Procedures for offline grid nesting

in regional ocean models

Evan Mason1, Jeroen Molemaker2 , Alexander F. Shchepetkin2 ,

Francois Colas2 , James C. McWilliams2, and Pablo Sangrà1
(1) Facultad de Ciencias del Mar, Universidad de Las Palmas de Gran Canaria, 35017,
Las Palmas de Gran Canaria, Spain
(2) Institute of Geophysics and Planetary Physics, University of California Los Angeles,
Los Angeles, California, USA

Abstract

One-way offline nesting of a primitive-equation regional ocean numerical model (ROMS)

is investigated, with special attention to the boundary forcing file creation process. The
model has a modified open boundary condition which minimises false wave reflections,
and is optimised to utilise high-frequency boundary updates. The model configuration fea-
tures an prevously computed solution which supplies boundary forcing data to an interior
domain with an increased numerical resolution. At the open boundaries of the interior grid
(the child) the topography is matched to that of the outer grid (the parent), over a narrow
transition region. A correction is applied to the normal baroclinic and barotropic veloci-
ties at the open boundaries of the child to ensure volume conservation. It is shown that
these steps, together with a carefully constructed interpolation of the parent data, lead to a
high-quality child solution, with minimal artefacts at the boundaries.
Sensitivity experiments provide information about the robustness of the model open
boundary condition to perturbations in the surface wind stress forcing field, to the perturba-
tion of the volume conservation enforcement in the boundary forcing and, to perturbation
of the vertical density structure in the boundary forcing. This knowledge is important when
extending the nesting technique to include external data from alien sources, such as ocean
models with physics and/or numerics different from ROMS, or from observed climatolo-
gies of temperature and salinity.

1 Introduction

In numerical oceanic and atmospheric modelling, downscaling is a practical solution to the problem
of resolving wide-ranging spatial and temporal scales of motion under the constraint of finite computing
power and storage (e.g., Penven et al., 2006; Auclair et al., 2006; Cailleau et al., 2008). Downscaling
involves the embedding, or nesting, of multiple model grids one within the other, each successive grid
having a higher resolution than its parent. Such grid configurations may be run independently (offline
nesting) or synchronously (online nesting). The important distinction between the two is the frequency
of the boundary updates, which is higher for the online case. Higher frequencies are desirable but, in
the offline case, judicious choice of the updating interval allows the resolution of features of interest.

Preprint submitted to Elsevier Science 6 October 2009

Both approaches therefore yield realistic high-resolution solutions permitting study of local problems:
the large-scale mean circulation and its associated variability (mostly mesoscale, plus seasonal and
possibly inter-annual contributions) is provided by and passed down from the outermost parent, while
the innermost child generates the high-resolution mesoscale and submesoscale data for the physical
problem at hand.
Downscaling is a one-way process, in that information is passed solely downstream, from parent to
child. Upscaling is the reverse process, where information is passed from a child solution back up to the
parent. Two-way nesting brings together downscaling and upscaling, so that there is a dual exchange
of information: parent-to-child and child-to-parent. This technique is necessarily online, requiring addi-
tional software tools beyond the model code itself to facilitate the information exchange.
In regional oceanic modelling, the AGRIF Fortran (Adaptive Grid Refinement in Fortran; Debreu et al.,
2008) package is one such tool in use. AGRIF has been incorporated into, amongst others, the ROMS
(Regional Ocean Modelling System; Shchepetkin and McWilliams, 2005, 2009) and NEMO-OPA
(Madec et al., 1998; Madec, 2008) hydrodynamic models, permitting synchronous multi-level down-
scaled simulations. An evaluation of the system is presented by Penven et al. (2006), who conducted
nested ROMS experiments for the U.S. West Coast. Cailleau et al. (2008), in a comparative study of
different nesting techniques, report the use of two-way OPA nesting using AGRIF. These authors also
introduce a more sophisticated coupling method, the Schwarz domain decomposition method, which is
shown to improve on the two-way solution, albeit at a high computational cost. A recent development is
a ROMS-AGRIF two-way setup, described by Debreu et al. (2009). See also Debreu and Blayo (2008)
for a review of the algorithms used in two-way embedding.
However, the objective of this paper is to show that one-way offline nesting does still have a role
to play in regional ocean modelling, despite the emergence of more advanced methods. Offline nesting
permits great flexibility in terms of grid dimensions and orientation. Grids may be rotated, enabling
boundaries to be aligned with coastlines to maximise the oceanic extent of the computational domain.
Furthermore, timestep and grid-resolution ratios between each grid and its parent may be varied.
Downscaling is often sufficient to provide an adequate physical posing of a problem, so that two-way
nesting becomes excessively laborious. In this respect, there is still no clear idea of when one approach
or the other may be expected to make an important difference.

1.1 2D versus 3D boundary forcing methods

ROMS permits the use of either 2D or 3D boundary forcing. For 3D boundary forcing, the parent
data 3D variables (i.e., baroclinic velocities and tracers) used at the child boundaries are stored in three-
dimensional arrays on the child grid. 2D variables (barotropic velocities and sea surface height) are
stored in 2D arrays. The bulk of these data, i.e., those within the interior of the domain, are redundant as
only data at the boundary itself, plus a several-grid-cell-deep variable-strength relaxation (or nudging)
layer, are used. (Within the nudging layer the model variables are restored towards the 3D boundary
forcing values, e.g., Marchesiello et al., 2001.) In practical terms, this has forced modellers to restrict the
size of child domains and, more importantly, the temporal frequency of the boundary forcing. Without
these restrictions the storage size of 3D boundary forcing files becomes prohibitive, making it difficult
to provide boundary information which resolves the mesoscale, let alone the submesoscale.
The use of 2D boundary forcing files, where only information at the boundaries is used (i.e., no
interior nudging layer), is clearly more efficient in terms of disc storage than 3D. Using this approach
together with adequate computing resources, (combinations of) long solutions (multiple model years,

2
Fig. 1. Example of western boundary rim currents in the v-component of the mean annual model velocity of the
7.5-km parent (L0) solution, which covers the northeast Atlantic. Red lines demark the boundary of the child (L1)
domain discussed throughout the present text. The locations of the Canary Islands and capes mentioned in the text
are marked.

see for example, Sec. 4) on large grids (order of 1000×1000×80 gridpoints, e.g., McWilliams et al.,
2009) with high-resolution (in time, ¡ 1 day; and space, ¡ 100 m) boundary inputs, are feasible. However,
to fully profit from the use of high-resolution boundary input, it has been necessary to modify the ROMS
barotropic-velocity open boundary condition (OBC), and switch to an upwind advection scheme for
tracers at the open boundary. More details are given in Sec. 2.

1.2 Rim currents

The volumetric nudging used in the 3D boundary forcing method is not possible with the 2D method;
here we choose not to explore any sensitivity to the loss of this option. However, nudging is often cred-
ited with reducing so-called rim currents, which are a primary error mode that may occur in ocean model
simulations. Figure 1 shows an example of a persistent rim current: In the v-component of the annual
mean velocity from a 50-year UCLA-ROMS (see Shchepetkin and McWilliams, 2009) simulation of
the northeast Atlantic Ocean at 7.5-km resolution (see Mason et al., 2009 for details of this solution), a
narrow strip of anomalous velocities (positive and negative) is seen along the western boundary (note

3
that the strong positive velocity in the northwest corner is likely to be a realistic feature, Mason et al.,
2009).
Rim currents are probably a result of mismatches between the boundary forcing and the evolving
child solution, and also the ability of the chosen OBC to ameliorate mismatches. Mismatches may occur
in the stratification, particularly within the mixed layer; they may also arise if there are differences in
the parent and child surface forcings, switching from one wind product to another for example; and
also if volume conservation is not enforced in the boundary velocities. Rim currents, if present, show
up most readily by viewing just the tangential component of the velocity along a boundary as in Figure
1. However, because we have largely minimised rim currents in our nested solutions, in the following
sections we shall show them using relative vorticity, which is especially sensitive to velocity anomalies.
We have developed a methodology to facilitate downscaling within a ROMS framework. The need for
such a methodology has come about because of recent development in the UCLA-ROMS code that invite
the use of high-frequency boundary forcing. In the following sections we present results from a ROMS
configuration centred on the Canary Island archipelago in the northeast Atlantic, prepared using the new
methodology. In Sec. 2 we briefly describe the ROMS model that provides the numerical framework
for our experiments and the regional configuration that we have designed, and we also introduce the
downscaling techniques that have been developed for the experiments. Results are presented in Sec. 3,
which demonstrate the effectiveness of the procedures we use. In Sec. 4 we test the model’s sensitivity
to perturbations in the boundary and surface forcing data. Sec. 5 concludes with a discussion of the
implications of our results.

2 Numerical methods and configuration

2.1 The ROMS model

ROMS is a primitive-equation, free-surface model which uses orthogonal curvilinear coordinate sys-
tem in horizontal directions and a generalized terrain-following coordinate in vertical. For the purpose of
computational efficiency the code utilizes the natural time-scale separation of barotropic and baroclinic
processes by employing mode-splitting algorithm which solves vertically-integrated barotropic momen-
tum equations using a much smaller time step with a specially designed fast-time-averaging procedure
to prevent aliasing of processes unresolved by the longer baroclinic time step, and, at the same time,
maintain all necessary conservation properties (Shchepetkin and McWilliams, 2005). For this study we
use UCLA variant of ROMS kernel (Shchepetkin and McWilliams, 2009).
From the very beginning ROMS was intended to be a regional model, which means that a numerical
side-boundary forcing algorithm must be developed to supply external data at the open boundary of
a limited-area domain. This algorithm must be mathematically compatible with the rest of the model,
which leads to an overall design somewhat unique to the particular variant of ROMS code we use.

2.1.1 Overview of numerical boundary condition algorithms

The task of the OBC is to supply the external information needed by the limited-area model domain
at the boundaries, whilst, at the same time, allowing information generated within the model domain
to exit throught the boundaries with minimal artificial influence onto the interior solution. The problem
is mathematically non-trivial, because it is impossible to predict a priori which boundary plays the
role of inflow, and which is outflow, and, furthermore, the same segment of open boundary may be

4
both inflow and outflow for different physical processes. Also, because of mode splitting used my the
main time stepping algorithm of ROMS, the boundary conditions must be formulated in terms of model
variables, i.e., separately for the barotropic and baroclinic modes. In the approach we chose the external
data are provided along single rows at the open boundaries, i.e., there is no volumetric nudging over an
interior boundary layer; for this approach to be successful, the performance of the OBC and the quality
of the external data are critical. Lateral viscosity and diffusion are, however, allowed to decrease away
from the boundary within a so-called sponge layer, in order to smooth minor inconsistencies between
the evolving model solution and the external data. The values and profiles of viscosity and diffusivity
within the sponge layers values must be prescribed.
For the barotropic mode, the free surface and normal velocity components utilize a Flather-type
condition Flather (1976), which is based on radiation and the prescription of characteristic variables
(Riemann invariants; Blayo and Debreu, 2005). This new numerical implementation warrants a more
detailed description, which we provide below. For the tangential velocity, an upstream advection scheme
is used for both outflow and inflow. This uses a fully 2D corner algorithm involving all nearest points.
The local advection velocity is used for the phase speed. In the case of inflow, external data provided by
a boundary file are used for the upstream donor-cell values. For outflow, the external data are ignored.
There are no user-definable variable parameters (i.e., nudging time-scales) associated with the barotropic
mode boundary condition.
The baroclinic mode utilizes, for the normal velocity components of the momentum equations, an
Orlanski-type (Orlanski, 1976) condition with adaptive strong (weak) nudging for inflow (outflow).
Parameters for the nudging strength must be defined by the user. For the tangential velocity, an upstream
advection scheme, with local advection velocity as phase speed, is used. In the case of inflow, the
external data are used for the upstream donor-cell value. For outflow, the external data are ignored. There
are no tunable parameters to be defined. Tracers use an upstream advection scheme scheme similar to
that for the tangential velocity above. Again, no tunable parameters are involved.

2.1.2 Flather boundary conditions for staggered grids

The original rationale for the Flather boundary condition comes from the assumption that the domi-
nant physical processes are surface gravity waves, and the desire to allow incoming sutface gravity waves
specified by the external data enter the computational domain, while at the same time achieve unimpeded
radiation for the outgoing waves, Chapman (1985). In practice this situation occurs only in purely tidal
simulations in a small domain (hence it is physically acceptable to assume that the sea-surface field is
controlled entirely by the lateral boundaries while neglecting the effect of gravitational tidal potential).
Furthermore, ROMS as a split-explicit model averages barotropic processes over one baroclinic time
step effectively filtering out unresolved barotropic time scales, while slower barotropic processes, e.g.,
tides, behave in such a way that if the internal solution instantaneously (on the baroclinic time scale)
adjusts to externally imposed boundary conditions. Thus, the ultimate goal of Flather condition is to
achieve the best possible matching between the model field with the external data for free-surface eleva-
tion and barotropic (vertically integrated) normal velocity component in situation when straightforward
clamping of both of them would result in over-specification, while if taken alone, neither primitive vari-
able can be used to specify a well-posed boundary problem. In contrast with the purely wave and tidal
cases, posing of the problem in a regional configuration requires specification of large-scale flow across
the open boundaries. Previously in the case of rigid-lid models the barotropic component was simply
prescribed by imposing Dirichlet-type boundary condition on barotropic streamfunction, Barnier et al.
(1998); Marchesiello et al. (1998); de Miranda et al. (1999). This is possible because the free-surface

5
wave motions are excluded completely, while the remaining barotropic dynamics is governed by elliptic-
type equations, which needs all-around specified boundary conditions for mathematical well-posedness.
Physically this means specification of externally prescribed flow, which is subject to the constraint of
zero net flux for volume conservation.
In the case of long-term, limited area simulation using a free-surface model the goal is essentially
the same – to constrain the model solution to be close to the externally defined flow at the boundaries
– while the method of achieving it is different: it is no longer possible to simply impose the normal ve-
locities as a hard constraint. There are two reasons why (i) doing so would not guarantee proper integral
value of net flux, resulting in drift of horizontally averaged free surface elevations, and (ii) appearance
of barotropic non-damped wave field trapped inside the model domain (recall that the advective veloci-
ties are usually small in comparison with external gravity wave speed, so prescribed boundaries would
effectively behave as perfectly reflective rigid walls. To address both of these issues Marchesiello et al.
(2001) proposed a combination of Orlanski-type radiation boundary condition applied to each prog-
nostic boundary individually along with differential relaxation toward external data (with empirically
chosen time scales; overall this can be classified as soft constraint) and integral flux constraint to avoid
net volume gain or loss. A similar approach, except that radiation boundary algorithm is applied to the
difference between the model solution and external data was explored by Perkins et al. (1997). Although
both are viable approaches, in practice success or failure depends on careful choice of user-defined pa-
rameters, often requiring elaborate empirical tuning to avoid artifacts associated with open boundaries.
We therefore motivated to explore an alternative approach which would not require such tuning.
The Flather-type characteristic boundary conditions are derived as follows

∂t u = −g · ∂x ζ, ∂t ζ = −h · ∂x u, (2.1)

and the possibility to rewrite the above in terms of characteristic variables,


r
g  ∂t ℜ+

+ c · ∂x ℜ+ = 0 q
±
ℜ = u± ·ζ hence c= gh, (2.2)
h ∂

tℜ
−
− c · ∂x ℜ− = 0

so that ℜ+ and ℜ− propagate to the right and to the left independently from each other with a known
+
phase speed, c. This means that ℜ− = ℜ− +
0 (x + ct) = const and ℜ = ℜ0 (x − ct) = const along their
respective characteristics, x ± ct = const. The boundary conditions for ℜ+ and ℜ− , taken individually,
are self-obvious: one must specify a value at the incoming side, while there is no need for a boundary
condition at the outgoing side, hence,

Left side: Right side:

q
g
ℜ+ = u(ext) + h
· ζ (ext) ℜ+ from free radiation condition (2.3)
q
g
ℜ− from free radiation condition ℜ− = u(ext) − h
· ζ (ext)

after which ℜ+ and ℜ− are transformed back to the original variables u, ζ. In principle, Eq. (2.3)
provides a method of solution for (2.1) not only near the boundaries, but also within the domain: one
can discretise the latter using a non-staggered grid with u, ζ co-located; point-wise transform them into
(2.2); and then solve it as two independent advection problems for ℜ+ , ℜ− . This approach is known as
the Riemann solver for systems of equations such as, in the simplest cases, (2.1), and they are widely
used for gas dynamical simulations, especially for shocks.

6
 However, virtually all ocean models use horizontally-staggered grids, where cells for the free surface
and velocity components are not co-located, and time step (2.1) directly using the original variables.
This, however, cannot be applied to the boundaries, where the problem is well-posed only for ℜ+ , ℜ− ;
but, at the same time, translation from u, ζ to ℜ+ , ℜ− is obscured because of the different locations of
the variables. Nevertheless, the ocean modeling literature has numerous examples of the use of Flather
boundary conditions, especially in the context of tidal simulations. In virtually all cases the dilemma
of placing u, ζ at different locations has been addressed by applying an ad hoc interpolation scheme to
translate each variable to the location of the other; see e.g., Chapman (1985) for a review. This procedure
has several drawbacks, such as excessive reflections, and the imposition of an additional restriction on
time step size relative to the stability limit of the main scheme. Below we describe a numerical algorithm
for Flather boundary conditions which is suitable for the staggered C-grid.
Assuming that the ghost-points of the free surface elevation, ζ, are located a half-grid-interval away
from the normal velocity points, u, the purpose of the algorithm is to impose the normal velocity in such
a way that Equation 2.3 is respected. This translates into the following four steps:
• Radiate out (details are below) u, ζ independently from each other to a common location at each new
time step:
ր u∗ = ũn+1
j+1/2
n+1
ր ζ ∗ = ζ̃j+ 1/2

• Construct the outgoing characteristic variable, ℜ+ or ℜ− depending on the side, using u∗ , ζ ∗ . E.g., on
the right-side boundary it is
g ∗
r
+ ∗
ℜ =u + ·ζ
h
where g is gravity and h is depth.
• Prescribe the incoming characteristic variable from external data, e.g., on the right-side

g (ext)
r
− (ext)
 
− (ext)
ℜ = ℜ =u − ·ζ
h

• Translate back
ℜ+ + ℜ− u∗ + u(ext) g ζ ∗ − ζ (ext)
r
un+1
j+1/2 = = + ·
2 2 h 2
q  
The above differs from the original Flather condition, u = u(ext) + hg ζ ∗ − ζ (ext) . Overall, apart
from the radiate out step, the above algorithm follows the characteristic method of Blayo and Debreu
(2005). Note that ℜ+ and ℜ− never appear explicitly in the code. Boundary conditions for ζ (i.e., setting
ζ at ghost points at a half-grid outside the boundary row of u-points) are needed only by the radiation
scheme; they are not needed outside of the Flather OBC and are therefore auxiliary. Since the time step
∆t for the barotropic mode is always expected to be limited by the stability criterion for the explicit
stepping, a simple explicit radiation scheme is sufficient to compute u∗ ,

u∗ = un+1 n n
j+1/2 = (1 − c̃) uj+1/2 + c̃uj−1/2 , (2.4)

√ restriction. Above c̃ is the normalized phase speed, i.e., the

without causing any additional time step
Courant number, c̃ = ∆t · c/∆x = ∆t gh/∆x < 1.
For the free surface, ζ, the situation is more complicated because it is located half-way between
u-points, and an explicit scheme would be stable only until c̃ ≤ 1/2. One can devise a switched explicit-

7
 a) b)
Fig. 2. Geometric explanation of computing ζ ∗ and u∗ in Eqs. (2.4), (2.5), and (2.7). The open boundary is
represented by vertical dashed line going through normal velocity point uj+1/2 , so that the interior of the domain
is on the left from it. The inclined line passing through ζ ∗ and u∗ represents characteristic for the outbound
Riemann invariantℜ+ . Since the time step is expected to be limited by the stability criterion c̃ < 1, u∗ can be
always computed by an explicit scheme via (2.4). This is not the case for ζ ∗ because of its placement: Eq. (2.5)
uses one of the two pairs of points, either ζjn , ζj+1
n or ζjn , ζjn+1 , depending on which segment is crossed by the
characteristic (denoted by the red dot on the left panel). If c̃ = 1/2, then ζ ∗ =ζjn , which results in slow-growing
instability due to the non-dissipative nature of interpolation in this case. The right panel illustrates how to avoid
using single point if c̃ = 1/2 by taking value of ζ ∗ from the point where the characteristic intersects the parabolic
the segment starting at −c∗ .

implicit scheme,

n 1 1 1
    
n

 ζ j + c̃ + ζj+1 − c̃ , if c̃ < ,
2 2 2





n+1
ζ ∗ = ζ̃j+ 1/2 =

(2.5)
ζjn + ζjn+1 (2c̃ − 1)


 1
if c̃ > ,



2c̃ 2

n
which in its turn may need boundary values for the free surface at their natural location, ζj+1 . These are
computed using an auxiliary radiation boundary condition,

ζjn+1 = (1 − c̃) ζj+1

n
+ c̃ζjn . (2.6)

The sketch in Figure 2a illustrates the placement of all variables.

Unfortunately the above algorithm suffers from a numerical instability when c ≈ 1/2. This instability
is rather unusual because it occurs only within a narrow band of Courant numbers, and is associated with
the fact that, if c̃ = 1/2, the outcome of the algorithm above is ζ ∗ = ζjn without any interpolation, and
therefore dissipation. This instability can be eliminated by avoiding single-point values when c̃ = 1/2.

8
To do so, we modify the algorithm as follows:

n 1 1
    
n
 ζj 2 + c̃ + ζj+1 − c̃ , if c̃ < c̃0 ,


2









c̃0 2
 "  #
 n 1 c̃0

   
∗
ζ = ζj + c̃0 2 − − 1− (2.7)
 2 c̃ c̃ 
1 c̃ c̃0

  
0
n , if c̃ > c̃0 ,

+ζj+1 − 2−



2 c̃ c̃



c̃0 2


n+1

+ζj 1−



c̃
 √ 
where c̃0 = 1/ 2 + 2 . This scheme, shown in Figure 2b, is always stable 1 . The newly-modified
Flather OBC has been shown to return a reflected wave with an amplitude of just 1% of that of the
original signal reaching the boundary.

2.2 The ROMS configuration

The experiments are based on a regional model configuration of the Canary Basin in the subtropical
northeast Atlantic Ocean. Figure 3a shows the child grid (hereafter L1) topography, embedded within
that of the parent grid (hereafter L0). The L1 domain includes the Canary Island archipelago that lies
within the path of the equatorward-flowing Canary Current (CanC). A large portion of the Canary up-
welling region is captured, from Cape Blanc (20.80◦ N) to Cape Ghir (30.65◦ N). In general, it may be
expected that the direction of flow at the open boundaries will correspond to the local characteristics
of the subtropical gyre (e.g., Hernńdez-Guerra et al., 2005; Machı́n et al., 2006). Hence, at the northern
boundary inflow is expected to predominate and, at the south, outflow, both related to the CanC. At the
western boundary, an inflow in the north is associated with the eastward-flowing Azores Current (AzC),
while a weaker outflow to the south is due to an offshore branch of the CanC. Downstream of the Canary
Islands, high mesoscale variability in the form of island-generated cyclonic and anticyclonic eddies is
prevalent (e.g., Tejera et al., 2002; Sangrà et al., 2005). Filaments associated with the coastal upwelling
interact with these eddies (e.g., Arı́stegui et al., 1994; Sandulescu et al., 2006).
The L1 grid has 332×534 grid points, making physical dimensions of 996×1602 km2 . The horizontal
resolution is ∆x = 3 km. The domain is rotated clockwise 28.5◦, so that the closed eastern boundary is
aligned with the African coast. The southern, western and northern boundaries are open.
The parent L0 is an extensive domain that includes the entire Canary Basin (Figure 1), extending
between ∼12-45◦ N and 5-40◦ W. L0 has horizontal resolution ∆x = 7.5 km. This implies a relatively
small grid refinement coefficient between L0 and L1 of 2.5 (e.g., Debreu and Blayo, 2008). More about
the preparation of L0 may be found in Mason et al. (2009).
L0 and L1 both contain 32 vertical sigma-levels, with hc = 120 m and a surface stretching factor of
θs = 6 maintaining high vertical resolution throughout the surface layers (i.e., boundary layers) of the

1 The hole instability, ∼0.48< c <∼0.52, was first encountered by Xavier Capet in a problem with realistic
topography. It manifests itself as a blow-up at the boundary with the particular place dependent on the setting of
the time step, and no blow-up occurs c̃ < 1/2. Because c̃ depends on the topography, and topography varies from
very shallow near the coast to deep, once the time step is large enough, there is always place at the boundary
where c̃ ≈ 1/2, resulting in instability. We are able now to reproduce it within an idealized wave problem.

9
Fig. 3. (a) Child (L1) model domain (dashed black line) and topography of the Canary upwelling region, shown
embedded within the parent (L0) domain (shown in Figure 1). (b) Zoom of the topography over the white box in
(a), demonstrating the matching of L1 topography to that of L0 in the boundary region. Child (parent) isobaths
are shown in white (black). Isobaths are at 100, 500, 1250 and 2500 m.

domains. θ is a refinement parameter that determines the magnitude of stretching of the vertical grid
in either the surface (θs ) or bottom layers (θb ). We use no refinement at the bottom (θb = 0). hc is a
depth above which the vertical grid spacing of the sigma layers becomes (a) nearly uniform and, (b)
nearly independent of local depth, h, as long as h ≫ hc. The model bathymetries are taken from (a)
the ETOPO2 2-minute topography of Smith and Sandwell (1997) for L0 and, (b) the GEBCO 1-minute
topography of Hunter and Macnab (2003) for L1. To prevent aliasing when interpolating to the model
grids, the raw topographies are smoothed using a Gaussian filter of width twice the size of each raw data
point. The interpolated topographies are then further smoothed by iteratively applying a filter to reduce
the r-factor to below 0.2 (r = △h/2h; e.g., Haidvogel and Beckmann, 1999). All depths shallower than
15 m (5 m) in L0 (L1) are reset to 15 m (5 m). See Sec. 2.3.1 for more on the preparation of the L1 grid.
The two grids both employ the same climatological forcing at the surface. Wind stress is taken from
the QuikSCAT-based Scatterometer Climatology of Ocean Winds (SCOW, Risien and Chelton, 2008).
Heat fluxes and precipitation come from the Comprehensive Ocean-Atmosphere Dataset (COADS;
da Silva et al., 1994; Woodruff et al., 1998), with a mild sea surface temperature (SST; 9-km Pathfinder,
Kilpatrick et al., 2001) and sea surface salinity restoring (Barnier et al., 1995). The forcing files are cre-
ated using the tools described by Penven et al. (2008). The L1 initial and lateral boundary forcing files
are prepared following the procedures laid out in Sec. 2.3. See Mason et al. (2009) for more on the
preparation of the L0 initialisation and boundary files.
An L1 base case (BC) solution is run for 10 years, with averages of the outputs saved every 2 days.
The baroclinic time-step is 540 seconds. At runtime there are two input parameters which influence

10
boundary behavior: the sponge and the baroclinic normal velocity nudging timescale, M3. We apply a
weak sponge layer at the boundaries, with the viscosity coefficient, Ah , decaying from 10 m2 s−1 at the
boundary to zero at the inner part of the sponge. Such a low value was chosen because we are interested
in velocity anomalies at the boundary, which a stronger sponge may smooth away. The eventual use of
small sponge values was anticipated by Marchesiello et al. (2001): high-resolution absolute velocities at
the boundaries may minimise differences between parent and evolving child solution to such an extent
that the sponge is largely redundant. We set the velocity nudging coefficient M3 = 0.1, 10 days for
the respective incoming and outgoing flows. These values were chosen considering a typical advective
timescale: for our 3-km grid with maximum velocities of ∼0.4 m s−1 , we obtain 3000/0.4 = 7500
seconds ≈ 0.1 days for the inflow. Given the upwind advection scheme used by ROMS, the outflow
parameterisation is less critical than the inflow, and so we set it to 10 days.

2.3 Downscaling methodology

In this section we describe how the interpolation of the model prognostic variables from a parent
ROMS solution to child boundary and initial files, is accomplished.

2.3.1 Preparation of the child grid

Advance preparation of the child grid, beyond the description given in Sec. 2.2, is important to
facilitate the interpolation. At higher resolutions steep slopes, such as at continental shelf edges, islands,
ridges and seamounts, are better resolved than at lower resolutions. This means that depth mismatches
between child and parent grids may be significant in these regions and become a problem when they
lie along open boundaries, because volume conservation is difficult to enforce. Figure 3b illustrates
the potential for such a problem. The child isobaths (white) in the interior of the L1 domain are seen
to diverge from the parent isobaths (black). In order to prevent such a situation occurring near to the
boundaries, topography mismatches are minimised by applying (e.g., Penven et al., 2006):

hchild = αhchild + (1 − α) hiparent , (2.8)

where hchild is the child-grid bottom topography, and hiparent is the parent topography interpolated to
the child grid. α is a parameter that ranges from 0 at the lateral boundary to 1 over a distance d (typically
10% of the domain) inside the domain.
We also stress that, when defining the child landmask close to the open boundaries, it is important
to pay attention to details of the parent landmask. In Figure 4a we show hypothetical landmasking at
rho-points (ROMS uses the Arakawa-C grid; Arakawa, 1966) on the southeastern corner of the child
grid. The coastline (in black) may lead us to naively assume Cape Timiris (on the southern boundary at
∼16.4◦ W) to be separated from land by a 9-km open channel, so that we would construct our landmask
accordingly. The landmask that we in fact use in our experiments is shown in Figure 4b. The parent mask
and full coastline are included, revealing the absence of a channel and, furthermore, that the parent is
masked all along the child boundary to as far as one child grid cell to the west of the cape. We therefore
choose to put a corresponding mask into our child grid. This procedure is followed at all open boundaries
where landmasking is required.
These steps, that attempt to match child boundaries as closely as possible to those of the parent, work
to minimise the need for extrapolation during the interpolation procedure.

11
Fig. 4. Schematic diagrams showing the landmasking (darker blue) just north of Cape Timiris on the southeastern
corner of the child (L1) domain. In (a) we show a grid where the landmasking is based only on the visible coastline
(shown in black), so that an open channel is created. The L1 base case grid is shown in (b); here we include the
parent mask (darker grey) to illustrate the need for correspondence between the parent and child masks, hence
leading to the omission of the open channel.

Fig. 5. (a) Salinity section at the northern boundary taken from the BC boundary forcing file at day 800 (year 2
month 6 day 20). (b) The difference between BC salinity in (a) and the transformed C4 salinity, see Sec. 4.3.
2.3.2 Interpolation
The interpolation process is broken down into two stages. For a given time step there is, first, a hori-
zontal stage where the parent variables (both 2D and 3D) are interpolated to the horizontal coordinates

12
(longitude, latitude) of the child domain. In a second, vertical stage for each 3D variable, at each hor-
izontal coordinate in the new matrix, a two-step vertical interpolation transforms the data from parent
sigma-coordinates to z-coordinates, and then to child sigma-coordinates. At this point we have the full
set of 2D and 3D prognostic variables on the child grid.
To ensure volume conservation, a global barotropic velocity correction is applied to both the baro-
clinic and barotropic velocities at the open boundaries. The correction, Ū⊥corr , is calculated as
Z Z
Ū⊥corr = − Ū⊥ · h/ h, (2.9)
Γ Γ
where Ū⊥ is the barotropic normal velocity at the open boundaries, Γ is a line integral along the
R

open boundaries, and h is the water depth.

Finally, slices are taken from the open boundary sides of each matrix and written to the boundary file.
This procedure is applied for each successive time record within a specified time range that is available
in the parent solution.
For the present experiments, the L0 solution provides 3-day averages of the prognostic variables. The
resulting L1 boundary file contains data at the three open boundaries (north, south, west) of the L1 grid,
at the 3-day frequency of L0. As an example, Figure 5 shows a summer salinity section at the northern
boundary. The prominent features are a high salinity surface layer, typical of subtropical eastern North
Atlantic Central Water and, centred at about 900 m, a second high salinity layer which corresponds to
Mediterranean Water (MW).

2.3.3 Initialisation file extrapolation

An L1 initialisation file containing the full set (2D and 3D) of prognostic variables is built using
the first time record extracted from the parent. Here, the essential steps described above are followed.
However, in the interior where depth mismatches between parent and child topography do occur, an
extrapolation step is sometimes required before the vertical interpolation: If, at a particular point in the
respective grids, there is no parent variable at or deeper than the child depth, we search horizontally
for the nearest suitable point where the required information is available. This information is then used
for the vertical interpolation. This is a relatively simple extrapolation scheme. Recently, Auclair et al.
(2006) have demonstrated that crude extrapolation techniques can lead to erroneous currents near steep
topography. They propose an optimal scheme to avoid such errors. However, because we use extrapola-
tion only when preparing the initialisation file, our solutions will suffer any such effects only in the very
early stages of the model run.

3 Results

In this section we show a series short-term and long-term averages from our base case (BC) solution
of surface fields of SST and the vertical component of the relative vorticity, ω. These demonstrate the
quality of the agreement between our L0 and L1 solutions using the methodology described above.

3.1 Instantaneous surface fields

The efficacy of the boundary condition and the boundary preparation procedure is demonstrated in
Figure 6. Two-day average (i.e., quasi-instantaneous) surface fields in summer of the model (a) SST
and (b) ω (normalised by the Coriolis parameter, f ), are shown. The L1 solution is superimposed upon

13
Fig. 6. 2-day child (L1) averages of (a) SST and (b) surface relative vorticity (ω) in the summer of model year 2,
superimposed upon the parent (L0) field which is nearest in time (half a day). Dashed lines mark the L1 boundary.

the temporally-coincident L0 solution (not shown in its entirety). The dashed-line box indicates the
boundaries of the L1 grid.
Figure 6a shows a strong upwelling front punctuated by filaments that extend a few hundred kilome-
ters offshore. The most marked of these filaments is that off Cape Ghir at 30.65◦N (e.g., Pelegrı́ et al.,
2005). Further south, filaments are seen reaching the Canary Islands, where they facilitate exchange
of nutrients and organic matter between the African coast and the archipelago (e.g., Bécognee et al.,
2006; Sandulescu et al., 2006). The eddy-rich wake downstream of the islands peaks in late summer; an
anticyclonic eddy is seen just south of the island of Gran Canaria at 28◦ N. In the offshore region the
strongest SST gradients occur along the northern L1 boundary. These are likely related to the Azores
front, which is associated with the zonal AzC located at about 34◦ N (e.g., Le Traon and De Mey, 1994).
Along the open model boundaries there is little evidence in the SST of discontinuities between the L0
and L1 solutions.
A sterner test of the boundary behaviour is provided by considering the ω field in Figure 6b. L1 ω,
owing to the increased grid resolution, has a noticeably greater amplitude than that of L0. Agreement
between the two solutions is particularly close at the northern (inflow) boundary. Along the western
and southern boundaries there are several instances of discontinuities, with elevated ω appearing to be
trapped along the L1 boundary: for example, between 23.0 ◦ -23.5◦ N there is a region of high cyclonic ω
in L1 that is not seen in L0. It is noteworthy that the best agreement between L0 and L1 occurs at inflow
regions along the open boundary.
Nevertheless, the correspondence between the solutions is remarkable given that we use one-way
nesting where there is no feedback to the parent domain. Animations produced with time series of ω,

14
Fig. 7. Child (L1) 7-year annual averages of (a) SST and (b) vorticity (ω) for the base case experiments, superim-
posed upon the equivalent averages from the parent (L0). Dashed lines show the L1 boundary.
plotted similarly to Figure 6b, show that the quality of agreement at the boundaries is persistent.

3.2 Mean surface fields

Figure 7 shows L1 7-year annual averages at the surface of SST and normalised ω, superimposed
upon corresponding L0 mean fields. The SST field in Figure 7a is typical of eastern boundary upwelling
regions, with a band of cooler upwelled surface waters along the coast. The coldest SSTs are found in
the vicinity of Cape Ghir at 30.65◦ N. A frontal region associated with the AzC is seen at 34◦ N. The
agreement between L0 and L1 mean SST all along the boundaries is good.
The annual mean surface ω comparison in Figure 7b is less clear than the SST. Averaging tends to
minimise the signal, so that low values are found in both solutions, particularly in the offshore regions
where there is less variability. Along the southern half of the western boundary there is a thin band
of trapped cyclonic vorticity, indicative of transient normal-velocity mismatching; along the northern
half, there is a wider band of anticyclonic activity. To the southwest of the Canary Islands zonal bands
of alternating cyclonic and anticyclonic vorticity can be distinguished. These are associated with the
Canary Island eddy field (e.g., Sangrà et al., 2009). The most intense band is anticyclonic with a width
of ∼1◦ , centred at ∼25.5◦ N; its tail at ∼23.5◦ W appears to have a spuriously high magnitude. Close
to the coast, however, there is good agreement between the L0 and L1 vorticity. Trapped against the
coast is a strip of intense anticyclonic vorticity, which is bounded by a wider ribbon of cyclonic flow.
A closeup of the annual mean coastal ω distribution in the southeast of the domain at the shallow Banc
d’Arguin is seen in Figure 8: (a) shows just the L0 vorticity, (b) shows the L1 vorticity overlain onto L0.
The distribution of vorticity is essentially the same in the parent and child solutions, including along the

15
Fig. 8. 7-year annual-mean surface vorticity (ω) at the southeastern corner of the child (L1) domain. In (a) only
the parent (L0) ω is shown, while in (b) L1 ω is superimposed upon L0. The dashed black line demarks the L1
boundary. The inshore region shown is the Banc d’Arguin where the water depth is shallow (∼ 20 m).

boundary.

4 Perturbation experiments

In this section we describe four sensitivity experiments where we attempt to break the BC solution
described in Sec. 3. In the first two experiments (C1 and C2) we introduce a small modification to
the surface wind forcing. In the third experiment (C3) we perturb the volume conservation imposed
by our methodology. In the final experiment (C4) the depth of the thermocline in the boundary files
is artificially lowered by up to 20 m. Table 1 provides a summary of the experiments. A comparison
between BC and C1, C3 and C4 is given in Figure 9, which shows annual averages (26 years for BC,
7 years for C1, C3 and C4) of surface ω (normalised by f ). The aim of the sensitivity experiments is
to explore the robustness of the boundary condition under a variety of credible less-than-perfect forcing
scenarios, such as may be encountered when downscaling from an Ocean General Circulation Model
(OGCM) to ROMS.
Figure 9a shows the BC surface ω field embedded within that of L0. This is a long-term mean, 26
years, which contains less noise in comparison with the 7-year mean of Figure 7. The largest vorticity
signals are found all along the coast, positive at the coastal boundary and negative just offshore. The
Canary Island wake constitutes the other dominant signal, traces of which extend as far as the western
boundary at ∼25◦ N.

16
Fig. 9. Annual mean surface relative vorticity (ω) fields from (a) the base case experiment, (b) case 1, (c) case 3
and, (d) case 4. See Table 1 for a summary of the cases. The L1 boundary is intentionally not demarked.

4.1 Modified wind experiments (C1, C2)

Two experiments were run where the surface forcing wind stress field was modified with respect to
BC. For C1 (C2), 0.02 N m−2 was subtracted from the u-component (v-component) of the wind stress
for each month of the surface forcing climatology. Figure 9b shows the C1 surface ω. The patterns are
generally similar to BC, with largest values associated with the coastal region and the Canary Island
wake. The further reaches of the wake (∼19◦ W), however, have considerably larger ω magnitudes than

17
Table 1
Details of the sensitivity experiments (C1-C4) in relation to the base case, BC. QC is QuikCOW, the surface wind
stress product described in Sec. 2.2. ds signifies L1 boundary data downscaled from L0.

Wind stress Volume conservation Boundary thermocline

BC QC Yes (temp, salt) ds

C1 QCu − 0.02 N m−2 Yes (temp, salt) ds
C2 QCv − 0.02 N m−2 Yes (temp, salt) ds
C3 QC (v, v̄)[N,S] + 0.5 cm s−1 (temp, salt) ds
C4 QC Yes (temp, salt) ds ⇓ 20 m

BC. At the boundaries, there are conspicuous strips of anomalous positive ω, in particular at the western
boundary. The C2 surface ω field is not shown as it is not markedly different in character from C1.

4.2 Barotropic flux perturbation experiment (C3)

In this experiment the baroclinic and barotropic normal velocities (v, v̄) applied at the northern and
southern boundaries are altered by adding a constant value of 0.5 cm s−1 , effectively removing the vol-
ume conservation imposed during boundary file creation. The effect of this perturbation is clearly seen
in the mean surface ω field along the western boundary in Figure 9c, in the form of a large anticyclonic
standing eddy. The latitudinal position of this eddy, which has a radius of ∼1.5◦ , is that of the incoming
AzC. Along almost the entire western boundary, C3 surface ω is strongly positive, exceeding that of the
wind cases. In addition, a wide strip of negative ω runs parallel to the positive boundary anomaly. The
overall pattern of ω in the interior is similar to BC.

4.3 Mixed layer depth perturbation experiment (C4)

In the final perturbation experiment, temperature and salinity in the boundary forcing is lowered in
depth, by up to a maximum of 20 m in the region of the thermocline. The depth change was done by
setting hc = 80 (Sec. 2.2) during the boundary file creation process. Figure 5b illustrates the transforma-
tion by showing the difference between a BC and modified C4 salinity section at the northern boundary.
The largest differences, positive and negative, correspond to the two high salinity layers seen in Figure
5a, the surface and the MW depth layer. Below about 2500 m differences tend to zero. The intention be-
hind this modification is to introduce into the ROMS solution, a small but significant horizontal density
gradient within the region of, and normal to, the open boundaries.
The C4 annual mean surface ω is shown in Figure 9d. The effect of the thermocline perturbation is
most visible along the western boundary, in the form of a narrow intermittent strip of positive vorticity
with a similar magnitude to that in C2. In the interior, the far-field island wake vorticity is somewhat
greater than it is for BC.

18
Fig. 10. 8-year time-series of monthly mean western boundary tangential barotropic velocities from BC (blue), C3
(green) and the boundary forcing file (red). From BC and C3, the mean is taken over the first 3 rows (i-direction)
of the velocity matrix.
4.4 Discussion

Our choice of perturbations was motivated by the ongoing need to go beyond a ROMS to ROMS
downscaling, i.e., we are interested in OGCM (and also climatology as in Figure 1) to ROMS. In these
situations, all of the above perturbations may come into play: the OGCM is likely to employ a different
vertical mixing scheme from ROMS, and to be forced with a different wind stress product. For this rea-
son we chose perturbations (0.02 N m−2 and 20 m, respectively) comparable to the expected differences
that may be found between typical OGCM and child ROMS values. We note, however, that our wind
perturbations are constants, so that we are not considering wind stress curl differences (between ROMS
and OGCM) which may be significant. Our results show that ROMS is relatively robust in response to
the perturbations, although the mean solutions presented do show some degradation (Figures 9b, d).
Of the perturbations (C1-C4), our methodology is only able to “fix” C3. We show in Figure 9c
that, for C3 where the volume conservation is broken, the mean solution is significantly altered. The
noise at the western boundary, in particular the standing eddy, is persistent in the instantaneous records
(not shown). Figure 10 compares 8-year time-series from BC and C3 of the monthly mean tangential
barotropic velocity (v̄) over a three-point-wide strip along the western boundary:
Ni XNj
1 X
V̄ = v̄ij , (4.1)
Ni × Nj i=1 j=1

where Ni = 3 and Nj = 534. V̄ is also computed from the boundary forcing file with Ni = 1. The figure
shows that the effect of perturbation C3 is rapid: within a couple of months from the initialisation, C3
V̄ is strongly positive and does not approach zero thereafter, indicative of a significant anomaly at the

19
Fig. 11. Zoom over the Canary Islands of long-term annual mean surface fields of u-velocity, v-velocity and SST
from BC (left panel) and their differences from C3 (BC - C3, right panel). The variables are (a, b): u, △u; (c, d):
v, △v; (e, f): SST, △SST.

boundary (i.e., rim current). BC V̄ , however, oscillates around the zero mark, as does the boundary V̄ .
The amplitude ranges for both cases are similar, that for the boundary is slightly smaller. Plots for C1,
C2 and C4, not shown, lie roughly within the same range as BC.

20
 A further finding of these experiments answers the question: does the quality of the boundary data
matter? Modellers place boundaries far away from a region of interest in an attempt to minimise the
effects of small boundary artefacts on the interior. Figure 11 shows comparisons of 7-year annual means
from BC and C3 of surface fields in the vicinity of the Canary Islands, the centre of the L1 domain.
Figures 11a, c, e show, respectively, u- and v-velocity components and SST from BC. Differences in
these variables between BC and C3 are shown in Figures 11b, d, f. The figures show that velocity
differences frequently exceed 0.05 m s−1 for both u and v, despite the distance from the boundary.
Similarly for SST, the ∆SST field shows that BC SST offshore of the shelf edge is cooler than C3 by
more than 0.25◦ C. These differences strongly suggest that remote forcing does have a global impact
upon the model solution, such that the steps taken in our methodology (Sec. 2.3) are justified.

5 Summary

High quality ROMS solutions with minimal artifacts at the boundaries are achievable using one-way
ROMS to ROMS offline-nesting techniques, provided that suitable care is taken in the boundary file
creation process. The main components of our boundary forcing methodology are accurate interpolation
to child grid-points of parent data, matching of the parent-child topography and land-mask, and a normal
velocity correction at the child boundaries to ensure volume conservation (Sec. 2.3).
A major advantage of a boundary forcing that is strictly 2-dimension as opposed to a 3-dimensional
nudging region is the ability to provide high-frequency boundary information for the inner grid over
an extended period of time. This allows the inner grid to be forced with an outer solution that contains
meso- or even submesoscale variance in the tracer and velocity field. Such an nesting strategy provides
an unique and convenient approach to study phenomena such as frontal instabilities and submesoscale
eddies (Thomas and Ferrari, 2008; Capet et al., 2008b) in an realistic environment in the higher resolu-
tion inner grid (Capet et al., 2008a; McWilliams et al., 2009). When averaging the outer solution to, for
instance, monthly means (Dong et al. (2009) one removes the mesoscale variability that is essential for
the submesoscale variability to be generated correctly.
However, limiting the boundary forcing to a 2-dimensional field of tracers and velocities may lead
to artifacts at the boundaries of the numerical domain. The combination of method that we propose in
this paper makes it possible to do so while avoiding such spurious behavior.
A limitation of the current paper is that we still have an incomplete understanding of which (integral)
properties of the parent solutions are essential to transfer accurately to the child grid. To further our un-
derstanding we performed a number of perturbation experiments. From those experiments it appears that
the consideration of volume conservation across the open boundaries of the child grid is important. A
small correction to the velocities at the boundaries leads to significant tangential velocities at the domain
boundaries (Figure 9). Other types of perturbations lead to responses in the solution of the inner grid
that are less pronounced. However, our experience with similar experiments for different parent-child
configurations are somewhat inconclusive. What this means is that we are currently able to obtain high
quality solutions when downscaling a ROMS solution to ROMS configuration. Our ability to downscale
arbitrary data such as observations or solutions computed by different models such as OGCMs is cur-
rently not as advanced. In these latter cases, our methodology is not as capable as evidenced by our
parent (L0) solution (see the rim currents in Figure 1) which was forced with a monthly observation
based climatology.
Finally, we show that small changes in boundary forcing data may lead to global differences in the
solution throughout the numerical domain. (See Figure 11). This implies that the argument that the

21
boundaries are outside of the domain of interest and therefore not important is not valid. This may
seem obvious, but unfortunately, still sometimes used when presenting regional ocean solutions with
boundary artifacts.

Acknowledgments

Evan Mason is supported by the Spanish Government through project RODA CTM2004-06842.
ROMS development at UCLA is supported by the Office of Naval Research (currently grant N00014-
08-1-0597). We thank Florian Lemarié for his revision and comments on the manuscript.

References

Arakawa, A. (1966). Computational design for long-term numerical integration of the equations of fluid motion:
Two-dimensional incompressible flow. Part I . J. Comput. Phys., 1(1):119–143.
Arı́stegui, J., Sangrà, P., Hernńdez-Leon, S., Cantón, M., Hernńdez-Guerra, A., and Kerling, J. L. (1994). Island-
induced eddies in the Canary Islands. Deep-Sea Research, 49(10):1087–1101.
Auclair, F., Estournel, C., Marsaleix, P., and Pairaud, I. (2006). On coastal ocean embedded modeling. Geophys.
Res. Lett., 33:L14602.
Barnier, B., Marchesiello, P., de Miranda, A. P., Molines, J.-M., and Coulibaly, B. (1998). A sigma-coordinate
primitive equation model for studying the circulation in the south Atlantic. Part I: Model configuration with
error estimates . Deep Sea Res., 45:543–572.
Barnier, B., Siefridt, L., and Marchesiello, P. (1995). Thermal forcing for a global ocean circulation model using
a three-year climatology of ECMWF analyses. J. Marine Syst., 6(4):363–380.
Bécognee, P., Almeida, C., Barrera, A., Hernńdez-Guerra, A., and Hernńdez-León, S. (2006). Annual cycle of
clupeiform larvae around gran canaria island, canary islands. Fisheries Oceanography, 15(4):293–300.
Blayo, E. and Debreu, L. (2005). Revisiting open boundary conditions from the point of view of characteristic
variables. Ocean Modelling, 9(3):234–252.
Cailleau, S., Fedorenko, V., Barnier, B., Blayo, E., and Debreu, L. (2008). Comparison of different numerical
methods used to handle the open boundary of a regional ocean circulation model of the bay of biscay. Ocean
Modelling, 25(1-2):1–16.
Capet, X., Campos, E., and Paiva, A. (2008a). Submesoscale activity over the argentinian shelf. Geophys. Res.
Lett., 35(L15605). doi:10.1029/2008GL034736.
Capet, X. J., McWilliams, J. C., Molemaker, M. J., and Shchepetkin, A. F. (2008b). Mesoscale to submesoscale
transition in the California Current System. Part I: Flow structure, eddy flux, and observational tests. Journal
of Physical Oceanography, 38(1):29–43.
Chapman, D. C. (1985). Numerical treatment of cross-shelf open boundaries in a barotropic coastal ocean model.
Journal of Physical Oceanography, 15(8):1060–1075.
da Silva, A. M., Young, C. C., and Levitus, S. (1994). Atlas of Surface Marine Data 1994, volume 1-5. NOAA
Atlas NESDIS 6-10, U.S. Government Printing Office.
de Miranda, A. P., Barnier, B., and Dewar, W. K. (1999). Mode waters and subduction rates in high-resolution
south atlantic simulation. Journal Marine Research, 58:213–244.
Debreu, L. and Blayo, E. (2008). Two-way embedding algorithms: a review. Ocean Dynamics, pages 415–428.
Special Issue on Multi-Scale Modelling: Nested Grid and Unstructured Mesh Approaches.
Debreu, L., Marchesiello, P., and Penven, P. (2009). Two-way embedding algorithms for a split-explicit free-
surface ocean model. Ocean Modelling, page In preparation.
Debreu, L., Vouland, C., and Blayo, E. (2008). AGRIF: Adaptive grid refinement in Fortran. Computers and
Geosciences, 34(1):8–13.

22
Dong, C., Idica, E., and McWilliams, J. (2009). Circulation and multiple-scale variability in the Southern Califor-
nia Bight. Prog. Oceanogr., 82:168–190.
Flather, R. A. (1976). A tidal model of the north-west European continental shelf. Memoires de la Societe Royale
des Sciences de Lige, 6(10):141–164.
Haidvogel, D. B. and Beckmann, A. (1999). Numerical Ocean Circulation Modeling. Imperial College Press.
Hernńdez-Guerra, A., Fraile-Nuez, E., López-Laatzen, F., Martı́nez, A., Parrilla, G., and Vélez-Belchı́, P. (2005).
Canary current and north equatorial current from an inverse box model. J. Geophys. Res., 110:C12019.
Hunter, P. and Macnab, R. (2003). The GEBCO Digital Atlas published by the British Oceanographic Data Centre
on behalf of IOC and IHO. North Atlantic region.
Kilpatrick, K., Podesta, G. P., and Evans, R. (2001). Overview of the NOAA/NASA Advanced Very High Resolu-
tion Radiometer Pathfinder algorithm for sea surface temperature and associated matchup database. J. Geophys.
Res., 106(C5):9179–9197.
Le Traon, P. Y. and De Mey, P. (1994). The eddy field associated with the Azores Front east of the Mid-Atlantic
ridge as observed by the Geosat altimeter. J. Geophys. Res., 99(C5):9907–9924.
Machı́n, F., Hernńdez-Guerra, A., and Pelegrı́, J. L. (2006). Mass fluxes in the canary basin. Progress in Oceanog-
raphy, 70(2-4):416–447.
Madec, G. (2008). NEMO Ocean Engine (Note du Pole de Modélisation). Technical Report 27, Institut Pierre-
Simon Laplace (IPSL), France. ISSN 1288-1619.
Madec, G., Delecluse, P., Imbard, M., and Lévy, C. (1998). Opa 8, ocean general circulation model reference
manual. internal report. Technical report, LODYC/IPSL, Paris.
Marchesiello, P., B., B., and de Miranda A. P. (1998). A sigma-coordinate primitive equation model for studying
the circulation in the south atlantic. part ii: Meridional transports and seasonal variablity. Deep Sea Research,
45:543–572.
Marchesiello, P., McWilliams, J. C., and Shchepetkin, A. (2001). Open boundary conditions for long-term inte-
gration of regional oceanic models. Ocean Modelling, 3(1-2):1–20.
Mason, E., Colas, F., Molemaker, M. J., Shchepetkin, A. F., Sangrà, P., and McWilliams, J. C. (2009). Seasonal
variability of the Canary Current System. In preparation.
McWilliams, J. C., Colas, F., and Molemaker, M. J. (2009). Cold filamentary intensification and oceanic surface
convergence lines. Geophys. Res. Lett., 36(L18602). doi:10.1029/2009GL039402.
Orlanski, I. (1976). A simple boundary condition for unbounded hyperbolic flows. J. Comput. Phys., 21:251–269.
Pelegrı́, J. L., Marrero-Dı́az, A., Ratsimandresy, A., Antoranz, A., Cisneros-Aguirre, J., Gordo, C., Grisolı́a,
D., Hernńdez-Guerra, A., Lı́z, I., Martı́nez, A., Parrilla, G., Pérez-Rodrı́guez, P., Rodrı́guez-Santana, A., and
Sangrà, P. (2005). Hydrographic cruises off northwest africa: The canary current and the cape ghir region.
Journal of Marine Systems, 54(1-4):39–63.
Penven, P., Debreu, L., Marchesiello, P., and McWilliams, J. C. (2006). Evaluation and application of the ROMS
1-way embedding procedure to the central California upwelling system. Ocean Modelling, 12(1-2):157–187.
Penven, P., Marchesiello, P., Debreu, L., and Lefevre, J. (2008). Software tools for pre- and post-processing of
oceanic regional simulations. Environmental and Modelling Software, 23:660–662.
Perkins, A. L., Smedstad, L. F., Blake, D. W., Heburn, G. W., and Wallcraft, A. J. (1997). A new nested boundary
condition for a primitive equation model. J. Geophys. Res., 102:3483–3500.
Risien, C. M. and Chelton, D. B. (2008). A global climatology of surface wind and wind stress fields from eight
years of QuikSCAT scatterometer data. J. Phys. Oceanogr. Early Online Releases.
Sandulescu, M., Hernńdez-Garcia, E., López, C., and Feudel, U. (2006). Kinematic studies of transport across an
island wake, with application to the canary islands. Tellus, 58(5):605–615.
Sangrà, P., Pascual, A., Rodrı́guez-Santana, A., Machı́n, F., Mason, E., McWilliams, J. C., Pelegrı́, J. L., Dong,
C., Rubio, A., Arı́stegui, J., Marrero-Dı́az, A., Hernńdez-Guerra, A., Martı́nez-Marrero, A., and Auladell, M.
(2009). The Canary Eddy Corridor: a major pathway for long-lived eddies in the subtropical North Atlantic.
Deep Sea Res. Accepted for publication.
Sangrà, P., Pelegrı́, J. L., Hernńdez-Guerra, A., Arregui, I., Martı́n, J. M., Marrero-Dı́az, A., Martı́nez, A., Ratsi-

23
 mandresy, A. W., and Rodrı́guez-Santana, A. (2005). Life history of an anti-cyclonic eddy. J. Geophys. Res.,
110:C03021.
Shchepetkin, A. F. and McWilliams, J. C. (2005). The regional oceanic modeling system (ROMS): a split-explicit,
free-surface, topography-following-coordinate oceanic model . Ocean Modelling, 9(4):347–404.
Shchepetkin, A. F. and McWilliams, J. C. (2009). Correction and commentary for “Ocean Forecasting in terrain-
following coordinates: formulation and skill assessment of the Regional Ocean Modeling System” by Haidvo-
gel et al, J. Comp. Phys. 227, pp. 3595-3624. J. Comput. Phys. in press.
Smith, W. H. F. and Sandwell, D. T. (1997). Global sea floor topography from satellite altimetry and ship depth
soundings. Science, 277(5334):1956 – 1962.
Tejera, A., Garcı́a-Weil, L., Heywood, K., and Cantón-Garbı́n, M. (2002). Observations of oceanic mesoscale
features and variability in the Canary Islands area from ERS-1 altimeter data, satellite infrared imagery and
hydrographic measurements . Intern. J. Remote Sensing, 23(22):4897–4916.
Thomas, L. and Ferrari, R. (2008). Friction, frontogenesis, frontal instabilities and the stratification of the ocean
surface mixed layer. J. Phys. Oceanogr., 38:2501–2518.
Woodruff, S. D., Diaz, H. F., Elms, J. D., and Worley, S. J. (1998). COADS Release 2 Data and metadata
enhancements for improvements of marine surface flux fields. Phys. and Chemist. of the Earth, 23(5):517–
526.

24