Reconstructing surface ocean carbon flux from physical parameters using deep learning methods

The global ocean carbon sink is a critical component of the Earth’s climate but current models are limited in their predictive capability because of the high computational cost that a biogeochemical model, required to simulate air-sea CO₂ fluxes, entails. In this study, we investigate the ability of deep learning (DL) methods to reconstruct monthly air-sea CO₂ fluxes using only the physical output of an ocean circulation model, thereby exploring a data-driven alternative to a costly biogeochemical model. We used a collection of global simulations from the ocean biogeochemistry model NEMO-MOPS at a horizontal resolution of 0.25°, which differ in their atmospheric forcing components. The simulations span 61 years (1958-2018), providing a long, high-resolution dataset that captures substantial interannual to decadal variability. Our objectives are threefold: (1) to assess how accurately DL models can reconstruct CO₂ fluxes from physical variables alone, (2) to evaluate the generalization of these models across unseen years and forcing regimes, and (3) to identify the relative importance of physical drivers and their temporal lags in predicting air-sea CO₂ exchange. To this end, we train a point-wise Long Short-Term Memory (LSTM) network augmented with a temporal attention mechanism, which enables dynamically weight information from different time steps, to predict present-month CO₂ fluxes. To this end, we use eight physical predictors from the current month and the preceding five months. Standard regression metrics indicate an overall accurate reconstruction even though extreme CO₂ outgassing events are often underestimated. Seasonal and interannual variations are mostly well reconstructed across different ocean regimes. Spatial patterns are also well reconstructed, even though the DL model is trained only with local features (not including latitude and longitude information). This is a promising result in terms of generalizing to other physical settings, which we aim to test in future experiments. We finally interpret the learned relationships, by computing the Shapley values to quantify the contribution of each physical driver across time lags. Overall, our work highlights the potential of combining DL based techniques and explainable AI as a scalable and transparent complement to traditional Earth system modeling for studying ocean carbon cycle dynamics.