GraphNeT Graph neural networks for neutrino telescope event reconstruction

2025-05-06 0 0 990.77KB 6 页 10玖币

侵权投诉

GraphNeT: Graph neural networks for neutrino

telescope event reconstruction

Andreas Søgaard

1¶

, Rasmus F. Ørsøe

, Leon Bozianu

, Morten

Holm1, Kaare Endrup Iversen1, Tim Guggenmos2, Martin Ha

Minh 2, Philipp Eller 2, and Troels C. Petersen 1

1Niels Bohr Institute, University of Copenhagen, Denmark

2Technical University of Munich, Germany

¶Corresponding author

16 September 2022

Summary

Neutrino telescopes, such as ANTARES (ANTARES Collaboration, 2011b),

IceCube (IceCube Collaboration, 2012,2017), KM3NeT (KM3NeT Collaboration,

2016), and Baikal-GVD (Baikal-GVD Collaboration, 2018) has the science goal of

detecting neutrinos and measuring their properties and origins. Reconstruction

at these experiments is concerned with classifying the type of event or estimating

properties of the interaction.

GraphNeT

(Søgaard et al., 2022) is an open-source python framework aimed

at providing high quality, user friendly, end-to-end functionality to perform

reconstruction tasks at neutrino telescopes using graph neural networks (GNNs).

GraphNeT

makes it fast and easy to train complex models that can provide

event reconstruction with state-of-the-art performance, for arbitrary detector

conﬁgurations, with inference times that are orders of magnitude faster than

traditional reconstruction techniques (IceCube Collaboration, 2022b).

GNNs from

GraphNeT

are ﬂexible enough to be applied to data from all neutrino

telescopes, including future projects such as IceCube extensions (IceCube-Gen2

Collaboration, 2017,2021;IceCube-PINGU Collaboration, 2014) or P-ONE

(P-ONE Collaboration, 2020). This means that GNN-based reconstruction can

be used to provide state-of-the-art performance on most reconstruction tasks

in neutrino telescopes, at real-time event rates, across experiments and physics

analyses, with vast potential impact for neutrino and astro-particle physics.

arXiv:2210.12194v1 [astro-ph.IM] 21 Oct 2022

Statement of need

Neutrino telescopes typically consist of thousands of optical modules (OMs) to

detect the Cherenkov light produced from particle interactions in the detector

medium. The number of photo-electrons recorded by the OMs in each event

roughly scales with the energy of the incident particle, from a few photo-electrons

and up to tens of thousands.

Reconstructing the particle type and parameters from individual recordings

(called events) in these experiments is a challenge due to irregular detector

geometry, inhomogeneous detector medium, sparsity of the data, the large

variations of the amount of signal between diﬀerent events, and the sheer number

of events that need to be reconstructed.

Multiple approaches have been employed, including relatively simple methods

(ANTARES Collaboration, 2011a;IceCube Collaboration, 2022a) that are robust

but limited in precision and likelihood-based methods (Aartsen & others, 2014;

Abbasi et al., 2013;AMANDA Collaboration, 2004;ANTARES Collaboration,

2017;Chirkin, 2013;IceCube Collaboration, 2022a,2014,2021b) that can attain

a high accuracy at the price of high computational cost and detector speciﬁc

assumptions.

Recently, machine learning (ML) methods have started to be used, such as

convolutional neural networks (CNNs) (IceCube Collaboration, 2021a;KM3NeT

Collaboration, 2020) that are comparably fast, but require detector data being

transformed into a regular pixel or voxel grid. Other approaches get around the

geometric limitations, but increase the computational cost to a similar level as

the traditional likelihood methods (Eller et al., 2022).

Instead, GNNs can be thought of as generalised CNNs that work on data with

any geometry, making this paradigm a natural ﬁt for neutrino telescope data.

The

GraphNeT

framework provides the end-to-end tools to train and deploy

GNN reconstruction models.

GraphNeT

leverages industry-standard tools such

pytorch

(Paszke et al., 2019),

PyG

(Fey & Lenssen, 2019),

lightning

(Falcon

& The PyTorch Lightning team, 2019), and

wand

(Biewald, 2020) for building

and training GNNs as well as particle physics standard tools such as

awkward

(Pivarski et al., 2020) for handling the variable-size data representing particle

interaction events in neutrino telescopes. The inference speed on a single GPU

allows for processing the full online datastream of current neutrino telescopes in

real-time.

Impact on physics

GraphNeT

provides a common framework for ML developers and physicists that

wish to use the state-of-the-art GNN tools in their research. By uniting both

user groups,

GraphNeT

aims to increase the longevity and usability of individual

code contributions from ML developers by building a general, reusable software

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：

GraphNeT:GraphneuralnetworksforneutrinotelescopeeventreconstructionAndreasSøgaard1{,RasmusF.Ørsøe2,LeonBozianu1,MortenHolm1,KaareEndrupIversen1,TimGuggenmos2,MartinHaMinh2,PhilippEller2,andTroelsC.Petersen11NielsBohrInstitute,UniversityofCopenhagen,Denmark2TechnicalUniversityofMunich,Germany{Corresp...

展开>> 收起<<

GraphNeT Graph neural networks for neutrino telescope event reconstruction.pdf

共6页,预览2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！

GraphNeT Graph neural networks for neutrino telescope event reconstruction

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: