LSCE/CEA Saclay, CEA-CNRS, Bat. 709 - Orme des Merisiers, F-91191 Gif-sur-Yvette, France

We have made preliminary simulations in a three-dimensional (3-D) ocean model as a means to estimate the time required for CO₂ injected into the deep ocean to be lost back to the atmosphere. Injections of 0.1 Pg C yr-1 were made at seven sites simultaneously. Deep-ocean injection was carried out for 100 years from the present time under postindustrial conditions. After injection, fossil emissions were continued for another 700 years. Separate simulations were made for permanent sequestration (i.e., storage without return to the atmosphere), injection at 1500 m, and injection at 3000 m. Two hundred years after the start of injection, permanent sequestration is only 80% efficient. The reason is that permanent sequestration results in lower atmospheric CO₂, relative to the control run, and thus a lower air-sea CO₂ flux. Relative to permanent sequestration, the 1500-m injection is 82% as efficient and the 3000-m injection is 98% as efficient, after 200 years. The capacity of individual sites to retain injected CO₂ varied greatly for injection at 1500 m but not at 3000 m. For the 1500 m case, our model predicts that injection just offshore of Tokyo is much more efficient than injection offshore of New York. That result is just the opposite of what has been found in previous work with another 3-D ocean model. Comparison of observed and modeled ¹⁴C suggests that for Tokyo, the estimates from the two models may bracket real ocean behavior. Further model validation and comparison will help discern the key regions and processes where model improvements are most necessary.

To help reduce the rate of increase of atmospheric CO₂, one could in principle divert CO₂ emissions from coastal power plants to the deep ocean and thereby short-circuit the naturally slow air-sea exchange of CO₂ (Marchetti, 1977). However two key scientific questions remain to be answered before such a practice could ever be implemented. Accurate answers to these questions are needed in order to fully address other crucial concerns—socioeconomic, political, legal, and technological—all of which will play a role in future decisions that governments must make concerning possible mitigation strategies. The first of these scientific questions is, "how extensively would deep-ocean injection affect the marine biota in the vicinity of the injection site?" The second question is, "if fossil CO₂ is to be injected into the ocean, how long might it stay isolated from the atmosphere?"

We focus on the second question, using the only means available, simulations in a global-scale ocean model. Previous studies have addressed this same question, first with relatively simplistic box models (Hoffert et al., 1979) and then with an ocean general circulation model (OGCM) from the Max Planck Institut für Meteorologie in Hamburg (MPI) (Bacastow and Stegen, 1991; Bacastow et al., 1995; Bacastow and Dewey, 1996). A good summary of previous work is provided by Dewey et al. (1996). OGCM's describe ocean circulation in three dimensions. Thus they can be used to make assessments concerning the dependence of injection site location and depth upon the effectiveness of the ocean to retain sequestered CO₂. Unfortunately, model estimates can be biased due to model simplifications, omissions, and uncertainties. Substantial effort is required in regards to model validation and comparison before one can lend confidence to predicted results.

Here we describe the first steps towards comparing simulations in two different OGCM's. For that, we have made CO₂ injection simulations in a recently developed OGCM, from the LODyC (Laboratoire d’Océanographie Dynamique et de Climatologie) in Paris. In this report, we have two objectives: (1) to begin to estimate the effect of model uncertainties on estimates of the long term retention of injected CO₂, and (2) to test methodologies that are planned to be used in a more detailed model comparison, involving at least seven different 3-D ocean models of diverse origins.

Planning for the more involved comparison is underway in the framework of a project known as the GOSAC (Global Ocean Storage of Anthropogenic Carbon), an international effort during 1998 to 2000, which is jointly funded by the EC Environment and Climate Programme and the IEA Greenhouse Gas R&D Programme. GOSAC includes efforts to compare model estimates of ocean sequestration of CO₂. GOSAC also includes more fundamental research objectives to study the ocean's carbon-cycle and to validate models with available tracer measurements. These fundamental objectives fit within the framework of the Ocean Carbon-Cycle Model Intercomparison Project (OCMIP, see http://www.ipsl.jussieu.fr/OCMIP/), a project of the Global, Anlysis, Interpretation, and Modeling Task Force (GAIM) of the International Geosphere-Biosphere Programme (IGBP).

The GOSAC comparison will provide uncertainty estimates as an envelope of model predictions for a given ocean injection scenario. The diversity of ocean models involved in GOSAC offers greater confidence that model predictions will bracket real ocean behavior; however, model bias cannot be excluded. To help locate model biases and assess uncertainties, GOSAC will compare measured distributions of relevant tracers, including ¹⁴C and CFC's, to corresponding model simulations.

For this study, we have used a 3-D model from IPSL (Institute Pierre Simon Laplace) which combines carbonate chemistry and ocean circulation. For circulation, we use a 3-D tracer transport model (offline) which is driven by dynamic fields of advection and eddy diffusion determined from a global version of the OGCM known as OPA (Océan Parallelisé). OPA is developed at LODyC in Paris (Madec and Imbard, 1996). OPA uses a C grid (Arakawa, 1972), resolved vertically by 30 layers (varying in thickness from 10 m at the surface to about 500 m at depth). Horizontal resolution averages 1.5^o x 2.0^o. The grid is contorted in the northern hemisphere so that its northern singularity is pushed over Asia, thereby allowing larger grid spacing in the Arctic Ocean, and thus longer time steps.

The OGCM employs surface boundary conditions which combine heat and fresh water fluxes with restoring towards observed monthly temperature T and seasonal salinity S (Levitus, 1982), using a restoring time constant that depends on mixed layer depth (48 days^-1 for a 40 m mixed layer). The OGCM also restores to observed T and S in the interior ocean, when below the surface mixed layer. That forcing normally has a restoring time of constant of 1 year; however, it is gradually relaxed when near the equator, in the high latitudes, and near any land-ocean boundary. Such forcing is termed semi-diagnostic. A special feature of OPA, important for CO₂ sequestration simulations, is that vertical turbulence is determined via a prognostic model of turbulent kinetic energy. Horizontally, eddy diffusion of tracers is defined a priori, with an explicit diffusion coefficient of 2000 m² s^-1. Passive tracers are advected offline with the MPDATA scheme using two iterations to correct for numerical diffusion (Marti, 1992). The relatively high resolution and the choice of the advection scheme make simulations computer intensive.

To speed simulations, our offline model runs were accelerated using degradation D2 of DEGINT (Aumont et al., 1998a). With an average resolution 6^o x 8^o, D2 is a smoothed version of the original offline model. Thus our results are preliminary. To further conserve computing resources, we have made several simplifications. First, our offline simulations were made without marine biota. To test the importance of that simplification, we made abiotic vs. biotic simulations for the case of permanent sequestration. Including ocean biota is much more costly, but changes results only slightly (<5%), in confirmation of previous work by Bacastow and Dewey (1996). Secondly, we transported only DIC within the ocean. Thus we neglected potential changes in alkalinity, based on the work of Archer et al. (1998) which suggests that injection of CO₂ would not cause substantial dissolution of CaCO₃ sediments for at least several centuries.

For air-sea gas exchange, we used the standard boundary conditions specified for the first phase of OCMIP: (1) the gas transfer coefficient from Wanninkhof's (1992) quadratic function of wind speed, with chemical enhancement; (2) wind speeds from remotely sensed SSMI data (Boutin and Etcheto, 1998); and (3) fractional sea ice cover, which inhibits gas exchange (Walsh, 1978; Zwally et al., 1983). Also following OCMIP guidelines, carbonate chemistry equilibria constants and CO₂ solubility were as given by DOE (1994).

Before adding the anthropogenic perturbations, we allowed our model to come to steady-state (global air-sea CO₂ flux < 0.05 Pg C yr^-1) with atmospheric pCO₂ at 278 ppm (Aumont et al., 1998b). We define that state as pre-industrial and set the year to be 1800. Subsequently, emissions from Bacastow and Dewey’s (1996) logistics function (Fig. 1) were injected into our model's 1-box atmosphere, whose mass is defined by the constant of 2.123 Pg C per ppm of atmospheric CO₂. Our 1-box atmosphere also exchanges CO₂ with all surface ocean grid boxes.

From the pre-industrial state (1800), we submitted the model to 200 years of anthropogenic emissions according to Bacastow and Dewey (Fig. 1). Then at year 2000, we initialized four different simulations: (1) Control run¾ same emissions scenario continued; (2) Permanent sequestration run¾ same emissions scenario, but 0.6 Pg C yr^-1 was also removed from the atmosphere during the next 100 years (2000-2100), and permanently isolated (i.e., no loss was allowed back to the atmosphere); (3) 1500 m run¾ same as the permanent sequestration run but we also added to the ocean, over the next 100 years, 0.1 Pg C yr^-1 as CO₂ at 1500 m at each of 7 sites; and (4) 3000 m run¾ same as for the shallower injection run, but injection was made instead at 3000 m. Thus for each of the two ocean injection runs, we added a total of 0.7 Pg C yr^-1 into the ocean; however, only 0.6 Pg C yr^-1 was removed from the atmosphere. That is, we assigned a 17% penalty as a means to represent the energy required to separate, transport, and pump CO₂ to depth. For all four runs during the final 700 years, emissions followed the Bacastow and Dewey (1996) scenario, but without injection. Our injection sites include (1) Bombay, (2) Bay of Biscay, (3) Jakarta, (4) New York, (5) Rio de Janeiro, (6) San Francisco, and (7) Tokyo. At each site, both the 1500-m and the 3000-m CO₂ injections were made in the nearest ocean grid box, within 1000 km from the coast, which had at least a depth of 3000 m.

For each injection simulation we used eight separate DIC tracers: one to track each plume of DIC extending from each of the seven injection sites, and an eighth tracer to track the invasion of CO₂ from the atmosphere. Such an approach allows one to study site dependent effects on DIC, pCO₂, the air-sea CO₂ flux, and pH. Furthermore it simplifies modeling and analysis and lowers computer time requirements. Our tests indicate that the carbon chemistry nonlinearities are small enough that differences between a 1-tracer approach and our 8-tracer approach (summing linearly) are negligible. For instance, atmospheric concentrations predicted by the multi-tracer vs. single-tracer approaches are nearly indistinguishable (maximum difference 0.15 ppm).

Figure 1: (a) Fossil emissions scenario (Bacastow and Dewey, 1996) used to force all simulations in this study. (b) Atmospheric pCO₂ for the control run (solid) and the second derivative of atmospheric pCO₂ with respect to time (dots).

Figure 2: (a) Reduction in atmospheric CO₂ relative to the control run due to permanent sequestration (solid), ocean injection of 0.7 Pg C yr-1 for 100 years at 1500 m (dots), and analogous injection at 3000 m (dashes). (b) The total efficiency (reduction in atmospheric CO₂ / reduction in emissions) for the same three scenarios (same line patterns). Note that injection year zero corresponds to calendar year 2000.

Although permanent sequestration does not directly affect the ocean, it results in lower levels of atmospheric CO₂ relative to the control run (Fig. 2). Thus, the air-to-sea flux of anthropogenic CO₂ is lower, so that by indirect effects, permanent sequestration becomes less than 100% efficient. For instance, the removal of 60 Pg C from the atmosphere during the 100-year injection period is equivalent to about a 28 ppm reduction in atmospheric CO₂. However, the maximum reduction found even in the permanent sequestration case is only about 23 ppm (Fig. 2a). During the century-long injection period, the total efficiency of permanent sequestration drops rapidly for the first 50 years to a low of 77% (Fig. 2b). During the next 50 years, the efficiency climbs back to 82%. The early minimum follows the inflection point in the rate of change of atmospheric CO₂ (Fig. 1b), which is controlled by the fossil emissions scenario. Subsequent to injection, consistently lower atmospheric CO₂ drives further reduction in ocean CO₂ uptake (around -4% per century) relative to the control.

As expected, injection at 1500 m is less efficient than injection at 3000 m (Fig. 2a). Injection at 1500 m remains 94% as efficient as permanent storage by the end of the 100-year injection period; however, over the next 300 years its relative efficiency declines around 8% per century faster, relative to permanent sequestration. Injection at 3000 m remains about 98% efficient, relative to permanent sequestration, after the first 200 years. Over the next 300 years, the decline in efficiency for the 3000-m injection is intermediate between that for permanent sequestration and injection at 1500 m.

Figure 3a illustrates (1) how the ocean's capacity to retain injected CO₂ depends upon injection site location and (2) how the relative efficiency between sites changes with time. For injection at 1500 m, six of the sites contribute equally to reducing atmospheric CO₂ at the end of the 100-year injection period. However at the same time, the seventh site, New York, retains only about 2/3 of the amount of injected CO₂ as do the others. After another two centuries until the end of the simulation, the sites can be classed in 3 categories: (1) high efficiency (San Francisco, Bay of Biscay, and Tokyo), (2) lower efficiency (New York, Rio de Janeiro, and Jakarta), and (3) intermediate efficiency (Bombay).

The sites at Rio de Janeiro and Jakarta lose CO₂ more rapidly because both are relatively closer to the Southern Ocean where there is extensive communication between surface and deep waters. Virtually all of the CO₂ injected at those sites is lost from the Southern Ocean. Interestingly, two of the high efficiency stations (San Francisco and Bay of Biscay) are on the eastern margins of the Atlantic and the Pacific, which do not have the intense recirculation currents that are typically found on western margins (e.g., the Gulf Stream and the Kuroshio). In contrast is Tokyo, which is also in the high efficiency category, but is on the western margin of the Pacific. Strikingly different are the low results for New York, on the western margin of the Atlantic.

Figure 3b shows that for the 3000-m injection, storage efficiencies are greater and inter-site differences are smaller. Up to about 200 years after the beginning of the injection period, all sites retain nearly 100% of their injected CO₂, relative to permanent sequestration. Subsequently differences between sites grow, but they remain much less than those for the 1500-m injection. The most efficient sites (Tokyo, San Francisco, Bay of Biscay, and New York) are located within the higher latitudes of the northern hemisphere; the least efficient sites (Bombay, Jakarta, and Rio de Janeiro) are nearest to or actually within the Southern Hemisphere. Most of the injected CO₂ is lost from the Southern Ocean.

Figure 3: The amount of CO₂ injected at (a) 1500 m and (b) 3000 m, which is retained by the ocean at each of the seven injection sites: Bombay (solid), Bay of Biscay (short dashes, Jakarta (long dashes), New York (widely spaced dots), Rio de Janeiro (dash-dot), San Francisco (dash-dot-dot), and Tokyo (closely spaced dots).

Figure 4: Reduction in atmospheric CO₂ relative to the control run with respect to (a) the cumulative effect of injection at two sites in both the MPI (thick line) and the IPSL models (thin line); and (b) comparison of the same two models at each of the two sites: New York (solid) and Tokyo (dashes).

We have compared results from this study with those from previous work. Within the ocean in the immediate vicinity of the injection sites, the dissolved inorganic carbon (DIC) in our model increases by more than 200 m mol kg-1 relative to the no sequestration case (control). These perturbations represent about 10% of the background concentrations which are present naturally. Yet they are much larger than the maximum perturbations of the order of 10 m mol kg-1 found previously for an analogous injection scenario in the MPI model (Dewey et al., 1996). On the other hand, perturbations of DIC found near injection sites in this study are of comparable magnitude to those found when using the higher resolution Mesoscale Dispersion Model (Dewey et al., 1996); yet, for the latter study CO₂ injection rates were about 10 times smaller. Both model resolution and the advection scheme are important in determining local CO₂ perturbations.

We were able to further compare results because we chose not only the same emissions and injection scenarios but also two of the same injection sites, New York and Tokyo, as did Dewey et al. (1996). Figure 4a shows that when the total reduction in atmospheric CO₂ is summed up for both sites, differences between the two models are negligible for the first 200 years after the start of injection. Subsequently, model estimates diverge with the IPSL model retaining about twice as much injected CO₂ as the MPI model after the next 300 years. Global differences between the two models could derive in part from different vertical model resolution. With 15 vertical layers for the MPI model, Bacastow and Dewey (1996) and Dewey et al. (1996) made the injection in the model layer that extends from 900 to 1500 m. In the IPSL model, which has 30 layers, the "1500-m" injection is effectively deeper, between 1211 and 1613 m. Additionally, global differences may be due in part to the vertical eddy diffusivity K_Z which differs substantially between MPI and IPSL models. In the MPI model, the K_Z comes from the numerical diffusion associated with the upstream advection scheme. In the IPSL model, K_Z is determined prognostically from surface wind stress and heat fluxes according to the Turbulent Kinetic Energy model of Gaspar (1990). Unlike other approaches, the TKE model is able to capture high values of K_Z in the mixed layer along with the sharp transition to low values in the thermocline, as is observed.

Despite overall agreement between both models over the first 200 years (Fig. 4a), site-by-site comparison reveals substantial model differences during that time and afterwards. The MPI and IPSL models predict opposite tendencies in regards to the capacity of New York and Tokyo sites to retain injected CO₂ (Fig. 4b). In the MPI model, injections just off New York and Tokyo are equally efficient for the 100-year injection period, but subsequently the injection off New York is much more efficient. For the IPSL model, the New York injection is much less efficient than that off Tokyo, throughout the simulation. During the first 200 years, the IPSL model predicts that the Tokyo injection is more efficient than the MPI prediction for either Tokyo or New York. For the same period, the IPSL model predicts that injection off New York is less efficient than MPI predictions for either site.

To determine how realistic a given model's predictions are, we need additional information. Of most interest for perturbations in CO₂ are ocean data for transient tracers that are transferred between the atmosphere and ocean by gas exchange. One such tracer is bomb ¹⁴C, which was produced during the atmospheric nuclear testing of the 1950's and 1960's. At its peak in 1963, the atmospheric level of ¹⁴C was nearly double that during the pre-nuclear era. Subsequently, atmospheric ¹⁴C declined, much of it being absorbed by the ocean. Because of its recent nature, bomb ¹⁴C can be used to help diagnose problems with the predicted near-surface circulation in ocean models. Bomb ¹⁴C simulations in both the MPI and IPSL models were compared to oceanic measurements of bomb ¹⁴C (Broecker et al., 1995) during OCMIP (Orr, 1996). Of particular interest is the east-west GEOSECS section at about the latitude of Tokyo. That section reveals that at the Tokyo injection site, both models manifest excessive upwelling (vertical advection) of deep waters from about 1000 m to the surface. Such a model artifact probably drives too much of the injected CO₂ out of the ocean. Excessive upwelling appears more widespread in the MPI model. Another OCMIP validation offers complementary information because it concerns the deep ocean. Model-data comparisons for pre-industrial or "natural" ¹⁴C suggest that in West Pacific the MPI model performs well, whereas the IPSL model suffers from excessive formation of intermediate and deep waters. Our spatial analysis of the CO₂ plume injected near Tokyo suggests that of the two artifacts, it is the excess formation of deep waters that dominates in the IPSL model. In summary, it seems likely that the MPI model loses too much of its CO₂ injected near Tokyo whereas the IPSL model retains too much.

It is too early to determine how realistic the two model predictions are for the Atlantic. Natural ¹⁴C in the West Atlantic GEOSECS section shows generally good agreement in both models, but this section is not immediately adjacent to the injection site. Inadequate spatial coverage of the ¹⁴C data set becomes an issue when interested in point-source ocean injection simulations. However, new and existing ¹⁴C data will be more fully exploited during OCMIP. Moreover, validations during OCMIP will also focus on two other transient tracers, CFC-11 and CFC-12, for which a much denser network of measurements is available. Meanwhile, we have tried to compare other relevant information between models. It appears that the larger efficiency of the MPI model to retain CO₂ injected off New York stems from its northward transport of the plume to deeper convective mixed layers in the high latitudes of the North Atlantic (Dewey et al., 1996). Convective mixed layers in the same region of the IPSL model are substantially shallower.

Large local differences between two 3-D ocean models indicate that it is indeed worthwhile to pursue more detailed model comparison and validation, as planned during OCMIP and GOSAC. These studies will help pinpoint the crucial processes and regions for which our understanding must be improved. Such efforts will lead to improved model-predicted circulation fields, which should eventually provide more trustworthy estimates of the ocean's retention time of injected CO₂. However, this study as well as previous work have been forced to make certain simplifications, as necessary in any modeling effort. Perhaps one of most important simplifications is to neglect the effect of climate change on ocean circulation. Climate-induced shifts in ocean circulation could result in substantially different retention times for injected CO₂. This uncertainty requires further research.

We appreciate discussions initiated by B. Ormerod, J. Sarmiento, and P. Freund, which incited this study. We thank B. Ormerod, H. Drange, J. Davison, B. Bacastow, C. S. Wong, G. Nihous for suggestions regarding ocean injection sites, depths, and rates. We are grateful to J. Davison for discussions throughout this work and R. Matear, P. Monfray, and C. Le Quéré for comments. Funding was provided by the EC Environment and Climate Programme (ENV4-CT97-0495) and the IEA GHG R&D Programme. However, this work does not necessarily reflect the opinions of either agency. Supercomputing resources were provided by grants from the CEA in Grenoble and the CNRS (IDRIS) in Orsay.