Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Novel machine learning applications at the LHC

Javier M. Duarte

TL;DR

ML is addressing core LHC challenges by enabling improved classification, fast detector simulation, data-driven unfolding, and broad anomaly detection. The paper surveys results across ATLAS, CMS, ALICE, and LHCb, highlighting architectures such as ParticleNet, ParT, GloParT, and UParT, along with robust training and calibration schemes. Key contributions include substantial gains in jet-tagging performance, fast calorimeter-shower simulation via VAEs and GANs, OmniFold-based unbinned particle-level unfoldings for Z+jets, and diverse anomaly-detection approaches like VAE-QR and AXOL1TL. Together, these methods reduce computational costs, expand measurable observables, and enhance sensitivity to new physics while addressing systematic uncertainties and robustness.

Abstract

Machine learning (ML) is a rapidly growing area of research in the field of particle physics, with a vast array of applications at the CERN LHC. ML has changed the way particle physicists conduct searches and measurements as a versatile tool used to improve existing approaches and enable fundamentally new ones. In these proceedings, we describe novel ML techniques and recent results for improved classification, fast simulation, unfolding, and anomaly detection in LHC experiments.

Novel machine learning applications at the LHC

TL;DR

ML is addressing core LHC challenges by enabling improved classification, fast detector simulation, data-driven unfolding, and broad anomaly detection. The paper surveys results across ATLAS, CMS, ALICE, and LHCb, highlighting architectures such as ParticleNet, ParT, GloParT, and UParT, along with robust training and calibration schemes. Key contributions include substantial gains in jet-tagging performance, fast calorimeter-shower simulation via VAEs and GANs, OmniFold-based unbinned particle-level unfoldings for Z+jets, and diverse anomaly-detection approaches like VAE-QR and AXOL1TL. Together, these methods reduce computational costs, expand measurable observables, and enhance sensitivity to new physics while addressing systematic uncertainties and robustness.

Abstract

Machine learning (ML) is a rapidly growing area of research in the field of particle physics, with a vast array of applications at the CERN LHC. ML has changed the way particle physicists conduct searches and measurements as a versatile tool used to improve existing approaches and enable fundamentally new ones. In these proceedings, we describe novel ML techniques and recent results for improved classification, fast simulation, unfolding, and anomaly detection in LHC experiments.
Paper Structure (6 sections, 10 figures)

This paper contains 6 sections, 10 figures.

Table of Contents

  1. Introduction
  2. Improved classification
  3. Fast simulation
  4. Unfolding
  5. Anomaly detection
  6. Summary and outlook

Figures (10)

  • Figure 1: The $c$ and light jet rejection of the different ATLAS flavor tagging algorithms over time in Monte Carlo simulation (left) ATLAS-PLOTS-FTAG-2023-01. The jet soft drop mass distribution after two-prong and light-flavor ParticleNet tagger selections, illustrating the prominent $W(qq)$ resonance (right) CMS-PAS-EXO-24-007.
  • Figure 2: Full set of training jet classes for GloParT (left). Distributions of the $b\overline{b}$-candidate jet mass in the VBF signal region (center). Upper limits on the inclusive $HH$ production cross section as a function of $\kappa_{2V}$ (right).
  • Figure 3: The ROC curves for UParT trained with R-NGM or nominal samples and evaluated with R-NGM or nominal samples (left). Median of the raw regressed jet energy response for UParT and ParticleNet (right).
  • Figure 4: Transformer architecture for particle identification in ALICE (left). Layered blocks are applied separately to each vector in a set. Single blocks are applied to their input as a whole. Domain-adversarial neural network training setup (right).
  • Figure 5: Flow chart of a CENNT (upper left) and SANNT (lower left). Negative log of the profile likelihood $-2\Delta \ln L$ as a function of $r_s$, taking into account (red) all and (blue) only the statistical uncertainties in $\Delta r_s$ (right). The results as obtained from CENNT (SANNT) are indicated by the dashed (solid) lines.
  • ...and 5 more figures