Presenter: Tristan Horner (Woods Hole Oceanographic Institution)

Description:
Barium (Ba) is a trace metal that exhibits a similar marine distribution to the major nutrient silicon (Si, as silicic acid). This similarity underpins a number of proxies whereby measurement of Ba stands in for Si. However, the global extent of—and spatial variability within—the Ba–Si relationship is unknown owing to the relative sparsity of dissolved Ba concentration (d[Ba]) measurements in seawater. A global assessment of the Ba–Si relationship thus requires developing a model that can predict d[Ba] in regions of the ocean where there are currently few observations. Here, we simulated the global oceanographic distribution of d[Ba] at 1x1-degree resolution using a Gaussian process machine learning regression model. Machine learning models ideally suited to this application because they do not require explicit parameterizations of the processes governing marine Ba cycling. Model development and testing comprised three stages. First, we ingested datasets from GEOTRACES that were complete for 12 core features that would serve as predictors for simulating d[Ba] (geospatial information, nutrients, surface chlorophyll, etc.). Second, the optimal feature set for predicting d[Ba] was identified by creating, training, and testing 2¹² models, corresponding to each possible combination of core features. Third, global d[Ba] was simulated by feeding data from the World Ocean Atlas into select predictor models whose accuracy was assessed via cross-validation against withheld d[Ba] data from the Indian Ocean. This assessment reveals that several of the machine learning models can accurately simulate d[Ba], which we use to explore regional variations in the Ba–Si relationship and to constrain the total Ba budget of the ocean.