require- Burchard and Baumert (1995)) and Canuto et al. (2001)
ments: They mix too little (see Martin 1985), they do are given. This should help the reader to better un-
not generally reproduce nonlocal processes [D’Alessio derstand these functions. In section 8g, the exact for-
et al. (1998) try to solve this problem by introducing mulations for the k and the e equation are given, as
nonlocal fluxes empirically on top of their statistical derived from the Navier–Stokes equations and the
local model], they rarely consider rotation, and they do Reynolds decomposition.
not sufficiently consider breaking surface [the breaking
surface wave parameterization suggested by Craig and
Banner (1994) has not yet been successfully imple- 2. Basic mean flow equations
mented into two-equation models] or internal waves The basic model assumption for obtaining the so-
(generally, internal waves are reduced to background called Reynolds equations is that any flow property
diffusivity in ocean models). X may be decomposed into a mean and a fluctuating
In this paper we carry out a comparative analysis of part:
four algebraic second-moment closure models and their
quasi-equilibrium versions. Their characteristics are fur- X 5 X 1 X. ˜ (1)
thermore compared to empirical data for the turbulent Here, X is called the ensemble average [see, e.g., Lesieur
Prandtl number, the Monin–Obukhov similarity, and (1997)] and will alternatively be denoted by ^X&. It is
length scale ratios. assumed that ^X X̃& 5 0, ^X̃& 5 0, ^X & 5 X̃, and ^XY & 5
Furthermore, we demonstrate how these closure XY (see e.g., Haidvogel and Beckmann 1999).
schemes can be embedded into a two-equation turbu- After application of this Reynolds decomposition and
lence model (here the well-known k–e model) in such the boundary layer approximation, the Navier–Stokes
a way that mixing in stratified shear flows is well re- equations and transport equations for active tracers
produced. The concept of steady-state Richardson (temperature T and salinity S), can be formulated as the
numbers for homogeneous shear layers has already so-called Reynolds equations for mean quantities:
been used by Burchard and Baumert (1995) for cali-
brating an empirical buoyancy parameter in k–e mod- ]t u 1 ]x (u 2 ) 1 ]y (u y ) 1 ]z (u w ) 1 ]z ^ũw̃& 2 ]z (n ]z u )
els. This approach is modified here in the same way 1
that Burchard (2001) recently did for the k–kL model 52 ] p 1 f y,
by Mellor and Yamada (1982). The well-known en- r0 x
trainment experiment by Kato and Phillips (1969) is ]t y 1 ]x ( y u ) 1 ]y ( y 2 ) 1 ]z ( y w ) 1 ]z ^ỹ w̃& 2 ]z (n ]z y )
used here as a simple performance test for these sec-
ond-moment closures, all embedded into a k–e model. 1
All models are furthermore applied for simulating the 52 ] p 2 f u,
r0 y
free convection experiment by Willis and Deardorff
(1974) and the results are compared to recent large ]t T 1 ]x (u T ) 1 ]y ( y T ) 1 ]z (w T ) 1 ]z ^w̃T&
˜ 2 ]z (n9]z T )
eddy simulation (LES) data by Mironov et al. (2000).
In a final performance test, the second-moment clo- 1
5 ] I,
sures suggested by Kantha and Clayson (1994), which cp r 0 z
is a slight improvement of the models of Mellor and
Yamada (1982) and Galperin et al. (1988), and Canuto ]t S 1 ]x (u S ) 1 ]y ( y S ) 1 ]z (w S ) 1 ]z ^w̃S&
˜ 2 ]z (n 0]z S )
et al. (2001) are compared by means of applying them 5 0. (2)
to the well-known mixed layer dataset OWS Papa
(northern Pacific). This clearly shows the improve- In these equations, u, y , and w are the x (eastward), the
ments achieved by the closure of Canuto et al. (2001). y (northward), and the z (upward) velocity components,
The paper is structured as follows: After presenting respectively, and t is time; w is calculated diagnostically
the mean flow equations (section 2), the procedure for from the incompressibility condition
various second-moment closures is discussed (section ] xu 1 ] yy 1 ] zw 5 0. (3)
3). Two-equation models are introduced in section 4
and their stationary solutions are investigated in sec- The pressure p is hydrostatic with
tion 5. Idealized and ocean mixed layer studies are
] zp 1 gr 5 0 (4)
presented in sections 6a–c. Final conclusions are dis-
cussed in section 7. In section a of the appendix our with gravitational acceleration g and density r. Earth

rotation is considered through the Coriolis frequency f tioned algebraic rules for the ensemble averaging of
5 2v sin(f) with the earth’s angular velocity v and tensors. The exact forms for the Reynolds stress and the
latitude f (positive for Northern Hemisphere). In the heat flux equation can be found, for example, in Launder
temperature (or local heat balance) equation, further et al. (1975) or Canuto (1994).
terms are solar radiation I (in W m 22 ) in the water column The equations are presented here in a closed form
(generally calculated from the given surface radiation as with the exception of third-order correlators occurring
an exponentially decreasing function with depth), the spe- on the left-hand sides causing turbulent transport of the
cific heat capacity of seawater Cp , and a mean density second-order correlators. The three approaches dis-
r 0 . The molecular diffusivities for momentum, temper- cussed here differ mainly in the manner how the pres-
ature, and salinity are given by n, n9, and n0, respectively. sure–strain correlators P ij 5 ^ũ i] j p̃& 1 ^ũ j] i p̃& and P Ti
Together with an equation of state giving density as 5 ^T̃] i p̃& are parameterized. All approaches neglect ro-
function of T, S, and p and suitable boundary conditions, tational terms, although they are known to be important
Eq. (2) is the physical basis for most marine models for free convection (see Mironov et al. 2000).
ranging from estuarine, over coastal, shelf sea, basin, In the following the Reynolds stress, the heat flux,
and global ocean scale. and the temperature variance equations are given as
The ocean circulation modeler might miss terms for prognostic equations (sections 3a and 3b), then an al-
horizontal mixing in Eq. (2). They do not appear here gebrazation is discussed (section 3c) and finally the
due to the boundary layer approximation. All mesoscale boundary layer approximation is applied (section 3d).
activity is already included in the equations above
through the advective terms. Mesoscale activity only a. Reynolds stress equation
needs to be parameterized because of the generally rath-
er coarse horizontal resolution in ocean models and has After neglecting viscous and rotational effects and
therefore to be considered as part of the discretization. parameterizing pressure–strain correlators, the Reynolds
In contrast, vertical turbulent transport needs to be pa- stress equations can be closed in the following form
rameterized due to the Reynolds decomposition and the (see Canuto et al. 2001):
hydrostatic approximation even for the (theoretical) lim- ]t ^ũ i ũ j & 1 ]l (u l ^ũ i ũ j & 1 ^ũ i ũ j ũ l &)
it of arbitrarily fine model resolution. Therefore, in hy-

1 2
drostatic ocean models, vertical and horizontal mixing « 2
must not be parameterized with the same model ap- 5 2c1 ^ũ i ũ j & 2 dij k (R1)
k 3

1 2
The main differences between all these models are due 2 2
to boundary conditions, coordinate transformations, nu- 1 (1 2 c2 ) Pij 2 dij P 1 dij P (R2)
3 3
merical aspects, and, of course, the closure assumptions
for the unknown second-order correlators ^ũw̃&, ^ỹw̃&,
1 (1 2 c )1B 2 d B2 1 d B
2 2
^w̃T̃&, and ^w̃S̃&, which are the Reynolds stresses, the tur- 3 ij ij ij (R3)
bulent heat flux, and the turbulent salinity flux. 3 3
For the derivation of these fluxes, we apply a major 2 c4 kSij (R4)
simplification that is used by almost all turbulence clo-
sure models: during the closure procedure, only one 2 c5 Z ij (R5)
active tracer is considered. Afterward, the derived flux
parameterization for one active tracer is applied to the 2 dij « (R6). (5)
other tracers in analogy. By means of this simplification, 3
consideration of correlators such as ^T̃S̃& is avoided. Here, the following definitions are used:
Pij 5 2]l u i ^ũ l ũ j & 2 ]l u j ^ũ l ũ i &,
3. Second-moment closure
Bij 5 bi ^ũ j T&
˜ 1 bj ^ũ i T&
˜ (6)
Here, three different approaches for closing the sec-
ond-order moments ^ũw̃&, ^ỹ w̃&, and ^w̃T̃& will be pre- are production due to shear and buoyancy with b1 5
sented. They have been published by b 2 5 0, and b 3 5 2g] T r/r, where the traces define the
shear and the buoyancy production:
R Kantha and Clayson (1994, hereafter KC),
R Burchard and Baumert (1995) [based on the work of 1 1
P 5 Pll , B 5 Bll . (7)
Rodi (1980) and Hossain (1980); hereafter RH], and 2 2
R Canuto et al. (2001 hereafter CA). Further definitions are shear
All of these approaches are based on the exact forms 1
of the Reynolds stress and the heat flux equation. These Sij 5 (]i u j 1 ]j u i ) (8)
forms can be derived from the Navier–Stokes and the
local heat balance equations by applying the aforemen- and

TABLE 1. Parameters for the Reynolds stress closure models.

Model c1 c2 c3 c4 c5 c1T c2T c3T c4T cT c0m

KC 2.98 0.0 0.0 0.32 0.0 3.70 0.7 0.2 0.0 1.23 0.094
RH 1.8 0.6 0.5 0.0 0.0 3.0 0.33 0.33 0.0 1.6 0.121
CA 2.5 0.984 1.6 0.533 0.416 5.97 0.6 0.33 0.4 1.44 0.077

and Yamada (1982) chose the same value for c1T that
1 2 1 2
2 2
Z ij 5 Vil ^ũ l ũ j & 2 dlj k 1 Vjl ^ũ l ũ i & 2 dil k (9) KC did, but used c 2T 5 c 3T 5 0.
3 3 Finally a dynamic equation of the temperature vari-
with vorticity ance ^T̃ 2 & has to be used:
1 ]t ^T˜ 2 & 1 ]j (u j ^T˜ 2 & 1 ^ũ i T˜ 2 &)
Vij 5 (]j u i 2 ]i u j ). (10)
5 22^ũ i T&]j T (T1)
The parameters c1 , . . . , c 5 have been introduced for
parameterization of the pressure–strain correlators. 1 « ˜2
22 ^T & (T2). (12)
Setting them all to zero would result in the neglect of cT k
these terms. The single terms have the following mean- Here the terms are production by mean gradient (T1)
ing: return to isotropy (R1) (see Rotta 1951), shear and dissipation (T2), which is often denoted as x. No
production (R2), buoyancy production (R3), noniso- basic difference can be found in the three approaches
tropic contribution due to shear (R4), and contribution considered in this paper, and the parameter c T varies
due to vorticity (R5). For the physical relevance of the only little (see Table 1).
terms (R4) and (R5), see Shih and Shabbir (1992) and
Canuto (1994). The empirical parameters are given in
Table 1. It can be seen that the model KC neglects c. Algebrazation
pressure–strain effects due to shear and buoyancy pro-
duction and vorticity, the model RH neglects effects The stability functions discussed in this paper are
of terms (R4) and (R5), and only the model CA con- derived from the dynamic equations (5), (11), and (12)
siders all terms. The calibration of these empirical pa- by means of two different approaches. The basic ap-
rameters is referenced in the respective publications of proach is to neglect time variation and advective and
the KC, the RH, and the CA models. It should be noted turbulent transports of Reynolds stresses ^ũ i ũ j &, heat
that in the original version of the RH model, wall prox- fluxes ^ũ i T̃&, and the autocorrelation ^T̃ 2 &. In order to
imity functions have been included in the parameters retain the transport of turbulent kinetic energy k, the
(see Hossain 1980). These are neglected here. TKE equation multiplied with (⅔)d ij has to be subtracted
from the Reynolds stress equation (5) first, and then the
left-hand side of the resulting transport equation for
b. Heat flux and temperature variance equation ^ũ i ũ j & 2 (⅔)d ij k is set to zero. Furthermore, the left-
Formulating the heat flux and the temperature vari- hand sides of the heat flux equation and the equation
ance equation on the same closure level as the Reynolds for ^T̃ 2 & are set to zero as well.
stress equation given above results in Mellor and Yamada (1974, 1982) used this approach
for their so-called level 2½ model. They justify the ne-
]t ^ũ i T&
˜ 1 ]j (u j ^ũ i T&
˜ 1 ^ũ i ũ j T&)
glect of transports for ^ũ i ũ j & and ^ũ i T̃& by a scaling pro-
« ˜ cedure where terms are ordered by their degree of de-
5 2c1T ^ũ i T& (H1) viation from isotropy. This results in their level 3 model.
k The additional neglect of transport terms in the ^T̃ 2 &
2 (1 2 c2T )^ũ j T&]
˜ j u i 2 ^ũ i ũ j &]j T (H2) equation then allows for a complete algebraization.
In the present paper, a nonequilibrium version of the
1 (1 2 c3T )bi ^T˜ 2 & (H3) model suggested by Kantha and Clayson (1994) is de-
rived by applying the above discussed algebraization
1 c4T ^ũ i T&V
˜ ij (H4). (11)
procedure to Eqs. (5), (11), and (12). This is an exten-
In (11), the right-hand side terms have the following sion of the Mellor and Yamada (1974, 1982) level 2½
physical meaning: return to isotropy (H1), production model in which the parameterization of the pressure
by mean gradients (H2), production by buoyancy (H3), correlations in the heat flux equation are simpler, be-
and contribution due to vorticity (H4). Here, the KC cause of c 2T 5 c 3T 5 0. The quasi-equilibrium model
and RH models neglect the (H4) term whereas CA con- by Kantha and Clayson (1994) is achieved from the
siders all terms (see Table 1). The closure by Mellor nonequilibrium model by applying the equilibrium con-

dition P 1 B 5 « to the resulting equations (see section of the eddy viscosity and the eddy diffusivity principle:
The same approach for algebraization is used by Can-
uto et al. (2001). After this procedure, a linear system k2
^ũw̃& 5 2cm ] u, (15)
of algebraic equations for ^ũ i ũ j &, ^ũ i T̃&, and ^T̃ 2 & is ob- « z
tained for the nonequilibrium version of the model by k2
Kantha and Clayson (1994) and the model by Canuto ^ỹ w̃& 5 2cm ]z y , (16)
et al. (2001).
˜ 5 2c9m k ]z T
A different, less rigorous procedure for the algebrai- ^w̃T& (17)
zation of the Reynolds stress equation has been sug- «
gested by Rodi (1976). The basic assumption is that the with the eddy viscosity nt and diffusivity n9t , respectively:
ratio of transport of Reynolds stresses to the transport
of TKE [see Eq. (A37)] is equal to the ratio of Reynolds
stresses to TKE: k2 k2
nt 5 cm , nt9 5 c9m . (18)
]t ^ũ i ũ j & 1 ]l (u l ^ũ i ũ j & 1 ^ũ i ũ j ũ l &) ^ũ i ũ j & « «
5 . (13)
P1B2« k This reflects the relation of Kolmogorov (1942) and
Prandtl (1945), which assumes that eddy viscosity and
A similar approach is then used for the heat flux equation: diffusivity are proportional to a velocity scale and a length
scale of turbulence. Here k1/2 is the velocity scale and
]t ^ũ i T&
˜ 1 ]l (u l ^ũ i T&
˜ 1 ^ũ l ũ i T˜ )& 1 ^ũ i T&
5 . (14) k 3/2
P1B2« 2 k L 5 cL (19)
This alternative approach has been first applied by Rodi
(1980) to the Reynolds stress equation and by Hossain with c L 5 (c m0 ) 3/4 a macrolength scale for energy con-
(1980) to the heat flux equation. taining eddies, calculated by means of the Taylor (1935)
The transport equation for the temperature fluctuation scaling. In (19) c m0 is the value for c m resulting from B
autocorrelation ^T̃ 2 & is algebraized in the same manner 5 0 and P 5 «. (see Table 1).
as for the previously discussed models simply by setting All the information on second-order correlators is
the transport to zero. Application of this method for now contained in the rather complex, nondimensional
algebrazation to the transport equations (5), (11), and stability functions c m and c9m . Despite their differences,
(12) leads to a nonlinear system of algebraic equations these stability functions depend for all models on only
for ^ũ i ũ j &, ^ũ i T̃&, and ^T̃ 2 &. two nondimensional parameters, the shear number and
the buoyancy number, respectively:

d. Boundary layer approximation k2 2 k2 2

aM 5 S , aN 5 N . (20)
The solution of the systems of algebraic equations «2 «2
for ^ũ i ũ j &, ^ũ i T̃&, and ^T̃ 2 & discussed above is simplified The different stability functions are given in section 3e.
by applying the so-called boundary layer approximation It should be noted that up to this point no assumption
first. This is justified since the resulting closure schemes about the calculation of the turbulent dissipation rate «
are designed for use in estuarine and ocean models has been made.
where the aspect ratio, that is, the ratio between vertical
and horizontal scales, is sufficiently small. The bound-
ary layer approximation is realized by setting all hori- e. Nonequilibrium stability functions
zontal gradients inside the algebraic system of equations
for ^ũ i ũ j &, ^ũ i T̃&, and ^T̃ 2 & to zero. For easier notation, Here the sets of stability functions are given for the
(ũ1 , ũ 2 , ũ 3 ) is set to (ũ, ỹ , w̃), (u 1 , u 2 , u 3 ) is set to (u , models of KC (Kantha and Clayson 1994), HR (Rodi
y , w ), and (]1 , ] 2 , ] 3 ) is set to (] x , ] y , ] z ). Consequently, 1980; Hossain 1980; Burchard and Baumert 1995), and
after applying the continuity equation (3), ] zw 5 0 re- CA (Canuto et al. 2001). An alternative set of stability
sults as well. For the resulting relations for P ij , B ij , S ij , functions also proposed by Canuto et al. (2001) (from
and V ij , see section b of the appendix. here on denoted by CB) is given as well. They are
After this boundary layer approximation, an algebraic displayed as functions of a M and a N in Figs. 1–4.
system of equations for the correlators ^ũ 2 &, ^ỹ 2 &,^w̃ 2 &, At this stage, the closure of the second-order moments
^ũw̃&, ^ỹw̃&, ^ũỹ&, ^ũT̃&, ^ỹT̃&, ^w̃T̃&, and ^T̃ 2 & is obtained is practically finished. It should be noted however that
for each model. This system of equations is linear for all these sets of stability functions should be limited by
the models KC and CA and nonlinear for the model RH. certain constraints in order to assure positivity of the
Despite the different structures of the systems of equa- eddy viscosity n t and diffusivity n9t and of the velocity
tions, they all result in the simple, well-known relations variances ^ũ 2 &, ^ỹ 2 &, and ^w̃ 2 &; see Eqs. (A10–A12).

1) MODEL OF KANTHA AND CLAYSON (1994) of the same form as the stability functions originally
suggested by Mellor and Yamada (1982) with the ex-
In Kantha and Clayson (1994), only the quasi-equi- ception that c 2 and c 3 are nonzero now.
librium version of the stability functions is given. How- In our notation (15)–(17), this set of stability func-
ever, full versions can also be derived, which are then tions may be formulated as

0.1682 1 0.03269a N
cm 5
1 1 0.4679a N 1 0.07372a M 1 0.01761a N a M 1 0.03371a N2
0.1783 1 0.01586a N 1 0.003173a M
c9m 5 . (21)
1 1 0.4679a N 1 0.07372a M 1 0.01761a N a M 1 0.03371a N2

The exact form of (21) in terms of the empirical pa- been solved in numerical models by using the value for
rameters (see Table 1) contained in Eqs. (5), (11), and (P 1 B)/e 2 1 on an old time level. In this form, the
(12) is given in section d of the appendix. stability functions have been presented first by Rodi
(1980) and Hossain (1980).
However, because of (P 1 B)/« 5 c m a M 2 c9ma N ,
2) MODEL OF RODI (1980), HOSSAIN (1980), AND these equations for c m and c9m can be expressed as im-
BURCHARD AND BAUMERT (1995) plicit functions of a M and a N . The evaluation procedure
that has been suggested by Burchard and Baumert
In contrast to the models of Kantha and Clayson (1995) is first solving for (P 1 B)/« 2 1 and then
(1994) and Canuto et al. (2001), the model of Rodi inserting that value into the formulations for c m and
(1980) and Hossain (1980) in the version of Burchard c9m. This procedure is discussed in detail in the section
and Baumert (1995) results in stability functions c m and e of the appendix.
c9m, which not only depend on a M and a N , but addition-
ally depend on the nondimensional term (P 1 B)/« 2
1, that is, on the degree of deviation from local tur- 3) MODEL OF CANUTO ET AL. (2001)
bulence equilibrium. This is a consequence of the spe- The stability functions as they result from the closure
cific closure concept used in that model; see Eqs. (13) assumptions carried out by Canuto et al. (2001) are as
and (14). Traditionally, these stability functions have follows:

0.1070 1 0.01741a N 2 0.00012a M

cm 5
1 1 0.2555a N 1 0.02872a M 1 0.008677a N2 1 0.005222a N a M 2 0.0000337a M2
0.1120 1 0.004519a N 1 0.00088a M
c9m 5 . (22)
1 1 0.2555a N 1 0.02872a M 1 0.008677a N2 1 0.005222a N a M 2 0.0000337a M2

Despite the higher complexity of the transport equations exact form can be found in section f of the appendix.
for Reynolds stresses and heat fluxes due to consider- In their paper, Canuto et al. (2001) give another set of
ation of more terms for the pressure-strain correlators, stability functions derived on the ground of different
these stability functions are structurally similar to those assumptions. They will be denoted by CB and are of
of Kantha and Clayson (1994). the following form:
These stability functions will be denoted by CA. The

0.1270 1 0.01526a N 2 0.00016a M

cm 5
1 1 0.1977a N 1 0.03154a M 1 0.005832a N2 1 0.004127a N a M 2 0.000042a M2
0.1190 1 0.00429a N 1 0.00066a M
c9m 5 . (23)
1 1 0.1977a N 1 0.03154a M 1 0.005832a N2 1 0.004127a N a M 2 0.000042a M2

FIG. 1. Complete version of stability functions c m and c9m displayed as functions of the nondimensional
buoyancy number a N and shear number a M according to Kantha and Clayson (1994). The white area
indicates that the functions are not defined there. The bold line marks the equilibrium state P 1 B 5 «.

f. Quasi-equilibrium stability functions For stable stratification, laboratory and LES data for
Quasi-equilibrium is defined as the state where pro- the turbulent Prandtl number Pr 5 c m /c9m are compared
duction and dissipation of turbulent kinetic energy are to the turbulent Prandtl number computed by the quasi-
balanced; that is, P 1 B 5 «. This can be transformed equilibrium stability functions (see Fig. 6). All the func-
to the relation tions are within the uncertainty of the data. Only the
KC quasi-equilibrium stability functions do not reach
c m a M 2 c9m a N 5 1. (24) high turbulent Prandtl numbers due to their relatively
This quasi-equilibrium state has often been used for small critical gradient Richardson number.
simplifying stability functions depending on both a M Another way of displaying the quasi-equilibrium sta-
and a N . The most well-known example is the work of bility functions is to transform them into the Monin–
Galperin et al. (1988) where relation (24) has been used Obukhov similarity form. This has already been sug-
for improving the performance of the stability functions gested by various authors (see Mellor and Yamada 1982;
proposed by Mellor and Yamada (1982), which have Kantha and Clayson 1994). Monin and Obukhov (1954)
been proven to be numerically unstable (see Deleersnij- found for the atmospheric boundary layer nondimen-
der and Luyten 1994). Galperin et al. (1988) found by sional relations between fluxes of momentum and heat
applying the scale analysis introduced by Mellor and and gradients of velocity and density, respectively.
Yamada (1974) that it is not a model inconsistency to Based on the Monin–Obukhov length, L M 5 2u*3 /(kB),
prescribe P 1 B 5 e only for the stability functions but gradients of momentum and buoyancy can be expressed
still retain the full dynamic TKE equation. as follows:
The bold lines in Figs. 1–4 indicate the quasi-equi-

1 2 1 2
librium states for the four sets of stability functions u* z9 B z9
]z u 5 FM , ]z b 5 2 FH (25)
discussed in this paper. The stability functions suggested kz9 LM kz9u* LM
by Galperin et al. (1988) are expressed as functions of
with the distance from the boundary z9 and buoyancy
a N . Relation (24) allows us also to express the stability
b 5 2g(r 2 r 0 )/r 0 .
functions depending on the gradient Richardson number
With the friction velocity
Ri 5 a N /a M , which has often been done for further
analyzing the stability functions as shown in Fig. 5 for k2
the four sets of stability functions. The maximum value u*2 5 cm |] u|,
« z
of Ri that can be reached in quasi-equilibrium is called
the ‘‘critical’’ Richardson number Ri c . For the models the buoyancy flux
discussed here, they have the values shown in Table 2. k2
It can be seen that the model of Kantha and Clayson B 5 2c9m ]b
(1994) already suppresses turbulence for stratifications « z
around Ri 5 0.2, whereas the other models allow for and the macro length scale L 5 kz9, the variables z 5
mixing at Richardson numbers significantly higher. z9/L M , F M , and F H can be expressed as

FIG. 2. Same as Fig. 1 but for the Rodi (1980) and Hossain (1980) [version of Burchard and Baumert
(1995)] stability functions.

c9m a N a 1/4 are contained in the stability functions as a ratio, defin-

z 5 (cm0 ) 3/4 · , FM 5 (cm0 ) 3/4 M
, ing a timescale t 5 k/« and in the relations of Prandtl
cm3/2 a M3/4 cm1/2
and Kolmogorov as ratio k 2 /«. One straightforward so-
cm1/2 a1/4
M lution of the remaining closure problem is therefore to
FH 5 (cm0 ) 3/4 , (26) construct prognostic equations for these two quantities.
For both k and «, exact transport equations can be de-
where k 5 0.4 is the von Kármán constant and Pr 5 rived from the Navier–Stokes equations after applying
F M /F H 5 c m /c9m the Prandtl number. Figure 7 shows the ensemble averaging [see Eqs. (A37) and (A39) in the
Monin–Obukhov similarity functions F M and F H as appendix].
functions of z in comparison to empirical curves by The k equation has already been discussed, it can also
Businger et al. (1971). The stability functions by Kantha be derived by taking the trace of the Reynolds stress
and Clayson (1994) are clearly the closest to the em- equation (5):
pirical curves, based on measurements in the atmo-
spheric boundary layer. It should be noted that this is ] t k 1 ] z F(k) 5 P 1 B 2 «. (27)
due to a tuning of the parameters c 2 and c 3 . Mellor and For closing the dissipation rate equation (A39), a
Yamada (1982) achieved a good agreement with the number of closure assumptions are needed (see section
Businger et al. (1971) data with their relatively simple g of the appendix):
closure by setting c 2 5 c 3 5 0. It is difficult to explain
why more complex parameterizations of the pressure– «
strain correlations cause worse agreement with the Mon- ]t « 1 ]z F(«) 5 (c«1 P 1 c«3 B 2 c«2 «) (28)
in–Obukhov theory. But it might be due to the fact that
this theory is valid only close to boundaries where some with c «1 5 1.44 and c «2 5 1.92. It has been shown by
of the closure assumptions could be wrong. Another Burchard and Baumert (1995) that c «3 is not an inde-
reason could be that the Businger et al. (1971) mea- pendent parameter but depends on the definition of a
surements are taken at the atmosphere where in com- steady Richardson number. Deviating from other au-
parison to water different molecular viscosities and dif- thors, Burchard and Baumert (1995), Burchard et al.
fusivities are present. This could be particularly signif- (1998), Burchard and Petersen (1999), and Baumert and
icant at strongly stable stratification with large positive Peters (2000) postulate that c «3 should be a negative
values of z. number. This is further discussed in section 5. However,
for unstable stratification, c «3 5 1 is used in order to
provide a source term for the dissipation rate in shear-
4. Two-equation models
free convective regimes (see Rodi 1987).
In the second-moment closure presented here as for- The resulting expressions for the diffusive fluxes F(k)
mulations for the eddy viscosity and eddy diffusivity and F(«) are third-order correlators for which closure
[see Eq. (18)], the turbulent kinetic energy k and its assumptions have to be made as well. For simplicity, we
dissipation rate « still occur as free parameters. They adopt here the eddy viscosity principle, which leads to

F(k) 5 2n k] z k, F(«) 5 2n «] z «, (29) cm c«2 2 c«1

Ri 5 Ri st 5 · . (32)
where the diffusivities are chosen as n k 5 n 1 n t and c9m c«2 2 c«3
n « 5 n 1 n t /s « . Here, the logarithmic law for constant The steady-state Richardson number therefore depends
stress layers leads to the Schmidt number for dissipa- on the empirical parameters c «1 , c «2 , and c «3 in the «
tion: equation and on the actual stability function chosen. In
k2 contrast to c «1 and c «2 , the buoyancy-flux-related pa-
s« 5 0 1/2 , (30)
(cm ) (c«2 2 c«1 ) rameter c «3 has never directly been determined by lab-
oratory experiments. Figure 8 shows how c «3 and Ri st
which is of the order of unity and varies for different are related to each other for the various stability func-
stability functions with c m0 . It should be noted that non- tions. This figure may be compared to Fig. 3 by Bur-
local parameterizations for F(k) have been suggested in chard and Baumert (1995) where the flux Richardson
the literature, which should improve the model perfor- number was considered instead. For steady-state Rich-
mance especially for convective regimes (see Canuto et ardson numbers below 0.35, negative values for c «3 have
al. 1994; D’Alessio et al. 1998). We did not use such to be used for all sets of stability functions. Suitable
models here since they are either very complex (Canuto values for Ri st will be calibrated in section 6 by means
et al. 1994) or based on bulk parameters such as the of an idealized wind mixing experiment.
Deardorff convective velocity scale w* (D’Alessio et An expression equivalent to (32) has recently been
al. 1998). This has the consequence that the turbulent suggested by Burchard (2001) for the k–kL model by
transport of TKE under free convection is underesti- Mellor and Yamada (1982). The surprising result was
mated by approximately a factor of 2; see section 6b). that for reasonable values for the steady-state Richard-
Boundary conditions for k and « are based on the son number of Ri st ø 0.2 the coefficient was far from
assumptions of a constant stress layer and local equi- the value suggested in the original paper. With this new
librium (P 5 «) near the boundaries and a macrolength calibration, a strict limitation of the length scale under
scale of L 5 k(z 1 z 0 ) with the roughness length z 0 . stable stratification, which was needed for the old ver-
We use here flux boundary conditions as suggested by sion, is no longer necessary.
Burchard and Petersen (1999), which performed opti-
mally in terms of numerical approximation and physical
behavior (e.g., decrease of eddy viscosity near the stress- b. Length scales
free surface).
The so-called structural equilibrium of the system of
It should be noted that the dissipation rate equation
might be replaced by a transport equation for the product Eq. (31) is reached when the timescale t 5 k/« is in
kL as suggested by Mellor and Yamada (1982). The steady state. It can be shown that—other than the total
dissipation rate would then have to be calculated by equilibrium discussed in section 5a—the solution for
using Eq. (19). There is however an ongoing discussion (31) tends to the structural equilibrium for all Richard-
whether or not the proportionality factor used in (19) son numbers. Baumert and Peters (2000) recently
should be a constant or depending on stratification (see showed how a k–« model equipped with an empirical
Mellor 2001). closure for the stability functions could reproduce data
found for ratios of relevant turbulent length scales. We
repeat this for the four algebraic second-moment clo-
5. Stationary solutions in homogeneous shear sures presented here.
layers The two length scales considered here are the Oz-
For some basic investigations of turbulence models midov scale
it is instructive to consider idealized flows far away from

1 2
boundaries with constant shear and stratification. This
leads to the concept of homogeneous shear layers. Math- LO 5 (33)
ematically formulated, this concept leads to vanishing
diffusion terms in all turbulence equations such that they and the buoyancy scale
become ordinary differential equations only depending k 1/2
on time t. After denoting k̇ 5 ] t k and «˙ 5 ] t «, the Lb 5 . (34)
equations for k and « may be written as N
« Another important length scale is the Ellison scale, de-
k̇ 5 P 1 B 2 «, «˙ 5 (c«1 P 1 c«3 B 2 c«2 «). (31) fined as L E 5 2r̃/] zr̃ (with the density fluctuation r̃),
which is often set equal to the Thorpe scale and is related
to the macrolength scale L [see Eq. (19)] as follows (see
a. Steady-state Richardson number Baumert and Peters 2000):
If k̇ and «˙ are zero, then the total equilibrium of the
k–« model is reached and the following relation, which LE 5 (35)
is a precondition for the steady state, can be derived: 2C(cm0 ) 3

FIG. 3. Same as Fig. 1 but for the Canuto et al. (2001) (version A) stability functions.

with C 5 1.4. When assuming structural equilibrium have to be compared to empirical curves based on lab-
for the k–« model in the homogeneous shear layer ap- oratory experiments; see Baumert and Peters (2000):
proximation, then the ratios L E /L O and L E /L b can be
expressed as follows: ø 4.2Ri 3/4 , ø 1.6Ri 1/2 . (37)
LE 1 3/4 LE 1 1/2
5 a , 5 a . (36) Figure 9 shows this functional relationship. It can be
LO 2C N Lb 2C N
clearly seen that all curves except those resulting from
In order to calculate these ratios for the algebraic sec- the KC model are in fairly good agreement with the
ond-moment closures presented here, a relation between empirical curves.
a M and a N is obtained after deriving a dynamic equation
for t and then setting ṫ 5 0. This allows for displaying
6. Mixed layer studies
L E /L O and L E /L b as functions of the gradient Richardson
number. For this calculation, the parameter c «3 has to An idealized wind entrainment experiment and an
be prescribed, which we take for a steady-state gradient idealized free convection experiment have been chosen
Richardson number of Ri st 5 0.25 (see Table 2). These here in order to test the k–« model together with the

FIG. 4. Same as Fig. 1 but for the Canuto et al. (2001) (version B) stability functions.

FIG. 5. Quasi-equilibrium versions of stability functions displayed as functions of the gradient Richardson number Ri
(a) Model of Kantha and Clayson (1994), (b) model of Rodi (1980) and Hossain (1980), (c) model A of Canuto et al.
(2001), (d) model B of Canuto et al. (2001). For the se versions of the Kantha and Clayson (1994) model, a steady-state
Richardson number of Ri st 5 0.225 was used.

stability functions of Kantha and Clayson (1994) (KC), applied to the OWS Papa data and their results compared
Rodi (1980) and Hossain (1980) (RH), and Canuto et to temperature profile observations. It should be noted
al. (2001) (CA, CB). Furthermore, the quasi-equilibrium that the scope of this paper is to test mixed layer models
version of the KC model and the full CA model are to be implemented into three-dimensional ocean models.
We therefore do not intend to compare here the model
performance to measurements of turbulent quantities for
TABLE 2. Critical Richarson number Ri c and c«3 . The latter is based
on a steady-state gradient Richardson number of Rist 5 0.25, but for which data are usually available only on short time-
the model KC, Rist 5 0.225 was used. scales.
Model Ric c «3
a. Wind entrainment
KC 0.235 20.404
RH 0.615 20.444 A wind entrainment experiment is used in order to
CA 0.847 20.629 test the assumptions on the stationary gradient Rich-
CB 1.02 20.566
ardson number Ri st . The wind entrainment experiment

FIG. 8. Stationary gradient Richardson number Ri st as function of

c 3 for the different sets of stability functions: Models of Kantha and
Clayson (1994) (KC), Rodi (1980) and Hossain (1980) (RH), version
FIG. 6. Prandtl number, calculated from the quasi-equilibrium ver- A of Canuto et al. (2001) (CA), and version B of Canuto et al. (2001)
sion of stability function displayed as function of the gradient Rich- (CB).
ardson number Ri. KC: Model of Kantha and Clayson (1994), RH:
model of Rodi (1980) and Hossain (1980), CA: model A of Canuto
et al. (2001), CB: model B of Canuto et al. (2001). The laboratory D m (t) 5 1.05u*s N 21/2
0 t1/2 , (38)
data are from Rohr (1985), the LES data are from Schumann and
Gerz (1995). The small bullets mark the maximum values of the where u*s is the surface friction velocity and N 0 the
Prandtl numbers due to the critical Richardson number Ri c . constant initial Brunt–Väisälä frequency. Following
several authors (see, e.g., Deleersnijder and Luyten
carried out here is inspired by the laboratory experiment 1994; Burchard et al. 1998) we transform this laboratory
of Kato and Philipps (1969). In this experiment, a mixed experiment to ocean dimensions with u*s 5 10 22 m s 21
layer induced by a constant surface stress penetrates into and N 0 5 10 22 s 21 .
a stably stratified fluid with density increasing linearly First of all, the concept of the steady-state gradient
downward from the surface. The water depth is assumed Richardson number Ri st can be validated by means of
to be infinite. Price (1979) suggested a solution for the this entrainment experiment. The stability functions CA
evolution of the mixed layer depth D m based on a con- and CB are applied with values for Ri st ranging from
stant Richardson number 0.2 to 0.8. Figure 10 shows the evolution of the en-

FIG. 7. Monin–Obukhov similarity functions F M and F H as calculated from the quasi-equilibrium

stability functions of Kantha and Clayson (1994) (KC), Rodi (1980) and Hossain (1980) (RH), Canuto
et al. (2001) (CA,CB). The empirical curves are taken from Businger et al. (1971).

TABLE 3. Entrainment depth for various mixed layer models after

3 days of constant cooling with 100 W m22 , with an initially stable
temperature gradient of 0.18C m21 and a constant salinity of 35 psu.

Model Entrainment depth (m)

k–e model, KC 12.5
k–e model, RH 11.9
k–e model, CA 12.2
k–e model, CB 12.4
KPP 13.0
Convective adjustment 11.2

trainment depth, defined as the distance of the lower-

most point below the surface with a turbulent kinetic
energy of k . 10 25 J kg 21 . It can be clearly seen that
Ri st determines the entrainment rate. For low values of
Ri st , a situation with Ri . Ri st is reached earlier with
the consequence that turbulence is decaying and the
entrainment is reduced. For high values of Ri st , Ri ,
Ri st holds over nearly the entire mixed layer and there-
fore mixing and entrainment is enhanced. It can be seen
from this experiment that Ri st 5 0.25 seems to be a
reasonable value for the steady-state Richardson num-
ber. With this choice, we can now fix c «3 for each set
of stability functions (see Table 2). The value of Ri st 5
0.25 cannot be reached by the model of Kantha and
Clayson (1994). We therefore use Ri st 5 0.225 for their
FIG. 9. The ratios L E /L O and L E /L b as functions of the gradient
Richardson number Ri, where L E is the Ellison, L O the Ozmidov, and model, which corresponds to the value for c «3 given in
L b the buoyancy scale. Shown are curves resulting from the Kantha Table 2. Figures 11 and 12 show results for the mixed
and Clayson (1995) (KC), the Rodi (1980) and Hossain (1980) (RH), layer depth evolution and profiles of eddy viscosity and
the Canuto et al. (2001) (CA, CB) second-moment closures. These diffusivity and turbulent kinetic energy at 30 hours after
are compared to the empirical estimates from Eq. (37).
the onset of surface stress. Eight different sets of sta-
bility functions have been used for these simulations,
namely the models KC, RH, CA, CB and their quasi-
equilibrium versions. For all simulations, the dissipation

FIG. 10. Development of the mixed layer depth (deepest point with k . 10 25 J kg 21 ) for the simulation of the Kato–
Phillips experiment. Model results for the complete versions of the models A (left) and B (right) of Canuto et al.
(2001) for various values of Ri st .

FIG. 11. Development of the mixed layer depth (deepest point with k . 10 25 J kg 21 ) and profiles of
mixing coefficients and TKE after 30 h for the simulation of the Kato–Phillips experiment. Model results
for the complete versions of the (a) model of Kantha and Clayson (1994), (b) Rodi (1980) and Hossain
(1980), (c) model A of Canuto et al. (2001), and (d) model B of Canuto et al. (2001). For all model runs,
Ri st 5 0.25 has been chosen, with the exception of the KC model, where Ri st 5 0.225 was used.

rate equation (28) has been used with the values for c «3 and therefore produces useless results. This has already
from Table 3. It can be seen from Fig. 11 that the non- been reported by Deleersnijder and Luyten (1994) for
equilibrium version of the KC model [it should be noted the similar model of Mellor and Yamada (1982). In
that it is the quasi-equilibrium version suggested by contrast to this, the quasi-equilibrium version of the KC
Kantha and Clayson (1994)] tends to strong oscillations model performs well; the empirical and simulated mixed

FIG. 12. Development of the mixed layer depth (deepest point with k . 10 25 J kg 21 ) and profiles of
mixing coefficients and TKE after 30 h for the simulation of the Kato–Phillips experiment. Model results
for the quasi-equilibrium versions of the (a) model of Kantha and Clayson (1994), (b) Rodi (1980) and
Hossain (1980), (c) model A of Canuto et al. (2001), and (d) model B of Canuto et al. (2001). For all model
runs, Ri st 5 0.25 has been chosen, with the exception of the KC model, where Ri st 5 0.225 was used.

layer depth are very close to each other (see Fig. 12). mixed layer experiment do unfortunately not exist.
It is however strange that the profile of turbulent kinetic However, in a large eddy simulation study of a similar
energy shows a maximum in the lower part of the mixed experimental setup (but with consideration of rotation)
layer. This effect has already been demonstrated by Bur- carried out by Moeng and Sullivan (1994) such a local
chard et al. (1998). Turbulence measurements for this maximum of k is not visible.

FIG. 13. Free convection experiment of Willis and Deardorff (1974). Profiles of the normalized temperature profile (T 2 Tmax )/T , normalized
temperature flux ^w̃T̃&/w T , and the normalized autocorrelations ^ũ 2 &/w*2 , ^w̃ 2 &/w*2 , and ^T̃ 2 &/T 2* and dissipation rate «/B 0 calculated by using
different models for the *stability
* functions: Kantha and Clayson (1994) (KC), Rodi (1980) and Hossain (1980) (RH), Canuto et al. (2001)
version A (CA), and Canuto et al. (2001) version B (CB). The model simulations are compared to LES simulations by Mironov et al. (1999).

The other sets of stability functions at show (i) a Bryan (1969), the depth is here the height of the ho-
perfect fit with the empirical curve of Price (1979), (ii) mogenized layer] and the KPP model [see Large et al.
the expected monotone decrease of turbulent kinetic en- (1994), the value has been estimated from their Fig. 1].
ergy down from the surface, and (iii) a numerically As expected, all depths for the models presented here
stable performance. are between the latter two depths. The least deepening
(11.2 m) is provided by the convective adjustment
scheme, which does not perform any active entrainment
b. Free convection
in terms of steepening the buoyancy gradient below the
Although strong convective events occur in the ocean convective boundary layer. The models presented here
in only a few areas, they are important for the ocean mix deeper (11.9–12.5 m) due to their capability of
circulation and it is therefore desirable that they are reproducing active entrainment. The fully empirical,
sufficiently reproduced by turbulence closure models. nonlocal KPP model provides further deepening (13.0
A free convection simulation similar to the laboratory m) of the convective boundary layer.
experiment carried out by Willis and Deardorff (1974) Furthermore, after 3 days of cooling profiles of var-
will be presented here. The scenario simulated here is ious quantities are shown, normalized by the Deardorff
the same as that used by Large et al. (1994). By means convective velocity scale w* 5 (B 0 H)1/3 , the tempera-
of a constant negative surface heat flux of 100 W m 22 , ture scale T* 5 (n9t ] z T) | z50 /w*, and the surface buoy-
a convective boundary layer is entrained into a stably ancy flux B 0 versus z/H. Here H is the depth of the
stratified ocean with a surface temperature of 228C and entrainment layer the base of which is defined as the
a temperature gradient of 18C per 10 m. Shear and ro- height with the minimum heat flux. The results for mean
tation are not present. For this free convection simu- temperature, heat flux, dissipation rate, and the vari-
lation recent LES data are available (see Mironov et al. ances ^ũ 2 &, ^w̃ 2 &, and ^T̃ 2 & [see Eqs. (A10)–(A12)] are
2000). shown in Fig. 13. Three shortcomings of the model are
In Table 3, the entrainment depth (position of mini- obvious:
mum normalized turbulent heat flux) for all experiments
after 3 days of cooling is given. Comparison is made 1) Countergradient fluxes occur over a large portion
to a simple scheme with convective adjustment [see (roughly 20.8 # z/H # 20.4) of the convective

FIG. 15. Relative heat content of the water column at OWS Papa
from Mar 1961 to Mar 1962. Bold line: 5 day means of heat content
FIG. 14. Budget of the TKE equation calculated from the CA as calculated from measured temperature profiles. Thin line: 5 day
model compared to LES data from Mironov et al. (1999). Diss: means of heat content as calculated from surface heat fluxes.
turbulent dissipation rate; Buoy: buoyancy production; Turb: tur-
bulent transport. For the LES data, turbulent and pressure transport
are added together here. All terms are normalized with the surface
buoyancy flux B 0 , the height is normalized with the entrainment
layer depth. Kantha and Clayson 1994; D’Alessio et al. 1998). As
for any realistic oceanic test case, other factors than the
choice of the mixed layer model also play an important
role for the agreement between the model results and
boundary layer, which is not included in the model
the measurements. First of all, the momentum and heat
because of its inherent downgradient approximation.
fluxes at the sea surface are never available as direct
Therefore, temperature profiles of LES and turbu-
observations but are calculated using bulk formulae.
lence closures are principally different. This suggests
Measurements such as fractional cloud cover are never
that nonlocal processes are important here.
exact. Furthermore, horizontal advection of heat and
2) The height of the active entrainment layer is under-
salt, neglected in one-dimensional water column mod-
estimated by the turbulence closure models. This can
els, can strongly influence the measured profiles of tem-
best be seen in the profiles of ^w̃T̃& and ^T̃ 2 &. Potential
perature and salinity. Maybe most important, the bulk
reasons could here be the downgradient approxi-
formulae for the parameterization of cross-surface flux-
mation for the TKE flux and the fact that turbulent
es of momentum, heat, and freshwater are strictly em-
transport of ^w̃T̃& and ^T̃ 2 & is neglected. It can be
seen from Fig. 14 that the TKE-diffusion term is
How the bulk formulae for surface heat and momen-
qualitatively reproduced by the model (here CA), but
tum fluxes have been used here is discussed in detail in
underestimated by a factor of approximately 2.
Burchard et al. (1999). The relative heat content of the
3) The profile of ^ũ 2 & near the surface suggests that the
upper 250 m of the water column from temperature
turbulent transport of this quantity should not be
profiles and surface heat fluxes between March 1961
neglected here.
and March 1962 is shown in Fig. 15. Until the beginning
It should be noted that Canuto et al. (1994) obtained of November 1961 (around day 310), the agreement
good agreement between LES data and simulation re- between the curves is sufficient enough for allowing for
sults for a free convection experiment with a full Reyn- one-dimensional simulation. Afterward, cold water is
olds closure model using dynamic transport equations horizontally advected, a process described in detail by
for ^w̃T̃&, ^T̃ 2 &, ^w̃ 2 &, ^ũ 2 &, and « including complex al- Large et al. (1994). For mixing below the thermocline,
gebraic closure schemes for all relevant third-order mo- an internal wave and shear instability parameterization
ments. It should be the aim of future work to find a as suggested by Large et al. (1994) has been used.
reasonable compromise between this complex model by Figure 17 shows results of the model simulations with
Canuto et al. (1994) and the two-equation models pre- the stability functions of Canuto et al. (2001) in com-
sented here, in terms of both efficiency and predict- parison to measured temperature profiles (Fig. 16). The
ability. overall temperature evolution is well simulated by this
model. A more detailed comparsion between measure-
ments and two different model simulations of temper-
c. Wind and convective mixing in the open sea
ature profiles is shown in Fig. 18. Besides the afore-
For the northern Pacific, long-term observations of mentioned Canuto et al. (2001) model (CA) the simu-
meteorological parameters and temperature profiles are lations are carried out here additionally with the quasi-
available. OWS Papa at 508N, 1458W has the advantage equilibrium version of Kantha and Clayson (1994)
that it is situated in a region where the horizontal ad- (KC). Until day 210, the agreement between both sim-
vection of heat and salt is assumed to be small. Various ulations and the observations is fairly good. Then,
authors used these data for validating turbulence closure around day 240, the models predict a too shallow mixed
schemes (Denman 1973; Martin 1985; Large et al. 1994; layer, obviously due to erroneous surface fluxes or

FIG. 16. Temperature evolution for OWS Papa in the northern Pacific Ocean from Mar
1961 to Mar 1962 from CTD measurements.

strong advective events such as downwelling (see Fig. tha and Clayson (1994), who had (while applying a kL
15), where a mismatch between the heat content of the instead of an « equation) to use a background diffusivity
water column and the accumulated surface heat fluxes five times higher in order to predict the SST realistically.
is evident around day 240. It can be seen as well that This leads however to a thermocline too diffusive com-
the KC model predicts a slightly shallower mixed layer pared to measured temperature profiles (see Burchard
than the CA model. This has the consequence that the et al. (1999)).
KC model overpredicts the SST during summer (days
210–280 see Fig. 19). Until day 280, the rms error for
7. Discussion and conclusions
SST between both simulations and the observations is
rather small, 0.368C for the Canuto et al. (2001) and The concept of the steady-state gradient Richardson
0.338C for the Kantha and Clayson (1994) model. How- number Ri st is applied here for adjusting mixed layer
ever, the SST evolution strongly depends on the internal models such that they realistically predict the mixed
wave parameterization, and thus these rms errors are not layer depth and consequently the sea surface tempera-
discriminative of the quality of the turbulence models. ture in the ocean. The simulations of the wind entrain-
It should be noted that we have used exactly the pa- ment experiment suggest a value of Ri st 5 0.25. Al-
rameters of Large et al. (1994) for this, other than Kan- though the empirical relation for the mixed layer sug-

FIG. 17. Temperature evolution for OWS Papa in the northern Pacific Ocean from Mar
1961 to Mar 1962. Results of the simulation with the version A of the Canuto et al. (2001)
model with the stationary gradient Richardson number set to Rist 5 0.25.

FIG. 18. Measured and simulated temperature profile at station OWS Papa during spring and summer 1961. The
simulations were carried out with the quasi-equilibrium version of the Kantha and Clayson (1994) model with the
stationary gradient Richardson number set to Rist 5 0.225 and the version A of the Canuto et al. (2001) model with
the stationary gradient Richardson number set to Rist 5 0.25.

gested by Price (1979), Eq. (38), cannot be used for because of its critical gradient Richardson number of
calibrating Ri st in a strict way, it can be concluded from Ri c 5 0.235. We therefore used Ri st 5 0.225 only for
Fig. 10 that it should be between 0.2 and 0.3. According this model, and the predicted mixed layer depth is in
to Schumann and Gerz (1995), Ri st , 0.25 should hold. good agreement with the curve of Price (1979). It should
By analyzing laboratory data from Rohr (1985) obtained be noted again, that the KC model is a slight improve-
from homogeneously shear-layered saltwater flow, they ment of the models of Mellor and Yamada (1982) and
conclude that Ri st 5 0.16 6 0.06, a value significantly Galperin et al. (1988), which compute Ri c 5 0.19. The
smaller than the value we suggest. The discrepancy wind entrainment experiment shows that the stability
could be explained by the argument that mixed layer functions of Kantha and Clayson (1994) do not perform
dynamics are quantitatively different than homogeneous well. This is due to inherent numerical instabilities (full
shear layers due to the nonnegligible vertical fluxes of version) and an unrealistic local maximum of turbulent
turbulent quantities. More efficient parameterizations kinetic energy right above the pycnocline (quasi-equi-
for turbulent transport of TKE than the downgradient librium version). A further weakness of the KC model
approach used here could assumedly result in more re- became apparent when the ratios of Ellison to Ozmidov
alistic estimates for Ri st . length scale and Ellison to buoyancy length scale had
Eight different sets of stability functions are tested been compared to empirical curves for structural equi-
here, namely the full and the quasi-equilibrium versions librium turbulence. These ratios were far larger than
of Kantha and Clayson (1994) (KC), Rodi (1980) and those for the other models, which all performed suffi-
Hossain (1980) [version of Burchard and Baumert ciently well. For the case of free convection, all models
(1995), RH], and versions A and B of Canuto et al. predict significant deviations from recent LES data. The
(2001) (CA, CB). The analysis of the quasi-equilibrium major deficiencies are possibly due to the neglect of
versions of the stability functions shows that the model turbulent fluxes of second moments, which can in prin-
of Kantha and Clayson (1994) cannot reach Ri st 5 0.25 ciple be added to the models presented here.

for vertical mixing even at high Richardson numbers.

(iii) They proved in this investigation a high predict-
ability since they produced good results for OWS Papa
without tuning them to these scenarios. (iv) They are
rather economical with only two additional equations
for turbulent quantities plus evaluating the ratio of two
polynomials. (v) They were numerically robust since
the time step was Dt 5 300 s with a vertical resolution
of Dz 5 1 m for OWS Papa and no numerical insta-
bilities could be seen. The ability of k–« models to
produce high-order approximations also for coarse ver-
tical resolutions has already been shown by Burchard
and Petersen (1999). In order to achieve this, a special
numerical treatment of the dissipation rate near the sur-
face is needed (see Burchard and Petersen 1999). Ocean
general circulation models have already been coupled
to two-equation turbulence models a decade ago [see
Rosati and Miyakoda (1988), who applied the Mellor
and Yamada (1982) k–kL model for a World Ocean cir-
culation study]. Applications of k–« models in OGCMs
have to the knowledge of the authors not been made
FIG. 19. SST at OWS Papa from Mar 1961 to Mar 1962. (a) yet. With the ongoing growth of computer resources,
Measured SST and SST simulated with the quasi-equilibrium ver-
sion of the Kantha and Clayson (1994) model with the stationary the extensive use of more complex turbulence closure
gradient Richardson number set to Ri st 5 0.225. (b) Measured SST schemes will be more feasible.
and SST simulated with the version A of the Canuto et al. (2001) In contrast to the properties (i)–(v) of the stability
model with the stationary gradient Richardson number set to Ri st 5 functions by Canuto et al. (2001), those of Rodi (1980)
0.25. and Hossain (1980) [version of Burchard and Baumert
(1995)] definitely do not fulfil (iv), computational
economy, due to their implicit formulation. As already
Two models are finally used for simulations of ocean mentioned above, the quasi-equilibrium version of the
mixed layer dynamics: the quasi-equilibrium version of Kantha and Clayson (1994) stability functions do not
the KC model and the full version of the CA model. fulfil (ii), physical soundness, due to the local TKE
The result is that the CA model mixes slightly deeper maximum in the entrainment experiment, and predict
than the KC model, maybe due to the smaller steady- ratios of relevant turbulent length scales for structural-
state gradient Richardson number chosen for the latter equilibrium turbulence that are far from observations.
model. Too shallow mixed layers computed by so-called The full version of the Kantha and Clayson (1994) as
differential mixed layer models (defined in contrast to well as the Mellor and Yamada (1982) stability func-
bulk models that average over the entire mixed layer) tions did not even allow for a numerically stable com-
have been reported by Martin (1985), who used, among putation.
others, the OWS Papa data for comparing various mod-
els. He explained this phenomenon by a too small crit- Acknowledgments. The work of Hans Burchard has
ical Richardson number and could obtain acceptable been funded through a Habilitation grant by the Deut-
model results at OWS Papa by increasing the critical sche Forschungsgemeinschaft (German Research Foun-
Richardson number to Ri c 5 0.3. dation) and the project PROVESS (MAS3-CT97-0025)
It should however be noted that it is mainly the of the MAST-III program of the European Commission.
steady-state Richardson number that determines the The work of Karsten Bolding has been carried out in
growth or decay of turbulence rather than the critical the framework of the project PROVESS as well. Both
Richardson number, which just sets an upper limit to authors are supported by the CARTUM project, a con-
the gradient Richardson number. certed action of the MAST-III program of the European
Coming back to the five paradigms for good turbu- Commission (MAS3-CT98-0172). The research for this
lence closure mentioned in the introduction—(i) deri- manuscript has been motivated by a visit of Vittorio
vation from second-moment transport equations, (ii) Canuto (New York) at the Institute for Oceanography
physical soundness, (iii) high predictability, (iv) com- in Hamburg, Germany, in September 1999. We are
putational economy, and (v) numerical robustness—the grateful to him for presenting us the manuscript Canuto
Canuto et al. (2001) stability functions are those that et al. (2001) directly after submission to J. Phys. Ocean-
best fulfill these requirements. (i) They consider more ogr. Helmut Baumert (Wedel, Germany) supported us
terms for the pressure–strain relations than all other with numerous valuable hints and comments on this
models. (ii) They are physically sound since they allow manuscript. The authors are further indebted to Manuel

Ruiz Villarreal (Santiago de Compostela, Spain), Pierre- which the stability functions depend are here conse-
Philippe Mathieu (Reading, United Kingdom), and quently expressed as
Georg Umgiesser (Venice, Italy) for their support in the
k2 2 k2 2
framework of the General Ocean Turbulence Model aM 5 S , aN 5 N . (A4)
(GOTM, online at, which has «2 «2
been used for the presented numerical calculations. And With the Mellor and Yamada (1974, 1982) notation for
we would like to express that we highly appreciated the these nondimensional parameters,
suggestions of two anonymous referees.
L2 2 L2 2
gm 5 S , gh 5 N , (A5)
Mathematical Details q2 q2

a. Notation the conversion between a M , a N , and g m , g h is of the

following form:
Different notations have always been a major thresh-
old for comparing different turbulence parameteriza- cL2 cL2
gm 5 a , gh 5 a . (A6)
tions. For calculating the eddy viscosity n t and eddy 2 M 2 N
diffusivity n9t , we use the following form (see section
By means of (A3) and (A6), the sets of stability func-
tions presented in sections 3e(1)–(3) may be trans-
k2 k2 formed to the other notation.
nt 5 cm , nt9 5 c9m , (A1)
« « It should be noted that further definitions of stability
functions are often used such as n t 5 c mÏkL (see Bur-
which automatically results from the algebraic Reynolds chard et al. 1998) and n t 5 2c m k 2 /« (see Canuto et al.
stress closure. In other models such as Mellor and Ya- (2001)). Another reason for confusion is often the dif-
mada (1974, 1982), an alternative formulation is used: ferent use of the parameter c «3 in the dissipation rate
equation (28). In this paper, the right-hand side of that
K M 5 S M qL, K H 5 S H qL (A2) equation reads as («/k)(c «1 P 1 c «3 B 2 c «2 «), but often
the form («/k)[c «1 (P 1 c «3 B) 2 c «2 «] (Rodi 1980; Bur-
with q 2 5 2k and K M and K H being other notations for chard and Baumert 1995) is used. This, of course will
the eddy viscosity and diffusivity n t and n9t , respectively; lead to different values for c «3 .
S M and S H correspond to the stability functions c m and
c9m with the conversion
b. Production, shear, and vorticity after boundary
cm c9m layer approximation
SM 5 , SH 5 (A3)
Ï2cL Ï2cL
Here, the matrices P ij , B ij , S ij , and V ij resulting from
with c L from (19). the boundary layer approximation (see section 3d) are
The nondimensional shear and buoyancy numbers on given:

 22]z u^ũw̃& 2]z u^ỹ w̃& 2 ]z y ^ũw̃& 2]z u^w̃ 2 &

 
Pij 5 2] z u^ ỹ w̃& 2 ]z y ^ũw̃& 22]z y ^ỹ w̃& 2]z y ^w̃ 2 & (A7)
 
 2]z u^w̃ 2 & 2]z y ^w̃ 2 & 0 

 0 0 b ^ũT&
˜ 
 ˜ 
Bij 5  0 0 b ^ỹ T& (A8)

b ^ũT&
˜ b ^ỹ T&
˜ 2b ^w̃T&˜ 
 0 0 ]z u   0 0 ]z u 
1  1 
Sij 5  0 0 ]z y  , Vij 5  0 0 ]z y  . (A9)
2  2 
]z u ]z y 0  2]z u 2]z y 0 

c. Autocorrelations validation (see section 6b). The autocorrelation terms

Although not needed for closure of the turbulent are often measured in the field or obtained in idealistic
transport terms for momentum and heat, it is possible situations by large eddy simulations or direct numerical
to compute the autocorrelations (or variances) of ve- simulations. For three closure procedures, these auto-
locity and temperature. They can be used for model correlators can be written as

2 k E 2 nt (]z u ) 2 2 E 3 nt (]z y ) 2 2 B
^ũ 2 & 5 k 1 E1 , (A10)
3 «
1 2
1 1 E4 1 21
« «
2 k 2E 3 nt (]z u ) 2 1 E 2 nt (]z y ) 2 2 B
^ỹ 2 & 5 k 1 E1 , (A11)
3 «
1 2
1 1 E4 1 21
« «
2 k (E 2 2 E 3 )nt (]z u ) 1 (E 2 2 E 3 )nt (]z y ) 2 2 2B
^w̃ 2 & 5 k 2 E1 , (A12)
3 «
1 2
1 1 E4 1 21
« «
^T˜ 2 & 5 cT nt9 (]z T ) 2 , (A13)

where the empirical parameters E1 , . . . , E 4 are given A1

in Table A1. Exact expressions for these parameters are s0 5 4 (1 2 3C1 ),
given in the appendix sections d–f.
It can be easily seen that (A10–A12) are consistent A1 A 2
with the definition of turbulent kinetic energy k [see Eq. s1 5 16 {(1 2 3C1 )[3B2 (1 2 C3 ) 1 12A1 ]
A36]. It should be noted that the well-known stability
functions suggested by Galperin et al. (1988) are in- 2 3[4A1 1 3A 2 (1 2 C2 )]},
cluded in the analysis of this paper as well. They are
obtained from the KC model by assuming quasi-equi- A2 A1 A22 A12 A 2 C1
s2 5 4 , s3 5 144 , s4 5 288 ,
librium and setting c 2 5 c 3 5 0. In the paper by Galperin B1 B13 B13
et al. (1988), (A36) does apparently not hold, because
a residual term remains. However, it can be shown that A2 A12
t1 5 12 [B (1 2 C3 ) 1 7A1 ], t 2 5 24 ,
this residual term is proportional to P 1 B 2 «, which B12 2 B12
vanishes with the quasi-equilibrium assumption.
A12 A 2
t 3 5 288 [23A 2 (1 2 C2 ) 1 B2 (1 2 C3 )],
d. Exact form of the Kantha and Clayson (1994) B14
stability functions A1 A22
The full form of the stability functions derived from t 4 5 432 [B2 (1 2 C3 ) 1 4A1 ]. (A16)
the parameterization by Kantha and Clayson (1994),
which is given by means of rounded numbers in Eq. The parameters A1 , A 2 , B1 , B 2 , C1 , C 2 , and C 3 , are taken
(21) is given here in the exact form. This allows us to from Kantha and Clayson (1994) and relate to the pa-
change parameterizations and carry out sensitivity stud- rameters given in Table 1 as follows:
ies with these formulas:
B1 B1
s0 1 s1 a N A1 5 5 0.92, A2 5 5 0.74,
cm 5 , (A14) 6c1 6c1T
1 1 t1 a N 1 t 2 a M 1 t 3 a N a M 1 t 4 a N2
B1 5 16.6, B2 5 0.5B1 cT 5 10.1,
s2 1 s3 a N 1 s4 a M
c9m 5 . (A15)
1 1 t1 a N 1 t 2 a M 1 t 3 a N a M 1 t 4 a N2
1 2
1 1 6A
C1 5 1 2 1/3 2 1 ø 0.008,
Here the parameters are defined as follows: 3 B1 A1 B1
C2 5 c2T , C3 5 c3T . (A17)
TABLE A1. Parameters for calculating the variances ^ũ 2 &, ^ỹ 2&,
^w̃ 2 &, and ^T̃ 2 &. The coefficients for calculating the autocorrelators in
Eqs. (A10)–(A12) are given as
Model E1 E2 E3 E4
KC 0.224 2.0 1.0 0.0 A1
RH 0.185 1.6 0.8 0.556
E1 5 4 , E 2 5 2, E 3 5 1,
CA 0.160 1.093 0.027 0.0
CB 0.135 1.362 0.033 0.0 E 4 5 0. (A18)

e. Exact form of the Rodi (1980) and Hossain (1980) 8

stability functions k 5 5 2 (c1 2 1){[(1 2 a 5 )a 3 (a 3 1 a10 a N ) 2 a 7 a 9 a N ]a M
As already mentioned, the stability functions of Rodi
2 a 7 a N2 2 a1 a 3 a N}, (A20)
(1980) and Hossain (1980) in the version of Burchard
and Baumert (1995) are implicit and therefore require
numerical methods for solution. First of all, the largest with
solution of the following fifth-order equation has to be
found. We apply a simple Newton iteration, and need 1
less than ten iterations in order to receive the solution a1 5 c1 2 1, a 3 5 c1T 2 , a 5 5 c2 ,
within an accuracy of 10 210 , if a good initial value for
(P 1 B)/e is found, for which we use 0.25a M 1 2a N : c7 5 1 2 c3 , a 9 5 1 2 c2T , a10 5 (1 2 c3T )cT ,

1 2 1 2 1 2
5 4 3
1 k1 1 k2 a11 5 2 c3 . (A21)
« « « 2

1k 1
« 2
1k 1
« 2

1k 50 (A19) After finding a solution for (P 1 B)/e, the following

3 4 5
parameters have to be solved:
with the coefficients
  21
k1 5 2a1 1 2(2a 3 1 a10 a N ), 1 2 c3T
A 5 1 1 cT a N , (A22)

5 6
k 2 5 4a 3 (a 3 1 a10 a N ) 1 4a1 (2a 3 1 a10 a N ) 1 a12  1 P B 
 c 1T 1
2 «
1 21

8 2  
1 a11 a N 1 2a 7 a N 2 (1 2 a 5 )a 5 a M ,
3 3 1 2 c3 1
D 5 11 aN,

1 2 1 2 5 6 5 6
16 1 1 P B 1 P B
k 3 5 a11 a1 1 a 3 a N 1 4a 7 a1 1 a 3 1 a10 a N a N c1 1 1 2 1 c1T 1 1 21
3 2 2 « « 2 « «
4 (A23)
2 a 5 (1 2 a 5 )(2a 3 1 a10 a N )a M
1 8a 3 a1 (a 3 1 a10 a N ) 1 2a12 (2a 3 1 a10 a N ) 3
2 c3
4 2 1
1 2 E 5 11 Aa N ,
8 1 1
2 (c1 2 1) a M (1 2 a 5 ) 2 a N ,
5 6 5 6
3 P B 1 P B
3 4 2 c1 1 1 2 1 c1T 1 1 21
16 « « 2 « «
k 4 5 a1 a 3 a11 a N 1 4a1 a 7 (a 3 1 a10 a N )a N
3 (A24)
8 16 1 2 c3
2 a 3 a 5 (1 2 a 5 )(a 3 1 a10 a N )a M 1 a 7 a11 a N2 H512
3 3 1 2 c2
8 1 2 c2T
1 a 5 a 7 a 9 a M a N 1 4a12 a 3 (a 3 1 a10 a N ) 3
Aa N .
2 (c1 2 1)
c1T 1 5
1 P B
1 2 1 c1T 1
2 « «
1 P B
1 21
2 « « 5 6
2 [ 1
3 (1 2 a 5 )(2a 3 1 a10 a N )a M 2 a1 1 a 3 a N ,
2 1 2 ] The anisotropy ^w̃ 2 &/k is then obtained from

c1 2 1

^w̃ 2 & 2
c1 1 5
1 21
« 6
5 . (A26)
k 3 2 c2 1 2 c2 H
E2 a
5 6 5 6
3 P B P B
c1 1 1 2 1 c1 1 1 21
« « « «

Finally, these parameters can be used for calculating the 4

stability functions c m and c9m: t 0 5 3L 25 , t1 5 L 5 (7L 4 1 3L 8 ) ,

1 2 c2 H ^w̃ 2 & 4
t 2 5 [L 25 (3L 23 2 L 22 ) 2 0.75(L 26 2 L 27 )] ,
cm 5 , (A27) t0
5 6
P G D k
c1 1 1 21
« « 16
t 3 5 L 4 (4L 4 1 3L 8 ) ,
A ^w̃ &
2 t0
c9m 5 . (A28)

5 6
1 P G k t 4 5 {L 4 [L 2 L 6 2 3L 3 L 7 2 L 5 (L 22 2 L 32 )]
c1T 1 1 21
2 « « 16
1 L 5 L 8 (3L 32 2 L 22 )} ,
The coefficients for calculating the autocorrelators in t0
Eqs. (A10)–(A12) are given as
t 5 5 0.25(L 22 2 3L 23 )(L 26 2 L 27 ) (A33)
2 t0
(1 2 c3 )
3 1 2 c2 with
E1 5 , E2 5 2 ,
c1 1 2 c3 4 1 1 2 c2 1 2 c2 1 c5
L1 5 , L2 5 , L3 5 ,
1 2 c2 1 15 c1 2c1 2c1
E3 5 , E4 5 . (A29)
1 2 c3 c1 c3 2 1
L4 5 , L 5 5 2c1T , L 6 5 1 2 c2T ,
5 1
f. Exact form of the Canuto et al. (2001) stability L 7 5 1 2 c2T , L 8 5 cT . (A34)
functions 3 3
It should be noted that c 4T 5 ⅔c 2T .
Only for version A of the Canuto et al. (2001) stability The coefficients for calculating the autocorrelators in
functions, the exact form is given here: equations (A10)–(A12) are given as
4 1 L 2 1 3L 3 L2
s0 1 s1 a N 1 s 2 a M E1 5 L 4 , E2 5 , E3 5 ,
cm 5 , (A30) 3 2 L4 L4
1 1 t1 a N 1 t 2 a M 1 t 3 a N2 1 t 4 a N a M 1 t 5 a M2
E 4 5 0. (A35)
s4 1 s5 a N 1 s6 a M
c9m 5 , (A31)
1 1 t1 a N 1 t 2 a M 1 t 3 a N2 1 t 4 a N a M 1 t 5 a M2
where g. Exact equations for k and «
The turbulent kinetic energy k, defined as the kinetic
2 energy per unit mass of the velocity fluctuations is de-
s0 5 1.5L 1 L 25 ,
t0 fined as

[ 1
s1 5 2L 4 (l 6 1 L 7 ) 1 2L 4 L 5 L 1 2 L 2 2 L 3
3 1 2
k 5 ^ỹ i2 &,
(the unit is usually J kg 21 ). A transport equation for k

1 1.5L 1 L 5 L 8
] 8
can be derived directly from the exact transport equation
for the Reynolds stresses (see, e.g., Sander 1998):

s2 5 2L 5 ,
s4 5 2L 5 ,
s5 5 2L 4 ,
t0 1 7 8
1 1
]t k 1 ]j ỹ j k 1 ỹ j ỹ i2 2 n ]j k 1 ^ỹ j p̃&
2 r0 2
s6 5
L (3L 23 2 L 22 ) 2 0.5L 5 L 1 (3L 3 2 L 2 )
3 5
5 2^ỹ j ỹ i &]j y i 2
| |
^ỹ r̃& 2 n ^(]j ỹ i ) 2 &.
r0 3
| |

1 0.75L 1 (L 6 2 L 7 )
] 8

For the dissipation rate,


« 5 n^(] j ỹ i ) 2 &


(the unit is usually W kg 21 ), which appears as a sink ——, A. Howard, Y. Cheng, and M. S. Dubovikov, 2001: Ocean
on the left-hand side of the k equation, an exact equation turbulence. Part I: One-point closure model momentum and heat
vertical diffusivities. J. Phys. Oceanogr., 31, 1413–1426.
can be derived as well: Craig, P. D., and M. L. Banner, 1994: Modeling wave-enhanced tur-
bulence in the ocean surface layer. J. Phys. Oceanogr., 24, 2546–

1 2
n 2559.
]t « 1 ]j ỹ j « 1 ^ỹ j n (]j ỹ i ) 2 & 2 n ]j « 1 2 ^] ỹ ] p̃&
r0 i j i D’Alessio, S. J. D., K. Abdella, and N. A. McFarlane, 1998: A new
second-order turbulence closure scheme for modeling the oce-
g anic mixed layer. J. Phys. Oceanogr., 28, 1624–1641.
5 22n 2 ^(]ij ỹ l ) 2 & 2 2n ] ^ỹ ] r̃& 2 2n ^]j ỹ i ]l ỹ i ]j ỹ l & Deleersnijder, E., and P. Luyten, 1994: On the practical advantages
r0 j 3 j of the quasi-equilibrium version of the Mellor and Yamada level
| |
| |
| 2.5 turbulence closure applied to marine modelling. Appl. Math.
P« B« «« Model., 18, 281–287.
Denman, K. L., 1973: A time-dependent model of the upper ocean.
2 2n ]j y l ^]j ỹ i ]l ỹ i & 2 2n ]l y i ^]j ỹ i ]j y l & J. Phys. Oceanogr., 3, 173–184.
Galperin, B., L. H. Kantha, S. Hassid, and A. Rosati, 1988: A quasi-
2 2n ]jl y i ^ỹ l ]j ỹ i &. (A39) equilibrium turbulent energy model for geophysical flows. J.
Atmos. Sci., 45, 55–62.
It is evident that both equations (A37) and (A39), are Haidvogel, D. B., and A. Beckmann, 1999: Numerical Ocean Cir-
culation Modelling. Series on Environmental Science and Man-
not affected by Coriolis rotation and that pressure fluc-
agement, Vol. 2, Imperial College Press, 318 pp.
tuations do not act as sources or sinks for k and «, but Hossain, M. S., 1980: Mathematische Modellierung von turbulenten
only transport these quantities as advective or diffusive Auftriebsströmungen. Ph.D. dissertation, University of Karls-
transports. The right-hand side of the transport equation ruhe, Karlsruhe, Germany, 145 pp.
for k can be left unchanged; however, the right-hand Kantha, L. H., and C. A. Clayson, 1994: An improved mixed layer
model for geophysical applications. J. Geophys. Res., 99,
side of the « equation needs some drastic empirical as- 25 235–25 266.
sumptions to be closed. In principle, it is assumed that Kato, H., and O. M. Phillips, 1969: On the penetration of a turbulent
the sources and sinks of « (after scaling them with the layer into stratified fluid. J. Fluid Mech., 37, 643–655.
turbulent timescale k/«) are proportional to those in the Kolmogorov, A. N., 1942: The equations of turbulent motion on an
k equation [see Eq. (28)]. The empirical coefficients c «1 , incompressible fluid. Izv. Akad. Nauk. USSR, Ser. Fiz., VI (1–
2), 56–58.
c «2 , and c «3 are then derived from laboratory experi- Large, W. G., and P. R. Gent, 1999: Validation of vertical mixing in
ments (freely decaying grid turbulence for c «2 ) and the an equatorial ocean model using large eddy simulations and
log-law (c «1 , see Rodi 1980). The meaning of c «3 has observations. J. Phys. Oceanogr., 29, 449–464.
first been discussed in detail by Burchard and Baumert ——, J. C. McWilliams, and S. C. Doney, 1994: Oceanic vertical
(1995) and is further investigated in this paper (see sec- mixing: A review and a model with nonlocal boundary layer
parameterization. Rev. Geophys., 32, 363–403.
tion 5). Launder, B. E., 1975: On the effect of a gravitational field on the
turbulent transport of heat and momentum. J. Fluid Mech., 67,
REFERENCES ——, and D. Spalding, 1972: Mathematical Models of Turbulence.
Baumert, H., and H. Peters, 2000: Second-moment closures and Academic Press, 169 pp.
length scales for weakly stratified turbulent shear flows. J. Geo- ——, G. J. Reece, and W. Rodi, 1975: Progress in the development
phys. Res., 105, 6453–6468. of a Reynolds-stress turbulence closure. J. Fluid Mech., 68, 537–
Bryan, K., 1969: A numerical model for the study of the world ocean. 566.
J. Comput. Phys., 4, 347–376. Lesieur, M., 1997: Turbulence in Fluids. 3d ed. Kluwer Academic,
Burchard, H., 2001: Note on the q 2 l equation by Mellor and Yamada 515 pp.
(1982). J. Phys. Oceanogr., 31, 1377–1387. Martin, P. J., 1985: Simulation of the mixed layer at OWS November
——, and H. Baumert, 1995: On the performance of a mixed-layer and Papa with several models. J. Geophys. Res., 90, 903–916.
model based on the k–e turbulence closure. J. Geophys. Res., Mellor, G. L., 1989: Retrospect on oceanic boundary layer modeling
100, 8523–8540. and second moment closure. Parameterization of Small-Scale
——, and O. Petersen, 1999: Models of turbulence in the marine Processes: Proc ‘Aha Huliko’ a Hawaiian Winter Workshop,
environment—A comparative study of two-equation turbulence Manoa, HI, University of Hawaii, 251–271.
models. J. Mar. Syst., 21, 29–53. ——, 2001: One-dimensional, ocean surface layer modeling: A prob-
——, ——, and T. P. Rippeth, 1998: Comparing the performance of lem and a solution. J. Phys. Oceanogr., 31, 790–803.
the Mellor–Yamada and the k–e two-equation turbulence models. ——, and T. Yamada, 1974: A hierarchy of turbulence closure models
J. Geophys. Res., 103, 10 543–10 554. for planetary boundary layers. J. Atmos. Sci., 31, 1791–1806.
——, K. Bolding, and M. R. Villarreal, 1999: GOTM—A general ——, and ——, 1982: Development of a turbulence closure model
ocean turbulence model. Theory, applications and test cases. for geophysical fluid problems. Rev. Geophys., 20, 851–875.
European Commission Rep. EUR 18745 EN, 103 pp. Mironov, D. V., V. M. Gryanik, C.-H. Moeng, D. J. Olbers, and T.
Businger, J. A., J. C. Wyngaard, Y. Izumi, and E. F. Bradley, 1971: H. Warncke, 2000: Vertical turbulence structure and second-
Flux profile relationships in the atmospheric surface layer. J. moment budgets in convection with rotation: A large-eddy sim-
Atmos. Sci., 28, 181–189. ulation study. Quart. J. Roy. Meteor. Soc., 126, 477–515.
Canuto, V. M., 1994: Large eddy simulation of turbulence: A subgrid Moeng, C.-H., and P. P. Sullivan, 1994: A comparison of shear and
model including shear, vorticity, rotation and buoyancy. Astro- buoyancy-driven planetary boundary layer flow. J. Atmos. Sci.,
phys. J., 428, 729–752. 51, 999–1022.
——, F. Minotti, C. Ronchi, M. Ypma, and O. Zeman, 1994: Second- Monin, A. S., and A. M. Obukhov, 1954: Basic laws of turbulent
order closure PBL model with new third-order moments: Com- mixing in the ground layer of the atmosphere. Akad. Nauk SSSR
parison with LES data. J. Atmos. Sci., 51, 1605–1618. Geofiz. Inst. Tr., 151, 163–187.

Prandtl, L., 1945: Über ein neues Formelsystem für die ausgebildete Rotta, J. C., 1951: Statistische Theorie nichthomogener Turbulenz.
Turbulenz. Nachr. Akad. Wiss., Goettingen, Math.-Phys. Kl., Z. Phys., 129, 547–572.
p. 6. Roussenov, V., E. Stanev, V. Artale, and N. Pinardi, 1995: A seasonal
Price, J. F., 1979: On the scaling of stress-driven entrainment ex- model of the Mediterranean Sea general circulation. J. Geophys.
periments. J. Fluid Mech., 90, 509–529. Res., 100, 13 515–13 538.
Rodi, W., 1976: A new algebraic relation for calculating the Reynolds Sander, J., 1998: Dynamic equations and turbulent closures in geo-
stresses. Z. Angew. Math. Mech., 56, T219–T221. physics. Continuum Mech. Thermodyn., 10, 1–28.
——, 1980: Turbulence models and their application in hydraulics. Schumann, U., and T. Gerz, 1995: Turbulent mixing in stably stratified
International Association for Hydraulic Research, Delft, Neth- shear flows. J. Appl. Meteor., 34, 33–48.
Shih, T. S., and A. Shabbir, 1992: Advances in modeling the pressure
erlands, 104 pp.
correlation terms in the second moment equations. Studies in
——, 1987: Examples of calculation methods for flow and mixing Turbulence, T. B. Gatsky, S. Sarkar, and C. G. Speziale, Eds.,
in stratified flows. J. Geophys. Res., 92, 5305–5328. Springer-Verlag, 91–128.
Rohr, J. J., 1985: An experimental study of evolving turbulence in Taylor, G. I., 1935: Statistical theory of turbulence, Parts i–iv. Proc.
uniform mean shear flow with and without stable stratification. Roy. Soc. London, 151A, 421–478.
Ph.D. dissertation, University of San Diego, 271 pp. Willis, G. E., and J. W. Deardorff, 1974: A laboratory model of the
Rosati, A., and K. Miyakoda, 1988: A general circulation model for unstable planetary boundary layer. J. Atmos. Sci., 31, 1297–
upper ocean simulation. J. Phys. Oceanogr., 18, 1601–1626. 1307.