Content uploaded by Eric P. Chassignet

Author content

All content in this area was uploaded by Eric P. Chassignet on Sep 06, 2019

Content may be subject to copyright.

)ORULGD6WDWH8QLYHUVLW\/LEUDULHV

Follow this and additional works at DigiNole: FSU's Digital Repository. For more information, please contact lib-support@fsu.edu

2018

Ocean Circulation Modeling for

Operational Oceanography: Current

Status and Future Challenges

Julien Le Sommer, Eric P. Chassignet and Alan J. Wallcraft

Terms and Conditions: Readers may view, browse, and/or download material for temporary copying purposes only, provided

these uses are for noncommercial personal purposes. Except as provided by law, this material may not be further

reproduced, distributed, transmitted, modified, adapted, performed, displayed, published, or sold in whole or in part,

without prior written permission from GODAE OceanView.

289

CHAPTER 12

OceanCirculationModelingfor

OperationalOceanography:Current

StatusandFutureChallenges

Julien Le Sommer1, Eric P. Chassignet2, and Alan J. Wallcraft2

1Univ. Grenoble Alpes, CNRS, IRD, IGE, Grenoble, France; 2Center for Ocean-Atmospheric Prediction

Studies (COAPS), Florida State University, Tallahassee, FL, USA

This chapter focuses on ocean circulation models used in operational oceanography, physical oceanography

and climate science. Ocean circulation models are a particular branch of ocean numerical modeling that

focuses on the representation of ocean physical properties over spatial scales ranging from the global scale

to less than a kilometer and time scales ranging from hours to decades. As such, they are an essential build-

ing block for operational oceanography systems and their design receives a lot of attention from operational

and research centers.

Introduction

cean modeling is an important branch of operational oceanography. Ocean numerical

models are an essential building block of global and regional operational oceanography

systems. In this context, ocean circulation models are used in conjunction with data

assimilation for extrapolating both in space and in time the available satellite and in situ oceanic

observations in order to build a physically consistent estimate of the ocean state and its evolution.

Ocean modeling is a relatively recent discipline in the field of oceanography. The underlying

principles of the algorithmic formulation of ocean circulation models were first proposed in the

1960s by Bryan (1969; see also McWilliams [1996] for an historical review). Since then, the

continuous increase in computing power has allowed ocean circulation models to provide ever more

meaningful and comprehensive descriptions of the ocean circulation. Ocean modeling is now

recognized as an essential supplement to more traditional scientific methodologies in oceanography.

The variety of oceanic physical processes accounted for in ocean circulation models has also notably

broadened over past decades. The original scope of ocean circulation models was to describe

oceanic properties and physical processes at scales significantly larger than the mesoscale

(horizontal scales on the order of 100 km and time scales on the order of three months). But the

increase in computing power and the improved physical consistency of their formulation now allow

ocean circulation models to resolve routinely oceanic flows down to the submesoscale (horizontal

Le Sommer, J., E.P. Chassignet, and A.J. Wallcraft, 2018: Ocean circulation modeling for operational oceanography: Current

status and future challenges. In "New Frontiers in Operational Oceanography", E. Chassignet, A. Pascual, J. Tintoré, and J.

Verron, Eds., GODAE OceanView, 289-306, doi:10.17125/gov2018.ch12.

O

290 JULIEN LE SOMMER ET AL.

scales on the order of 10 km; Chassignet and Xu [2017]) and to describe internal wave and internal

tides at a global scale (Shriver et al., 2012).

The range of uses and applications of ocean circulation models has diversified over the past

several decades. Ocean circulation models are now used standalone for simulating ocean

circulation; when integrated in a data assimilation framework, they may be used for producing

short-range ocean forecasts and for ocean reanalysis; coupled to atmospheric circulation models,

they may be used for seasonal to decadal forecasts; and fully integrated into Earth System Models,

they are crucial to climate modeling. Scientists also use ocean circulation models as the

experimental tool of choice for improving our mechanistic understanding of the ocean.

The variety of oceanic physical processes accounted for in ocean circulation models has also

broadened in recent decades. Ocean circulation models were initially used to describe oceanic

properties and physical processes at scales significantly larger than the mesoscale (horizontal scales

on the order of 100 km and time scales on the order of three months). But the increase in computing

power and improved physical consistency of their formulation now allow ocean circulation models

to resolve routinely oceanic flows down to the submesoscale (horizontal scales on the order of 10

km; Chassignet and Xu [2017]) and to describe internal wave and internal tides at the global scale

(Shriver et al., 2012; Savage et al., 2017).

However, despite the maturity and broad range of applications of ocean circulation models

achieved to date, ocean modelers still face notable challenges. Running ocean circulation models at

increased model resolution does not solve the problem of subgrid scale closures (unresolved

processes), and may even question some of the underlying assumptions and algorithms of current

ocean circulation models (hydrostatic vs. non hydrostatic, for example). The modularity of modern

geoscientific models requires the design of robust and physically rational approaches for coupling

ocean circulation models with other model components. Most importantly, there is a growing

concern about the necessity to describe explicitly how model uncertainty propagates in geoscientific

modeling systems. This is why ocean circulation model design is still a very active field of research

and will likely remain so well into the future.

In this chapter, we aim to: (i) briefly describe the principles that underpin the formulation of

ocean circulation models, (ii) review current skills of ocean circulation models, and (iii) present

what we believe are the new frontiers in ocean circulation model design. Obviously, given the

complexity of the issues at stake and the amount of energy that is put into ocean model development,

this chapter will only briefly touch upon the above objectives. Our goal is therefore not to cover

these questions thoroughly, but rather to provide an entry point to the science of ocean models. The

second section of this chapter focuses on the science and applications of ocean circulation models;

the third section describes ongoing research avenues geared toward improving the representation of

physical processes in ocean circulation models; and the fourth section describes the ongoing

paradigm shift in ocean modeling towards a more probabilistic description of oceanic flows. More

general considerations regarding the future of ocean modeling are presented in the conclusion.

OCEAN CIRCULATION MODELING FOR OPERATIONAL OCEANOGRAPHY: CURRENT

STATUS AND FUTURE CHALLENGES 291

OceanCirculationModels:Scope,Usage,&Fundamentals

Ocean modeling is a branch of numerical modeling that focuses on representation of the physical

mechanisms governing the evolution of ocean physical properties, namely T, S, u, v, and w, where

T is temperature, S is salinity, and u, v, and w are the horizontal and vertical components of the

velocity V. A good understanding of a range of physical processes is required in order to build a

numerical model that is capable of faithfully representing the ocean circulation. Without this

physical understanding, it is easy to derive erroneous conclusions since a numerical model is

constructed using discretized equations of motion. Indeed, direct numerical simulations (DNS) of

the ocean (i.e., numerical representation of the smallest turbulent scale, i.e. the Kolmogorov length

scale, which is on the order of 1 cm (Smyth et al., 2001)), is not possible with present-day

computers. The largest simulation achievable today is on scales of 10 m (Yeung et al., 2015).

At present, we cannot represent these small scales; we can only achieve a truncated

representation of the ocean and this will remain the case for the foreseeable future. The spatial and

temporal scales that one can currently represent strongly depends on the model configuration and

application. High resolution operational oceanography requires accurate depiction of upper ocean

structure and mesoscale features such as eddies and meandering fronts. Accurate sea level

representation is crucial for coastal models (response to wind, tides, and surface pressure) and

seasonal-to-interannual forecasts require a good representation of the upper ocean mass field and

the coupling to an active atmosphere. On global and basin scales, high horizontal resolution (1/10o

to 1/25o and very rarely to 1/50o; Chassignet and Xu [2017]) is mostly used for ocean “weather”

and seasonal-to-decadal variability (Fig. 12.1). The emphasis is on short integrations (years to

decades) and most models of this class are coupled to a sea-ice model but are stand-alone and use

prescribed atmospheric fields. Coarser resolution (1/4o to 1o) is principally used for climate

applications (Griffies et al., 2000). The emphasis for this class of models is on long integrations

with fully coupled ocean-ice-atmosphere models.

Figure 12.1. Example of submesoscale permitting modeling – March 1st surface vorticity over the Gulf Stream

in a1/50° North Atlantic domain (~1.5 km horizontal grid spacing). Adapted from Chassignet and Xu (2017).

292 JULIEN LE SOMMER ET AL.

In practice, ocean models consist of a numerical solution to a set of partial differential equations

(PDEs) describing the ocean dynamics. These PDEs are based on an approximated version of

Navier-Stokes equations adapted to our regimes of interest (see the review papers by Griffies [2004]

and Griffies and Treguier [2013] for details). An ocean numerical model typically consists of 20,000

to 200,000 lines of code, usually written in FORTRAN, and can easily take up to 10 years of

community development to be fully functional.

The first ocean numerical model was put forward by Bryan (1969) following in the footsteps of

numerical weather prediction. The first discretized equations for an atmospheric application were

solved manually by Richardson (1922) and numerically by Charney et al. (1950). The latter

consisted of a 15 by 18 grid (∆x = 736 km) and demonstrated that numerical weather prediction was

possible. The latest configurations use horizontal resolution ∆x on the order of 1 km (Chassignet

and Xu, 2017). The sheer size of the ocean problem is such that it will always require the latest

generation of supercomputers, and ocean/climate models have traditionally been one of the biggest

users of computer resources. This means that close collaborations with computer scientists are

essential to ensure that the numerical codes run efficiently and take full advantage of the computing

architecture, as this changes constantly and ocean modelers need to be flexible and ready to adopt

new approaches. The main limitation is not computational speed of the processors, but access to

memory and latency in reading/writing on disk drives (I/O). This limitation is not likely to change

in the near future since supercomputer development is closely linked to the performance of

commodity chips, which are not well-adapted to ocean applications (i.e., GPUs) and do not facilitate

memory access.

There is a wide range of applications for ocean models. Without data assimilation, they are

mostly used to scientifically rationalize the observed ocean by testing mechanisms underlying

observations via idealized or realistic configurations. With data assimilation, they are used to

perform hindcasts in order to understand past evolution and to perform forecasts that can be used

for societal applications (Chassignet and Verron, 2006; Dombrowsky et al., 2009; Schiller and

Brassington, 2011; Bell et al., 2015). Observational data via data assimilation sets the stage for

model state estimates and forecasts (Chassignet et al., 2009). The quality of the estimates and the

forecast will depend on the ability of an ocean numerical model to faithfully represent the resolved

dynamics of the ocean and the parameterized subgrid scale physics. It is therefore important to

realize that even using an infinite amount of data to constrain the initial conditions of an ocean

model will not necessarily improve the forecast when using a poorly performing ocean numerical

model.

It is also important to realize that while numerical models are necessary to understand ocean

dynamics, they cannot represent reality because of limitations in computational power, incomplete

understanding of subgrid scale parameterizations, poorly known forcing fields, and poorly

understood interactions with other components of the earth’s system such as the atmosphere and

sea ice. Observations are also not an accurate representation of reality because of the many space

and time gaps in the observations that give us only limited information about the ocean’s state, its

variability, trends, and possible instabilities and regime shifts.

OCEAN CIRCULATION MODELING FOR OPERATIONAL OCEANOGRAPHY: CURRENT

STATUS AND FUTURE CHALLENGES 293

In summary, while numerical models allow for hypothesis testing and experimentation, one

needs to understand their limitations in order to use them to their full potential. Good modelers are

aware of the strengths AND weaknesses of their models, and the main difficulty is the quantification

of the truncation errors introduced by the discretization of the Navier-Stokes equations.

Truncation errors arise from the discretization of PDEs derived from the Navier-Stokes

equations, which usually assume that the ocean is incompressible and make the spherical

approximation (assumes the earth is a sphere), the thin-shell approximation (allows to neglect

variations of the local rotation rate with respect to depth, therefore simplifying the treatment of the

Coriolis acceleration), the Boussinesq approximation (neglects variation of relative density in the

horizontal momentum equations), and the hydrostatic approximation (allows to neglect the vertical

acceleration).

The primitive Navier-Stokes equations are

with Uh is the horizontal velocity vector (u,v); U, the three-dimensional velocity (u,v,w); f, the

Coriolis parameter; p, the pressure; g, the gravity; , the density; T, the temperature; S, the salinity;

D, the dissipation; and F, the forcing. Temperature and salinity here implicitly refer to conservative

temperature and absolute salinity as defined in the Thermodynamic Equation Of Seawater - 2010

(TEOS-10; McDougall and Barker, 2011). The equation of seawater relates density to temperature,

salinity and pressure:

Several other equations are required for setting boundary conditions at the ocean surface and at the

seafloor. For instance, ocean circulation models enforce the following conditions at the ocean

surface:

- Kinematic boundary condition

294 JULIEN LE SOMMER ET AL.

- Surface pressure condition

- Air-sea fluxes of freshwater and heat

- Air-sea fluxes of momentum

where is the sea surface height, E, the evaporation; P, the precipitation; , the wind stress; K, the

diffusivity; A, the viscosity; and Q, the heat flux. Finally, one needs to use the turbulent closure

hypothesis to close the system by assuming that the nonlinear terms involving correlations of

variables at unresolved small scales can be fully prescribed from knowledge of the large scales.

RepresentationofPhysicalProcessesinOceanCirculationModels

In this section, we review some of the issues that arise as we strive to improve the representation of

key physical processes in the ocean model component of operational systems. This includes

broadening the spectrum of explicitly represented processes and improving the representation of

unresolved processes in ocean circulation models.

Resolvedversusunresolvedphysicalprocessesinoceancirculationmodels.

As discussed in the previous section, ocean circulation models do not account explicitly for the

entire range of scale interactions that control ocean circulation (Fig. 12.2). Because of the turbulent

nature of oceanic flows, ocean circulation at a given scale is indeed fundamentally dependent on

oceanic motions at scales ranging from global (of order 10,000 km) to dissipative (of order 1 cm;

Fig. 12.3). But the finite grid resolution of a particular ocean model configuration constrains the

spectrum of scales of motions that are explicitly represented in the model solution. Consider, for

instance, ocean mesoscale eddies, which are known to play a fundamental role in shaping ocean

circulation (McWilliams, 2008) and their representation in ocean models, which has motivated a

large number of studies. Even the most high-end global ocean circulation models with grid

resolutions down to just a few kilometers (Chassignet and Xu, 2017) are still not able to fully capture

the dominant length scales of mesoscale variability at high latitudes (Hallberg, 2013). Scale

interactions involving oceanic mesoscale eddies are therefore not explicitly represented at high

latitudes in these models (Fig. 12.4).

OCEAN CIRCULATION MODELING FOR OPERATIONAL OCEANOGRAPHY: CURRENT

STATUS AND FUTURE CHALLENGES 295

Figure 12.2. Range of ocean physical processes (adapted from Griffies and Treguier [2013]).

Figure 12.3. Space and time scale of ocean physical processes.

296 JULIEN LE SOMMER ET AL.

Therefore, a practical challenge for most applications of ocean circulation models is to describe

the broadest possible spectrum of oceanic scales of motion at a given computational cost. The

overall computational budget of an application is constrained by the available computing resources

and by the expected time-to-solution, the latter constraint being particularly limiting in operational

applications that deliver near real-time products. The model grid of ocean model components of

operational systems is also usually constrained by other practical factors. The same model grid is

often used for several years because of the effort required to tune the physical model configuration,

calibrate the data assimilation components, and prepare the downstream data production chains.

Given these constraints, the optimal design of ocean model components in operational systems

requires a robust and rational understanding of a priori what controls the spectrum of resolved

process in ocean circulation models besides the model grid resolution. Developing this

understanding of what controls the spectrum of resolved scales in ocean model should rely on

idealized process studies, model inter-comparison exercises and confrontation of model solutions

with observations.

A commonly accepted practical criteria for deciding whether a flow feature is resolved in an

ocean model is that there should be no less than 2 grid points spanning the feature (see, for instance,

Griffies and Treguier [2013]). This is supported by the notion of effective resolution, defined by

Skamarock (2004) as the smallest resolved scale that is not significantly affected by numerical

dissipation due to discretization errors. The effective resolution of an ocean circulation model is

about 6 to 10x, where x is the horizontal grid resolution, but the validity of the concept of

effective resolution is generally restricted to model configurations with small enough grid

resolutions and using high order discretization schemes (Soufflet et al., 2016).

Although effective resolution allows us to characterize the smallest scale of motion that could

ideally be described without numerical error in an ocean model, it is generally impossible to

distinguish unambiguously between resolved and unresolved physical processes in ocean

circulation models. Indeed, only the physical processes with dynamics that do not depend on scales

smaller than the effective resolution are unambiguously represented in a given ocean model

configuration. In practice, a lot of physical processes of interest involve scales between the model

grid scale and the model effective resolution. In such cases, although the model is a priori unable to

fully account for the dynamics, the model solutions may still exhibit some properties of the

misrepresented physical process. A good example of such a situation is discussed in a recent study

by Uchida et al. (2017), who found signatures of submesoscale mixed layer instabilities in their

1/10 resolution model although their model grid resolution is a priori not able to adequately capture

these dynamics.

Howtobroadenthespectrumofresolvedprocessesinocean

circulationmodels?

Using computationally efficient numerical discretization schemes with good mimetic properties is

a first approach for broadening the spectrum of resolved processes in ocean circulation models at a

given computational cost. The notion of computational efficiency refers to the accuracy of model

OCEAN CIRCULATION MODELING FOR OPERATIONAL OCEANOGRAPHY: CURRENT

STATUS AND FUTURE CHALLENGES 297

solution relative to the computational cost of numerical integration. The notion of mimetic

properties refers to the ability of a discrete model to emulate the properties of the underlying

continuous mathematical model, for instance as conservation laws and geometrical symmetries.

Ocean circulation models with grid resolutions approaching 1 km can be sensitive to numerical

discretization schemes (Ducousso et al., 2018). This is because ocean model solutions are generally

more nonlinear and less controlled by diffusion in this range of resolution. The current trend in

ocean circulation modeling is, therefore, to use high order discretization schemes in order to

increase model effective resolution (Soufflet et al., 2016).

Multiscale modeling approaches coming from coastal ocean modeling also provide a framework

for locally increasing the effective resolution according to a particular scientific or operational

objective and at a given computational cost. Multiscale modeling methods, which allow us to locally

refine the grid mesh, are now commonly used in large-scale ocean circulation modeling.

Applications of multiscale approaches in ocean circulation modeling can be classified according to

their grid mesh topology into structured mesh methods (for instance as orthogonal curvilinear grids

used by most ocean circulation models), unstructured mesh methods (e.g., see Piggott et al., 2008;

Danilov et al., 2017; Ringler et al., 2013) and block structured mesh methods (see Blayo and

Debreu, 1999; Jablonowski et al., 2006). Although many research efforts have focused on

unstructured mesh ocean circulation modeling over recent years, the design of robust discretization

schemes for large-scale ocean models on triangular or hexagonal meshes remains challenging

(Danilov, 2013). And it should also be noted that it is still unclear how to define objective criteria

for adjusting meshes and, therefore, optimally use the potential for mesh refinement and mesh

adaptability of multiscale modeling approaches in large-scale ocean circulation models.

Improving the performance of ocean circulation models on modern high performance

computing (HPC) platforms is also essential for broadening the spectrum of resolved scales. The

current trend in HPC is toward massively parallel machines with heterogeneous multicore

architectures (Giles and Reguly, 2014). However, while modern HPC platforms can deliver a peak

performance in the Petaflop/s range, existing ocean circulation models are unable to exploit this

potential. The computational intensity, i.e., floating point operations per memory access, of ocean

models is very low because they are dominated by stencil operations (typical of discretized PDEs).

So, they typically run at ~5% of the system’s peak speed. Also, ocean models have a very small

vertical dimension, typically O(10), so they scale more like 2-D domains than 3-D domains. In

practice, 1/12o and 1/25o global ocean models might scale well to 8,000 and 30,000 cores,

respectively. Scalability is eventually limited by the communication overhead, load imbalance, and

latency inherent to spatial 2-D domain decomposition and by I/O overhead. Single processor

performance depends on computational intensity, which can be improved by using higher order

discretization schemes; but scalability can actually be reduced by better single processor

performance because relatively more time is spent in communications.

In the future, overcoming single processor and scalability bottlenecks will require sustained

collaborations between ocean modelers and computer scientists. Practical approaches will likely

involve an increase in hybrid parallel programming in order to more efficiently exploit memory

298 JULIEN LE SOMMER ET AL.

hierarchy and innovative algorithms for solving the set of PDEs that govern ocean dynamics, as for

instance parallelization in time in addition to spatial domain decomposition (Schreiber et al., 2017).

Looking ahead, because of the multiscale nature of oceanic flows and the needs of end-users of

operational oceanography, it is necessary for ocean circulation models to account explicitly for the

broadest range of physical scales possible. But this broadening of the spectrum of resolved scales

will not simply be a consequence of an increase in computing power. How efficiently ocean

circulation models will use computing resources will depend on the efficiency of the algorithms

used for translating ocean dynamical equations into practical computation. And foreseeing what

approach will be most used in the future for broadening the spectrum of resolved scales is not

straightforward, but it is arguable that HPC and algorithmic aspects will be crucial in future ocean

model development. New advances will most likely rely on a high level of collaboration between

ocean modelers, applied mathematicians, and computer scientists.

Figure 12.4. Grid resolution required to resolve the first baroclinic radius with 10 grid points. The figure shows

the local grid resolution (in kilometers) divided by 10 as an estimate the grid resolution required for a model

to effectively capture structures with wavelength close to the first baroclinic Rossby radius. The first baroclinic

Rossby radius is estimated from a CMEMS 1/12° global reanalysis. The estimate of the number of points

required for the effective resolution to match the first baroclinic Rossby radius is slightly more restrictive than

that proposed by Hallberg (2013) following the work by Soufflet et al. (2016).

Howtobetteraccountforunresolvedprocessesinoceancirculationmodels

Ocean circulation models need to allow for the important physical processes that are not accounted

for in their discrete approximation to the ocean dynamical equations. A common approach to the

OCEAN CIRCULATION MODELING FOR OPERATIONAL OCEANOGRAPHY: CURRENT

STATUS AND FUTURE CHALLENGES 299

representation of unresolved processes in ocean circulation models is to use a subgrid scale (SGS)

closure (aka physical parameterization). Typically, these do not focus on particular nonlinear terms

in the primitive equations, but rather on particular physical processes (that may affect several terms).

A rough classification of unresolved physical processes would include: (i) SGS closures for

balanced turbulence (eddy closures), (ii) SGS closures for processes in the ocean surface boundary

layer, (iii) SGS closures for interior mixing processes, and (iv) SGS closures for bottom boundary

layer processes. The design of SGS closures associated with physical processes usually involves

some combination of idealized process studies with models, realistic process studies with models,

targeted field experiments at sea, and (more rarely) lab experiments. Importantly, a SGS closure

should always target a particular ocean circulation model resolution because the information at the

resolved scale depends on the model resolution.

Depending upon the available information from the scales assumed to be resolved in the target

model, two categories of closures can be distinguished. The first is a RANS-type (cf Reynolds

average) closure, implicitly based on some sort of ensemble average (in practice emulated through

time average); for example, the Gent-McWilliams closure for mesoscale eddy transport (Gent,

2011). The other is a LES type closure (by analogy to Large Eddy Simulation) implicitly based on

some spatial average; for example, Leith’s (1968) momentum closure. LES-type subgrid closures

can further be subdivided into functional and structural closures. Functional closures consider the

bulk action of the subgrid terms on the resolved scales (e.g., Smagorinsly momentum closures

[Griffies and Hallberg, 2000]). Structural closures aim at estimating the best local approximation

of the unknown SGS term by constructing it from the known small-scale stuctures (e.g., closures

defined by Taylor series expansion). A recent trend in the design of SGS closures in ocean

circulation modeling deals with how to adapt them to the resolved physical scales of motions in the

model solution (Bachman et al., 2017). There is also a growing concern in ocean models design

regarding the complex interplay between numerical discretization errors and SGS closures,

especially in the finest resolution applications (Campin et al., 2011; Lemarie et al., 2012). Another

emerging approach to the representation of unresolved processes in ocean circulation models is to

couple ocean circulation models with third-party code components, which are specifically designed

to describe the evolution of a particular class of processes. Physical processes that are not explicitly

accounted for in ocean models have been extensively studied (e.g., wave modelling) and estimates

of the subgrid states associated with these unresolved processes must take into account non-local

effects in space and time (e.g., propagation and dissipation of internal waves, cf. IDEMIX [Olbers

and Eden, 2013]). It may also make sense to modularize, as much as possible, a complex closure so

that it can be used in different ocean models and thus become an independent model library or

component (e.g., https://github.com/CVMix). How to fully integrate multiple code components,

each describing different aspect of the dynamics, in operational systems with data assimilation is

likely to become a key research question in the future. A particularly important concern is associated

with the representation of the fine-scale feedbacks between the atmospheric and oceanic boundary

layers.

300 JULIEN LE SOMMER ET AL.

TowardsDatadriven,ProbabilisticOceanCirculationModels

Stochasticparameterizationsinoceancirculationmodels

Classical approaches in the design of SGS closures for ocean circulation models assumes that the

bulk effect of SGS fluctuations can be deterministically predicted from the resolved scales of

motions. A number of recent works have started to reconsider this assumption. These more recent

studies allow the closure to be only weakly constrained by the resolved scales. Carefully designed

stochastic parameterizations appear to be a promising approach to the representation of a large class

of unresolved SGS processes in ocean circulation models (Berloff, 2005; Palmer and Williams,

2010; Williams, 2012; Grooms and Majda, 2013; Porta Mana and Zanna, 2014; Cooper and Zanna,

2015; Brankart et al., 2015; Andrejczuk et al., 2016). A certain amount of randomness is introduced

with a stochastic procedure to account for the possible variability of fluxes at macroscales.

Stochastic parameterizations seem particularly well-suited to the representation of the cross-scale

exchanges of energy and momentum for SGS-balanced turbulence (in particular, energy

backscatter). Notable improvement of Gulf stream dynamics has also been obtained through the

stochastic representation of upscaling due to the nonlinear nature of the equation of the state of sea-

water (Brankart, 2013). More generally, the concept of introducing random perturbations in ocean

models is also supported by another rationale coming from the theory of non-linear systems. The

topology of the phase space of non-linear systems can indeed be rather complicated, showing

multiple possible local energy minima. Adding perturbations could allow a system to escape from

local potential wells and, therefore, explore more reliably the range of possible states.

The paradigm underpinning this idea of stochastic parameterization is actually not restricted to

the design of SGS closures and could also allow us to account for more general sources of

uncertainty in ocean circulation models. These include errors in their initial condition (particularly

at depth), errors in the forcing function (uncertain atmospheric fields, uncertain parameter in bulk

formula. etcetera), errors associated with subgrid closures (closures are imperfect models, with

usually weakly constrained parameters), errors from the physical approximation use in for their

continuous formulation (e.g., non-traditional Coriolis terms), and numerical discretization errors

that tend to accumulate over time (even high order schemes have non-zero errors). Overall,

uncertainty is a major property of ocean circulation models that physicists often tend to neglect,

although some modeling frameworks are now tackling this problem more explicitly (e.g., see

Brankart et al., 2015). At this point, it is arguable that further developing stochastic

parameterizations for ocean circulation models will require a better understanding of their impact

on the resolved scales of motion.

Towardprobabilisticoceancirculationmodelingthroughensemblesimulations

Ensemble modeling is a now common approach in operational forecast and climate modeling for

accounting for the inherent uncertainty of geoscientific model solutions. Ensembles are, for

instance, routinely used in operational systems to account for uncertainty in initial states, model

OCEAN CIRCULATION MODELING FOR OPERATIONAL OCEANOGRAPHY: CURRENT

STATUS AND FUTURE CHALLENGES 301

formulation, or forcing. Because they sample the space of possible states given such uncertainty,

ensemble approaches allow for a more objective comparison of model solutions with observations,

which is key to operational systems. More recently, ensemble modeling has also become an

experimental tool of choice for investigating the sources of oceanic variability in eddying regimes

(Sérazin et al., 2017). It should be stressed that ensemble modelling is a reasonable strategy in terms

of high performance computing, because in most cases ensemble runs are entirely independent of

each other and therefore 100% scalable.

The combination of ensemble and stochastic methods allows us to deal more objectively with

uncertainty in ocean models (Toth et al., 2003; Palmer, 2012; Brunton et al., 2016; Bessieres et al.,

2017). The purpose of ongoing efforts is to attempt to explicitly sample the probability distribution

of possible states given explicitly formulated uncertainties following a Bayesian approach. This

shift towards probabilistic ocean circulation modeling is arguably a paradigm transition in our field.

But it is still unclear how to deal with the daemon of dimensionality. Indeed, ensemble sizes are

usually constrained to several tenth of members for practical reasons while the parameter space they

are sampling is orders of magnitude larger. The emergence of probabilistic ocean circulation

modeling could bring several benefits to ocean model design and usage in the future. But

probabilistic ocean modeling also raises several technical and scientific challenges that ought to be

addressed in the future.

FinalRemarks

In this chapter, we have briefly described the fundamental principles that underpin the formulation

of modern ocean circulation models, shown some of their recent achievements, and discussed what

we believe are the future frontiers in ocean circulation model development. This chapter has, in

particular, illustrated how a modeling strategy proposed in the 1960s for solving the primitive

equations has yielded ocean circulation models that are now used for a wide range of applications

and that form a building block of modern operational oceanography. We have also presented the

view that future frontiers in ocean circulation modeling will depend upon (i) the computational

performance of ocean models, (ii) their ability to represent scale interactions either explicitly or

though parameterizations, and (iii) the representation of model uncertainty with stochastic and

ensemble approaches.

This chapter illustrates how the field of ocean circulation model design has reached its maturity

and now involves strong collaboration between different fields of expertise. Ocean circulation

model design is a very active field of research with entire scientific teams dedicated to developing

or improving ocean circulation models. The maturity of the field is arguably a consequence of the

high level of collaboration and merging of efforts among different groups involved in different

applications and aspects of ocean circulation models. A key driver for this collaborative approach

to ocean model development is this notion of seamless geoscientific modeling, which suggests that

the same numerical code can actually be used for a range of different applications covering a range

of different resolutions and dynamical regimes (Hurrell et al., 2009). Overall, this approach has

302 JULIEN LE SOMMER ET AL.

improved the robustness of ocean circulation models and the sustainability of the ocean model

development process over recent decades.

Although not been discussed much in this chapter, we would like to stress the importance of

ocean observing networks for improving ocean circulation models. Sustained ocean observations

from satellites and in situ networks are critical for routinely assessing the skills and limitations of

circulation ocean models over different timescales ranging from days to decades. There is also mu ch

to be learned from targeted field observations aimed at documenting specific oceanic processes in

order to improve their representation in models. For instance, recent experiments documenting fine-

scale ocean processes in the ocean surface boundary layer provided a wealth of information that can

be used for improving surface processes in ocean circulation models (Shcherbina et al., 2015;

Buckingham et al., 2016).

An aspect that we believe has also been critical in the continuous improvement of ocean

circulation models over recent decades is the shared vision in the ocean modeling community that

open source is the only sustainable approach to geoscientific model development. Ocean model

developers have generally been early adopters of modern practices in software development (e.g.,

version control, unit testing, continuous integration). All the major ocean circulation codes are also

distributed under open source licenses and therefore exposed to the scrutiny of other research

groups. Arguably, workflows in ocean modeling could be improved and made more reproducible,

but it is fair to recognize that this community has long been concerned with these issues. Building

more open and transparent data processing chains for pre- and post-processing is probably an

important next step to making ocean circulation modeling more robust and reproducible in the future

(Stodden et al., 2016).

In conclusion, we would like to raise what is likely to become a key issue in the future

development of ocean circulation models. With a broadening of scope and an increase in the number

of users, ocean circulation models have also grown in complexity. But the actual size of the

community of ocean model developers is still rather small. Furthermore, most of the research

questions that have been raised in this chapter would require strong and sometimes new

collaborations among ocean modelers, field oceanographers, process-oriented oceanographers,

applied mathematicians, and computer scientists. Sustained and proactive initiatives from major

funding agencies (one good example of such an initiative being the US CLIVAR Climate Process

Team scheme) will certainly be vital to the success of these interdisciplinary collaborations and to

bringing more early career scientists in to contribute to the science of ocean circulation models.

Acknowledgments

EPC and AJW are supported by the Office of Naval Research and the Naval Research Laboratory.

The authors also thank R. Bourdallé-Badie for providing Fig. 12.4.

OCEAN CIRCULATION MODELING FOR OPERATIONAL OCEANOGRAPHY: CURRENT

STATUS AND FUTURE CHALLENGES 303

References

Andrejczuk, M., F.C. Cooper, S. Juricke, T.N. Palmer, A. Weisheimer, and L. Zanna, 2016: Oceanic stochastic

parameterizations in a seasonal forecast system. Monthly Weather Review, 144 (5), 1867–1875.

Bachman, S.D., B. Fox-Kemper, and B. Pearson, 2017: A scale-aware subgrid model for quasi- geostrophic

turbulence. Journal of Geophysical Research: Oceans, 122 (2), 1529–1554, doi: 10.1002/2016JC012265.

Bell M.J., A. Schiller, P.-Y. Le Traon, N.R. Smith, E. Dombrowsky, and K. Wilmer-Becker, 2015: An

introduction to GODAE OceanView. Journal of Operational Oceanography, 8:sup1, s2-s11, doi:

10.1080/1755876X.2015.1022041

Berloff, P. S., 2005: Random-forcing model of the mesoscale oceanic eddies. Journal of Fluid Mechanics,

529, 71–95, doi:10.1017/S0022112005003393.

Bessieres, L., and coauthors, 2017: Development of a probabilistic ocean modelling system based on NEMO

3.5: Application at eddying resolution. Geoscientific Model Development, 10 (3), 1091–1106,

doi:10.5194/gmd-10-1091-2017.

Blayo, E., and L. Debreu, 1999: Adaptive Mesh Refinement for Finite-Difference Ocean Models: First

Experiments. Journal of Physical Oceanography, 29 (6), 1239–1250, doi: 10.1175/1520-0485(1999).

Brankart, J.-M., 2013: Impact of uncertainties in the horizontal density gradient upon low resolution global

ocean modelling. Ocean Modelling, 66, 64–76, doi:10.1016/j.ocemod.2013.02.004.

Brankart, J.-M., G. Candille, F. Garnier, C. Calone, A. Melet, P.-A. Bouttier, P. Brasseur, and J. Verron, 2015:

A generic approach to explicit simulation of uncertainty in the NEMO ocean model. Geoscientific Model

Development, 8 (5), 1285–1297, doi:10.5194/gmd-8-1285-2015.

Brunton, S.L., J.L. Proctor, and J.N. Kutz, 2016: Discovering governing equations from data by sparse

identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113 (15),

3932–3937, doi:10.1073/pnas.1517384113.

Bryan, K., 1969: A numerical method for the study of the circulation of the world ocean. Journal of

Computational Physics, 4 (3), 347–376, doi:10.1016/0021-9991.

Buckingham, C.E., A. C. Naveira Garabato, A. F. Thompson, L. Brannigan, A. Lazar, D. P. Marshall, A. J.

George Nurser, G. Damerell, K. J. Heywood, and S. E. Belcher, 2016: Seasonality of submesoscale flows

in the ocean surface boundary layer. Geophysical Research Letters, 43, 2118-2126,

doi:10.1002/2016GL068009.

Campin, J.-M., C. Hill, H. Jones, and J. Marshall, 2011: Super-parameterization in ocean model- ing:

Application to deep convection. Ocean Modelling, 36 (1-2), 90–101, doi:10.1016/j.ocemod. 2010.10.003.

Charney, J., A. Fjortoft, and J. von Neuman, 1950: Numerical integration of the barotropic vorticity equation.

Tellus, 2, 237-254, doi:10.1111/j.2153-3490.1950.tb00336.x.

Chassignet, E.P., and J. Verron (Eds.), 1998. Ocean Modeling and Parameterization. Kluwer Academic

Publishers, 451 pp.

Chassignet, E.P., and J. Verron (Eds.), 2006: Ocean Weather Forecasting: An Integrated View of

Oceanography. Springer, 577 pp.

Chassignet, E.P., H.E. Hurlburt, E.J. Metzger, O.M. Smedstad, J. Cummings, G.R. Halliwell, R. Bleck, R.

Baraille, A.J. Wallcraft, C. Lozano, H.L. Tolman, A. Srinivasan, S. Hankin, P. Cornillon, R. Weisberg, A.

Barth, R. He, F. Werner, and J. Wilkin, 2009: U.S. GODAE: Global Ocean Prediction with the HYbrid

Coordinate Ocean Model (HYCOM). Oceanography, 22 (2), 64-75.

Chassignet, E.P. and X. Xu, 2017: Impact of horizontal resolution (1/12° to 1/50°) on Gulf Stream separation,

penetration, and variability. Journal of Physical Oceanography, 47, 1999-2021, doi:10.1175/JPO-D-17-

0031.1.

Cooper, F. C., and L. Zanna, 2015: Optimisation of an idealised ocean model, stochastic parameterisation of

sub-grid eddies. Ocean Modelling, 88, 38–53, doi:10.1016/j.ocemod.2014.12.014.

Danilov, S., 2013: Ocean modeling on unstructured meshes. Ocean Modelling, 69, 195– 210,

doi:10.1016/j.ocemod.2013.05.005.

Danilov, S., D. Sidorenko, Q. Wang, and T. Jung, 2017: The Finite volumE Sea-ice Ocean Model (FESOM2).

Geoscientific Model Development, 10 (2), 765–789, doi:10.5194/gmd-10-765-2017.

Dombrowsky, E., L. Bertino, G.B. Brassington, E.P. Chassignet, F. Davidson H.E. Hurlburt, M. Kamachi, T.

Lee, M.J. Martin, S. Mei, and M. Tonani, 2009: GODAE systems in operation. Oceanography, 22, 80–95.

Ducousso, N., J. Le Sommer, J.-M. Molines, and M. Bell, 2018: Impact of the Symmetric Instability of the

Computational Kind on oceanic hindcasts at mesoscale and submesoscale permitting resolutions. Ocean

Modelling, submitted.

Gent, P.R, 2011: The Gent–McWilliams parameterization: 20/20 hindsight. Ocean Modelling, 39, 2-9.

doi:10.1016/j.ocemod.2010.08.002.

304 JULIEN LE SOMMER ET AL.

Giles, M.B., and I. Reguly, 2014: Trends in high-performance computing for engineering calculations.

Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 372

(2022), 20130 319–20130 319, doi:10.1098/rsta.2013.0319.

Griffies, S.M., C. Boening, F.O. Bryan, E.P. Chassignet, R. Gerdes, H. Hasumi, A. Hirst, A.-M. Treguier, and

D. Webb, 2000: Developments in Ocean Climate Models, Ocean Modelling, 2, 123-192.

Griffies, S.M., and R.W. Halberg, 2000: Biharmonic friction with a smagorinsky-like viscosity for use in large-

scale eddy-permitting ocean models. Mon. Weather Rev., 128, 2935–2946.

Griffies, S., 2004: Fundamentals of Ocean Climate Models. Princeton University Press.

Griffies, S.M., and A.M. Treguier, 2013: Ocean Circulation Models and Modeling. International Geophysics,

Vol. 103, Elsevier, 521–551.

Grooms, I., and A.J. Majda, 2013: Efficient stochastic superparameterization for geophysical turbulence.

Proceedings of the National Academy of Sciences, 110 (12), 4464–4469, doi: 10.1073/pnas.1302548110.

Hallberg, R., 2013: Using a resolution function to regulate parameterizations of oceanic mesoscale eddy

effects. Ocean Modelling, 72, 92–103, doi:10.1016/j.ocemod.2013.08.007.

Hurrell, J., G. A. Meehl, D. Bader, T. L. Delworth, B. Kirtman, and B. Wielicki, 2009: A Unified Modeling

Approach to Climate System Prediction. Bulletin of the American Meteorological Society, 90 (12), 1819–

1832, doi:10.1175/2009BAMS2752.1.

Jablonowski, C., M. Herzog, J. E. Penner, R. C. Oehmke, Q. F. Stout, B. Van Leer, and K. G. Powell, 2006:

Block-structured adaptive grids on the sphere: Advection experiments. Monthly weather review, 134 (12),

3691–3713.

Leith, C.E., 1968: Diffusion approximation for two-dimensional turbulence. Physics Fluids, 10, 1409–1416.

Lemarie, F., L. Debreu, A. Shchepetkin, and J. McWilliams, 2012: On the stability and accuracy of the

harmonic and biharmonic isoneutral mixing operators in ocean models. Ocean Modelling, 52-53, 9–35,

doi:10.1016/j.ocemod.2012.04.007.

McDougall, T. J. and P. M. Barker, 2011: Getting started with TEOS-10 and the Gibbs Seawater (GSW)

Oceanographic Toolbox, 28pp., SCOR/IAPSO WG127, ISBN 978-0-646-55621-5.

McWilliams, J.C., 1996: Modeling the oceanic general circulation. Annual Review of Fluid Mechanics, 28 (1),

215–248.

McWilliams, J.C., 2008: The nature and consequences of oceanic eddies. Geophysical Monograph Series, 177,

5–15, doi:10.1029/177GM03.

Olbers, D. and C. Eden, 2013: A global model for the diapycnal diffusivity induced by internal gravity waves.

Journal of Physical Oceanography, 43, 1759–1779, doi:10.1175/JPO-D-12-0207.1.

Palmer, T., and P.D. Williams, Eds., 2010: Stochastic physics and climate modelling. Cambridge University

Press, Cambridge, UK; New York.

Palmer, T. N., 2012: Towards the probabilistic Earth-system simulator: a vision for the future of cli- mate and

weather prediction. Quarterly Journal of the Royal Meteorological Society, 138 (665), 841–861,

doi:10.1002/qj.1923.

Piggott, M.D., G.J. Gorman, C.C. Pain, P.A. Allison, A.S. Candy, B.T. Martin, and M.R. Wells, 2008: A new

computational framework for multi-scale ocean modelling based on adapting unstructured meshes.

International Journal for Numerical Methods in Fluids, 56 (8), 1003–1015, doi:10.1002/fld.1663.

Porta Mana, P., and L. Zanna, 2014: Toward a stochastic parameterization of ocean mesoscale eddies. Ocean

Modelling 79, 1–20.

Richardson, L. F.,1922: Weather Prediction by Numerical Process. Cambridge University Press. (Reprinted

by Dover Publications, New York, 1965, with a new introduction by S. Chapman, xvi+236; 2d Edn. by

Cambridge University Press, 2007, with a new introduction by P. Lynch.).

Ringler, T.D., M. Petersen, R. Higdon, D. Jacobsen, P. Jones, and M. Maltrud, 2013: A multiresolution

approach to global ocean modeling. Ocean Modelling, 69, 211–232, doi:10.1016/j.ocemod.2013.04.010.

Schiller, A. and G.B. Brassington (Eds), 2011: Operational oceanography in the 21st century. Springer, ISBN

9789400703315.

Schreiber, M., S.P. Peixoto, T. Haut, and B. Wingate, 2017: Beyond spatial scalability limitations with a

massively parallel method for linear oscillatory problems. International Journal of High Performance

Computing Applications. doi:10.1177/1094342016687625.

Sérazin, G., A. Jaymond, S. Leroux, T. Penduff, L. Bessières, W. Llovel, B. Barnier, J.M. Molines, and L.

Terray: 2017. A global probabilistic study of the ocean heat content lowfrequency variability:

Atmospheric forcing versus oceanic chaos, Geophysical Research Letters, 44, 5580–5589,

doi:10.1002/2017GL073026.

Shcherbina, A.Y., M.A. Sundermeyer, E. Kunze, E.A. D’Asaro, G. Badin, D. Birch, A.-M. E.G. Brunner-

Suzuki, J. Callies, B.T. Kuebel Cervantes, M. Claret, B. Concannon, J. Early, R. Ferrari, L. Goodman,

R.R. Harcourt, J.M. Klymak, C.M. Lee, M.-P. Lelong, M.D. Levine, R.-C. Lien, A. Mahadevan, J.C.

OCEAN CIRCULATION MODELING FOR OPERATIONAL OCEANOGRAPHY: CURRENT

STATUS AND FUTURE CHALLENGES 305

McWilliams, M.J. Molemaker, S. Mukherjee, J.D. Nash, T. Özgökmen, S.D. Pierce, S. Ramachandran,

R.M. Samelson, T.B. Sanford, R.K. Shearman, E.D. Skyllingstad, K. Schafer Smith, A. Tandon, J.R.

Taylor, E.A. Terray, L.N. Thomas and J.R. Ledwell, 2015: The LatMix summer campaign: Submesoscale

stirring in the upper ocean. Bulletin American Meteorological Society, 1257-1279.

Shriver, J.F., B.K. Arbic, J.G. Richman, R.D. Ray, E.J. Metzger, A.J. Wallcraft, and P. G. Timko, 2012: An

evaluation of the barotropic and internal tides in a high-resolution global ocean circulation model. Journal

of Geophysical Research, 117, C10024, doi:10.1029/2012JC008170

Skamarock, W.C., 2004: Evaluating mesoscale NWP models using kinetic energy spectra. Monthly Weather

Review, 132, 3019–3032, doi:10.1175/MWR2830.1.

Smyth, W.D., J.N. Moum, and D.R. Caldwell, 2001: The efficiency of mixing in turbulent patches: Inferences

from direct simulations and microstructure observations, Journal of Physical Oceanography, 31, 1969–

1992, doi:10.1175/1520-0485(2001)031<1969:TEOMIT>2.0.CO;2.

Soufflet, Y., P. Marchesiello, F. Lemarie, J. Jouanno, X. Capet, L. Debreu, and R. Benshila, 2016: On effective

resolution in ocean models. Ocean Modelling, 98, 36–50.

Stodden V, M. McNut, D.H. Bailey, et al, 2016: Enhancing reproducibility for computational methods.

Science, 354:1240-1. doi:10.1126/science.aah6168.pmid:27940837

Toth, Z., O. Talagrand, G. Candille, and Y. Zhu, 2003: Probability and ensemble forecasts, in Forecast

Verification: a Practitioner’s Guide in Atmospheric Science, edited by: Jolliffe, I. and Stephenson, D. B.,

Wiley, UK, 137–163, 2003.

Uchida, T., R. Abernathey and K.S. Smith, 2017: Seasonality in ocean mesoscale turbulence in a high

resolution global climate model. Ocean Modelling, 118, 41-58.

Williams, P.D., 2012: Climatic impacts of stochastic fluctuations in air–sea fluxes. Geophysical Research.

Letters, 39, L10705, doi:10.1029/2012GL051813.

Yeung, P.K., Zhai, X.M., and Sreenivasan, K.R., 2015: Extreme events in computational turbulence.

Proceedings of the National Academy of Sciences of the United States of America, 112(41), 12633–12638,

doi:10.1073/pnas.1517368112.