The Elephant in the Room: Fluid Dynamics in the age of Machine Learning

"With four parameters I can fit an elephant, and with five I can make him wiggle his trunk," said John von Neumann in a powerful exhortation that physical models should contain only a handful of parameters. A century later, we seem 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 frame a response within a Bayesian framework, in which physical principles such as conservation of mass and momentum are expressed as high quality prior information with quantified uncertainties. I will show how Bayesian inference becomes computationally tractable when combined with adjoin methods, and demonstrate this through assimilation of 3D Flow-MRI data directly into CFD, and selection of models in an acoustic problem.