An advection-based neural network for reconstructing mesoscale surface currents from tracers

Ocean mesoscale dynamics play a major role in phytoplankton productivity and other physically driven ecological processes. Large-scale surface patterns of chlorophyll and sea surface temperature are well-described thanks to remote sensing observations. However, the role of submesoscale to mesoscale dynamics in chlorophyll distribution remains unclear. Through nonlinear scale interactions, horizontal surface circulation shapes mesoscale tracer variability, but the limited characterization of surface currents at these scales complicates the evaluation of mesoscale-driven ecological variability. This limitation has been partially addressed using high-resolution ocean color and sea surface temperature datasets, which capture fine-scale tracer variability and allow the possibility of inferring surface circulation through an inverse approach. In the Bay of Biscay, these mesoscale features have more relevance due to complex bathymetry and a large-scale poleward current, generating slope water oceanic eddies (SWODDIES). Although these dynamical features have been described, the increased availability of high resolution tracer fields provides a source to resolve mesoscale surface circulation in this region.

In this study, we explore the capabilities of a neural network to reconstruct ocean surface currents and mesoscale features based on the advection visible in tracer fields and auxiliary variables such as altimetry products. We implemented a convolutional encoder-decoder architecture trained with an advection-based cost function that minimizes the difference between advected tracer fields and their observed distributions, while incorporating additional constraints to enforce physical boundaries and flow properties such as divergence. For training the network, in this stage we used model daily fields from the Irish-Iberian Biscay (IBI-MFC) reanalysis and non-assimilative biogeochemical hindcast, testing different configurations with chlorophyll, temperature and salinity fields from 1993 to 2019. The network was evaluated by reconstructing zonal and meridional fields for the year 2020. The results show consistent patterns of both velocity components along selected transects, with an RMSE of 0.039 m/s for the total velocity magnitude compared to a mean velocity of 0.059 m/s obtained from independent observations. This model provides an alternative for combining high-resolution tracer observations with physics-informed constraints with the aim of reconstructing surface circulation at mesoscale.