Machine-Learning Compression for Particle Physics Discoveries Jack H. Collins1 Yifeng Huang2 Simon Knapen34 Benjamin Nachman35 and Daniel Whiteson2

2025-05-02 0 0 493.34KB 9 页 10玖币

侵权投诉

Machine-Learning Compression for Particle Physics

Discoveries

Jack H. Collins1, Yifeng Huang2, Simon Knapen3,4, Benjamin Nachman3,5, and Daniel Whiteson2

1SLAC National Accelerator Laboratory

2Department of Physics and Astronomy, University of California, Irvine

3Physics Division, Lawrence Berkeley National Laboratory

4Berkeley Center for Theoretical Physics, University of California, Berkeley

5Berkeley Institute for Data Science, University of California, Berkeley

jackadsa@gmail.com,yifengh3@uci.edu,smknapen@lbl.gov,

bpnachman@lbl.gov,daniel@uci.edu

Abstract

In collider-based particle and nuclear physics experiments, data are produced at

such extreme rates that only a subset can be recorded for later analysis. Typically,

algorithms select individual collision events for preservation and store the complete

experimental response. A relatively new alternative strategy is to additionally save

a partial record for a larger subset of events, allowing for later speciﬁc analysis

of a larger fraction of events. We propose a strategy that bridges these paradigms

by compressing entire events for generic ofﬂine analysis but at a lower ﬁdelity.

An optimal-transport-based

Variational Autoencoder (VAE) is used to automate

the compression and the hyperparameter

controls the compression ﬁdelity. We

introduce a new approach for multi-objective learning functions by simultaneously

learning a VAE appropriate for all values of

through parameterization. We

present an example use case, a di-muon resonance search at the Large Hadron

Collider (LHC), where we show that simulated data compressed by our

-VAE has

enough ﬁdelity to distinguish distinct signal morphologies.

1 Introduction

The rate and size of interaction events at modern particle and nuclear physics experiments typically

prohibits storage of the complete experimental dataset and require that many interaction events

be discarded in real time by a trigger system. For example, at the Large Hadron Collider (LHC),

collisions occur at a rate of 40 MHz, but the ATLAS and CMS experiments recording rates are

typically

O(kHz)

[1,2]. For selected events, the complete experimental response is preserved for

later analysis. When the scientiﬁc goals only require identifying events which contain rare and

easy-to-identify objects, such as high energy photons, the trigger system is highly efﬁcient. However,

this strategy leaves the vast majority of the events unexamined, including many with complex features

that are hard to quickly identify online or may not be rare.

An alternative approach to fully recording a small fraction of the events is to preserve a partial record

of a larger fraction [3–5]. This strategy has allowed access to lower-energy phenomena which occur at

higher rates, but the utility of these partial data records is limited. For example, a recent partial-event

analysis targets di-muon resonances [6], only recording the four-momenta of the two muons and a

small number of additional event properties for low-mass events that would otherwise be too high

rate for the full-event trigger system. This approach has the potential to make a major discovery, but

the lack of a full event record could make it challenging to diagnose such a discovery. To distinguish

arXiv:2210.11489v2 [hep-ph] 18 Dec 2022

between several competing hypotheses which might generate a peak in the di-muon spectrum would

require recording new data with a dedicated trigger, which is both time consuming and expensive.

We propose an approach that bridges the full and partial event paradigms automatically with machine

learning. This is accomplished by training a neural network to learn a lossy event compression with

a tunable resolution parameter. An extreme version of this approach would be to save every event

at the highest resolution allowable by hardware (see e.g. Ref. [7] for autoencoders in hardware).

We present a more modest version in which we envision full event compression which could run

alongside partial event triggers to expand their utility for a larger range of ofﬂine analyses. Our

approach uses a optimal transport-based Variational Autoencoder (VAE) following Ref. [8].

In a proof-of-concept study, we compress and record a sample of simulated interactions which are

similar to those analyzed in Ref [6], preserving information which would otherwise be lost. We show

that this additional information can be used to effectively discriminate between two signal models

which are difﬁcult to distinguish with only the muon kinematics. The overall structure of the proposal

is that ﬁrst, a signal is discovered in a trigger-level analysis such as this dimuon resonance search.

Subsequently, a compressed version of the hadronic event data, which has been stored alongside the

muons, can be used to rule out or favor candidate signal models.

2 Related Work

An alternative to compressing individual events is compressing the entire dataset online [9], which

is methodologically and practically more challenging. An alternative to saving events for ofﬂine

analysis is to look for new particles automatically with online anomaly detection [10–13]. While we

build our VAE on the setup from Ref. [8] using the Sinkhorn approximation [14,15] to the Earth

Movers Distance, other possibilities have been explored, such as using graph neural networks [16].

We leave a comparison of the power of different approaches to future work.

3β-parameterized Variational Autoencoder

We represent each collider event

as a point cloud of 3-vectors

{pT/HT, η, φ}

, where

and

are the

geometric coordinates of particles in the detector, and

their transverse momenta which correspond

to the weights in the point cloud. These are normalized for each event using

HT=PipT,i

. We build

an EMD-VAE [8,17,18] trained to minimize a reconstruction error given by an approximation to the

2-Wasserstein distance between collider events xand reconstructed examples x0, with loss function

L=hS(x, x0(z))/β +DKL(q(z|x)||p(z))ip(x).(1)

An encoder network maps the input

to a Gaussian-parameterized distribution

q(z|x)

on 256-

dimensional latent coordinates

. This network is built as a Deepsets/Particle Flow Network (PFN) [19,

20]. A decoder x0(z)maps latent codes zto jets x0, parameterizing a posterior probability

log p(x|z)∝S(x, x0(z))/β ,

where

S(x, x0(z))

is a sharp Sinkhorn [15,21–23] approximation to the 2-Wasserstein distance

between event

and its decoded

with ground distance given by

Mij = ∆R2

ij ≡(ηi−ηj)2+

(φi−φj)2

, and calculated using the same algorithm and parameters as in Ref [8]. This decoder

network is built as a dense neural network.

DKL(q(z|x)||p(z))

is the KL divergence between the

encoder probability

q(z|x)

and the prior

p(z)

, which we take to be a standard Gaussian. This KL

divergence can be expressed as a sum of contributions from each of the 256 latent space directions.

The details of the architecture is described in the Appendix.

The quantity

is typically taken to be a ﬁxed hyperparameter of the network [24] which controls the

balance between reconstruction ﬁdelity and degree of compression in the latent space. In this work,

we elevate

from a ﬁxed hyperparameter to an input [25] of both the encoder and decoder networks

1The authors are grateful to Jesse Thaler for this suggestion.

Note added post-publication: A similar idea was pursued in [26], which was submitted for publication

concurrently with this work. The implementation in their study differs from ours by using a hypernetwork to

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：

Machine-LearningCompressionforParticlePhysicsDiscoveriesJackH.Collins1,YifengHuang2,SimonKnapen3;4,BenjaminNachman3;5,andDanielWhiteson21SLACNationalAcceleratorLaboratory2DepartmentofPhysicsandAstronomy,UniversityofCalifornia,Irvine3PhysicsDivision,LawrenceBerkeleyNationalLaboratory4BerkeleyCenterfo...

展开>> 收起<<

Machine-Learning Compression for Particle Physics Discoveries Jack H. Collins1 Yifeng Huang2 Simon Knapen34 Benjamin Nachman35 and Daniel Whiteson2.pdf

共9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！

Machine-Learning Compression for Particle Physics Discoveries Jack H. Collins1 Yifeng Huang2 Simon Knapen34 Benjamin Nachman35 and Daniel Whiteson2

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: