Particle identification with machine learning in ALICE Run 3

Aug 5, 2026

ALICE Run 3 requires high-precision particle identification to enable detailed quark–gluon plasma measurements across momenta from roughly 100 MeV/c to 20 GeV/c. Traditional PID based on selection regions becomes inefficient when particle-species signals overlap. The work studies domain-adaptation neural networks to mitigate discrepancies between Monte Carlo simulations and experimental data, extending the approach with feature set embedding and attention for flexible training on varied detector-signal sets. Preliminary results show improved particle classification, integrated into the ALICE Run 3 Analysis Framework.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Introduction

ALICE and the need for accurate PID

Detectors used for PID

Limitations of traditional selection methods

Machine learning for PID

Feed-forward neural network approach

Training on Monte Carlo and probability output

Input choices and model specialization

Limitations from simulation, calibration, and reconstruction

arXiv :2309 .07768v1 [hep-ex] 14 Sep 2023 Particle identification with machine learning in ALICE Run 3 Maja Karwowska 1,2 , ∗ , Monika Jakubowska3 , Łukasz Graczykowski2 , Kamil Deja4,5 , and Miłosz Kasak4 1CERN – European Organization for Nuclear Research 2Faculty of Physics, Warsaw University of Technology 3Faculty of Electrical Engineering, Warsaw University of Technology 4Faculty of Electronics and Information Technology, Warsaw University of Technology 5IDEAS NCBR Abstract. The main focus of the ALICE experiment, quark–gluon plasma measurements, requires accurate particle identification (PID) . The ALICE subdetectors allow identifying particles over a broad momentum interval ranging from about 100 MeV/c up to 20 GeV/c. However, a machine learning (ML) model can explore more detector information. During LHC Run 2, preliminary studies with Random Forests obtained much higher efficiencies and purities for selected particles than standard techniques. For Run 3, we investigate Domain Adaptation Neural Networks that account for the discrepancies between the Monte Carlo simulations and the experimental data. Preliminary studies show that domain adaptation improves particle classification. Moreover, the solution is extended with Feature Set Embedding and attention to give the network more flexibility to train on data with various sets of detector signals. PID ML is already integrated with the ALICE Run 3 Analysis Framework. Preliminary results for the PID of selected particle species, including real-world analyzes, are discussed as well as the possible optimizations. 1 Introduction ALICE (A Large Ion Collider Experiment) [1] is one of the four big detectors located at the Large Hadron Collider (LHC) at CERN. ALICE studies the properties of quark–gluon plasma (QGP), a hot and dense state of matter, and the strong force that holds quarks together inside hadrons [2] . Detailed analysis of QGP requires accurate particle identification (PID), i.e., the ability to discriminate between different particle species produced during the collision. High PID precision distinguishes ALICE from other LHC experiments and allows for selecting a subset of particles required for specific analysis. The ALICE experiment is composed of several sub-detectors, some of which measure particle properties that can be used for identification. Figure 1 presents a scheme of the detector in Run 1 and Run 2 LHC data-taking periods. The three detectors particularly useful for PID are: Time Projection Chamber (TPC), Time-of-Flight (TOF), Transition Radiation Detector (TRD) . TPC is one of the most important ALICE detectors as it records 3D information of the trajectory of the particles, and ∗[e-mail: mkabus@cern.ch](e-mail: mkabus@cern.ch) Figure 1. Components of the ALICE detector in its Run 2 configuration [3] . particle ionization energy loss essential for particle identification. The Time-of-Flight detector measures particle travel time from the collision vertex to the detector, from which particle velocity and mass are calculated. TRD records transition radiation, that is, the emission of a few photons by traversing electrons, which helps in distinguishing electrons from other charged particles. With the signals recorded by the detectors described above, particles are chosen using selection criteria compared with theoretical calculations. The traditional method compares the number of standard deviations from the expected value for all detector signals. Particles falling outside the selection region are rejected. However, when the characteristics of different particle species overlap, combining information from multiple detectors becomes difficult. Choosing selection regions by trial and error is less effective in this case, lowering PID efficiency and limiting the statistical significance of the final analysis. These shortcomings can be addressed with more advanced classification methods such as Bayesian models or neural-net

FAQs

Why is particle identification crucial for ALICE Run 3 analyses?

Accurate PID is required to distinguish particle species produced in collisions, enabling reliable selection of particle subsets for quark–gluon plasma studies and improving analysis precision.

What problem do traditional PID selection regions face?

When different particle species overlap in detector characteristics, selecting regions by trial and error becomes less effective, reducing efficiency and limiting statistical significance.

How does the proposed domain-adaptation neural network improve PID?

It accounts for discrepancies between Monte Carlo simulations and experimental data, and preliminary studies indicate better particle classification; the method is further enhanced using feature embedding and attention.

Aug 5, 2026

Uploaded by Lucas Martin