Latest machine learning developments for the LHCb experiment

by Julian Garcia Pardinas (CERN)

Wednesday 10 Jul 2024, 11:00 → 12:00 Europe/Zurich

40/S2-D01 - Salle Dirac (CERN)

Starting from Run 3, the LHCb experiment has transitioned to a fully software-based trigger leveraging a heterogeneous GPU+CPU architecture, which offers unprecedented opportunities for physics exploration. Simultaneously, the LHCb detector has been fully upgraded to handle higher luminosities and collect a significantly larger volume of data. These advancements have introduced new challenges for event reconstruction and filtering in the trigger and for Data Quality Monitoring (DQM), and have emphasised the need for ultra-fast simulation techniques to produce ever-larger simulation samples. Modern machine learning (ML) techniques offer unique opportunities to address these challenges and to extend the reach of the experiment, leveraging the software and hardware upgrades. In this talk, I will begin by summarising the latest ML developments already implemented for LHCb in Run 3, including the use of Lipschitz neural networks in the trigger and the use of Generative Adversarial Networks (GANs) for fast simulation. Subsequently, I will discuss various ongoing R&D projects connected to trigger applications and to DQM. The first topic includes the use of Graph Neural Networks (GNNs) for track finding in the VELO detector and for a Deep-learning based Full Event Interpretation (DFEI), as well as the application of Normalised Autoencoders (NAE) for beyond-the-Standard-Model anomaly detection. The second topic will focus on a novel application of Reinforcement Learning from Human Feedback (RLHF) towards automating DQM tasks in the control room and in subsequent offline stages.

 

Julián García Pardiñas is a Senior Research Fellow at CERN, specialised in the intersection between machine learning and experimental particle physics. He serves as co-coordinator of the IML Working Group, representing the LHCb experiment. Julián received his PhD in Experimental Particle Physics at the University of Santiago de Compostela, after which he worked as a postdoctoral researcher at the University of Zürich and later as a Marie Skłodowska-Curie Postdoctoral Fellow at the University of Milano-Bicocca. His recent research focuses on the application of Graph Neural Networks for event reconstruction, leading the R&D project on Deep-learning based Full Event Interpretation (DFEI) for LHCb, and on the pioneering application of Reinforcement Learning from Human Feedback for Data Quality Monitoring in HEP experiments.

 

Coffee will be served at 10:30.

ML4LHCb_CERNSeminar_JGP_100724.pdf

Recording

Video preview