Abstract

This talk will show how to combine physics-based modelling with data-driven machine learning. The method assimilates data into physics-based models using Bayesian inference accelerated with adjoint methods. In other words, it takes a qualitatively-accurate physics-based model and renders it quantitatively-accurate by assimilating data.

If the physics of the problem is known, this method is better than assimilating data into a Neural Network. This is because the physics-based model requires less training data, is interpretable, and extrapolates to situations that share the same physics. This framework also rigorously compares physics-based models against each other, allowing the best model to be selected.

This work is inspired by David MacKay’s book on information theory, inference, and learning algorithms (https://www.inference.org.uk/itprnn/book.pdf). I will present applications in Magnetic Resonance Imaging of flows (flow-MRI) and thermoacoustic oscillations in rockets and aicraft engines. An overview of the talk can be found at https://mpj1001.user.srcf.net/MJ_inference.html.