A statistical field theory approach to physics-informed machine learning – report

The RSS North Eastern local group held an on-line seminar talk on ‘A statistical field theory approach to physics-informed machine learning’ on Tuesday 19 April, 2022 at 4pm (BST) via Zoom.

The speaker was Ilias Bilionis, associate professor of mechanical engineering at the School of Mechanical Engineering at Purdue University in the USA. Ilias Bilionis has established the Predictive Science Laboratory (PSL) whose mission is to create artificial intelligence technologies that accelerate the pace of engineering innovations with particular focus in application such as technical systems (eg electric machines, high-performance materials, medical devices) and sociotechnical systems (eg smart buildings, extra-terrestrial habitats).

lias Bilionis presented an overview of the research portfolio of his lab. Then he presented modern applications on sociotechnical systems. Mathematically, these problems are smoothing and calibration problems involving physical fields (eg strains, stresses, temperatures, pressures) that satisfy PDEs with potentially unknown initial and boundary conditions, parameters, random external excitations, or even missing physics, and hence they can be addressed by building predictive models combining existing physical knowledge with noisy data.

To address the statistical challenges, he proposed a novel unifying Bayesian framework inspired by statistical field theory which involves constructing a prior probability on the space of physical fields that assigns higher probability to fields that satisfy the physical equations.

Author: Aamir Khan

Load more