Our study uses a macroalgae model embedded within a coupled physical-biogeochemical ocean model. Ocean physics is represented by the Nucleus for European Modelling of the Ocean framework (NEMO; ). NEMO is composed of an ocean general circulation model coupled to a separate sea-ice model, the Sea Ice modelling Integrated Initiative (SI3; ). The NEMO domain used here (the extended ORCA1 grid; eORCA1) is at a horizontal resolution of approximately 1°, and uses a tripolar model grid and incorporates an equatorial band of enhanced resolution. Vertical space in NEMO is resolved into 75 z-levels that range in thickness from approximately 1 m at the surface to approximately more than 200 m at abyssal depths.

Ocean biogeochemistry is represented by the Model of Ecosystem Dynamics, nutrient Utilisation, Sequestration and Acidification (MEDUSA; ). MEDUSA is a dual size-class “intermediate complexity” ecosystem/biogeochemistry model that resolves nutrients and two size classes of phytoplankton, zooplankton, and detritus (NPZD) components. The smaller size class includes non-diatom phytoplankton, microzooplankton and small particles of slow-sinking detritus, while the larger size class includes diatoms, mesozooplankton and large particles of fast-sinking detritus (the latter modelled implicitly). The biogeochemical cycles of N, Fe, C, silicon, oxygen, and alkalinity are represented, coupled in both fixed and dynamic stoichiometric relationships.

Air–sea CO2 flux in NEMO-MEDUSA is calculated prognostically using standard equations for the partition of dissolved inorganic carbon (DIC) into dissolved CO2 and carbonic acid, bicarbonate, and carbonate ions. Partitioning is a function of local temperature, salinity, DIC, and total alkalinity, and uses the MOCSY-2.0 carbonate chemistry routines . The exchange of CO2 with the atmosphere is then calculated from the dissolved CO2 fraction, atmospheric CO2, temperature, salinity, and surface winds, following the gas transfer parameterisation of . Because gas transfer is finite and wind speed dependent, the rate at which surface ocean pCO2 anomalies equilibrate with the atmosphere is limited, introducing a natural lag between biological carbon drawdown and atmospheric CO2 uptake. A full description of MEDUSA can be found in , with the current version evaluated in .

The ocean model is forced at its surface boundary using a climatologically-adjusted version of the ERA-5 atmospheric reanalysis . Ocean physical fields and the biogeochemical fields of oxygen and nutrients (N and silicon) are initialised using the World Ocean Atlas 2023 . As simulations are initialised from a time-point prior to the reference period of the GLODAP climatology , dissolved inorganic carbon (DIC) and alkalinity fields are initialised from prior simulations of the model (in the UKESM1 framework; ) to minimise bias in the CO2 air–sea flux due to model equilibration. For other plankton tracers we use uniform nominal values. Riverine nutrient inputs are not included in this configuration, so nutrient supply is limited to oceanic initial conditions and internal cycling.

This initial state is then simulated for two cycles of the forcing period between 1976 to 2024 (2 × 49 years = 98 years). This is done so that surface biogeochemical properties reach quasi-equilibrium, minimising the drift during the experimental period when macroalgae are introduced (see Sect. ). Note that, since DIC has a strong secular signal during the 1976–2024 period caused by rising atmospheric pCO2, its concentration is reset to the 1976 UKESM1 state at the beginning of each forcing cycle, and the rising atmospheric pCO2 is still simulated during the simulation until the end of the cycle. The temperature increase within this simulated period is relatively small compared to late 21st-century warming used in other CDR studies (e.g. in ) and therefore show limited effects on the macroalgae and other biogeochemical tracers.

mCDR by macroalgae is simulated by adding a new aquaculture module to MEDUSA. This module effectively represents macroalgae as being cultivated on a fabricated floating structure that is anchored and immobile . As such, macroalgae is located on the ocean model's grid but it is not advected by ocean currents, although it interacts with MEDUSA's advected passive tracers. We assume that macroalgae cultivation started from the last 20 years of the control run (2004–2023), starting from a small initial biomass of 0.01 mmolNm-3. This initial value is consistent with other biological tracers in MEDUSA and serves as an “inoculation”, after which biomass grows in response to temperature, light and nutrient limitation.

Fundamentally, the macroalgae growth and loss processes are similar to those of phytoplankton with the key differences subsequently outlined. Unlike MEDUSA's phytoplankton, modelled macroalgae biomass is static within the model grid and has a restricted vertical extent, by default from approximately 5 to 30 m in the water column (10 model levels) following previous modelling studies and offshore cultivation experiment . This depth range is intended to represent large-scale offshore cultivation and differs from nearshore farming practices, where shallower deployments are common (e.g. ). In an initial simulation, we found negligible differences between cultivating macroalgae between 1–30 and 5–30 m (results not shown). Functionally, macroalgae growth is loosely based on that of MEDUSA's phytoplankton, with separate chlorophyll (Chl) and N biomass, and with its rate governed by temperature and the availability of light and dissolved inorganic nitrogen (DIN) and iron (DFe). No external nutrient inputs or pulsed fertilisation are applied. As a result macroalgae compete directly with phytoplankton for the same nutrient resources. Fig.  presents a schematic of the macroalgae model and its relationships with MEDUSA's dissolved tracers.

In terms of loss processes, while the simulated macroalgae incur non-harvest losses, they are assumed not to be explicitly grazed by zooplankton. However, their cultivation involves active harvesting, the precise mechanism and fate of which forms part of the experimental design described below.

Following the work of , we use multiple species in our macroalgae module with: two cold water species (Macrocystis and Saccharina) and two warm water species (Sargassum and Euchema). These taxa span a board range of thermal niches and cultivation context: cold water species usually occupies high latitude nutrient-rich upwelling system, commonly found in the North Atlantic aquaculture industry , while warm water species are predominantly cultivated in tropical and subtropical regions . All four are represented in identical functional from, with a number of shared parameters, but each species has different values for certain key parameters. These include optimal temperature range (between Topt1 and Topt2), half-saturation constant for DIN uptake (kDIN), and C:N ratio (ΘC:N) adopted from . Although these taxa differ in their real-world cultivation methods and physiological traits, they are represented using a common idealised cultivation framework to isolate first-order biogeochemical effects at large spatial scales. We acknowledge that growth rates, carbon content, and other physiological traits vary considerably across and within taxa in reality , but our choice rather reflects a functional representation designed to capture first-order biogeochemical effects at global scale. Our analysis therefore focuses on the sensitivity of large-scale biogeochemical responses to macroalgal cultivation.

Macroalgae growth is limited by temperature (TL), light (LL), and nutrient (NL), and its losses comprises of non-harvest loss Lossnoharv and harvesting loss Lossharv. Therefore the change in each macroalgae type, Mn is calculated by: 1dMndt=Growth-Lossnoharv-Lossharv

The temperature limitation term is similar to a Gaussian probability curve with flat peak, adopted from : 2TL=exp⁡β1(T-T1,opt)2,   T<T1,opt3TL=exp⁡β2(T-T2,opt)2,   T>T2,opt4TL=1,T=Topt

Where Topt is a 5 °C optimal temperature range for each macroalgae group. β1 and β2 are adjusted near the lower (β1) and upper temperature limits (β2) to reach zero, respectively. The temperature limitation function acts as a smooth growth filter rather than a binary on-off switch. As such, growth declines gradually as temperature departs from the optimal range for each represented taxon, ensuring more realistic transitions across latitude and season.

Light limitation is formulated in the same way as in phytoplankton: 5LL=μTLα^mI(μTL)2+α^m2I20.5

Where I is the irradiance and α^ is the initial slope of photosynthesis-irradiance curve, so that macroalgae with a high chlorophyll content have an elevated response to irradiance.

Macroalgae compete with phytoplankton to obtain nutrients (n), and nutrient uptake (nup) is formulated using Michaelis–Menten kinetics: 6nup=nkm+n

Unlike the default formula for phytoplankton, in the macroalgae default simulation, we assume that Fe is supplemented (Feup = =1) and only DIN is the limiting nutrient. However, we allow multiple nutrient limitation, by using the Liebig's law of minimum, described below (NL): 7NL=min(DINup,Feup)

In addition to harvesting, simulated macroalgae experience non-harvest losses, which represents biomass losses outside deliberate harvesting, such as physical erosion, aging, fragmentation, and implicit grazing processes. Non-harvesting loss are transferred directly to the slow detrital pool, where they are remineralised according to the standard MEDUSA slow sinking detritus formulation. As a result, processes such as direct DOC release during production are implicitly subsumed within the detrital pathway rather than treated explicitly: 8Lossnoharv=m2maxMn

Where m2max is the rate of mortality.

For harvesting, once total integrated macroalgae biomass at a grid point has reached the target harvesting threshold (TW), it is harvested over the course of a single simulated day, using a linear loss term and parameter m1max. Otherwise, the macroalgae continues to grow. 9Lossharv=m1maxMn

In the default case, harvested macroalgae biomass is assumed to be baled-up and immediately sunk to the seafloor in the same water column, so that the associated carbon biomass is durably stored and isolated from the surface ocean and atmosphere. This process is formulated similarly to that of fast-sinking organic detritus in MEDUSA (see Sect. 2.3.10 of ), but with a different remineralisation length-scale, and without the influence of any biomineral ballasting. The value of the length-scale is set to considerably deeper than the seafloor (40 km) to represent this active baling and deposition that aims to ensure that the harvested biomass reaches the seafloor does not undergo substantial decay. The organic material reaching the seafloor in this way is added to MEDUSA's existing benthic reservoirs and is remineralised from these in the same manner as material arriving via the natural biological pump (i.e. slow remineralisation to DIN, DIC and DFe with the associated consumption of oxygen). This approach avoids cultivated macroalgae biomass undergoing remineralisation in the upper water column and increases the efficiency with which fixed CO2 is stored in the ocean (and removed from exchange with the atmosphere). Parameter values used in these equations are described in Table .

Note that MEDUSA resolves biological respiration and remineralisation for both phytoplankton and macroalgae at every depth level and timestep. Air–sea CO2 flux is therefore driven by the net change in surface DIC after all biological production, community respiration, grazing, and remineralisation processes operate, rather than by NPP alone. The net ecosystem production (NEP) signal is thus an emergent property of the model's coupled biological and chemical state.

After initialisation and a spin-up period (see Sect. ), we run the control model (NEMO-MEDUSA only) using ERA-5 forcing from 1976 until 2024. We assume that macroalgae cultivation started from the last 20 years of the control run (2004–2023). We will focus on quantifying CO2 sequestration, biomass harvested, and assessing how harvesting strategy would affect ocean biogeochemistry including nutrient cycles (N and Fe), phytoplankton net primary production (NPP) and plankton biomass, carbonate chemistry (air–sea CO2 flux), seafloor oxygen, and carbon pool distributions. The results shown in this manuscript are the average from 2015–2024.

To assess the biogeochemical impacts of large-scale macroalgae cultivation, we compare a baseline MEDUSA simulation with one incorporating dynamic macroalgae growth (MEDUSA + Macroalgae). Building on this framework, we conduct four sensitivity experiments to explore key intervention strategies and test modelling uncertainties:

varying the biomass threshold for harvesting and sinking

These experiments and shorthands are summarised in Table .

In all experiments we define the target biomass as the sum of local integrated macroalgae biomass in mmolNm-2. The default target biomass is 400 mmolNm-2 (500 gdryweight(DW)m-2), resulting in two harvests a year on average based on studies by , although warmer macroalgae species can be harvested up to 12 times a year . To assess the sensitivity of target biomass in harvesting, we vary the weight to 200 and 800 mmolNm-2. Macroalgae are not harvested until they reach these target biomasses, so will continue to grow until they do (and will continue to experience non-harvest losses). Harvesting involves decreasing the biomass of macroalgae by 90 % during the harvest day. Since Fe is one of the limiting nutrients for macroalgae, we prescribe Fe:N ratio of 5.1 × 10⁻³ and half saturation constant of 1.5 × 10⁻³ based on laboratory studies . However, in the default simulation (as well as all other experiments except for Fe limitation), we assume that macroalgae are only limited by DIN. This corresponds to an assumption that Fe is supplemented, following established farming practices , and that it is at non-limiting (replete) concentrations. When macroalgae is being remineralised, the added Fe will also be released.

To determine whether CDR is efficient within a certain region, we use two different CDR efficiency metrics; efficiency calculated using macroalgae NPP (CDReffNPP) and harvested macroalgae biomass (CDReffHarv). CDReffNPP can be calculated as the proportion between macroalgae NPP and the additional air–sea CO2 flux within each grid cell (CDR flux) : 10CDReffNPP=CO2fluxmacroalgae-CO2fluxcontrolMacroalgaeNPP×100%

If annual average macroalgal NPP associated carbon is equivalent to the annual average of additional atmospheric CO2 uptake within a grid cell then the CDR efficiency would be 100 %, as it is assumed that the macroalgae-induced DIC deficit within the grid cell has driven the CO2 uptake. The calculation of CDReffNPP is therefore a net measure, incorporating macroalgal production, phytoplankton feedback, and air–sea carbon equilibration collectively . It should be noted that we use annual average macroalgae NPP, to account for macroalgae seasonality, whereas the modelling study by uses prescribed maximal macroalgal NPP.

To calculate how much carbon is assumed to be durably stored, we calculate CDR efficiency using the proportion between harvested macroalgae biomass and the CDR flux : 11CDReffNPP=CO2fluxmacroalgae-CO2fluxcontrolMacroalgaeHarvest×100%

In the default simulation, macroalgae occupies 51.7 × 10⁶ km2, accounting for approximately 14.35 % of the global ocean surface. Macroalgae were primarily distributed across nitrogen-rich surface waters and major upwelling zones, including the Equatorial Pacific, North Pacific, North Atlantic, Southern Ocean, and coastal regions such as Chile, Argentina and Namibia (Fig. a). The average annual macroalgae NPP of 111.1 gCm-2yr-1 (40.34 PgCyr-1) was similar in magnitude to the average phytoplankton NPP from the control run (47.7 PgCyr-1). The most dominant macroalgae species are Saccharina and Sargassum in high and low latitudes, respectively (see Fig. S1 in the Supplement), with the former contributing the most to macroalgae NPP (36.11 PgCyr-1). The annual harvest yield that is sunk to the deep ocean reached 12.5 PgCyr-1 (135.59 tDW (tonne Dry Weight) km-2yr-1), 31.1 % of the macroalgae NPP, with a high proportion harvested in the North Atlantic, Southern Ocean, and along the coast of Chile (Figs. b–d and ). The spatial patterns of macroalgae, NPP, and nutrients shown here are emergent outcomes of open-ocean biogeochemical conditions within a coarse-resolution global model, and should be interpreted as indicating potential cultivation areas and not as plausible regions for macroalgae farms. Many regions identified as productive for macroalgae would present extreme technological, logistical, and societal challenges.

Different cultivation protocols and model assumptions affect macroalgae outcomes. Varying the harvesting threshold changes NPP magnitude while preserving the general spatial pattern (Figs. S2 and S3 in the Supplement). Harvest 800 led to less harvested biomass (-8.7 %), higher area of coverage (+6.28 %), and higher macroalgae NPP (+49 %). Harvest 200 decreases macroalgae coverage (-5.43 %), reduced total macroalgae NPP (-30 %), and modestly raised the harvested biomass (+13 %) due to more frequent harvest (see Fig. ). This resulted in a higher harvest efficiency (i.e., harvested biomass as a fraction of macroalgae NPP, see Fig. S3b in the Supplement) compared to the default simulation (Harvest 400). Other protocols also affected macroalgae production, Fig. . Extraction shows slight decreases of macroalgae NPP and area (by -1.37 %). Introducing Fe limitation (i.e. stopping Fe supplementation), dramatically suppressed macroalgae growth and harvest was almost entirely eliminated (-99.2 %), and the cultivated area shrank by 47.95 %. Macroalgae NPP also declined by 72.8 %, especially in regions where Fe concentration is low, such as the Equatorial and North Pacific (see Fig. S2f in the Supplement). Higher non-harvesting loss also significantly reduced macroalgae area (44.6 %), NPP and harvest.

Under default simulation conditions assuming no Fe limitation, global-scale macroalgae cultivation enhanced net air–sea CO2 flux by approximately 11.0 PgCyr-1 (equivalent to 40.3 PgCO2yr-1; Fig. ), while without macroalgae, CO2 flux is only 1.4 PgCyr-1, indicating a substantial contribution to marine carbon dioxide removal (mCDR). The increase in CO2 flux under the default simulation is sufficient to limit warming to +2 °C (between 0.7–3.6 PgCyr-1 ). Although 31.1 % of primary production is being harvested and sunk, the CDReffNPP (the proportion of additional CO2 flux to macroalgae NPP) was only 27.3 % but CDReffHarv was 87.9 %. CDReffNPP also varied regionally. High-efficiency zones (CDReffNPP > 50 %) were simulated along the Southern Ocean, while low-efficiency areas (< 10 %) emerged in the central Equatorial Pacific, off the western coast of North Japan, and offshore South Africa (Fig. c). Areas off southern coast of Chile, exhibited net outgassing due to surface DIC accumulation (Fig. b). Since additional CO2 flux often occurs outside harvesting area (Fig. c and Fig. b), we do not show efficiency maps for CDReffHarv.

The mCDR efficiency is sensitive to cultivation protocol. Harvest 200 slightly decreased the CO2 flux, improved the CDReffNPP to 38.4 % (Macroalgae NPP of 27.95 PgCyr-1, CDR flux of 10.73 PgCyr-1), but slightly lowers CDReffHarv to 86.1 %, due to lower CDR flux. Whereas Harvest 800 reduced the CDR flux to 9.6 PgCyr-1 and lowered CDReffNPP to 16.0 %, but slight increases CDReffHarv to 89.4 %. In all threshold scenarios, the Equatorial Pacific consistently showed the lowest CDReffNPP, whilst modest improvements occurred in the Southern Ocean under the low-threshold scenario (Fig. S4b in the Supplement).

Further amendments to the cultivation strategies and modelling assumptions also influenced mCDR outcomes, Fig. . The High Loss experiment increases macroalgae NPP to 60.18 PgCyr-1 because of the regenerated nutrients from non-harvesting loss, but reduced global CO2 flux into the ocean to 7.3 PgCyr-1, because there are less macroalgae harvested and sunk, which reduces CDReffNPP to 16.0 %, but slightly increases CDReffHarv to 88.4 %. Extraction increased the global CO2 flux by 0.13 PgCyr-1, because of less remineralisation in shallow waters, which induced CO2 flux. This protocol also has the highest CDReffHarv of 91.7 % due to higher CDR flux. Simulating realistic Fe limitation constraints on macroalgae growth causes outgassing by contributing 0.2 PgCyr-1 to the atmosphere, indicating that scaled-up macroalgae cultivation without Fe fertilisation is not an effective mCDR technique (negative CDR efficiency, -1.6 %).

Macroalgae cultivation had significant effects on surface nutrient concentrations and oxygen levels, primarily in areas of cultivation and harvesting, summarised in Table . In the default simulation, surface DIN declined by 53.1 %. This decline was co-located with high macroalgae productivity. Because of the assumption of Fe supplementation, surface DFe increases by 90.26 % from non-harvesting loss. Furthermore, to grow macroalgae ubiquitously in the default run, on average 1.4 PgFeyr-1 was added to the ocean. The increased DFe occurs in areas occupied by macroalgae. DIN and DFe accumulation was also found near the seafloor at deposition sites, reflecting remineralisation of sunken biomass. Due to high Fe:N ratio of macroalgae, the increase in seafloor DFe can reach up to 3× the control simulation.

Altering the harvest threshold had a significant effect on surface nutrients (Table  and Fig. S5a–h in the Supplement). Harvest 200 shows slightly weaker DIN drawdown (-50.3 %), while Harvest 800 showed a slightly stronger DIN drawdown (-53.9 %). These experiments also produce higher and lower surface DFe compared to the control simulation (83.56 % and 102.84 % more DFe for Harvest 200 and Harvest 800, respectively). Among the alternative cultivation protocols, higher non-harvesting loss rate leads to less DIN decline (-42.5 %) and lower DFe increase (41.05 %). In contrast, the Extraction experiment removed the benthic DIN enrichment entirely, and shows a slight benthic DFe increase, while having similar surface impacts to the default run. For the Fe limited experiment, the surface DIN and DFe increases and decreases by 47.2 % and -30.7 %, respectively. At the seafloor, the DIN concentration declines by 0.48 % because of the lack of harvest but DFe increases by 1.64 % due to high Fe:N ratio.

Oxygen concentrations were affected by both macroalgae growth and biomass disposal. At the seafloor, oxygen losses reached 20.3 % on average globally, but can reach 70 % decline in the harvesting and deposition areas (see Fig. c). These areas can also become suboxic (Fig. k), which covers 7.9 % of the seafloor, compared to 0.5 % in the control simulation.

Impact on oxygen concentration is closely linked with harvesting threshold (Table  and Figs. S5o–u and S6 in the Supplement). Harvest 200 led to greater seafloor oxygen loss. Whereas Harvest 800 reduced oxygen depletion at the seafloor due to less deposition. Similarly, High Loss shows lower oxygen loss in the deep ocean, as less macroalgae is harvested (see Fig. ). In contrast, the Extraction experiment eliminated benthic oxygen loss. The inclusion of Fe limitation also shows very low benthic oxygen less compared to other experiments owing to suppressed harvest and sinking (Fig. S6f and g in the Supplement).

Large-scale macroalgae cultivation substantially altered surface nutrient availability and light conditions, leading to pronounced changes in the phytoplankton NPP. Note that these changes reflect competition for nutrients and light as represented in the model and provide an indicative estimate of the potential biogeochemical response, rather than a detailed prediction of realised ecosystem behaviour. In the default simulation, phytoplankton NPP fell by 49.78 % (-24.81 PgCyr-1), accompanied by declines of in phytoplankton biomass (45.40 %) and zooplankton biomass (47.16 %), with the largest reductions in nutrient-rich and upwelling regions (Fig. b, e, and h). There are also areas where both phytoplankton and zooplankton concentrations are increasing, such as the polar Southern Ocean (Fig. d, e, and f), which may occur due to the increase in DFe concentration (Fig. d–f). Adjusting harvesting threshold modified the severity but not the direction of these changes (Fig. 7c, d, j, and k in the Supplement): Harvest 200 slightly alleviated phytoplankton biomass and NPP losses but intensified zooplankton declines, while Harvest 800 worsened phytoplankton and NPP reductions but not as much zooplankton decline as the default run.

Other modelling considerations, such as imposing higher non-harvest loss would reduce less phytoplankton biomass, NPP, and zooplankton biomass. Extraction had little effect on plankton dynamics compared to the default, suggesting that surface processes dominate short-term responses. Imposing Fe limitation sharply constrained macroalgal growth and partly restored surface nutrients, especially DIN, yet phytoplankton and zooplankton showed net decline (Table ), although regions such as the Indian, Atlantic, and parts of North Pacific did exhibit increased phytoplankton and zooplankton concentrations (Fig. S7 in the Supplement). These results underscore that ecosystem-wide impacts persist even when macroalgae productivity is minimal.

The enhanced CO2 uptake in the default scenario was accompanied by major biogeochemical side-effects. Macroalgae cultivation reduced global phytoplankton NPP and zooplankton biomass by almost half the control model, which is caused by DIN depletion in surface waters. This decline exceeded 50 %, consistent with nutrient robbing seen in previous modelling studies . However it is also important to note that the macroalgae submodule includes only a simple loss term and does not explicitly resolve grazing or other higher trophic food-web feedbacks (e.g. in zooplankton grazing is explicitly represented). Additionally, MEDUSA only represents two phytoplankton functional types, which do not fully capture the diversity of nutrient acquisition strategies and ecological responses present in natural systems. As a result, the simulated reduction in phytoplankton NPP and zooplankton biomass reported here may represent an overestimation, because explicit herbivory on macroalgae could recycle nutrients and reduce the impact on phytoplankton NPP, as illustrated in the High Loss experiment (see Fig. 3). A more diverse representation of phytoplankton functional types could also alter the magnitude and spatial pattern of NPP responses. This makes the simulated phytoplankton response in this study a first-order biogeochemical estimate of potential competition effects.

Sustaining large-scale global macroalgal cultivation required Fe supplementation, which drove a ∼ 90 % increase in surface DFe and up to 3× increase at the seafloor due to the remineralisation of biomass deposition. Despite the DFe enrichment phytoplankton NPP biomass declined by nearly half (see Fig. , Table ) due to macronutrient robbing by macroalgae. An ocean Fe fertilisation modelling study also showed large-scale Fe fertilisation caused enhanced consumption of major nutrients in surface waters, and also reduces nutrients availability for lower latitudes, which offsets overall NPP . This suggests that mCDR approaches can risk redistributing ecological pressure, while also suppressing additional carbon removal from phytoplankton.

Deep-sea biogeochemistry will also be altered due to macroalgal deposition. In the default simulation, deep-sea oxygen losses can reach suboxic levels in deposition zones (Fig. j-l) which has been a concern in previous studies . Severe deoxygenation would favour smaller species, reduces large predators and bioturbation, and triggers faunal emergence or habitat avoidance, especially in the coastal benthos , which may alter natural carbon sequestration in the deep sea environment. Deposition also increases DIN at the seafloor (Table ) and the introduction of nutrients to the oligotrophic seafloor may also alter benthic species interactions . We note that these results are not a detailed characterisation of benthic ecosystem impacts. MEDUSA's representation is intentionally simple, and a full assessment of seafloor community responses, including benthic respiration dynamics and faunal impacts, would require dedicated benthic ecological models. However, these results may be used to identify large-scale signals and pattens that would motivate future benthic studies.

To examine how these impacts depend on cultivation and modelling design, we tested alternative harvesting threshold, extraction, loss rates, and nutrient limitations. Harvest timing and threshold can affect the optimal yield and capital expenditure . Lowering the harvest threshold improved CDR flux and efficiencies, both CDReffHarv and CDReffNPP, while slightly reducing biogeochemical disruption, and with similar CDR flux as the default run. A field study has showed that a more frequent harvest without re-seeding, would increase yield per meter growth line and reduced cultivation cost . However, it also resulted in a greater decline in zooplankton biomass than phytoplankton biomass (Table ) indicating potential downstream effects of macroalgae cultivation on higher trophic levels and, potentially, fisheries . Higher thresholds allow more macroalgae biomass to accumulate ahead of harvesting, leading to an increase in macroalgal NPP. However, as more of this NPP is lost to non-harvesting losses in the model, the associated CDR efficiency is actually lower compared to other harvest experiments (Fig. ). Further, by increasing the quantity of harvested material transferred to depth, higher thresholds additionally cause greater seafloor oxygen depletion, as well as a larger reduction in the NPP of phytoplankton (Table , Fig. k).

Other modelling and cultivation protocols can also significantly affect macroalgae production and CDR capacity. When non-harvesting loss is doubled, the macroalgae NPP is reduced by 2.20 PgCyr-1 and its CDR flux, by 3.7 PgCyr-1, making it less efficient (Fig.  a). High-loss experiments also lose less surface DIN compared to the default run because of higher remineralisation near the surface. In farmed macroalgae, non-harvesting loss due to falloff and frond erosion can reach more than 10 % of its growth rate . This simulation explores how CO2 flux and ocean biogeochemistry are affected when non-harvesting loss consume more of macroalgal biomass. When harvested biomass is extracted rather than sunk to the deep ocean, CDR efficiencies become slightly higher than the default simulation because of the lack of carbon leaking when depositing macroalgae at shallower depths. This simulation also shows minimal impacts at the seafloor due to the absence of deposition (Table ).

The CDR capacity of macroalgae is lower when Fe is not supplemented, since most areas of the open ocean will not be suitable for growing macroalgae . When Fe limitation is implemented, macroalgae NPP collapsed with a 73.7 % decrease in magnitude, aligning with recent work by , and simulate a net increase in CO2 flux to the atmosphere of 0.2 PgCyr-1. However, due to the positive bias in Fe concentration in MEDUSA, compared to the major decline of macroalgae harvest in the Southern Ocean that is simulated in , the decline in this study is not as dramatic. The Fe limitation experiment also simulated 30.7 % less DFe concentrations in surface waters than the control model (Table ), further depressing phytoplankton NPP in the Southern Ocean, Equatorial Pacific, and subpolar North Atlantic. This simulation emphasises that large-scale macroalgal CDR depends critically on Fe supplementation and the enhanced CO2 uptake simulated under Fe-replete conditions depends on nutrient subsidies.

In carbon dioxide removal, additionality denotes that an mCDR strategy must demonstrate more CO2 uptake relative to what would occur without it which will require model assessments due to the global scale of impacts and potential feedbacks between reservoirs . Assessing the additionality of macroalgae-based mCDR can be challenging because its CO2 uptake enhancement is tightly coupled to the redistribution of nutrients and phytoplankton NPP loss within the ocean. In our simulation large-scale cultivation suppressed phytoplankton NPP both within and beyond cultivation areas (Figs.  and S7 in the Supplement), and this can complicate the measurement of net carbon removed . In terms of macroalgal cultivation, even with Fe supplementation, the increase of macroalgae biomass is at the expense of natural biological carbon export . According to the framework of macroalgae cultivation may show even smaller additionality since the carbon cost for setting up offshore macroalgae farm can be more than half of the potential CDR .

Compared to abiotic approaches, such as coastal ocean alkalinity enhancement (OAE), our large-scale macroalgal cultivation uptake is less spatially concentrated. Using the same modelling framework, reported that 13.8 % of the shelf accounts for 50 % of the total extra CO2 flux, with warmer areas contributing higher CO2 uptake, indicating that CDR flux is dominated by a few high efficiency regions. In contrast, our default macroalgae simulation indicates that 28.9 % of the cultivation area is required to reach the same proportion (Fig. a). The regions driving this uptake coincide with areas of high macroalgae NPP and CDR flux (Figs. b and b) particularly subpolar Southern Ocean and coast of Chile, as shown in Fig. b, reflecting a more widespread uptake pattern. This difference arises from the tight nutrient constrain in macroalgae-cultivation, while in OAE, additional CO2 uptake is dependent on temperature, that enhances dissolution rate and the geography of the shelf. Furthermore, the rate of air–sea CO2 gas transfer, parameterised following as a function of local wind speed, introduces a natural lag between biological carbon drawdown and atmospheric CO2 uptake. DIC deficits that persist beyond a single growing season continue to drive CO2 uptake as surface waters are re-exposed to the atmosphere through seasonal mixed layer dynamics and circulation. This explains the regional variation in CDR efficiency shown in Fig. 4c. In the Equatorial Pacific, relatively high macroalgal NPP is associated with lower CDR efficiency. In contrast, the deeper mixed layer in the Southern Ocean keeps low-DIC water in contact with the atmosphere for longer, resulting in higher efficiency. Although regions where high cumulative CO2 share, CDR flux, and macroalgae NPP are located in the Southern Ocean, this area may not be suitable. Large-scale offshore macroalgae deployment would require extensive floating infrastructure, long-distance operations, and resilience to extreme wave and weather conditions . Such engineering and logistical considerations are not represented in the present model and remain outside the scope of this study.

Macroalgal extraction and product substitution can offer co-benefits such as substituting high-emission products with seaweed-based alternative. However, life-cycle analyses show that processing macroalgal products can offset much of the theoretical climate advantage . A recent economic model estimates that product substitution yields a net profit of ∼ USD 50 pertCO2 avoided, compared to a cost of USD 480 pertCO2 for deep-ocean deposition , and product substitution could cut regional emissions by up to 13 %, whereas direct sequestration contributes relatively little . Together these studies highlight the importance of robust monitoring, reporting, and verification frameworks that can verify additionality and durable sequestration.

Natural climate solutions, such as saltmarshes, mangroves, and seagrass ecosystems or coastal macroalgal cultivation, can deliver both carbon sequestration and a range of co-benefits, including coastal protection, nutrient retention, and biodiversity support . These systems provide relatively direct and more verifiable carbon storage through sediment burial, while also helping mitigate eutrophication. Such coastal approaches may offer more tractable mCDR deployment, even though at smaller scale. These findings reinforce that no single approach is likely to deliver the required scale of carbon removal.

Several model limitations highlight the need for cautious interpretation of the quantitative estimates. Although MEDUSA captures the patterns and concentration of macronutrients fairly well , the model tends to overestimates DFe concentrations in regions known for persistent Fe limitation . Additionally, our macroalgae module lacks variable stoichiometry , DOC release , and explicit macroalgal grazing and erosion . The model also does not represent calcifying epibionts that commonly colonise sargassum, and their calcification can offset CDR . Furthermore, MEDUSA does not resolve coccolithophores phytoplankton types, meaning that potential reduction in calcite ballasting due to the reduction in phytoplankton concentration, are not captured. These omissions would act to further reduce CDR efficiency.

The absence of riverine nutrient inputs and coastal-shelf processes also prevents realistic simulation of Sargassum dynamics and likely underestimates the productive potential of warm-water cultivation in regions such as Southeast and East Asia where terrestrial nutrient subsidies support high macroalgal biomass in nearshore systems. Although the four taxa represented in this study span a range of thermal niches, cold and warm water taxa are associated with different cultivation systems, whereby warm water species are usually farmed using single step methods through vegetative propagation and low-cost fixed or floating raft systems in nearshore environments while cold water species require multi-step propagation from spores, hatchery-based seeding, and deployment on longlines system . These differences are not captured in this idealised framework and may mean that the modelled growth potential presented here translates differently to real-world aquaculture output depending on the region in question.

Although MEDUSA includes deoxygenation and acidification processes, the simplicity of its benthic ecosystem model cannot represent the impacts of such mechanisms on the seafloor ecosystems. While some processes may be partially represented within MEDUSA’s remineralisation scheme, a targeted evaluation of acidification dynamics was beyond the scope of this study. Our results should be interpreted as a first order approximation of the biogeochemical consequences of large-scale macroalgae cultivation, and that representing these additional processes in future model developments will be important in estimating macroalgal CDR potential

Taken together, our results reinforce growing concerns that large-scale macroalgae cultivation may offer limited net climate benefits when ecological externalities and feedbacks are considered . Despite its theoretical potential, macroalgal CDR appears constrained by nutrient bottlenecks, competition with natural carbon pumps, and low CDR capacity compared to macroalgal NPP. These findings point to the need for regionally tailored deployment, modelling and observation monitoring strategies, and explicit additionality accounting if macroalgal cultivation is to play a role in future climate mitigation portfolios.