The impact of forcing and initialisation on physical-biogeochemical simulations of the Southern Ocean

The Southern Ocean is a major sink for anthropogenic carbon dioxide (CO₂). The pattern and magnitude are set by an interplay of biological activity and physical processes, which leads to strong variability of CO₂ fluxes on different spatial and temporal scales, and complex interactions and feedbacks. Observational coverage in the past has been too sparse to resolve all relevant processes, while the complexity of the processes involved is also a challenge for models. Here, we explore the potential of the NEMO-MEDUSA model for creating accurate reanalyses of the Southern Ocean CO₂ sink, investigating the influence of forcing and initial conditions. NEMO-MEDUSA forms the ocean component of the UK Earth System Model (UKESM1), which contributes coupled simulations to CMIP6. We first run ocean-only experiments forward from 1980, with surface forcing from the coupled runs replaced by ERA-interim fluxes. Subsequently, we replace the initial model fields of dissolved inorganic carbon (DIC) and alkalinity with climatologies derived from observations. We find that imposing realistic forcing increases the agreement between simulated and observed water column structure. In particular, intermediate and mode water masses are too buoyant in the UKESM1 runs but this is corrected with realistic forcing within 5-10 years, and remains improved. The seasonal cycle of air-sea CO₂ fluxes is also aligned with observations in terms of magnitude and phase, unlike the UKESM1, but the net CO₂ uptake is larger than observed. We improve this bias in CO₂ flux when initialising the ocean-only runs with climatologies of DIC and alkalinity, by providing more realistic magnitudes of these variables. Still, some regional differences to the climatologies persist over the simulation period. Our findings highlight the substantial impact of the choice of initial conditions on simulating the Southern Ocean CO₂ sink and its variability, but also show the need for further improvements to reduce model biases, which could be approached, e.g., with data assimilation.