Rejecting noise in Baikal-GVD data with neural networks I. KharukabG. RubtsovacG. Safronovad

2025-04-29 0 0 906.93KB 16 页 10玖币

侵权投诉

Rejecting noise in Baikal-GVD data with neural

networks

I. Kharuk,a,b G. Rubtsov,a,c G. Safronova,d

aInstitute for Nuclear Research of the Russian Academy of Sciences,

60th October Anniversary Prospect, 7a, Moscow, 117312, Russia

bMoscow Institute of Physics and Technology,

Institutsky lane 9, Dolgoprudny, Moscow region, 141700, Russia

cLaboratory of Cosmology and Elementary Particle Physics, Novosibirsk State University,

Novosibirsk, 630090 Russia

dJoint Institute for Nuclear Research,

Joliot-Curie 6, Dubna, Moscow Region, 141980, Russia

E-mail: ivan.kharuk@phystech.edu

Abstract: Baikal-GVD is a large (∼1 km3) underwater neutrino telescope installed in the

fresh waters of Lake Baikal. The deep lake water environment is pervaded by background

light, which is detectable by Baikal-GVD’s photosensors. We introduce a neural network

for an eﬃcient separation of these noise hits from the signal ones, stemming from the

propagation of relativistic particles through the detector. The model has a U-net-like

architecture and employs temporal (causal) structure of events. The neural network’s

metrics reach up to 99% signal purity (precision) and 96% survival eﬃciency (recall) on

Monte-Carlo simulated dataset. We compare the developed method with the algorithmic

approach to rejecting the noise and discuss other possible architectures of neural networks,

including graph-based ones.

arXiv:2210.04653v2 [astro-ph.IM] 9 Jul 2023

Contents

1 Introduction 1

2 Monte-Carlo simulations 3

3 Neural network architecture 4

3.1 Data representation 4

3.2 Neural network architecture 6

4 Results 8

5 Discussion and conclusion 10

A Graph neural network 12

1 Introduction

Baikal-GVD is a large-volume water-based neutrino telescope aimed at studying the

ﬂux of high-energy cosmic neutrinos and searching for their sources [1,2]. The experi-

ment is located in Lake Baikal, Russia, and, as of 2022, has an eﬀective working volume

of approximately 0.5 km3with respect to the high-energy neutrino-induced cascades. The

telescope is targeted to reach the volume of 1 km3by 2030 and is currently the largest neu-

trino telescope in the Northern Hemisphere. The location of Baikal-GVD and IceCube [3]

experiments makes them complimentary, as their data combined allows for comprehensive

full-sky astrophysical surveys. The four major neutrino telescopes, i.e. IceCube, Baikal-

GVD, KM3NeT [4], and ANTARES [5], cooperate within the Global Neutrino Network.

The basic components of Baikal-GVD’s detector are optical modules (OMs) accommo-

dating 10-inch (25 cm) high quantum eﬃciency photomultipliers (Hamamatsu R7081-100).

They are designed to register Cherenkov light produced by relativistic particles (originating

from cosmic neutrinos and air showers) in the eﬀective volume of the detector. OMs are

carried by strings (36 OMs per string) and are located at a depth of 750 to 1275 meters

with a 15-meter spacing. The strings are organized into clusters – approximately regular

heptagons with a string in the middle, see ﬁgure 1. In total, as of 2022, there are 11 clusters

with an average distance of 300 meters between the clusters and an average cluster radius

of 60 meters.

As OMs are located underwater, they are subject to the natural luminescence of

Baikal’s water and to the photoemission of molecules and atoms in excited states [6].

Corresponding random, uncorrelated activations of OMs constitute background (noise)

hits in the data. The charge deposition spectrum of these activations is well studied and

– 1 –

Figure 1. Schematic representation of the Baikal-GVD detector.

Figure 2. Normalized distributions of integral charges registered by OMs for all Monte-Carlo data

(solid, blue) and noise subsample (dotted, orange).

is presented in ﬁgure 2. Noise hits have the rate of 20–100 kHz (depending on depth and

season), and their charge deposition is of the order of 1 photo-electron (p.e.). On average,

noise hits constitute approximately 85–90% of the data collected for the analysis. In what

follows, we will refer to the non-noise hits as signal ones.

One can suppress the noise hits by introducing a signal-level threshold around several

p.e. This will, however, suppress some of the signal hits as well. This is undesirable, since

the reconstruction algorithms would beneﬁt from keeping as much signal hits as possible.

Therefore, eﬀective ﬁltering algorithms are essential.

In this paper we present a neural network for rejecting background hits in a standalone

cluster readout regime. The main features of the developed method, discussed in detail in

– 2 –

the main part of the paper, are the following:

•Representation of Baikal-GVD’s data in a form that makes direct use of the causal

structure of the events;

•Introduction of multiplicative Gaussian noise for proper simulation of the ﬂuctuations

of OMs readings;

•The use of a custom loss function that reduces time residuals of the hits identiﬁed as

signal ones by the neural network;

•Neural network is insensitive to the auxiliary hits, which are introduced to make data

representation uniform;

•Compared to the noise suppression algorithms developed by the Baikal-GVD collab-

oration [1,7], our neural network has better metrics and allows for a much faster

data analysis.

The developed method can also be readily extended to process the multi-cluster data.

The paper is structured as follows. In section 2we describe the Monte-Carlo simula-

tions of Baikal-GVD’s data. In section 3and appendix Adata representation and neural

network’s architecture are discussed. The results, including comparison with non-machine

learning methods, are presented in section 4. Finally, section 5concludes the paper.

2 Monte-Carlo simulations

Simulation of the data is performed via the Monte-Carlo method. Two types of in-

coming particles are considered: 1) muon neutrinos arriving from under the horizon, and

2) bundles of muons originating from the cosmic air showers. The energy spectrum and

incoming directions of arriving particles are chosen to coincide with the ones expected in

the experiment, see [8,9] for details. The procedure takes into account photon scattering

in Baikal’s water and full simulation of cosmic air showers’ evolution using QGSJET II-03

[10] and CORSIKA [11]. The noise rate and charge distribution are simulated so as to

mimic the experimentally expected detector conditions.

The triggering condition for identifying an event is the following: two adjacent OMs,

within the 100 ns time window, register signals that are at least 4.5 and 1.5 p.e. If this

condition is fulﬁlled, the data from all OMs exceeding the signal-level threshold (0.3 p.e.)

is collected for further analysis.

Registered signals (waveforms) are approximated by discrete hits (pulses). Each of the

hits is characterized by the following physical observables:

1. Time at which the hit was registered;

2. Integral charge (in p.e.) registered by OM;

3. Maximal amplitude of the registered signal.

– 3 –

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

RejectingnoiseinBaikal-GVDdatawithneuralnetworksI.Kharuk,a,bG.Rubtsov,a,cG.Safronova,daInstituteforNuclearResearchoftheRussianAcademyofSciences,60thOctoberAnniversaryProspect,7a,Moscow,117312,RussiabMoscowInstituteofPhysicsandTechnology,Institutskylane9,Dolgoprudny,Moscowregion,141700,RussiacLaborat...

展开>> 收起<<

Rejecting noise in Baikal-GVD data with neural networks I. KharukabG. RubtsovacG. Safronovad.pdf

共16页,预览4页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Rejecting noise in Baikal-GVD data with neural networks I. KharukabG. RubtsovacG. Safronovad

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: