The elephant in the room: fluid dynamics in the age of machine learning

Add to your list(s) Download to your calendar using vCal

• Matthew Juniper, CUED
• Friday 19 January 2024, 16:00-17:00
• MR2.

If you have a question about this talk, please contact Professor Grae Worster.

John von Neumann is often quoted as saying “with four parameters I can fit an elephant, and with five I can make him wiggle his trunk.” The implication seems to be that physical models should contain only a handful of parameters. A century later, however, we are happy to use physics-agnostic neural networks containing millions of parameters. What would von Neumann say? How should physical modellers respond?

In this talk, I will show that von Neumann’s quote is more nuanced than it sounds. I will then frame a response within a Bayesian framework, in which physical principles such as conservation of mass and momentum are treated as high quality prior information, with quantified uncertainty, expressed as PDEs or low order models. The information content of data can then be quantified and the likelihood of different candidate models can be compared after the data arrives. I will show how Bayesian inference becomes computationally tractable when combined with adjoint methods. I will demonstrate this through (i) assimilation of 3D Flow-MRI data into Finite Element CFD , and (ii) assimilation of acoustics data into low order models, which are then ranked by their likelihood given the data. The main message of the talk is “keep the physics in the model if you can.”

This talk is part of the Fluid Mechanics (DAMTP) series.

This talk is included in these lists:

• All CMS events
• All Talks (aka the CURE list)
• CMS Events
• CamBridgeSens
• Cambridge talks
• Combined External Astrophysics Talks DAMTP
• Cosmology, Astrophysics and General Relativity
• DAMTP Fluids Talks
• DAMTP info aggregator
• Fluid Mechanics (DAMTP)
• Interested Talks
• Life Science Interface Seminars
• MR2
• SJC Regular Seminars
• School of Physical Sciences
• Talks related to atmosphere and ocean dynamics and climate science
• Trust & Technology Initiative - interesting events
• bld31

Note that ex-directory lists are not shown.