Physics-Driven Machine Learning for Applications in Geophysical Fluid Dynamics
Name
champenois-bchamp-phd-meche-thesis-2026.pdf
Size
154.46 MB
Format
Adobe PDF
Checksum (MD5)
5bc740685cfb50d4f50d603e372add24
Author(s)
Champenois, Bianca
Advisor(s)
Sapsis, Themistoklis P.
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Modeling turbulent and geophysical fluid dynamics is challenging due to their nonlinear, multiscale, and chaotic nature. High-resolution simulations are computationally expensive, while real-world observations are sparse, noisy, and irregular. Bridging these gaps is critical for improving prediction and understanding of ocean and climate systems, particularly in the context of global warming and ocean acidification. This thesis develops machine learning approaches for geophysical fluid dynamics. The first part of this thesis presents a framework for producing real-time maps with uncertainty quantification of coastal ocean acidification in the Massachusetts and Cape Cod Bays, using surface in situ and satellite measurements. A temporal convolutional network trained on reanalysis data infers subsurface temperature and salinity from surface observations. These predictions are then used in region-specific Bayesian regression models to estimate dissolved inorganic carbon and alkalinity, which can be used to quantify aragonite saturation. The second part of this thesis introduces a data-driven method for Lagrangian data assimilation to predict ocean flow from sparse drifter trajectories. A convolutional autoencoder learns a reduced representation of simulation or reanalysis data, and Bayesian optimization in the latent space minimizes the discrepancy between observed trajectories and model predictions. The third part of this thesis develops an active sampling algorithm to improve the prediction of extreme events in learning tasks involving very large, high-dimensional datasets. The algorithm selects the most informative training points to reduce uncertainty and increase accuracy in the tails of the distribution, validated on a machine learning climate model. Together, these methods provide efficient, data-driven tools for integrating physics-based models with observations, reconstructing ocean states from limited measurements, and improving extreme event prediction. They demonstrate a pathway toward practical, real-time monitoring and prediction systems that support both scientific discovery and practical decision-making.
MIT Department
Massachusetts Institute of Technology. Department of Mechanical Engineering
Massachusetts Institute of Technology. Center for Computational Science and Engineering
Terms of Use
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
Copyright retained by author(s)
Persistent DSpace Link
