Deep transfer learning for global ocean surface pCO2 reanalysis

Ocean surface partial pressure of carbon dioxide ( p CO 2 ) is essential for understanding the global carbon cycle, monitoring ocean health, and predicting their future changes. However, the current observational network poses a challenge for global ocean carbon cycle monitoring. Recent studies have used physical oceanographic datasets and feed-forward neural networks to generate global p CO 2 reanalysis products, but they are lack of considering geographically varying physical variables- p CO 2 relationships. To address this, this study developed SNU- p CO 2 , a deep learning model integrating a climate model with transfer learning to generate a global ocean surface p CO 2 reanalysis. Given sparse observational data, SNU- p CO 2 was first trained on Community Earth System Model version 2 Large-Ensemble simulations to learn ocean surface p CO 2 dynamics, then fine-tuned with reanalysis fields and Surface Ocean CO 2 Atlas (SOCAT) observations. A conservative SOCAT-based 10-fold cross-validation showed that SNU- p CO 2 achieved competitive predictive accuracy relative to existing data-driven p CO 2 products, with an average Pearson correlation coefficient of 0.91 and a root mean square error of 13.57 μ atm. Independent validation against Lamont-Doherty Earth Observatory and SOCAT Flag E observations further confirmed the robust performance of SNU- p CO 2 , with RMSE values of 20.75 and 22.94 μ atm, respectively. In particular, SNU- p CO 2 exhibited superior performance in regions with sparse direct observations but relatively dense surrounding data, effectively leveraging the influences of the spatially adjacent observations. The model also reproduced physically consistent p CO 2 anomaly patterns associated with El Niño and La Niña, accurately capturing dominant physical drivers and phase-dependent processes across the equatorial Pacific. We additionally quantified the uncertainty of the estimated p CO 2 using the Monte Carlo Dropout. The uncertainty patterns aligned with RMSE distributions, enabling both the assessment of prediction confidence and the identification of regions requiring enhanced observational coverage and model refinement. This reanalysis supports studies of the carbon cycle, air-sea CO 2 fluxes, climate, and biogeochemical processes.