KSC2021 Community Forum 1 - Machine Learning for HEP

Tuesday 26 Oct 2021, 13:00 → 18:00 Asia/Seoul

Kihyeon Cho (KISTI)

• 국내외 고에너지물리 및 머신러닝 관련 연구자들이 참석하여, 슈퍼컴5호기를 활용한 고에너지물리 분야 머신러닝 연구 사례를 발표한다. 특히, 요사이 핫 이슈가 되고 있는 암흑물질에 관하여 현재 진행사항과 향후 전망에 대하여 논의한다.

13:00 → 13:30 21세기 입자물리학, 천체물리학, 우주론

2012년 힉스 입자의 발견을 통해 입자물리학의 표준모형 (The Standard Model)은 완성되었다. 그러나 표준모형과 일반상대성이론을 고려하였을 때 초기 우주의 급팽창, 우주의 후기 급팽창, 거대구조와 은하 형성 등 관측적 사실과 부합하지 못하며 암흑 성분을 추가로 필요하다는 것이 밝혀졌다. 이에 21세기 입자물리학은 천체물리학, 우주론 연구와 협력을 필요로 하며, 우주의 암흑 성분을 이해해야하는 당면 과제를 맞이하게 되었다. 본 발표에서는 관련 분야에서 현재 연구되고 있는 연구 현황을 물리학과 천체물리학 관점에서 소개하고, 미래의 발전을 위해 기대하는 컴퓨팅 파워와 전산 기법의 활용에 대해 논의한다.

Speaker: Prof. 박성찬 (연세대학교)

01.KSC2021_발표자료_박성찬교수.pdf

13:30 → 14:00 Light Dark Sector Searches at Beam-induced Neutrino Experiments

The sensitivity to dark matter signals at neutrino experiments is fundamentally challenged by the neutrino rates, as they leave similar signatures in their detectors. As a way to improve the signal sensitivity, we investigate a dark matter search strategy which utilizes the timing and energy spectra to discriminate dark matter from neutrino signals at low-energy, pulsed-beam neutrino experiments such as CCM, COHERENT, JSNS2. The dark matter candidate of interest comes from the relatively prompt decay of a dark sector gauge boson which may replace a Standard-Model photon, so the delayed neutrino events can be suppressed by keeping prompt events only. Furthermore, prompt neutrino events can be rejected by a cut in recoil energy spectra, as their incoming energy is relatively small and bounded from above while dark matter may deposit a sizable energy beyond it. We apply the search strategy of imposing a combination of energy and timing cuts to the existing CsI and LAr data of the COHERENT experiment as concrete examples, and report a mild excess beyond known backgrounds. We then investigate the expected sensitivity reaches to dark matter signals in our benchmark experiments.

Speaker: Prof. 박종철 (충남대학교)

02.KSC2021_발표자료_박종철교수.pdf

14:00 → 14:30 Dark Sector Studies at Belle and Belle II

Belle and Belle II are e+e- collision experiments using the KEKB and Super-KEKB asymmetric energy e+e- colliders at KEK, Japan. In the Belle/Belle II experiments, center of momentum energy is tuned to the mass of Upsilon(4S) resonance, which decays into a pair B mesons. The main goals of the two experiments are searching for new physics beyond the standard model by analyzing the rare and exotic decays of B and charm mesons as well as tau leptons. In addition, they also probe the dark sector both in the ISR-type processes and the B meson’s rare decay modes.

In the analyses of Belle/Belle II data, various machine learning tools are used for various purposes such as B-meson full reconstruction, flavor tagging, continuum suppression and for other specific functions. In this presentation, we explain the machine learning tools used in Belle and Belle II analysis. Models, training and testing, and some examples of applying Machine Learning tools will be presented.

Speaker: Mr 조성진 (연세대학교)

09.KSC2021_발표자료_조성진.pdf

14:30 → 15:00 Pros and cons of ultra-light dark matter

This is a brief review on the recent progress of fuzzy dark matter also known as BEC dark matter, wave dark matter, or ultra-light axion. We discuss cosmological constraints on this model and show how this model can explain mysteries of galaxies using numerical results.

Speaker: Prof. 이재원 (중원대학교)

04.KSC2021_발표자료_이재원교수.pdf

15:00 → 15:30 Break

15:30 → 16:00 Application of Real-time Machine Leaning Technology for HEP Data Analysis

Collecting the high quality datasets efficiently at High Luminosity LHC (HL-LHC) will be a challenge in the high pileup environment of 200 proton-proton collisions per beam crossing. To deal with the large size of the dataset from the HL-LHC, the Phase-2 Upgrade of the Level-1 (L1) trigger system at the CMS experiment is essential. We will present results on the missing transverse energy (MET) regression with machine learning technique for CMS Phase-2 L1 trigger. The MET is calculated based on the PF and PUPPI algorithms for the L1 correlator trigger. In this talk, the performance of the L1 MET trigger using real-time machine learning techniques will be presented.

Speaker: Prof. 문창성 (경북대학교)

KSC2021-Chang-Seong_Moon.pdf

16:00 → 16:30 양자알고리즘을 활용한 고에너지 가속기 실험 데이터 분석

양자어닐링 머신을 활용한 LHC 데이터 분석, 특히 입자들의 붕괴방식을 찾아내는 양자 알고리즘에 대한 최근 연구결과에 대해 발표한다. 이외 입자물리학에서 양자컴퓨터에 대한 활용 가능성에 대해 논의한다.

Speaker: Prof. 박명훈 (서울과학기술대학교)

05.KSC2021_발표자료_박명훈교수.pdf

16:30 → 17:00 Deep Learning to Search for rare processes at CMS experiment

There have been lots of developments in the data analysis at the Large Hadron Collider (LHC). The High-Luminosity LHC (HL-LHC) will be a challenging and interesting environment for the application of the Deep Learning. The size of the data will be increased by huge factors with very large number of particle multiplicities due to the multiple proton-proton interactions. Therefore, Very efficient Deep Learning architecture and computing environment will be one of the crucial element for the Deep Learning for the HL-LHC era. We use the Convolutional Neural Network (CNN) and Graph Neural Network (GNN) architecture to develop algorithms to discriminate SUSY events from the QCD backgrounds at CMS experiment in the HL-LHC scenario. The same algorithms are applied to search for the rare top quark production process with minimal changes. Training and evaluation of the Deep Learning model can be accelerated with the distributed training and evaluation at the high performance computing such as 5th gen. supercomputer at KISTI.

Speaker: Prof. 고정환 (경희대학교)

06.KSC2021_발표자료_고정환교수.pdf

17:00 → 17:30 SND@LHC Experiment

The SND@LHC (Scattering and Neutrino Detector at the LHC) is a newly approved experiment whose goal is to directly detect collider neutrinos at TeV energies for the first time and search for feebly interacting particles (FIPs) in an unexplored domain. Unlike the four big LHC experiments, the SND@LHC sits in a tangent line from the interaction point (IP) and therefore it can detect "forward moving particles" which will cover the high pseudo-rapidity region (7.2 < η < 8.7). The SND@LHC detector will be installed in TI18 tunnel, 480m away from ATLAS IP and positioned slightly off the beam axis on the opposite side of FASER. In order to identify neutrino interactions of the three flavours and search for FIPs via their scatterings on the target, the ECC (Emulsion Cloud Chamber) will be used together with electronic detectors. The ECC which is composed of nuclear emulsion films interleaved with tungsten plates can play the roles of both target and precision tracker. It displays the tracks of the particles produced from the neutrinos and FIPs scatterings, while the electronic tracking devices provide time stamps for these tracks. The tracking devices also measure the energy of the particles together with the downstream muon detector. It is expected to begin taking data when the LHC Run 3 starts up in 2022. In this talk, we will introduce the current status of the experiment and also the possible future plan of FPF (Forward Physics Facility) at LHC.

Speaker: Prof. 윤천실 (경상국립대학교)

08.KSC2021_발표자료_윤천실교수.pdf

17:30 → 18:00 Some features in multi-component SIMP scenarios

In this talk, I discuss multi-component SIMP scenarios in three different setups: dark QCD, effective field theory approach and a dark gauge model with U(1) that is spontaneously broken into Z2 X Z3. For the first two cases, one has to maintain some degree of mass degeneracy, whereas in the third model two components of DM can have very different mass scales.

Speaker: Prof. 고병원 (고등과학원 (KIAS))

03.KSC2021_발표자료_고병원교수.pdf