Initialized climate model simulations have proven skillful for near-term predictability of the key physical climate variables. By comparison, predictions of biogeochemical fields like ocean carbon flux, are still emerging. Initial studies indicate skillful predictions are possible for lead-times up to six years at global scale for some CMIP6 models. However, unlike core physical variables, biogeochemical variables are not directly initialized in existing decadal preciction systems, and extensive empirical parametrization of ocean-biogeochemistry in Earth System Models introduces a significant source of uncertainty. Here, we propose a new approach for improving the skill of decadal ocean carbon flux predictions using observationally-constrained statistical models, as alternatives to the ocean-biogeochemistry models. We use observations to train multi-linear and neural-network models to predict the ocean carbon flux. To account for observational uncertainties, we train using six different observational estimates of the flux. We then apply these trained statistical models using input predictors from the Canadian Earth System Model (CanESM5) decadal prediction system to produce new decadal predictions. Our hybrid GCM-statistical approach significantly improves prediction skill, relative to the raw CanESM5 hindcast predictions over 1990-2019. Our hybrid-model skill is also larger than that obtained by any available CMIP6 model. Using bias-corrected CanESM5 predictors, we make forecasts for ocean carbon flux over 2020-2029. Both statistical models predict increases in the ocean carbon flux larger than the changes predicted from CanESM5 forecasts. Our work highlights the ability to improve decadal ocean carbon flux predictions by using observationally-trained statistical models together with robust input predictors from GCM-based decadal predictions.

Coastal oceans may play an important role in regulating the concentration of carbon dioxide in the atmosphere. Quantification of carbon fluxes at this highly dynamic land-ocean interface will aid in monitoring, reporting, and verification for marine carbon dioxide removal. Here, we use a two-step neural network approach to generate basin-wide estimates from sparse observational data in the coastal Northeast Pacific Ocean at an unprecedented spatial resolution of 1/12° with coverage in the nearshore (0 - 25 km offshore). We compiled partial pressure of carbon dioxide (pCO2) observations as well as a range of predictor variables including satellite-based and physical oceanographic reanalysis products. With the predictor variables representing processes affecting pCO2, we created non-linear relationships to interpolate observations from 1998-2019. Compared to in situ shipboard and mooring observations, our coastal pCO2 product captures broad spatial patterns and seasonal cycle variability well. A sensitivity analysis identifies that the parameters responsible for the neural network’s ability to capture regional pCO2 variability agrees with mechanistic processes. Using wind speed and atmospheric CO2, we calculated air-sea CO2 fluxes. We report an anticorrelation between net annual air-sea CO2 flux and air-sea CO2 flux seasonal amplitude and suggest the relationship is driven by regional processes. We show the inclusion of nearshore net outgassing fluxes lowers the overall regional net flux. Overall, our results suggest that the region is a net sink (-0.7 mol m-2 yr-1) for atmospheric CO2 with trends indicating increasing oceanic uptake due to strong connectivity to subsurface waters.

Increasing anthropogenic CO2 emissions to the atmosphere are partially sequestered into the global oceans through the air-sea exchange of CO2 and its subsequent movement to depth, and this collective large-scale absorption is commonly referred to as the global ocean carbon sink. Quantifying this ocean carbon sink provides a key component for closing the global carbon budget which is used to inform and guide policy decisions. These estimates are typically accompanied by an uncertainty budget built by selecting what are perceived as critical uncertainty components based on selective experimentation. However, there is a growing realisation that these budgets are incomplete and may be underestimated, which limits their power as a constraint within global budgets. In this study, we present a methodology for quantifying spatially and temporally varying uncertainties in the air-sea CO2 flux calculations and data that allows an exhaustive assessment of all known sources of uncertainties, including decorrelation length scales between gridded measurements, and the approach follows standard uncertainty propagation methodologies. The resulting standard uncertainties are higher than previously suggested budgets, but the components are consistent with previous work, and they identify how the significance and importance of key uncertainty components change in space and time. For an exemplar method (the UEP-FNN-U method) the work identifies that we can currently estimate the annual ocean carbon sink to an accuracy of ±0.72PgCyr-1 (1 standard deviation uncertainty). Due to this method having been built on established uncertainty propagation and approaches, it appears applicable to all data-product assessments of the ocean carbon sink.

Initialized climate model simulations have proven skillful for near-term predictability of the key physical climate variables. By comparison, predictions of biogeochemical fields like ocean carbon flux, are still emerging. Initial studies indicate skillful predictions are possible for lead-times up to six years at global scale for some CMIP6 models. However, unlike core physical variables, biogeochemical variables are not directly initialized in existing decadal preciction systems, and extensive empirical parametrization of ocean-biogeochemistry in Earth System Models introduces a significant source of uncertainty. Here, we propose a new approach for improving the skill of decadal ocean carbon flux predictions using observationally-constrained statistical models, as alternatives to the ocean-biogeochemistry models. We use observations to train multi-linear and neural-network models to predict the ocean carbon flux. To account for observational uncertainties, we train using six different observational estimates of the flux. We then apply these trained statistical models using input predictors from the Canadian Earth System Model (CanESM5) decadal prediction system to produce new decadal predictions. Our hybrid GCM-statistical approach significantly improves prediction skill, relative to the raw CanESM5 hindcast predictions over 1990-2019. Our hybrid-model skill is also larger than that obtained by any available CMIP6 model. Using bias-corrected CanESM5 predictors, we make forecasts for ocean carbon flux over 2020-2029. Both statistical models predict increases in the ocean carbon flux larger than the changes predicted from CanESM5 forecasts. Our work highlights the ability to improve decadal ocean carbon flux predictions by using observationally-trained statistical models together with robust input predictors from GCM-based decadal predictions.

As a contribution to the Regional Carbon Cycle Assessment and Processes phase 2 (RECCAP2) project, we present synthesized estimates of Arctic Ocean sea-air CO2 fluxes and their uncertainties from 8 surface ocean pCO2-observation products, 18 ocean biogeochemical hindcast and data assimilation models and 6 atmospheric inversions. For the period of 1985−2018, the Arctic Ocean was a net sink of CO2 of 116 ± 4 TgC yr−1 in the pCO2 products and 92 ± 30 TgC yr−1 in the models. The CO2 uptake peaks in late summer and early autumn, and is low in winter when sea ice inhibits sea-air fluxes. The long-term mean CO2 uptake in the Arctic Ocean is primarily caused by steady-state fluxes of natural carbon (70 ± 15 %), and enhanced by the atmospheric CO2 increase (19 ± 5 %) and climate change (11 ± 18 %). The annual mean CO2 uptake increased from 1985 to 2018 at a rate of 31 ±13 TgC yr−1dec-1 in the pCO2 products and 10 ± 4 TgC yr−1dec-1 in the models. Moreover, 77 ± 38 % of the trend in the net CO2 uptake over time is caused by climate change, primarily due to rapid sea ice loss in recent years. Both, the mean CO2 uptake and the trend, is substantially weaker in the atmospheric inversions. Uncertainties across all estimates are large, in the pCO2 products because of scarcity of observations and in the models because of missing processes.

\justify Several methods have been developed to quantify the oceanic accumulation of anthropogenic carbon dioxide (CO$_2$) in response to rising atmospheric CO$_2$. Yet, we still lack a corresponding estimate of the changes in the total oceanic dissolved inorganic carbon (DIC). In addition to the increase in anthropogenic CO$_2$, changes in DIC also include alterations of natural CO$_2$. Once integrated globally, changes in DIC reflect the net oceanic sink for atmospheric CO$_2$, complementary to estimates of the air-sea CO$_2$ exchange based on surface measurements. Here, we extend the MOBO-DIC machine learning approach by \citeA{keppler_mapped_2020} to estimate global monthly fields of DIC at 1$^{\circ}$ resolution over the top 1500 m from 2004 through 2019. We find that over these 16 years and extrapolated to cover the whole global ocean down to 4000 m, the oceanic DIC pool increased close to linearly at an average rate of 3.2$\pm$0.7 Pg C yr$^{-1}$. This trend is statistically indistinguishable from current estimates of the oceanic uptake of anthropogenic CO$_2$ over the same period. Thus, our study implies no detectable net loss or gain of natural CO$_2$ by the ocean, albeit the large uncertainties could be masking it. Our reconstructions suggest substantial internal redistributions of natural oceanic CO$_2$, with a shift from the mid-latitudes to the tropics and from the surface to below $\sim$200 m. Such redistributions correspond with the Pacific Decadal Oscillation and the Atlantic Multidecadal Oscillation. The interannual variability of DIC is strongest in the tropical Western Pacific, consistent with the El Ni$\tilde{n}$o Southern Oscillation.