Bio:
Dr. Navarra is a postdoctoral researcher with a strong interdisciplinary background in ocean science, climate physics, and machine learning. Holding a PhD in Ocean Science and Engineering from Georgia Tech and currently affiliated with Princeton University, his research centers on understanding and modeling interactions between ocean physics, biogeochemical cycles, and marine ecosystems in a changing climate.
His expertise lies in applying advanced causal inference methods, Bayesian statistics, and Physics-Informed Neural Networks (PINNs) to key questions such as chlorophyll variability drivers, carbon export to the twilight zone, and fisheries dynamics. With a robust publication record in top journals, he aims to leverage data-driven and physics-constrained approaches to improve the predictive skill of Earth system models from physics to ecosystems.
Absract：
The ocean biological carbon pump is a fundamental driver of global biogeochemical cycles and ecological processes, most notably the sequestration of organic carbon. However, complex physical biological interactions and the inherent difficulty of comprehensive oceanic sampling result in significant spatial and temporal uncertainties in particle fields. In this study, we have reconstructed global particle size distributions (PSDs) by leveraging a supervised machine learning framework to extrapolate a sparse global compilation of in situ Underwater Vision Profiler 5 (UVP5) measurements.
We specifically demonstrate that Gaussian Process Regression (GPR) is a robust and highly effective approach for handling the sparse nature of optical particle datasets, providing a sophisticated means of mapping particle size distribution (PSD) parameters and biovolume of the water column. Our results reveal clear global patterns: polar regions exhibit elevated particle biovolume and flatter PSD slopes, signaling a dominance of larger, rapidly sinking particles.
By integrating these machine learning derived reconstructions with the PRISM physical model, we produce, depth-resolved estimates of flux or organic matter and remineralization fields coefficients. This hybrid approach significantly improves the characterization of spatial variability in critical, yet hard-to-observe, biogeochemical fields such as the vertical flux of organic carbon and remineralization rates. Our findings underscore the power of combining statistical learning with physical modeling to transform sparse observations into a predictive, global understanding of marine particle dynamics.
