Detall

Conferenciant: Prof. Dirk Husmeier (University of Glasgow).

Resum: I will discuss parameter estimation in complex physical systems. For illustration, I will be looking at two examples taken from fluid dynamics and soft-tissue mechanics, but the methods generalise to other physical systems as well. I will first discuss the standard statistical approach to parameter estimation, and explain why this is computationally challenging or even intractable for complex physical systems. I will then discuss an alternative method from machine learning, based on emulation and Gaussian processes. I will apply this framework to model calibration and parameter estimation, and will discuss forward versus inverse uncertainty quantification. If time permits, I will conclude with an outlook on very recent work on message passing and graph neural networks.