Summary
A significant portion of atmospheric CO2 emissions are absorbed by the ocean, resulting in a decreased pH that is harmful to marine life and ecosystems — a process known as ocean acidification (OA). However, monitoring and forecasting indicators of OA is difficult due to a lack of in-situ measurements, the nonlinearity of the dynamics, and the high costs of computational numerical models. We develop a data-driven framework to model properties that drive OA in the Massachusetts Bay and Stellwagen Bank. In the first step of the framework, we train a neural network on data from a historical physics-based numerical simulation to predict 3D temperature and salinity (x,y,z) from quantities at the surface (x,y). The relationship between 2D surface properties and 3D properties is captured through the in-depth modes obtained from principal component analysis. We use this trained model to estimate real-time 4D temperature and salinity from satellite and buoy surface measurements. Then, we use standard Bayesian regression methods to estimate region-specific relationships for total alkalinity (TA) and pH as a function of temperature and salinity. The model's performance is evaluated using withheld measurements at multiple depths. Furthermore, each step of the framework includes uncertainty quantification which can be used to plan future operations and optimally place measurement stations.
