Novel characterization and correction of ocean model biases using sparse observations: application to surface ocean radiocarbon

Many oceanic properties are still only observed with hydrographic cruises that produce sparse datasets in space and time. Estimation of spatial and temporal variations in these properties thus generally requires the use of ocean models, which may include biases. We construct a bias correction technique that applies machine learning to ocean model simulations and sparse observations, taking surface ocean measurements of radiocarbon in dissolved inorganic carbon as a case study. The technique first applies clustering to ocean model simulations to identify coherent regions of radiocarbon content and change over time. Then it uses the sparse observations available to characterize ocean model biases in the identified clusters. Fitting a particular shape to the temporal variation of the biases, it calculates a bias correction field that can be applied to the original model output. The method considers the distribution of available data to ensure there is enough data in each cluster to characterize the bias and it uses fuzzy clustering methods to avoid discontinuities in the bias correction field. We apply the method using three models, and test the method by analysis of pseudodata from the models. The results show reductions in biases in most regions but some cases of overcorrection. We use the bias-corrected surface ocean radiocarbon to calculate air-sea fluxes of radiocarbon and discuss the impact on ocean radiocarbon inventories and radiocarbon in atmospheric CO₂.