## 1. Introduction

Describing the global ocean circulation is still a problem despite more than 100 years of systematic observation and research in oceanography. The large-scale transport and the distribution of water masses can be derived from hydrographic data (temperature and salinity) using the geostrophic equations. The hydrographic datasets are spatially and temporally inhomogeneous and geostrophy can be used only to describe the vertical shear of the velocity field. To get absolute velocities it is often assumed that the flow velocity vanishes at some reference depth. The error made by this assumption is small near the surface where the velocities are larger than in the deep ocean where they have an order of 0.01 m s^{−1}. As a consequence a small shift of the reference level can produce errors of the latter order of magnitude. In local regions where the ocean is shallower than the reference depth the bottom depth has to be chosen as reference, for example.

Another possibility to calculate absolute velocities is the determination of reference velocities at the ocean surface. The velocities in the deep ocean can be calculated by vertical integration from the surface. As a second independent dataset, measurements from satellite altimetry can be used to determine the reference velocities.

Numerical models and assimilation techniques are combined for this purpose. The model states computed must fulfill the model equations and have to take into account the data. Strictly speaking, data assimilation is used for prognostic models. Diagnostic models where the model parameters are adapted to the data are called inverse models. An optimization approach is applied to fit the parameters. A so-called cost function is minimized subject to the model equality constraints. Basic concepts are described by Thacker (1988), Olbers (1989), and Wunsch (1996). Computations with this method were made, for example, by Schröter (1989). Schröter et al. (1993) and Vogeler and Schröter (1995, 1999) applied the method to regional eddy-resolving ocean models. Schlitzer (1993, 1995) developed a box model for the Atlantic, which is extended to the global ocean (de las Heras and Schlitzer 1999). Nechaev and Yaremchuk (1995) and Yaremchuk et al. (1998) presented an adjoint model for analyzing hydrographic sections. A global model to investigate the barotropic ocean circulation originated from Seiß (1996). Sloyan (1997) presented a box model for the Southern Ocean.

The models mentioned above are box models and/or use the finite-difference method (FDM). In the present model the finite-element method (FEM) is used. Originally the FEM was developed in the 1950s and 1960s to solve problems occuring in elastomechanics. Later the FEM was expanded for applications to many other physical processes that are described by partial differential equations. The theory of the FEM is described in detail, for example, by Zienkiewicz (1977) and Johnson (1990). There are successful applications in tidal models since the beginning of the 1970s. Examples are the works of Connor and Wang (1973), Walters and Werner (1989), and Le Provost and Vincent (1991). But there are only a few works that use the FEM in (global) circulation models. A first work was presented by Fix (1975) with a quasigeostrophic model. Haidvogel et al. (1980) compared the FEM with the FDM. For problems in fluid dynamics and particularly for convection–and advection–diffusion problems the FEM became popular in the 1980s. At that time different upwind methods for stabilization were developed. A basic work is given by Brooks and Hughes (1982). In recent years Myers and Weaver (1995) used the FEM in a diagnostic barotropic ocean model. Iskandarani et al. (1995) presented a spectral finite-element model for the shallow water equations. With a similar model Haidvogel et al. (1997) described the global circulation and Wunsch et al. (1997) investigated the dynamics of long-period tides. This model was also used by Taylor et al. (1997), Levin et al. (1997), and Curchitser et al. (1998, 1999). Further developments of finite-element models for the shallow water equations were presented by Behrens (1996, 1998), Le Roux et al. (1998), and Giraldo (1999, 2000a,b). Zhu et al. (1994) worked with data assimilation in a two-dimensional finite-element model for the shallow water equations.

Another finite-element circulation model is QUODDY (Ip and Lynch 1995), which is used for regional comprehensive coastal modeling (Lynch et al. 1996). QUODDY is based on the conventional three-dimensional shallow water equations. It is tide resolving to investigate shelf circulation. The meshes are created by prisms and a general terrain-following vertical coordinate is used. Recently, an inverse model version of QUODDY has been designed that uses an assimilation scheme such as the adjoint method (Lynch and Hannah 2001). Altogether applications of the FEM in oceanography—especially in three-dimensional inverse models—are still rare.

The major goal of this work is the development of an internally consistent three-dimensional adjoint model using finite elements to investigate the large-scale ocean circulation. The model solves the stationary advection–diffusion problem. A mesh is generated with tetrahedra. These elements are particularly suitable to represent the complex three-dimensional geometry given by the bottom topography. The mesh is totally unstructured so that any local refinement is possible. This is one of the great advantages of the FEM compared to the FDM. Another advantage is that the adjoint code can be created relatively easily.

The model describes a flow field in geostrophic balance and conserves mass and heat. The control variables of the inverse model are horizontal reference velocities at the ocean surface and sources (heat and freshwater fluxes) in the advection–diffusion equation. Assimilated data are among others climatologic hydrographic data and sea surface height (SSH) data from satellite altimetry.

The second major goal of this work is a first application of the model to the South Atlantic to study the large-scale circulation, the formation and spreading rates of water masses and the exchange of heat and salt. The investigations focus on the Weddell Sea.

The organization of this paper is as follows: the governing model equations and the strategy of the adjoint model are described in section 2. In section 3 the use of the FEM is presented. The application of the model and its results are discussed in sections 4 and 5. Finally, a summary and conclusions are given in section 6.

## 2. The inverse model

In this section the strategy of the model is presented. The model parameters are temperature *T,* salinity *S,* density *ρ,* velocities **u** = (**u**_{h}, *w*) = (*u,* *υ,* *w*), and sources *F*^{T} and *F*^{S} of the advection–diffusion equation for temperature and salinity. The density *ρ* is calculated from the temperature field and the salinity field with the equation of state. Using the density and a prescribed reference velocity the horizontal flow field **u**_{h} is computed and from that the vertical velocities *w* are determined using volume conservation. Solving the advection–diffusion equations with this flow field yields new temperatures and salinities. Data are assimilated with the adjoint method in the model. A cost function *J* fits the model to the data and measures the quality of the model solution. The model parameters are divided into independent parameters—the control variables of the cost function—and into dependent parameters. We solve for a set of parameters that minimize the value of the cost function. This set represents a certain circulation pattern (flow field) and the corresponding temperature and salinity fields that are both in geostrophic and advective–diffusive balance.

### a. Model equations

*ρ*depends on the temperature

*T*and the salinity

*S*of the seawater and is given by a highly nonlinear equation of state:

*E*

^{ρ}

*ρ*

*R*

*T,*

*S,*

*p*

*p*= 0.1Δ

*z*is the pressure in bars. A formula (so-called UNESCO formula 1981) for

*R*is presented by Gill (1982).

*u*and

*υ*are the horizontal velocities,

*g*is the acceleration of gravity and

*f*= 2Ω sin

*ϕ*is the Coriolis parameter. The geostrophic approximation represents the balance between the Coriolis force and the pressure gradient. Together with the hydrostatic balance ∂

*p*/∂

*z*= −

*gρ*the thermal wind relation

*u*

_{ref}and

*υ*

_{ref}at the surface results in

*z*≥ −

*h*

_{e})

**u**

_{h}(

*z*) =

**u**

_{ref}= (

*u*

_{ref},

*υ*

_{ref}) is valid. The vertical velocities are computed by vertical integration of the equation of continuity. The integration begins at the bottom with the kinematic boundary condition

*w*(−

*H*) =

**u**

_{h}· ∇

*H*and ends at the depth −

*h*

_{e}of the Ekman layer:

*K*

_{h}and

*K*

_{υ}are the horizontal and the vertical diffusion coefficients. The sources

*F*

^{T}and

*F*

^{S}are implemented to parameterize subscale processes, which are not resolved by the model because of its properties (e.g., local grid spacing).

Using this set of model equations to determine the deep circulation for a steady state is a common choice for inverse models (see Wunsch 1996).

*u*

_{ref}and

*υ*

_{ref}and the sources

*F*

^{T}and

*F*

^{S}:

*T,*the salinity

*S,*the density

*ρ,*and the vertical velocities

*w*constitute the set of dependent parameters

### b. Cost function

The value of the cost function *J* calculated from all independent and dependent parameters 𝗽 and 𝗾 represents the quality of the actual model state with respect to the data. The different terms contributing to the cost function penalize unwanted properties of the model solution. The goal is to minimize the value of the cost function by driving the model solution to the preferred model state. The model solution depends highly on the characteristics of the cost function.

*J*

_{uυ}] consider the gradient of the sea surface height

*ζ*;

*ζ*

_{dat}are measurements of SSH from satellite altimetry. Deviations between the SSH gradient and the corresponding reference velocities are penalized. The terms (11) and (12) lumped into

*J*

_{F}fit the sources

*F*

^{T}and

*F*

^{S}to

*F*

^{T}

_{dat}

*F*

^{S}

_{dat}

*J*

_{w}(13) penalizes deviations between the vertical velocities

*w*

_{i}at

*z*= −

*h*

_{e}and the curl of the wind field

*τ*

^{w}over

*f*. The last two terms

*J*

_{T}and

*J*

_{S}[(14) and (15)] ensure that the temperatures

*T*and salinities

*S*are close to the data

*T*

_{dat}and

*S*

_{dat}. The weighting matrices

*W*

_{ij}are diagonal. The diagonal elements are the inverse of the variances

*σ*

^{2}

_{kk}

### c. The adjoint model

_{new}and 𝗾

_{new}), are computed with a smaller value of the cost function

*J*than for the actual state with parameters 𝗽 and 𝗾. Finally the value of the cost functions achieves a minimum. A descent direction must be found in the phase space spanned by the control variables 𝗽 so that the value of the cost function decreases. For that, the gradient of the cost function ∇

*J*is provided to an optimization algorithm. In this work a variable-storage quasi-Newton algorithm from Gilbert and Lemaréchal (1989) is used. Minimizing the cost function is an optimization problem with equality constraints that are the model equations [(1)–(6)]. The model states have to obey the model equations exactly. A well-known method to solve such problems is the method of Lagrange multipliers (see Le Dimet and Talagrand 1986; Thacker 1988; Schröter 1989). The benefit is that the so-called Lagrangian function

*L*has a stationary point in the minimum of the cost function

*J.*The Lagrangian function

*L*is a function of all model parameters 𝗽 and 𝗾 and the unknown Lagrange multipliers

*λ*= [… ,

*λ*

_{j}, …]

^{T}. It reads

*E*

_{j}= 0 denote the model equations (1)–(6). Here

*n*

_{eq}is the number of model equations. The gradient of the Lagrangian function becomes zero at the minimum of the cost function (∇

*L*= 0). Then all partial derivatives of

*L*are vanishing:

*λ*

_{j}yield the model equations. The derivatives with respect to the the dependent parameters 𝗾 and the control variables 𝗽 (18) and (19) are called

*adjoint equations.*The minimum of the cost function has to fulfill all three equations. The optimization algorithm is an iterative process. First the Lagrange multipliers are computed from Eq. (18), which can be written in an

*n*

_{eq}-dimensional system of linear equations

*λ*

_{j}are substituted in Eq. (19). One obtains then the partial derivatives ∂

*L*/∂

*p*

_{i}, which are part of the input to the optimization procedure of Gilbert and Lemaréchal (1989).

The entire model algorithm contains the following steps.

Initialization of the model including the control variables 𝗽.

Forward model: Computation of the dependent parameters 𝗾 by solving the model equations (1)–(6).

Compare model parameters with data. Stop if optimum is reached.

Adjoint model:

(a) Compute the Lagrange multipliers from the adjoint Eq. (18).

(b) Compute the gradient ∇

_{p}*L*from the adjoint Eq. (19).(c) Provide the gradient ∇

_{p}*L*to the optimization algorithm and receive new improved independent parameters 𝗽_{new}.Back to step 2.

Following from the model algorithm the density of the *k*th iteration is calculated from the temperature and the salinity of the previous iteration [*ρ*^{k} = *R*(*T*^{k−1}, *S*^{k−1}, *p*)]. This fact results in a separation of the *n*_{eq}-dimensional system of linear equations (20) into six subsystems with dimensions not larger than the number of nodes of the model grid.

The initialization of the model algorithm includes the computation of an initial horizontal flow field **u**_{h} from the geostrophic equations using an initial density field *ρ* = *R*(*T*_{dat}, *S*_{dat}, *p*). The reference depth of the flow field is 3000 m. In regions shallower than 3000 m the local depth of the ocean is used as reference. In the Ekman layer (*z* ≥ −*h*_{e}) Ekman velocities are added to **u**_{h}.

## 3. The finite-element model

Physical processes and mathematical models of engineering are often described by partial differential equations. Only very simple cases can be treated analytically. In general it is necessary to apply numerical methods for finding approximate solutions. A common one is the finite element method, which was developed first in the 1950s and 1960s to treat problems in elastomechanics. Later the method was expanded to a lot of other physical problems.

In any numerical model the given continuous problem with infinitely many degrees of freedom is transformed to a discrete problem with a finite number of degrees of freedom. In the classical finite-difference method, the derivatives of a differential equation are replaced by difference quotients to obtain the discrete equations. In finite elements the discretization originates from a variational principle and uses a low order expansion of the unknown fields. The goal is that the solution of the discrete system of equations yields a sufficiently good approximation of the solution of the original continuous problem. Using the FEM to create a model to solve a given partial differential equation contains four basic steps (Johnson 1990): 1) variational formulation of the given problem, 2) discretization of the problem with the FEM, 3) solving the discrete problem, and 4) implementation of the model on a computer.

The theory and the basic concepts of the FEM are described in detail in several books like those of Zienkiewicz (1977) and Johnson (1990). Here we discuss our use of finite elements and we describe the variational formulation and the discretization (section 3a) including stabilization (section 3b) for the stationary advection–diffusion problem (6).

A great advantage of finite-element models is that unstructured meshes can be used. The FEM is very useful to treat problems on complex geometries, which is much more difficult to deal with using finite differences. Unstructured grids are very suitable to describe the bottom topography and the coastlines of the oceanic basins. Also, any local mesh refinement is possible without increasing the number of nodes too much. Three-dimensional meshes can be generated by tetrahedra, prisms, parallelepipeds, or curvilinear elements.

The variational principle includes integration by parts that result in a reduction of the order of the highest derivative to only half of the order as the derivatives of the original partial differential equation. Thus, the elements and their transition conditions can be chosen rather simply.

The parameter fields are defined continuously on the whole domain in finite elements as a superposition of the so-called basis functions. The basis functions are polynomial functions of low order. They are defined, respectively, on a single element in a way that they are equal to unity at one node of that element and that they are vanishing at all other nodes (see below).

An additional item of interest for inverse models is that the adjoint of the discrete equations given by the weak formulation of the variational principle is easily derived. In this work tetrahedra (Fig. 1) are the basic elements used in a totally three-dimensional unstructured mesh. They are much more flexible in 3D mesh generation than prisms, for example, and they have easy basis functions and transition conditions. Of course, the ratio between the horizontal and the vertical diameter of the tetrahedra reaches an order of magnitude up to *O*(100)–*O*(1000). However, the ratio between the scales of the horizontal and the vertical terms in the advection–diffusion Eq. (6) has the same order of magnitude. A vertically stratified mesh using a 2d unstructured mesh that is copied to each depth-level or a terrain-following vertical coordinate mesh such as in QUODDY (Lynch et al. 1996) is not quite as flexible as a 3D unstructured mesh for local refinement. In addition, a terrain-following vertical coordinate may have problems with calculating the pressure gradient at steep changes of the topography across the shelf breaks.

### a. The standard Galerkin method for advection–diffusion problems

*C*=

*C*(

**x**,

*t*) like temperature

*T*or salinity

*S*is written

**u**is the velocity field and

*K*is the diffusion coefficient. Here

*f*is a source term on the right-hand side and

*t*is the time. In the case of incompressible fluids (∇ ·

**u**= 0) and steady states, this equation simplifies to

*K*=

*K*

_{h}=

*K*

_{υ}. In the application of the model (see section 4) the coefficients are different and values of

*K*

_{h}= 5 × 10

^{4}m

^{2}s

^{−1}and

*K*

_{υ}= 5 × 10

^{−2}m

^{2}s

^{−1}, respectively, are used.

_{1}∪ Γ

_{2}of Ω are supposed as

*ϕ*and integration over the domain Ω

**u**

*C,*

*ϕ*

*K*

*C,*

*ϕ*

*f*

*ϕ*

_{1}[

*ϕ*(

**x**) = 0 on

**x**

*ε*Γ

_{1}, homogeneous boundary conditions], the weak formulation reads

*C*(

**x**)

*ε*

*H*

^{1}

_{0}

**u**

*C,*

*ϕ*

*K*

*C,*

*ϕ*

*f*

*ϕ*

*q,*

*ϕ*

_{Γ2}

*ϕ*(

**x**)

*ε*

*H*

^{1}

_{0}

*H*

^{1}

_{0}

In the next step the weak formulation (27) will be approximated by finite dimensional subspaces of *H*^{1}_{0}

*C*

_{h}

*εV*

_{h}⊂

*H*

^{1}

_{0}

*V*

_{h}

*ϕ ∈ H*

^{1}

_{0}

*ϕ*

_{h|τ}

*∈ P*

_{1}

*τ*

*τ ∈ T*

_{h}

**u**

*C*

_{h}

*ϕ*

_{h}

*K*

*C*

_{h}

*ϕ*

_{h}

*f*

*ϕ*

_{h}

*q,*

*ϕ*

_{h}

_{Γ2}

*T*

_{h}is the partition of Ω in subdomains (finite elements)

*τ*and the elements are tetrahedra

*P*

_{1}denotes the space of linear functions defined on the elements

*τ.*The finite series by linear functions is sufficient because only first derivatives occur in the weak formulation. The approximated

*C*

_{h}are

*C*

_{b}

*ε*

*H*

^{1}

_{0}

*N*is the number of grid points given by the elements

*τ.*The functions

*ϕ*

_{i}form a finite basis

*ϕ*= {

*ϕ*

_{1},

*ϕ*

_{2}, … ,

*ϕ*

_{N}}, which span the subspace

*V*

_{h}. They are called basis functions and satisfy the homogeneous boundary conditions. Now, any

*ϕ*

_{h}

*εV*

_{h}can be written as a linear combination of the

*ϕ*

_{i}:

*τ*can be defined by

*ϕ*

_{i}

**x**

*a*

_{i}

*x*

*b*

_{i}

*y*

*c*

_{i}

*z*

*d*

_{i}

*ϕ*

_{i}

**x**

_{j}

*δ*

_{ij}

**x**

*ετ*(see appendix). The four corner points of the tetrahedra

*τ*are denoted by the index

*i,*

*j*= 1, … , 4. The so-called global basis functions

*ϕ*

_{h}are expressed by the so-called local basis functions

*ϕ*

_{i}.

*ϕ*

_{h}by the local basis functions

*ϕ*

_{i}, Eq. (29) can be written as a system of

*N*(index

*i*) ordinary differential equations with

*N*unknown coefficients

*C*

_{j}(index

*j*):

**u**. It is well known that a main problem of the FEM is the proper choice of discrete spaces to guarantee the property of conservation for the discrete problem. In Eq. (33) the tracer

*C*is approximated by piecewise linear basis functions. Using piecewise constant velocities

**u**= (

*u,*

*υ,*

*w*) in conjunction with piecewise linear tracers

*C*is a well-known approximation (Thomasset 1981; Brezzi et al. 1992). First, because this set of velocities is nondivergent within each element and it simplifies the stabilization scheme (see next section). Second, it was demonstrated (Brezzi et al. 1992) that different stabilization methods result in the well-known “streamline diffusion” stabilization method, which is described below. In addition, this low-order approximation minimizes the number of control variables 𝗽 (7) and dependent model parameters 𝗾 (8), which is an important task to limit the requirements of memory and cpu time of the adjoint model. The piecewise constant velocities used in the advection–diffusion Eq. (6) are derived by projection from the piecewise linear velocities (Thomasset 1981) given by the thermal wind equations (4) and the equation of continuity (5). The diffusion coefficient

*K*is also chosen piecewise constant on the elements.

**c**

**f**

**f**is denoted the load vector, and

**c**= (

*C*

_{1}, … ,

*C*

_{j}, … ,

*C*

_{N})

^{T}is the unknown solution vector.

### b. Stabilization

The FEM used to be unpopular for treating advection–diffusion (convection–diffusion) problems. The reason was that numerical instabilities may occur for advection-dominated flows. This is also a well-known problem when using finite differences. As a consequence various stabilization techniques (upwind schemes) were developed. In this subsection we will have a brief look on some common techniques, which turn out to be equivalent to each other for the case considered here. For more details we refer to the cited literature.

_{τ}describes the relation between the advective and the diffusive term of the advection–diffusion equation (22) in the local element

*τ*:

*h*

_{τ,h}is the horizontal and

*h*

_{τ,υ}the vertical diameter of the element

*τ.*The flow is advective/convective dominated for large Péclet numbers.

*ϕ̃*

*ϕ*+

*ψ.*The

*ϕ*are continuous but the

*ψ*can be discontinuous over the boundaries of the elements. They assume the part of the upwind corrections. Following the weak formulation of the SGM the discrete equations with the upwind corrections read:

*a,*

*b*)

_{τ}= ∫

_{τ}

*ab*

*d*Ω. The summations have to be carried out over all elements

*τ.*These terms represent a kind of artificial diffusion. For the functions

*ψ*

_{h}Brooks and Hughes (1982) introduced

*ξ*depending on the local Péclet number Pe

_{τ}(Hughes et al. 1986; Franca et al. 1992):

*ξ*(Pe

_{τ}) show that there is no or only a small effect of the stabilization scheme for small Péclet numbers (Pe

_{τ}≲ 1).

*ϕ*

_{h}of the function

*ψ*

_{h}[Eq. (37)] is substituted by the residuum of the advection–diffusion equation applied to the test function

*ϕ*

_{h}. The function

*ψ*

_{h}is written then

*ϕ*

_{h}is piecewise linear. In this special case

*ψ*

_{h}of Eq. (38) is equivalent to the

*ψ*

_{h}of Eq. (37) if the parameter

*Ã*

_{τ}=

*A*

_{τ}is used.

Brezzi et al. (1992) presented another rather new stabilization technique, which uses the so-called bubble functions. However, they showed as well that in the case of piecewise linear tracers and piecewise constant velocities this method is also equivalent to SUPG and GLS.

## 4. An application to the South Atlantic

The model presented in sections 2 and 3 is first applied to the South Atlantic. The model domain extends from 70°W to 20°E and from 30° to 74°S. Our interest is focused on the Weddell Sea, which is embedded in the model domain.

### a. The mesh

The mesh is unstructured and contains 13 431 nodes and 65 623 tetrahedra. The grid spacing is ≤2° horizontally and ≤750 m in the vertical direction. There are prescribed levels in 50, 100, 200, and 400 m. Normal to the bottom the grid spacing is smaller than 500 m. The mesh is generated with the Intersection-Based Grid (IBG) generator written by I. Schmelzer (1995) using ETOPO5 data (National Geophysical Data Center 1988) for topography. The IBG generator is a grid based generator and works with a combined Octree/Delaunay method. Figure 2 shows the bottom topography and the open boundaries of the model. The mesh yields a sparse stiffness matrix 𝗦 with rank 13 431 and *nz* = 182 103 nonzero elements.

### b. Data and boundary conditions

The climatologic data for temperature *T*_{dat} and salinity *S*_{dat} used in the model are data provided by the World Ocean Circulation Experiment (WOCE) Hydrographic Programme (WHP) Special Analysis Center (Gouretski and Jancke 1998). Wind stress is taken from Hellerman and Rosenstein (1983). The annual mean is used for the stationary model.

The dataset for the sea surface height *ζ* (SSH) was produced from measurements of satellite altimetry. The mean sea surface—originating from the Collecte Localisation Satellites/Service Hydrographique et Océanographic de la Marine (Hernandez and Schaeffer 2000)—is referenced to the EGM96 geoid (Lemoine et al. 1998). To reduce small-scale noise of the geoid model a filtering technique according to Wahr et al. (1998) is used. The averaging kernel of the filter has a radius of 4.5°. Filtered SSH data are presented in Fig. 3.

The advection–diffusion equation (6) is solved for the temperature (*C* = *T*) and the salinity (*C* = *S*) on the model domain. The boundary conditions for *T* and *S* are ∂*T*/∂*n* = ∂*S*/∂*n* = 0 at the bottom. Along the open boundaries *T* = *T*_{dat} and *S* = *S*_{dat} is prescribed. In the regions with outflow between 0°E and South Africa along the northern boundary at 30°S (Benguela Current) and between 40° and 60°S along the eastern boundary at 20°E [Antarctic Circumpolar Current (ACC)] ∂*T*/∂*n* = 0 and ∂*S*/∂*n* = 0 are set as outflow conditions. The horizontal and the vertical diffusion coefficients are prescribed with *K*_{h} = 5 × 10^{4} m^{2} s^{−1} and *K*_{υ} = 5 × 10^{−2} m^{2} s^{−1}. The total volume transports across the open boundaries are prescribed too. Through the boundary along 70°W 130 Sv (1 Sv ≡ 10^{6} m^{3} s^{−1}) flow into the South Atlantic. This value is taken from Reid (1989). The same amount leaves the model domain at 20°E. The total transports across the northern and southern open boundaries vanish. Constraining the transport of the ACC to investigate the circulation in the Weddell Sea is a common practice. For example, Beckmann et al. (1999) also prescribed the ACC transport through the Drake Passage with a value of 130 Sv.

### c. Solving the systems of linear equations

The model computations were performed on a Cray J90 computer. The system of linear equations (34) arising from the finite-element discretization of the advection–diffusion equation and the calculation of the Lagrange multipliers are solved using the GMRES solver (see section 2c) of the SITRSOL routine. The SITRSOL routine is part of the scientific library for Cray systems. It provides iterative solvers for real general sparse systems using preconditioned conjugate-gradient-like methods. ILU(*m*)-factorization is used for preconditioning. The (integer) parameter *m* denotes the level of fill-in for the factorization. The amount of fill-in in the lower and the upper triangular matrices 𝗟 and 𝗨 is given by 2*mnz.* To improve the performance of the GMRES solver and to reduce the requirements of memory for the fill-in elements the Metis package (Karypis and Kumar 1998) is used for reordering the coefficient matrix 𝗦. The high quality of the reordering produced by the Metis package for our problem was demonstrated earlier by Dobrindt and Frickenhaus (2000).

## 5. Results

We now discuss the results of the model application to the South Atlantic. The model solution found is distinguished by reference velocities *u*_{ref}, *υ*_{ref} close to those calculated from the gradient of the SSH data and by temperatures *T* and salinities *S* close to the data *T*_{dat} and *S*_{dat}. The differences between model values and hydrographic data are in the range of the errors of the hydrographic data. The deviations of the reference velocities from their equivalent are very small. First we have a look on the spreading of different water masses. Afterward we discuss the circulation pattern of the model solution, the volume transports, and the exchange of heat and salt.

### a. Water masses

The classification of water masses depending on the neutral density *γ* is the same as in Sloyan (1997): surface water (SW) is defined by *γ* < 26.5. For the Antarctic Intermediate Water (AAIW) 26.5 ≤ *γ* < 27.6 is valid. The deep water (27.6 ≤ *γ* < 28.2) is composed of the North Atlantic Deep Water (NADW) and the Circumpolar Deep Water (CDW). The CDW flows as a part of the Antarctic Circumpolar Current through the Drake Passage into the South Atlantic. The Antarctic Bottom Water (AABW) has a neutral density of *γ* ≥ 28.2. The upper limit for the AABW is in agreement with Orsi et al. (1999). The AABW, which is formed in the basin of the Weddell Sea [Weddell Sea Deep Water (WSDW) and Weddell Sea Bottom Water (WSBW)] is denser (*γ* ≥ 28.27). But the WSDW and WSBW is modified by the lower Circumpolar Deep Water (LCDW) when it is transported to the north by the deep western boundary currents. Therefore *γ* ≥ 28.2 is a reasonable value for the AABW in the South Atlantic.

Figures 4–7 show the spreading of surface water, intermediate water, deep water, and bottom water in the South Atlantic along the Greenwich meridian, 20°W, 40°W, and 30°S. The solid lines represent the model solution while the distribution computed from the hydrographic data is shown in dashed lines. The agreement of the model and the data is very good. Surface water occurs only in the upper 400 m north of circa 40°S. The *γ* = 27.6 isosurface extends up to ∼1300 m at 30°S and reaches the ocean surface between 56° and 60°S. The lower boundary of the deep water ranges from the ocean bottom at 30°S up to 1000-m depth at the southern boundary of the model domain. The AABW is found in the Weddell Sea below 1000 m and spreads through the deep channels of the Argentine Basin and the Cape Basin at 30°S.

### b. The large-scale circulation

Figure 8 presents the reference velocity field (*u*_{ref}, *υ*_{ref}). The surface circulation is smooth because it reflects the structure of the SSH data. The upper-level flow field is similar to that circulation pattern. Comparisons of the circulation of the South Atlantic with that from literature (e.g., Reid 1989; Peterson and Stramma 1991) show much agreement. The ACC flows through the Drake Passage in the Scotia Sea and turns to the northeast. When leaving the Scotia Sea the ACC shifts back to flow in an eastern direction. The ACC flows out of the model domain at the eastern boundary between 40° and 56°S. The subtropical gyre enters into the model domain across the northern boundary east of the Brazilian coast and flows as the South Atlantic Current (SAC) parallel to the ACC. In the region around 6°E the SAC turns to the north and leaves the model domain.

In section 4 we have already mentioned that our major attention is directed to the Weddell Sea where the circulation forms the Weddell gyre. The upper-level Weddell gyre is shown in Figs. 8 and 9. The southward flowing part into the Weddell Sea occurs east of 6°E. Then the current turns to the west and flows along the coast of Antarctica. The center of the gyre is at (64°S, 3°W). There is a wide part of the gyre flowing to the north with an intensification near the Antarctic Peninsula. The ACC–Weddell gyre boundary runs along 58°S. This agrees with the schematic representation of the large-scale upper-level currents in Fig. 1 of Peterson and Stramma (1991). Figures 10, 11, and 12 present the pattern of the Weddell gyre at depths of 1112, 3962, and 4675 m, respectively. The gyre keeps its structure from the upper layers down to the bottom. The beginning of the western boundary current that transports WSBW and WSDW, respectively, from the Weddell Sea Basin to the north is recognizable at 3962-m depth. At the depth of 4675 m the Weddell Sea Basin is separated from the Argentine Basin and the Cape Basin. The circulation pattern of the Weddell gyre agrees at all depth levels with that from Reid (1989). The velocities of the Weddell gyre decrease from 3 to 4 cm s^{−1} near the surface to less then 2 cm s^{−1} at the bottom. These are typical values compared to Fahrbach et al. (1994a,b).

In the upper-level circulation (Fig. 8) the Falkland/Malvinas Current separates from the ACC in the Scotia Sea. This western boundary current flows along the coast of South America to the north. The Brazil Current and the so-called confluence region (Gordon 1989) near 40°S where the Brazil Current meets the Falkland/Malvinas Current is not marked. This was expected because the SSH data contain only large-scale features yielding only a smooth flow field and because of the mesh resolution in the present model version. Near South Africa the Benguela Current and the Agulhas retroflection are only weakly represented because of the smooth SSH data. Nevertheless the Agulhas retroflection has its maximum east of 20°E. North of the Weddell gyre the circulation is dominated by the ACC and the subtropical gyre. Figure 13 shows the horizontal flow field of the whole model domain at 3962 m. The Weddell gyre at this depth was already presented in Fig. 11. There is a current from the Weddell Sea basin into the Argentine Basin that transports bottom water from the Weddell Sea into the Argentine Basin. Along the way the bottom water is modified by the LCDW. But the flow field in the Argentine Basin is dominated by a counterclockwise flow pattern. The reason is that the northern component of the reference velocities depending on the smooth SSH data are not large enough to produce a northward flow at the bottom.

Comparisons of the model circulation with the flow fields produced by other numerical models show also agreement. The upper-level circulation is similar to that from England and Garcon (1994) and Barnier et al. (1996). In the work of Barnier et al. (1996) a similar westward current in the northern part of the Argentine Basin at 4375-m depth is presented.

### c. Transports

As mentioned before the total volume transports across the open boundaries are prescribed. However, the transports of the different water masses described in 5a are model results. They are presented in Fig. 14. Through the Drake Passage 38.2 Sv intermediate water (IW), 87.7 Sv deep water (DW), and 4.1 Sv bottom water (BW) are flowing into the South Atlantic. Also 0.2 Sv SW, 44.1 Sv IW, 70.3 Sv DW, and 15.4 Sv BW are exported into the Indian Ocean across 20°E. Across the northern open boundary of the model at 30°S 4.1 Sv SW and 20.4 Sv AAIW are transported to the north, which is compensated by a southward flow of 23.6 Sv NADW. The transport of 0.9 Sv AABW to the south is due to the deep flow field in this region, which has the wrong direction as we have discussed above. A closer inspection shows its characteristic as modified bottom water with 28.2 ≤ *γ* < 28.27. The northward spreading of denser bottom water with *γ* ≥ 28.27 extends up to about 35°S, which is in agreement with Orsi et al. (1999). There is 13.0 Sv of this water mass formed in the Weddell Sea. These are exported totally across 20°E. The formation rate of AABW with a neutral density of *γ* ≥ 28.2 over the whole model domain is 10.4 Sv. Rintoul (1991) estimated a formation rate of 9 Sv in his standard case and 13 Sv in the “warm-water-path” case. Other values from previous studies are 12.3 Sv (de las Heras and Schlitzer 1999, study B), 5 Sv (Matano and Philander 1993), and 1 Sv (Sloyan and Rintoul 2001). In the work of Macdonald (1995) no bottom water production is detected. Additional estimates of AABW formation rates of previous studies are in the range from 2 to 15 Sv (see Table 3 in Orsi et al. 1999).

These transports result in a heat transport of 0.64 PW across 30°S to the north. This value is compared with the heat transports presented in previous studies of Hastenrath (1980, 1982), Fu (1981), Semtner and Chervin (1988, 1992), Webb et al. (1991), Rintoul (1991), Matano and Philander (1993), Schlitzer (1995), Macdonald (1995), de las Heras and Schlitzer (1999) and Sloyan and Rintoul (2001)(see Fig. 15). The heat transports across 70°W and 20°E amount to 1.06 and 1.25 PW (PW = 10^{15} W). The ratio between the transports of intermediate and deep water together with the heat transports across the open boundaries compares well with the warm-water-path case of Rintoul (1991). In that study the total volume transports across 32°S, 68°W, and 20°E are likewise 0, 129, and 130 Sv. The heat transport across 32°S is assumed to be 0.69 PW (the value given by Hastenrath 1982), which results in a transport of 22 Sv NADW across 30°S to the south. The heat transports across 68°W and 20°E are 1.3 and 1.12 PW. The export of bottom water into the Indian Ocean amounts to 16 Sv. A transport of surface water into the South Atlantic across this section like the 13 Sv in Rintoul's work is absent in our model results.

The transports of salt in our model are 4628 kt s^{−1} eastward across 70°W, 4631 kt s^{−1} eastward across 20°E, and 13 kt s^{−1} across 30°S to the south. The heat and salt transport across the southern open boundary at 74°S are nearly vanishing.

### d. Vertical velocities and sources

Figure 16 shows the vertical velocity field *w* at 100-m depth. The order of magnitude of *w* is 10^{−6} m s^{−1}. This is in good agreement with vertical velocities presented by Olbers and Wenzel (1989) and Schlitzer (1995). For clarity only the zero and the 10 × 10^{−7} m s^{−1} contour lines are printed. Regions with *w* < −20 × 10^{−7} m s^{−1} occur southwest of South Africa and east of 35°W in the Weddell Sea. West of 35°W *w* increases up to 60 × 10^{−7} m s^{−1}. Near the coast of South America we find a small area around 45°S where the vertical velocity is in the range between ±80 × 10^{−7} m s^{−1}. The maximum value of 140 × 10^{−7} m s^{−1} occurs at the Antarctic Peninsula.

The distribution of the sources *F*^{T} and *F*^{S} of the advection–diffusion equation (6) yields a pattern with no regular structure. Changes of sign are frequent. The strength of the sources increases near the bottom especially in the vicinity of steep gradients in the model topography. In such regions tetrahedra have small volumes to get a mesh close to the topography. As an example Fig. 17 shows the sources field along 20°W.

## 6. Summary and conclusions

In this study we have presented a new adjoint model with finite elements applied to a stationary advection–diffusion problem. The usage of the finite-element method is still rare in ocean modeling. As shown in section 3 the FEM is very suitable to treat problems on complex geometries given by the coastlines and the bottom topography. It provides the possibility to work with structured or unstructured grids and local refinement. The weak formulation of the variational principle (standard Galerkin method, section 3a) reduces the order of the partial differential equation by integration by parts, so that one can use simple basis functions. Different stabilization techniques for advective problems are well known (section 3b). The adjoint equations are given by the transpose of the stiffness matrix and by derivatives of the stiffness matrix.

We presented a first application of the new model to the South Atlantic with the major attention directed to the Weddell Sea. We assimilated climatological hydrographic data and SSH data from satellite altimetry in the model. The transports of the model solution across the open boundaries agree with the values from literature. The heat transport of 0.64 PW across 30°S to the north connected with a southward transport of 23.6 Sv NADW corresponds to the warm-water-path case of Rintoul's study (1991). The formation rate of bottom water with a neutral density of *γ* ≥ 28.2 in the model domain amounts to 10.4 Sv. The upper-level circulation corresponds to that from literature. Similarly the circulation of the Weddell gyre is in good agreement over the full depth of the Weddell Sea to that from literature. Bottom water formed in the Weddell Sea is transported to the east and to the north where it is modified by the lower CDW of the ACC.

The southward transport of 0.9 Sv bottom water (28.2 ≤ *γ* < 28.27) at 30°S is in disagreement with literature. The main reason is that in this region the northern components of the reference velocities fitted to the gradient of the SSH data are too small to yield a northward transport at the bottom. Another important reason is the resolution of the mesh near the bottom at 30°S. It is somewhat coarser than in the Weddell Sea and there can be also some influence from open boundary effects. There are possibilities for mesh refinement but they are always limited by the large memory requirements of adjoint models. Parallelizing the model code can help here in the future.

The sources *F*^{T} and *F*^{S} of the advection–diffusion equation were introduced to parameterize subscale processes that are not resolved by the mesh and by the diffusion coefficients *K*_{υ} and *K*_{h}. Constant diffusion coefficients on the whole mesh created by tetrahedra with a large range of volumes are describing the processes on the elements only roughly. In the further development of the model diffusion coefficients that are varying spatially could be suitable.

## Acknowledgments

We would like to thank I. Schmelzer from the Weierstraß Institute of Applied Analysis and Stochastics in Berlin, Germany, who let us use the 3D version of the IBG grid generator and we thank two anonymous reviewers who provided many useful suggestions that helped to improve the manuscript.

## REFERENCES

Barett, R., and Coauthors. 1994:

*Templates for the Solution of Linear Systems: Building Blocks for Iterative Methods*. SIAM, 124 pp.Barnier, B., Marchesiello P. , and de Miranda A. P. , 1996: Modelling the ocean circulation in the South Atlantic: A strategy for dealing with open boundaries.

*The South Atlantic: Present and Past Circulation,*G. Wefer et al., Eds., Springer-Verlag, 189–304.Beckmann, A., Hellmer H. , and Timmermann R. , 1999: A numerical model of the Weddell Sea: Large scale circulation and water mass distribution.

,*J. Geophys. Res.***104**(C10) 23375–23391.Behrens, J., 1996: Adaptive Semi-Lagrange-Finite-Elemente-Methode zur Lösung der Flachwassergleichungen: Implementierung und Parallelisierung. Rep. 217 on Polar Research, Alfred Wegener Insitute, Bremerhaven, Germany, 167 pp.

Behrens, J., 1998: Atmospheric and ocean modeling with an adaptive finite element solver for the shallow-water equations.

,*Appl. Numer. Math.***26****,**217–226.Brezzi, F., Bristeau M-O. , Franca L. P. , Mallet M. , and Rogé G. , 1992: A relationship between stabilized finite element methods and the Galerkin method with bubble functions.

,*Comput. Methods Appl. Mech. Eng.***96****,**117–129.Brooks, A. N., and Hughes T. J. R. , 1982: Streamline upwind/Petrov–Galerkin formulations for convection dominated flows with particular emphasis on the incompressible Navier–Stokes equations.

,*Comput. Methods Appl. Mech. Eng.***32****,**199–259.Connor, J. J., and Wang J. D. , 1973: Finite-element modeling of hydrodynamic circulation.

*Numerical Methods Fluids Dynamics,*Pentech Press, 355–387.Curchitser, E., Iskandarani M. , and Haidvogel D. B. , 1998: A spectral element solution of the shallow-water equations on multiprocessor computers.

,*J. Atmos. Oceanic Technol.***15****,**510–521.Curchitser, E., Haidvogel D. B. , and Iskandarani M. , 1999: On the transient adjustment of a mid-latitude abyssal ocean basin with realistic geometry: The constant depth limit.

,*Dyn. Atmos. Oceans***29****,**147–188.de las Heras, M., and Schlitzer R. , 1999: On the importance of intermediate water flows for the global ocean overturning.

,*J. Geophys. Res.***104**(C7) 15515–15536.Dobrindt, U., 1999: Ein Inversmodell für den Südatlantik mit der Methode der finiten Elemente. Ph.d. thesis, Alfred Wegener Institute for Polar and Marine Research, Bremerhaven, and University of Bremen, Germany, 131 pp.

Dobrindt, U., and Frickenhaus S. , 2000: Partitioning, reordering and solving an advection diffusion problem formulated in finite elements. Rep. 98, Dept. of Physics, Alfred Wegener Institute, Bremerhaven, Germany, 32 pp.

England, M. H., and Garcon V. C. , 1994: South Atlantic circulation in a World Ocean model.

,*Ann. Geophys.***12****,**812–825.Fahrbach, E., Rohardt G. , Schröder M. , and Strass V. , 1994a: Transport and structure of the Weddell Gyre.

,*Ann. Geophys.***12****,**840–855.Fahrbach, E., Schröder M. , and Klepikov A. , 1994b: Circulation and water masses in the Weddell Sea.

,*Phys. Ice-Covered Seas***2****,**569–604.Fix, G. J., 1975: Finite element model for ocean circulation problems.

,*SIAM J. Appl. Math.***29****,**371–387.Franca, L., Frey S. L. , and Hughes T. J. R. , 1992: Stabilized finite element methods. I. Application to the advective–diffusive model.

,*Comput. Methods Appl. Mech. Eng.***95****,**253–276.Fu, L-L., 1981: The general circulation and meridional heat transport of the subtropical South Atlantic determind by inverse methods.

,*J. Phys. Oceanogr.***11****,**1171–1193.Gilbert, J. C., and Lemaréchal C. , 1989: Some numerical experiments with variable-storage quasi-Newton algorithms.

,*Math. Programm.***45****,**407–435.Gill, A. E., 1982:

*Atmosphere–Ocean Dynamics*. Academic Press, 662 pp.Giraldo, F. X., 1999: Trajectory calculations for spherical geodesic grids in Cartesian space.

,*Mon. Wea. Rev.***127****,**1651–1662.Giraldo, F. X., 2000a: Lagrange–Galerkin methods on spherical geodesic grids: The shallow water equations.

,*J. Comput. Phys.***160****,**336–368.Giraldo, F. X., 2000b: The Lagrange–Galerkin method for the 2D shallow water equations on adaptive grids.

,*J. Numer. Methods Fluids.***33****,**789–832.Gordon, A. L., 1989: Brazil–Malvinas confluence—1984.

,*Deep-Sea Res.***36****,**359–384.Gouretski, V. V., and Jancke K. , 1998: A new climatology for the World Ocean. WHP SAC Tech. Rep. 3, WOCE Rep. 162/98, WOCE Special Analysis Center, Max-Planck Institute, Hamburg, Germany, 23 pp. [Available online at http://www.dkrz.de/∼u241046.].

Haidvogel, D. B., Robinson A. R. , and Schulman E. E. , 1980: The accuracy, efficiency and stability of three numerical models with application to open ocean problems.

,*J. Comput. Phys.***34****,**1–53.Haidvogel, D. B., Curchitser E. , Iskandarani M. , Hughes R. , and Taylor M. , 1997: Global modeling of the ocean and atmosphere using the spectral element method.

*Numerical Methods in Atmospheric and Ocean Modelling,*C. A. Lin, R. Laprise, and H. Ritchie, Eds., Canadian Meteorology and Oceanographic Society, 505–531.Hastenrath, S., 1980: Heat budget of tropical ocean and atmosphere.

,*J. Phys. Oceanogr.***10****,**159–170.Hastenrath, S., 1982: On meridional heat transports in the World Ocean.

,*J. Phys. Oceanogr.***12****,**922–927.Hellerman, S., and Rosenstein M. , 1983: Normal monthly wind stress over the World Ocean with error estimates.

,*J. Phys. Oceanogr.***13****,**1093–1104.Hernandez, F., and Schaeffer P. , 2000: Altimetric mean sea surfaces and gravity anomaly maps intercomparisons. AVISO Tech. Rep. AVI-NT-011-5242-CLS, CLS Editor, Toulouse, France, 48 pp.

Hughes, T. R. J., Mallet M. , and Mizukami A. , 1986: A new finite element formulation for computational fluid dynamics. II. Beyond SUPG.

,*Comput. Methods Appl. Mech. Eng.***54****,**341–355.Hughes, T. R. J., Franca L. P. , and Hulbert G. M. , 1989: A new finite element formulation for computational fluid dynamics: VIII. The Garlerkin/least-squares method for advective–diffusive equations.

,*Comput. Methods Appl. Mech. Eng.***73****,**173–189.Ip, J. T. C., and Lynch D. R. , 1995: Comprehensive coastal circulation simulation using finite elements: Nonlinear prognostic time-stepping model. QUODDY3 user's manual. Rep. NML95-1, Dartmouth College, Hanover, NH, 45 pp.

Iskandarani, M., Haidvogel D. B. , and Boyd J. P. , 1995: A staggered spectral finite element method for the shallow water equations.

,*Int. J. Numer. Methods Fluids***20****,**393–414.Johnson, C., 1990:

*Numerical Solution of Partial Differential Equations by the Finite Element Method*. Cambridge University Press, 279 pp.Karypis, G., and Kumar V. , 1998: METIS—A software package for partitioning unstructured graphs, partitioning meshes and computing fill-reducing orderings of sparse matrices (version 4.0). University of Minnesota, Dept. of Computer Science/Army HPC Research Center, Minneapolis, MN, 44 pp.

Le Dimet, F. X., and Talagrand O. , 1986: Variational algorithms for analysis and assimilation of meteorological observations: Theoretical aspects.

,*Tellus***38A****,**97–110.Lemoine, F. G., and Coauthors. 1998: The development of the joint NASA GSFC and National Imagery and Mapping Agency (NIMA) Geopotential Model EGM96. NASA/TP-1998-206861, National Aeronautics and Space Administration Goddard Space Flight Center, 575 pp.

Le Provost, C., and Vincent P. , 1991: Finite element for modelling ocean tides.

*Tidal Hydrodynamics,*B. Parker, Ed., John Wiley and Sons, 41–60.Le Roux, D. Y., Staniforth A. , and Lin C. A. , 1998: Finite elements for shallow water equations ocean models.

,*Mon. Wea. Rev.***126****,**1931–1951.Levin, J. G., Iskandarani M. , and Haidvogel D. B. , 1997: A spectral filtering procedure for eddy-resolving simulations with a spectral element ocean model.

,*J. Comput. Phys.***137****,**130–154.Lynch, D. R., and Hannah C. G. , 2001: Inverse model for limited-area hindcasts on the continental shelf.

,*J. Atmos. Oceanic Technol.***18****,**962–981.Lynch, D. R., Ip J. T. C. , Naimie C. E. , and Werner F. E. , 1996: Comprehensive coastal circulation model with application to the Gulf of Maine.

,*Contin. Shelf Res.***16****,**875–906.Macdonald, A. M., 1995: Oceanic fluxes of mass, heat and freshwater: A global estimate and perspective. Ph.d. thesis, Massachusetts Institute of Technology/Woods Hole Oceanographic Instution Joint Program, 326 pp.

Matano, R. P., and Philander S. G. H. , 1993: Heat and mass balances of the South Atlantic Ocean calculated from a numerical model.

,*J. Geophys. Res.***98**(C1) 977–984.Myers, P. G., and Weaver A. J. , 1995: A diagnostic barotropic finite-element ocean circulation model.

,*J. Atmos. Oceanic Technol.***12****,**511–526.National Geophysical Data Center, 1988: Digital relief of the surface of the earth. Data Announcement 88-MGG-02, NOAA, Boulder, CO.

Nechaev, D. A., and Yaremchuk M. I. , 1995: Application of the adjoint technique to processing of a standard section data set: World Ocean Circulation Experiment section S4 along 67°S in the Pacific Ocean.

,*J. Geophys. Res.***100**(C1) 865–879.Olbers, D., 1989: A geometrical interpretation of inverse problems.

*Oceanic Circulation Models: Combining Data and Dynamics,*D. L. T. Anderson and J. Willebrand, Eds., Kluwer Academic, 79–93.Olbers, D., and Wenzel M. , 1989: Determining diffusivities from hydrographic data by inverse methods with applications to the circumpolar current.

*Oceanic Circulation Models: Combining Data and Dynamics,*D. L. T. Anderson and J. Willebrand, Eds., Kluwer Academic, 95–139.Orsi, A. H., Johnson G. C. , and Bullister J. L. , 1999: Circulation, mixing and production of Antarctic Bottom Water.

*Progress in Oceanography,*Vol. 43, Pergamon, 55–109.Peterson, R. G., and Stramma L. , 1991: Upper-level circulation in the South Atlantic Ocean.

*Progress in Oceanography,*Vol. 26, Pergamon, 1–73.Reid, J. L., 1989: On the total geostrophic circulation of the South Atlantic Ocean: Flow patterns, tracers and transports.

*Progress in Oceanography,*Vol. 23, Pergamon, 149–244.Rintoul, S. R., 1991: South Atlantic interbasin exchange.

,*J. Geophys. Res.***96****,**2675–2692.Saad, Y., and Schultz M. H. , 1986: GMRES: A generalized minimal residual algorithm for solving nonsymmetric linear systems.

,*SIAM J. Sci. Stat. Comput.***7****,**856–869.Schlitzer, R., 1993: Determining the mean, large-scale circulation of the Atlantic with the adjoint method.

,*J. Phys. Oceanogr.***23****,**1935–1952.Schlitzer, R., 1995: An adjoint model for the determination of the mean oceanic circulation, air–sea fluxes and mixing coefficients. Rep. 156 on Polar Research, Alfred Wegener Institute, Bremerhaven, Germany, 103 pp.

Schmelzer, I., 1995: 3D grid generation with contravariant geometry description. Ph.d. thesis, Weierstrass Institute of Applied Analysis and Stochastics, Berlin, Germany, 103 pp.

Schröter, J., 1989: Driving of non-linear time-dependent ocean models by observation of transient tracers—A problem of constrained optimization.

*Oceanic Circulation Models: Combining Data and Dynamics,*D. L. T. Anderson and J. Willebrand, Eds., Kluwer Academic, 257–285.Schröter, J., Seiler U. , and Wenzel M. , 1993: Variational assimilation of Geosat data into an eddy resolving model of the Gulf Stream extension area.

,*J. Phys. Oceanogr.***23****,**925–953.Seiß, G., 1996: Ein inverses modell der globalen barotropen zirkulation des ozeans. Ph.d. thesis, Rep. 72, Dept. of Physics, Alfred Wegener Institute, Bremerhaven, Germany, 102 pp.

Semtner, A. J., and Chervin R. M. , 1988: A simulation of the global ocean circulation with resolved eddies.

,*J. Geophys. Res.***93****,**15502–15522.Semtner, A. J., and Chervin R. M. , 1992: Ocean general circulation from a global eddy-resolving model.

,*J. Geophys. Res.***97****,**5493–5550.Sloyan, B., 1997: The circulation of the Southern Ocean and the adjacent ocean basins determined by inverse methods. Ph.d. thesis, Institute of Antarctic and Southern Ocean Studies, University of Tasmania, 468 pp.

Sloyan, B., and Rintoul S. R. , 2001: The Southern Ocean limb of the global deep overturning circulation.

,*J. Phys. Oceanogr.***31****,**143–173.Taylor, M., Tribbia J. , and Iskandarani M. , 1997: The spectral element method for the shallow water equations on the sphere.

,*J. Comput. Phys.***130****,**92–108.Thacker, W. C., 1988: Three lectures on fitting numerical models to observations. Tech. Rep. GKSS 87/E/65, GKSS-Forschungzentrum Geesthacht, 64 pp.

Thomasset, F., 1981:

*Implementation of Finite Element Methods for Navier–Stokes Equations*. Springer-Verlag, 160 pp.Vogeler, A., and Schröter J. , 1995: Assimilation of satellite altimeter data into an open ocean model.

,*J. Geophys. Res.***100**(C8) 15951–15963.Vogeler, A., and Schröter J. , 1999: Fitting a regional ocean model with adjustable open boundaries to TOPEX/Poseidon data.

,*J. Geophys. Res.***104**(C9) 20789–20799.Wahr, J., Molenaar M. , and Bryan F. , 1998: Time variability of the earth's gravity field: Hydrological and oceanic effects and their possible detection using GRACE.

,*J. Geophys. Res.***103**(B12) 30205–30299.Walters, R. A., and Werner F. E. , 1989: A comparison of two finite element models of tidal hydrodynamics using a North Sea data set.

,*Adv. Water Resour.***12****,**184–193.Webb, D. J., and Coauthors. 1991: Using an eddy resolving model to study the Southern Ocean.

,*Eos, Trans. Amer. Geophys. Union***72****,**169–174.Wunsch, C., 1996:

*The Ocean Circulation Inverse Problem*. Cambridge University Press, 442 pp.Wunsch, C., Haidvogel D. B. , and Iskandarani M. , 1997: Dynamics of the long-period tides.

*Progress in Oceanography,*Vol. 40, Pergamon, 81–108.Yaremchuk, M., Nechaev D. , Schröter J. , and Fahrbach E. , 1998: A dynamically consistent analysis of circulation and transports in the southwestern Weddell Sea.

,*Ann. Geophys.***16****,**1024–1038.Zhu, K., Navon I. M. , and Zou X. , 1994: Variational data assimilation with a variable resolution finite-element shallow-water equations model.

,*Mon. Wea. Rev.***122****,**946–965.Zienkiewicz, O. C., 1977:

*The Finite Element Method*. McGraw-Hill, 743 pp.

## APPENDIX

### The Basis Functions

*ϕ*

_{i}

*x,*

*y,*

*z*

*a*

_{i}

*x*

*b*

_{i}

*y*

*c*

_{i}

*z*

*d*

_{i}

*ϕ*

_{i}are defined on each tetrahedron. Let

*i,*

*j*= 1, 2, 3, 4 denote the corner points of the tetrahedron. Then each function

*ϕ*

_{i}is equal to 1 at the corner

*i*=

*j,*zero at all other corners, and linear in between. The coefficients

*a*

_{i},

*b*

_{i},

*c*

_{i}, and

*d*

_{i}are easily calculated from the coordinates

**x**

_{j}of the corner points

*j.*For instance the first basis function

*ϕ*

_{j=1}(

*x,*

*y,*

*z*) of the tetrahedron

*τ*

_{m}reads

The mesh: bottom topography and open boundaries

Citation: Journal of Atmospheric and Oceanic Technology 20, 3; 10.1175/1520-0426(2003)020<0392:AAOMUF>2.0.CO;2

The mesh: bottom topography and open boundaries

Citation: Journal of Atmospheric and Oceanic Technology 20, 3; 10.1175/1520-0426(2003)020<0392:AAOMUF>2.0.CO;2

The mesh: bottom topography and open boundaries

Citation: Journal of Atmospheric and Oceanic Technology 20, 3; 10.1175/1520-0426(2003)020<0392:AAOMUF>2.0.CO;2

SSH data in m

SSH data in m

SSH data in m

Water masses along the Greenwich meridian

Water masses along the Greenwich meridian

Water masses along the Greenwich meridian

Water masses along 20°W

Water masses along 20°W

Water masses along 20°W

Water masses along 40°W

Water masses along 40°W

Water masses along 40°W

Water masses along 30°S

Water masses along 30°S

Water masses along 30°S

Reference velocities (*u*_{ref}, *υ*_{ref}) in the South Atlantic with South America at the northwest corner, the Antarctic Peninsula at the southwest corner, Antarctica at the southeast corner, and South Africa at the northeast corner

Reference velocities (*u*_{ref}, *υ*_{ref}) in the South Atlantic with South America at the northwest corner, the Antarctic Peninsula at the southwest corner, Antarctica at the southeast corner, and South Africa at the northeast corner

Reference velocities (*u*_{ref}, *υ*_{ref}) in the South Atlantic with South America at the northwest corner, the Antarctic Peninsula at the southwest corner, Antarctica at the southeast corner, and South Africa at the northeast corner

The Weddell gyre in 100 m with the Antarctic Peninsula at the southWest and Antarctica at the southEast

The Weddell gyre in 100 m with the Antarctic Peninsula at the southWest and Antarctica at the southEast

The Weddell gyre in 100 m with the Antarctic Peninsula at the southWest and Antarctica at the southEast

The Weddell gyre at 1112 m

The Weddell gyre at 1112 m

The Weddell gyre at 1112 m

The Weddell gyre at 3962 m

The Weddell gyre at 3962 m

The Weddell gyre at 3962 m

The Weddell gyre at 4675 m

The Weddell gyre at 4675 m

The Weddell gyre at 4675 m

Horizontal flow field at 3962 m of the South Atlantic

Horizontal flow field at 3962 m of the South Atlantic

Horizontal flow field at 3962 m of the South Atlantic

Transports in Sv of surface water (SW, *γ* < 26.5), intermediate water (IW, 26.5 ≤ *γ* < 27.6), deep water (DW, 27.6 ≤ *γ* < 28.2), and bottom water (BW, *γ* ≥ 28.2) across the open boundaries of the model domain at 30°S, 74°S, 20°E, and 70°W; negative sign denotes transports to the west or to the south

Transports in Sv of surface water (SW, *γ* < 26.5), intermediate water (IW, 26.5 ≤ *γ* < 27.6), deep water (DW, 27.6 ≤ *γ* < 28.2), and bottom water (BW, *γ* ≥ 28.2) across the open boundaries of the model domain at 30°S, 74°S, 20°E, and 70°W; negative sign denotes transports to the west or to the south

Transports in Sv of surface water (SW, *γ* < 26.5), intermediate water (IW, 26.5 ≤ *γ* < 27.6), deep water (DW, 27.6 ≤ *γ* < 28.2), and bottom water (BW, *γ* ≥ 28.2) across the open boundaries of the model domain at 30°S, 74°S, 20°E, and 70°W; negative sign denotes transports to the west or to the south

Comparison of the heat transports from different studies across 30°S in the South Atlantic (PW = 10^{15} W)

Comparison of the heat transports from different studies across 30°S in the South Atlantic (PW = 10^{15} W)

Comparison of the heat transports from different studies across 30°S in the South Atlantic (PW = 10^{15} W)

Vertical velocities *w* in 10^{−7} m s^{−1} at 100 m of the South Atlantic

Vertical velocities *w* in 10^{−7} m s^{−1} at 100 m of the South Atlantic

Vertical velocities *w* in 10^{−7} m s^{−1} at 100 m of the South Atlantic

Sources *F*^{T} and *F*^{S} along 20°W. The zero contour line is dotted

Sources *F*^{T} and *F*^{S} along 20°W. The zero contour line is dotted

Sources *F*^{T} and *F*^{S} along 20°W. The zero contour line is dotted