An interpretable machine learning approach for alkalinity reconstruction in the Mediterranean Sea

Ocean acidification has significant impacts on marine ecosystems and human activities, and its understanding relies on an accurate characterization of the marine carbonate system, in which alkalinity plays a central role. We propose a Machine Learning (ML) approach based on Genetic Programming (GP) to model alkalinity and apply this framework to the surface layers of the Mediterranean Sea. Our framework produces interpretable equations that capture alkalinity typical patterns and its finer-scale variability by inferring its relation with key physical and biogeochemical variables. Results, supported by quantitative metrics and visual analyses, demonstrate that our method reliably reproduces the spatio-temporal variability of alkalinity with a high level of predictive accuracy when compared with in situ observations. Moreover, we use the derived alkalinity equations to produce gap-free 2D surface alkalinity maps using satellite data. The maps correctly capture spatial gradients, seasonal patterns, and riverine contributions, reinforcing the robustness of the proposed approach.