ODSL Forum: Topological Data Analysis for Jet Physics

by Dawson Thomas (Harvard)

Friday 20 Oct 2023, 14:00 → 15:00 Europe/Berlin

Basement Seminar Room (Origins Building)

Abstract: Particle physicists think of the universe as being made up of a certain set of fundamental building blocks. However, only a few of these particles are stable enough for direct observation. The rest decay into lighter particles, which themselves subsequently decay and radiate. The result is a collimated, messy spray of particles known as a jet. For any subsequent physics analysis, then, it is imperative that we can identify which particle initiated a given jet – a classification task known as jet-tagging. Consequently, jet tagging has emerged as one of the most lively areas for the development of machine learning within physics.

Inspired by favorable robustness and stability properties, we present an approach to jet physics which leverages methods from computational topology. Topological representations of jets not only respect fundamental physical symmetries, but also capture structure across multiple scales. In this talk, we explore the advantages these representations have to offer in both supervised and unsupervised settings. We’ll see that sparse taggers trained on topological representations can perform on par with deep, heavily parametrized approaches. We further introduce a distance in the space of jets defined by the 2-Wasserstein distance between their respective persistence diagrams and see how distance-preserving embeddings organize the jets along physically meaningful parameters in a way unique to our metric. Join us to see how these results give way to an exciting new perspective on data-driven collider phenomenology and how experimental particle physics stands to benefit from topological techniques.

Paper: https://ml4physicalsciences.github.io/2022/files/NeurIPS_ML4PS_2022_176.pdf