The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Differentiable Programming for Oceanography with Patrick Heimbach - #557
Jan 31, 2022
Patrick Heimbach, a professor at the University of Texas, dives deep into the intersection of machine learning and oceanography. He discusses the challenges of simulating ocean circulation and how machine learning can significantly improve model accuracy. The importance of differentiable programming in integrating observational data with physical models is highlighted. Heimbach also explores modular oceanographic modeling and how machine learning assists in analyzing ice sheet dynamics and calving processes, showcasing a bright future for these technologies.
AI Snips
Chapters
Transcript
Episode notes
Computational Oceanography
- Computational oceanographers simulate ocean circulation on computers, similar to weather forecasting.
- Ocean simulations are computationally challenging due to smaller scales than atmospheric models, requiring large supercomputers.
Ocean Model Uncertainties
- Ocean models have uncertainties in constitutive equations, empirical relationships for properties like viscosity.
- Subgrid scale parametrizations address unresolved smaller-scale processes, introducing another source of uncertainty.
Machine Learning for Subgrid Processes
- Machine learning can help by learning parameters from separate datasets or learning alternative functions altogether.
- This addresses computational expense and uncertainty in parameterized functions used for subgrid scale processes.