实验粒子物理计算研讨会（2025）

Session

人工智能和机器学习的应用

14 Jul 2025, 10:50

SJTU Xichang Center

7. A Transformer-based Jet Flavour Tagger at ATLAS

Bingxuan Liu (Shenzhen Campus of Sun Yat-sen University)

14/07/2025, 10:50

邀请报告

ATLAS最近发布了一个基于 Transformer架构的味标记算法，性能相较于上一代算法有了质的提升。此外，新算法还带了范式改变，直接利用喷注与径迹信息作为输入而不依赖于中间算法。校准结果显示新算法在数据中的效能与模拟类似，预期将为新一轮的物理分析结果产生重大影响。该报告将会详细介绍此算法的技术细节以及开发过程背后的经验得失，同时也会对未来标记算法的发展做出展望。

17. Advancing fully hadronic final state searches at the LHC: The case of HH(4b)

Congqiao Li (Peking University)

14/07/2025, 11:30

口头报告

We show that fully hadronic final state searches at the LHC—exemplified by the $HH(4b)$ flagship analysis—can achieve a 5–10× improvement in sensitivity by transitioning from jet-based methods to jet-free, event-level analysis powered by large-scale AI models. Existing approaches depend heavily on jet tagging, whose performance has plateaued despite sophisticated ML developments. Our strategy...

24. A holistic approach of event reconstruction and physics analysis

Manqi RUAN (Institute of High Energy Physics, Beijing, China)

14/07/2025, 13:30

邀请报告

Identification plays a central role in experimental particle physics, either identifying different physics objects or distinguishing signal events from backgrounds.
Benefit from the rapid development of AI tools, we propose a holistic approach for the identification task.
This approach takes all the reconstructable information as input and infers the relevant categories using sufficiently...

9. 正负电子希格斯工厂上一一对应的粒子流重建

Yuexin Wang (IHEP, CAS)

14/07/2025, 14:10

口头报告

粒子流重建是目前高能对撞机实验上广泛采用的重建方法，其终极目标是实现重建前后粒子的一一对应性并高效鉴别其种类。但在实际实验中，如探测器有限的接收度和空间颗粒度、重建算法的不完美等因素往往会打破这种一一对应性。本报告将以 CEPC 为例，介绍未来正负电子希格斯工厂上关于一一对应性粒子流重建性能的研究，通过使用 truth link 对重建粒子进行细致的分析和分类并结合机器学习技术，实现了假粒子的高效鉴别与去除，将双喷注末态希格斯玻色子的不变质量分辨率提升了15%。此外也同时实现了9种粒子的高效鉴别，对五种带电粒子（电子、$\mu$子、$\pi$介子、K介子、质子）和光子的鉴别效率可高达97%以上，对三种中性强子（长寿命中性K介子、中子、反中子）的借鉴效率也可达到75%-80%，极大拓展了传统粒子鉴别的类别。

22. Search for Long-lived Particles at Future Lepton Colliders Using Leptons and Jets

Xiang Chen

14/07/2025, 14:40

口头报告

Long-lived particles (LLPs) have a clear signature of physics beyond the Standard Model (BSM), characterized by displaced vertex or decay lengths corresponding to lifetimes of nanoseconds or longer. Lepton colliders provide a clean environment for LLP searches. This study employs full simulation data to investigate LLPs produced in Higgs decays via the process e+e−→ZH, where each LLP decays...