Correction of sea surface biases in the NEMO ocean general circulation model using neural networks

Abstract

The atmospheric forcing and the heat exchanges between the ocean and the atmosphere represent one of the major sources of uncertainty for numerical ocean reconstructions and predictions, together with inaccuracies in vertical mixing and solar radiation penetration. Air-sea heat fluxes may suffer from inaccuracies in meteorological fields, sea surface variables, and bulk formulations, which have a strongly nonlinear dependence on the ocean state. Here, state-dependent errors in heat fluxes are learned by artificial neural networks (ANNs) from a dataset of heat flux correction terms, derived in turn from previous sea surface temperature nudging experiments. The pre-trained model predictors include stationary fields, atmospheric forcing data, ocean state, and stratification indices. Variable importance scores emphasize the dependence of air-sea heat flux errors on wind forcing. The pre-trained heat flux correction model is then used to adaptively correct fluxes online, in a series of global ocean experiments performed with the NEMO version 4 (Nucleus for European Modelling of the Ocean) ocean general circulation model, augmented with ANN inference capabilities in Fortran90. Results indicate the positive impact of the correction procedure, beyond the training period, e.g. in independent observation-poor and -rich periods, leading to the same dynamic and subsurface signature as in nudging experiments. Prediction experiments also indicate the method's potential for use in operational forecast applications. The method may also be adopted in coupled long-term reanalyses, long-range predictions, and projections.