BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:The elephant in the room: Machine Learning directly into physical 
 models - Matthew Juniper
DTSTART:20260617T130000Z
DTEND:20260617T140000Z
UID:TALK242650@talks.cam.ac.uk
CONTACT:Anna Walczyk
DESCRIPTION:John von Neumann is often quoted as saying "with four paramete
 rs 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 ha
 ndful of parameters. A century later\, however\, we are happy to use physi
 cs-agnostic neural networks containing millions of parameters. What would 
 von Neumann say? How should physical modellers respond? \n\nIn this talk\,
  I will show that von Neumann's quote is more nuanced than it sounds. I wi
 ll then frame a response within a Bayesian framework\, in which physical p
 rinciples such as conservation of mass\, momentum\, and energy are treated
  as high quality prior information\, with quantified uncertainty\, express
 ed as PDEs or low order models. The information content of data can then b
 e quantified and the likelihood of different candidate models can be compa
 red after the data arrives. I will show how Bayesian inference becomes com
 putationally tractable when combined with adjoint methods. I will demonstr
 ate this through (i) assimilation of 3D Flow-MRI data in complex geometry 
 into Finite Element CFD\, (ii) assimilation of flame and acoustics data in
 to low order models of a thermoacoustic system\, 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."\n
LOCATION:LT6
END:VEVENT
END:VCALENDAR