by Dr Francesco Vaselli (Scuola Normale Superiore & INFN Pisa (IT))

→ Europe/Zurich
593/R-010 - Salle 11 (CERN)

593/R-010 - Salle 11

CERN

50
Show room on map
Description

Abstract

Generative machine learning has moved from mainly an interest of industry to a practical scientific tool that is reshaping how high-energy physics simulates, searches, and stores its data. This class introduces the field from the ground up and follows it to today's frontier. The first lecture builds intuition for the core model families, i.e., GANs, variational autoencoders, and normalizing flows, and surveys where they are already used in physics, from fast detector simulation to anomaly detection and unfolding. The second lecture develops flow matching, a recent and remarkably effective way to build generative models by learning the vector field of an ordinary differential equation, and connects it to diffusion models and the latest developments in the field. A hands-on session puts these ideas into practice, building flow matching from scratch for detector simulation. Prior experience with Python and PyTorch is recommended. No GPU required, everything runs on your laptop or Colab. You are welcome to attend as a small group if you prefer.

Lecturer Bio 

Francesco Vaselli is a physics PhD candidate at Scuola Normale Superiore and INFN Pisa, working with CERN on generative machine learning for high-energy physics. He is the lead developer of FlashSim, the first end-to-end ML-based fast simulation adopted by a major LHC experiment (CMS), and first author on the first proof of concept of flow matching to real-time anomaly detection at the LHC's 40 MHz trigger. His interests span surrogate ML models of detector response for fast simulation, anomaly detection, and data compression. He lectures regularly at the CERN School of Computing.