Abstract
A significant portion of atmospheric emissions is absorbed by the ocean, resulting in acidified seawater and altered carbonate composition that is harmful to marine life. Despite detrimental effects, assessing ocean and coastal acidification (OCA) is difficult due to the scarcity of in situ measurements and the high costs of computational modeling. We develop a parsimonious data-driven framework to model indicators of OCA and test it in the Massachusetts Bay and Stellwagen Bank, a region with fishing and tourism industries affected by OCA. First, we trained a neural network to predict in-depth fields for temperature and salinity using surface quantities from satellites and in situ measurements . The relationship between 2D surface and 3D properties is captured through the in-depth modes and coefficients obtained from principal component analysis applied to a high-resolution historical reanalysis data set. Next, we used Bayesian regression methods to estimate region-specific relationships for in-depth total alkalinity (TA), dissolved inorganic carbon (DIC), and aragonite saturation state as functions of temperature, salinity, and chlorophyll. Lastly, 4D daily field predictions are generated from surface measurements with a spatial resolution of 4 km horizontally and 45 sigma levels vertically. The model’s performance is evaluated using withheld measurements across depths, locations, and seasons with RMSEs of 1.59°C, 0.31 PSU, 37.54 mol, 39.40 mol, and 0.42 for temperature, salinity, TA, DIC, and , respectively, at one withheld location. The framework is useful for understanding OCA and includes uncertainty quantification for future planning and optimal sensor placement.
Plain Language Summary
About a quarter of carbon dioxide emissions in the atmosphere is absorbed by the oceans. When this carbon dioxide dissolves in seawater, it results in ocean acidification (OA). A useful indicator of OA is the saturation state of aragonite, a type of calcium carbonate used by organisms that form shells. However, understanding the effects of OA is difficult due to the lack of observations and the high cost and complexity of modeling. We present a data-driven approach to model carbonate chemistry using readily available observations from satellites and low-cost sensors. Given surface measurements of temperature, salinity, and chlorophyll, our machine learning model produces temperature, salinity, total alkalinity, dissolved inorganic carbon, and aragonite saturation state covering spatial (latitude, longitude, and depth) and temporal domains for these variables. Compared to withheld observations, our model achieved reasonable accuracy across many seasons and depths, a level of resolution not matched by other models for the same set of inputs. Our model is useful for monitoring, decision-making, and future planning.
Key Points
- We present a data-driven approach to rapidly model 4D carbonate chemistry fields given readily available surface observations
- By using training data from both physics simulations and field observations, the model achieves very high resolution with minimal inputs
- The step-by-step method can be reproduced in other regions or for new data, and includes uncertainty quantification for decision-making
Champenois B., Bastidas C., LaBash B. & Sapsis T. P., 2025. Data‐driven modeling of 4D ocean and coastal acidification in the Massachusetts and Cape Cod Bays from surface measurements. Journal of Geophysical Research: Biogeosciences 130(6): e2024JG008465. doi: 10.1029/2024JG008465. Article.
