What Machine Learning Can Do for Focusing Aerogel Detectors - Filtering Signal Hits and Event Reconstruction for FARICH

Aug 3, 2026

Upload by Ivy

Machine-learning approaches are explored for particle identification using the Focusing Aerogel Ring Imaging CHerenkov detector (FARICH) at the Super Charm-Tau factory. Because the detector placement makes proper cooling difficult, ambient background hits are captured at high rates and must be filtered to reduce data flow and improve particle velocity resolution. The work formulates noise filtering as binary classification and studies model performance in terms of efficiency and reduction of noise events. ResNet-18 is adapted with a two-channel representation combining photon coordinates and normalized hit times.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

AccessFree to download

Introduction

Machine learning motivation for FARICH noise filtering

Data processing and model architecture

Simulation inputs and two-channel image representation

ResNet-18 adaptation

Noise filtering

Binary classification formulation

Ground-truth labeling and class balance

Evaluation metrics

What Machine Learning Can Do for Focusing Aerogel Detectors Foma Shipilov1 , ∗ , Alexander Barnyakov2,3 , Vladimir Bobrovnikov2 , Sergey Kononov2,4 , and Fedor Ratnikov1 1NRU Higher School of Economics, Moscow, Russia 2Budker Institute of Nuclear Physics of Siberian Branch Russian Academy of Sciences, Novosibirsk, Russia 3Novosibirsk State Technical University, Novosibirsk, Russia 4Novosibirsk State University, Novosibirsk, Russia Abstract. Particle identification at the Super Charm-Tau factory experiment will be provided by a Focusing Aerogel Ring Imaging CHerenkov detector (FARICH) . The specifics of detector location make proper cooling difficult, therefore a significant number of ambient background hits are captured. They must be mitigated to reduce the data flow and improve particle velocity resolution. In this work we present several approaches to filtering signal hits, inspired by machine learning techniques from computer vision. 1 Introduction Reliable particle identification (PID) is a crucial component of modern physics experiments. Particle identification in the Super c-τ factory (SCTF) experiments will be provided by FARICH detector [1] . The use of a FARICH is under intensive discussion for the Spin Physics Detector (SPD) detector at NICA [2] . FARICH uses multilayer aerogel for Cherenkov ring proximity focusing (Fig. 1) . The detector may use both seedless real-time signal finder to produce fast trigger and mitigate noise background, and seeded off-line reconstruction mode for precise identification, however, SiPM properties and operating temperatures may result in a significant background hit rate f ∼1 MHz/mm2 (signal-to-background ratio ≈ 0.014), which necessitates the development of robust noise filtering techniques. Conventional pattern recognition noise rejection methods aim to eliminate background hits directly by calculating carefully crafted empirical statistics derived from the physical properties of the system [4], e.g. DELPHI RICH automates background removal by measuring photon hit times, the number of hits for each detector cell, and other parameters [5] . Statistical approaches to pattern recognition and reconstruction provide excellent precision and enable simple error estimation. However, such methods are only effective in moderate background conditions [5] due to the local nature of the statistics. This does not suit well for the heavy background environment of the FARICH. Additionally, statistical approaches require a particle track prior and may be difficult to incorporate new data sources into. ∗ e-mail: [foma@shipilov.ru](foma@shipilov.ru) © The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 ([https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/)). Figure 1: Left to right: possible integration of FARICH in SPD detector [2], focusing aerogel operation [3] . In contrast, Machine Learning (ML) techniques can extract high-level features from the input data and use them to efficiently solve various tasks. Many different data sources can be relatively easily incorporated, contributing to the overall robustness of the method. ML approaches proved to be useful across a diversity of nuclear physics research topics [6] . Early variants of neural networks had been developed specifically for use in high energy physics [7] . LHCb experiment uses neural networks for fast fake track rejection [8] and is planning to incorporate them in the RICH PID system [9] . ML has been applied to calibration and reconstruction of Cherenkov detectors with promising results [10] . Recently, object detection techniques from computer vision have been utilized in end-to-end data reconstruction pipeline for LArTPC neutrino imaging [4] . Object detection has also been adapted for sparse detector data in the object condensation pipeline [11] . In this work, we prese

FAQs

Why is noise filtering needed for the FARICH detector?

FARICH background hits are captured at high rates due to difficult cooling in its installation area. Filtering is required to reduce data flow and improve particle velocity resolution.

What data representation is used to train the model?

Photon coordinates and hit times from a Geant4 simulation are provided as a two-channel image: a bilinearly interpolated mask from (xc, yc) and a channel of normalized hit times tc for corresponding pixels.

How is the noise filtering task formulated for training?

The task is treated as binary classification of events with or without signal using a logistic regression objective, with labels set by computing a bounding box around the signal ellipse and requiring a minimum number of signal photons.

Aug 3, 2026

Upload by Ivy