Highlights
- Physics-guided ML forecasts surface pCO2 and pH along a Western Mediterranean VOS line.
- Day-ahead pCO2 is predicted with μatm-level RMSE; pH behaves nearly deterministically.
- Boosted trees and sequence models retain skill under strict, deployable forecast conditions.
- Explainable AI recovers dominant thermal control and air–sea CO2 gradient drivers.
- Improved pCO2 forecasts directly reduce uncertainty in air–sea CO2 flux estimates.
Abstract
We introduce a hybrid, physics-guided machine-learning system for forecasting and explaining surface marine carbonate changes along a fixed Volunteer Observing Ship route between Gibraltar and Barcelona from 2019 to 2024. The dataset includes more than 90 high-frequency transects collected under ICOS/SOOP standards, containing underway pCO2/fCO2, pH (measured and derived), sea-surface temperature, and salinity. After applying consistent quality control and harmonizing the data in time and space, we combine physics-based carbonate diagnostics—such as the thermal/non-thermal decomposition (FASS) and first-order Taylor attribution of temperature, salinity, total alkalinity, and dissolved inorganic carbon sensitivities—with time-aware models including linear regressions, boosted trees, and sequence networks (1-D CNNs and LSTMs) trained on historical windows. We evaluate generalization and uncertainty through chronological splits, leave-one-year-out tests, and year-wise bootstrap sampling. With all current predictors available, day-ahead pH and pCO2 predictions reach near-optimal skill; pH behaves almost deterministically, while pCO2 achieves RMSE on the order of a few μatm. Even under stricter forecast conditions without real-time carbonate chemistry, boosted trees and sequence models maintain strong performance by exploiting persistence and seasonal timing. Model-explanation tools (SHAP, partial dependence) recover the expected carbonate drivers, highlighting dominant thermal effects and key roles of seawater CO2 state and air–sea gradients. Spatial–temporal diagnostics reveal amplified summer pCO2 peaks in the Alboran/northern Morocco region and out-of-phase pH patterns. Predicted fields are converted to air–sea CO2 flux using standard solubility and gas-transfer formulations, and propagated uncertainties show that improving pCO2 accuracy directly reduces flux uncertainty. The resulting air–sea CO2 fluxes exhibit a pronounced seasonal cycle, with summer outgassing reaching several mmol m-2 d-1 and winter uptake of comparable magnitude along the transect, while interannual variability dominates over 2019–2024 and no statistically robust long-term trend is detected; typical flux uncertainties are on the order of 0.1–0.2 mmol m-2 d-1. Altogether, this delivers an explainable, uncertainty-aware system ready for deployment, linking forecast skill to process understanding and CO2 exchange in a climate-sensitive corridor.
Continue reading ‘Physics-guided machine-learning forecasting and analysis of carbonate changes in the surface Western Mediterranean’



