What information can a neural network learn from diverse physics datasets?

Tuesday 20 Oct 2026, 15:00 → 17:00 US/Central

WH11NE Sunrise

15:00 → 16:00 What information can a neural network learn from diverse physics datasets?

While scientific analyses require thoughtfully curated datasets to test a given hypothesis, modern machine learning methods thrive when given vast datasets with little to no filtering. How, then, should physicists curate input datasets and training strategies for machine learning models designed to facilitate scientific discoveries? Machine learning models have the potential to learn highly useful representations of physics concepts by synthesizing multiple heterogeneous data sources, but best practices for developing such models using many diverse inputs (e.g. across modalities, detectors, or even scientific domains) are still not fully understood. This talk will discuss my recent work with collaborators on multimodal scientific foundation models, evaluations of model transfer across physics detectors and disciplines, and information-theoretic approaches for combining wide-ranging information content into broadly generalizable embeddings.

Speakers: Prof. Mariel Pettee (University of Wisconsin--Madison), Mariel Pettee