AI+HEP in East Asia

23 Feb 2025, 15:00 → 28 Feb 2025, 17:00 Asia/Seoul

IBS

---

Notes:

• Please ignore any emails from 3rd party companies, as we do not have any contract.
• Connections to this website from overseas may be slow due to the IBS firewall. Sorry for the inconvenience.

---

 

This regional workshop aims to connect researchers in East Asia working in the interdisciplinary field of Artificial Intelligence and High Energy Physics (AI+HEP). The main topics covered include machine learning for particle theory, phenomenology and experiments, astrophysics and cosmology, as well as HEP tools for AI theory.

The workshop will have invited plenary talks, contributed presentations, and ample time for discussions. Both domain experts and those who are interested in exploring the field are welcome to participate, especially postdocs and graduate students. The goals are to foster a regional research community and to stimulate more collaborations. 

Workshop Venue

Room B438, CTPU seminar room, IBS main building

Invited Speakers:

• Cheng-Wei Chiang (National Taiwan University (NTU))
• Ahmed Hammad (KEK)
• Ji-hoon Kim (Seoul National University)
• Congqiao Li (Peking University)
• Vinicius Mikuni (NERSC, Berkeley Lab)
• Masahiro Morinaga (ICEPP, University of Tokyo)
• Myeonghun Park (Seoultech)
• Manqi Ruan (IHEP)
• Taikan Suehara (ICEPP, University of Tokyo)

Organizing Committee:

• Tianji Cai (SLAC)
• Sung Hak Lim (CTPU-PTC, IBS)
• Huilin Qu (CERN)

Advisory Committee:

• Mihoko M. Nojiri (KEK)
• David Shih (Rutgers)

Acknowledgment

This workshop is mainly funded and supported by CTPU, IBS in Korea.  The workshop is partly supported by 22H05113 “Foundation of Machine Learning Physics”, Grant in Aid for Transformative Research Areas, Japan.

Group Photo

Sunday, 23 February ¶

18:00 Pre-workshop discussions (organizers only)

Monday, 24 February ¶

08:50 Registration

1 Opening¶

Speaker: Sung Hak Lim (IBS CTPU-PTC)

AI+HEP_opening.pdf

Monday Talks¶

Convener: Sung Hak Lim (IBS CTPU-PTC)

2 Quantum Machine Learning for High Energy Physics¶

Speaker: Myeonghun Park (Seoultech)

MPARK.pdf

10:30 Coffee Break

3 Modern Deep Learning for LHC Physics: Personal Insights and Reflections¶

Speaker: Congqiao Li (Peking University)

25.02.24_AIHEP_ModernDL4LHC_CL.pdf

12:00 Lunch

Monday Talks¶

Convener: Sung Hak Lim (IBS CTPU-PTC)

4 Weak Supervision Techniques in Collider Physics¶

In this talk, we will first review CWoLa in the broad framework of weak supervision. Taking the Dark Valley model as an explicit new physics example, we demonstrate how CWoLa can be employed to train a neural network based purely on real data and show its performance in comparison with traditional methods.  By way of transfer learning and data augmentation, we illustrate how the performance can be further improved, along with discussions of trends and features of these methods.

Speaker: Cheng-Wei Chiang (National Taiwan University)

Weak Supervision Techniques in Collider Physics - Chiang.pdf

14:30 Coffee Break

5 How Can Machine Learning and Simulations Help Us Study Our Universe (and "Astronomically" Large Datasets)?¶

As astronomers and cosmologists grapple with the inherently "astronomical" size of their datasets, machine learning is rapidly being adopted in a variety of astrophysical applications.  In this talk, I will showcase some examples of how it can potentially revolutionize the way we study our Universe.  In one example, I will discuss how machine learning could be utilized to extract the fundamental parameters of our Universe from a large galaxy survey.  In another example, I will present a pipeline that estimates the baryonic (visible) properties of galaxies based only on their dark matter (invisible) content in a large dark matter-only simulation.

Speaker: Ji-hoon Kim (Seoul National University)

AI+HEP_IBS_022425.pdf

16:00 Free Discussions

18:00 Reception

Tuesday, 25 February ¶

09:00 Coffee Break

Morning Talks¶

Convener: Mihoko Nojiri (IPNS, KEK)

6 Status and prospects of DNN-based reconstruction for future Higgs factories¶

Speaker: Taikan Suehara (ICEPP, The University of Tokyo)

250224-hfreco-aihepasia-suehara.pdf

10:30 Coffee Break

7 Trilogy of event reconstruction for the future electron positron Higgs factory¶

Hadronic events are the bulk part of physics events at future electron-positron Higgs factories. For instance, 97% of ZH signal decays into the final state with jets, while the majority are actually full hadronic events. Therefore, an efficient reconstruction of those hadronic events is critical for the physics exploration at the future Higgs factory, and, actually the entire high-energy frontier.

Using Artificial Intelligence, we propose and realize a trilogy for the hadronic event reconstruction: firstly, jet origin identification that distinguishes jets originating from 11 different kinds of colored particles; second, one-one correspondence reconstruction that aims at efficiently reconstructing and identifying all the visible particles; and thirdly, color singlet identification that aims at distinguish the color singlet origin of each reconstructed particles, for example to identify from Z or Higgs boson a final state particle is generated at full hadronic Z events. We will present the current status of relevant performance studies, and discuss its impact on the physics measurements at future collider experiments. We’d also emphasize its impact on high-precision QCD measurements and studies.

Speaker: Manqi Ruan (IHEP)

Trilogy AI Asia-v2.pdf

12:00 Lunch

Afternoon Talks¶

Convener: Huilin Qu (CERN)

8 Overview of Machine Learning applications in JUNO¶

Jiangmen Underground Neutrino Observatory (JUNO), located in the southern part of China, will be the world’s largest liquid scintillator (LS) detector upon completion. Equipped with 20 kton LS, about 17612 20-inch PMTs and 25600 3-inch PMTs in the central detector (CD), JUNO will provide a unique apparatus to probe the mysteries of neutrinos, particularly the neutrino mass ordering puzzle. In recent decades Machine Learning has been more and more widely used in various neutrino experiments. If each PMT is viewed as a pixel, the JUNO CD can be regarded as a large spherical camera, providing a perfect scenario for the application of Machine Learning. This talk will present an overview of Machine Learning applications in JUNO, including reconstruction, particle identification and event classification, etc. These Machine Learning based methods not only provide alternative approaches complementary to the traditional ones, but also demonstrate huge potential on enhancing the performance of the JUNO detector.

Speaker: Wuming Luo (IHEP, CAS)

AIHEP25_IBS_luowm .pdf

9 New Potentials in Boosted-Jet Regimes through Inclusive Jet Model Pre-Training¶

Advanced boosted-jet taggers, such as those involving H->bb/cc or t/W/Z tagging, have significantly enhanced the sensitivity of many analyses at the LHC. Recently, an inclusive pre-trained large-R jet model has been successfully deployed in the CMS experiment. In this talk, we discuss two novel potentials brought by this technique. Firstly, the novel X->cb tagging technique can be utilized to measure the W->cb decay, opening a new avenue for constraining the CKM matrix element Vcb at the LHC, with performance surpassing traditional W->cb measurement methods. Additionally, the latent features of the pre-trained model can serve as new jet variables. When combined with other event-level features, it can facilitate high-performance event-level classifiers, leading to enhanced results in e.g., H->bb and HH->4b analyses.

Speaker: Leyun Gao (Peking University)

report.pdf

14:30 Coffee Break

10 From boosted H→WW* taggers to inclusively pre-trained jet models at the LHC¶

The SM and BSM searches via Lorentz-boosted jets are a key focus at the LHC, yet much of the potential phase space remains underexplored. In this talk, we first present the recent transformer-based tagger developed within CMS for SM H→WW* decays and demonstrate its superior performance. Building upon this, we introduce the next-generation Global Particle Transformer 3 (GloParT-3), designed to cover a broad phase space with 750 output nodes. We highlight its training philosophy, dubbed the Signature-Oriented Pre-training for Heavy-resonance ObservatioN (Sophon), and discuss the fine-tuning potential from its hidden layer scores. This new tagger not only improves tagging performance for boosted Higgs/Z/W/Top jets but also has advantages in detecting BSM resonances through direct tagging and fine-tuning, promising significant progress in jet tagging for future CMS analyses.

Speaker: Dawei Fu (Peking University)

HWW_to_pretrained_tagger_v2.pdf

11 Searching for Long-lived Particles at Future Lepton Colliders Using Deep Learning Techniques¶

This talk presents long-lived particle (LLP) searches in Higgs decays at future lepton colliders (e+e−→ZH) using deep learning techniques. Scanning LLP lifetimes from 0.001 to 100 ns and masses from 1 to 50 GeV, we find that the best sensitivity is achieved at 50 GeV and 1 ns, where deep neural networks, including CNNs and GNNs, reach up to 99% signal efficiency with zero Standard Model background. The Higgs branching ratio into LLPs can be constrained to $1.2\times {10}^{-6}$ with a dataset of $4\times {10}^6$ Higgs bosons, setting a state-of-the-art limit. Additionally, we explore the use of Local Contrastive Learning Machines (LCLMs) to further improve signal purity and reduce training uncertainty. These results demonstrate the power of deep learning in boosting LLP searches at future lepton colliders.

Speaker: Xiang Chen (Shanghai Jiao Tong University)

AI_for_hep_llp_search_20250225.pdf

12 Probing Light Fermiophobic Higgs in Type-I 2HDM with CNN-Based Diphoton Jet Substructure Analysis¶

In the Two-Higgs-Doublet Model (2HDM) Type-I, setting the coupling between a light CP-even Higgs boson and fermions to zero introduces a fermiophobic Higgs ℎ𝑓, which dominantly decays as ℎ𝑓 → γγ. Searching for ℎ𝑓 with a mass below 10 GeV presents a challenge, as conventional isolated diphoton methods become ineffective. This is due to the ℎ𝑓 → γγ decay producing highly collimated diphotons that merge into a single jet, making traditional photon-based searches impractical. To address this, we develop a machine learning model that leverages CNNs to analyze jet substructure, enabling the identification of diphoton jets and distinguishing them from QCD backgrounds. In this talk, I will introduce the 2HDM Type-I fermiophobic Higgs, discuss the unique challenges associated with detecting diphoton jets, and compare the effectiveness of cut-based and machine learning analyses in enhancing signal discrimination.

Speaker: Soojin Lee (Konkuk University)

IBS-AI+HEP-diphotonjet.pdf

Wednesday, 26 February ¶

09:00 Coffee Break

Morning Talks¶

Convener: Tianji Cai (SLAC)

13 Transformers for collider analysis¶

Attention-based Transformer networks have become increasingly prominent in collider analysis, delivering superior performance in tasks such as jet tagging. However, their high computational demands and substantial data requirements pose significant challenges. In this talk, I will explore the role of Transformer networks in LHC analysis, focusing on various attention mechanisms, including self-attention, cross-attention, and differential attention. Additionally, I will discuss different strategies for reducing network complexity while preserving high performance.

Speaker: Ahmed Hammad (KEK)

AI+HEP_Hammad.pdf

10:30 Coffee Break

14 Foundational Models in High Energy Physics: current trends and early results¶

Speaker: Vinicius Mikuni (NERSC, Berkeley Lab)

slides

12:00 Lunch

Afternoon Talks¶

Convener: Cheng-Wei Chiang (National Taiwan University)

15 A Tree-to-Token Paradigm for Jet Physics: Bridging High-Energy Phenomenology and Language Modeling¶

We propose a novel approach for more advanced analysis of jets, which serve as crucial objects of study in high-energy particle physics experiments. While conventional methods often treat jets as point clouds, our work focuses on the binary tree structure obtained during clustering. It explores ways to handle its constituents (such as tracks) using natural language processing language models. Specifically, we convert the binary tree derived from clustering into a bracketed representation, serialize it into a one-dimensional sequence, and then apply tokenization (quantization) to produce a data format suitable for training Transformer models. In this presentation, we will discuss the generation of the tree structure, the tokenization process, and, if time permits, the results of Transformer-based training, thereby demonstrating the potential of this novel perspective for jet analysis.

Speaker: Masahiro Morinaga (ICEPP, The University of Tokyo)

HEPAIAsia2025.v2.pdf

16 LLM-based AI assistant for HEP data analysis¶

The data processing and analyzing is one of the main challenges at HEP experiments, normally one physics result can take more than 3 years to be conducted. To accelerate the physics analysis and drive new physics discovery, the rapidly developing Large Language Model (LLM) is the most promising approach, it have demonstrated astonishing capabilities in recognition and generation of text while most parts of physics analysis can be benefitted. In this talk we will discuss the construction of a dedicated intelligent agent, an AI assistant at BESIII based on LLM, the potential usage to boost hadron spectroscopy study, and the future plan towards a AI scientist.

Speaker: Ke Li (Institute of High Energy Physics (IHEP), CAS)

AIHEP_KeLi_0226.pdf

15:00 Coffee Break

17 Grand Design of the Collider Space: How can AI and human work together to mine more physics from the LHC data?¶

As the Large Hadron Collider (LHC) generates hundreds of petabytes of data and even more with its high-luminosity upgrade, particle physics is entering a new era of data-driven discovery where machine learning (ML) techniques play a pivotal role. Alongside numerous task-specific ML algorithms, recent works have introduced foundation models excelling across diverse applications. At the heart of these ML models, especially general-purpose ones, is a geometric representation space for collider events that encodes the essential physics. Key questions then arise: How can we probe and refine the representation space for theoretical insights?
This talk presents a first step towards the construction and analysis of a collider space. I will introduce two metric structures, one inspired by the mathematical theory of optimal transport and the other grounded in the physical phase space. Such explicitly-defined metrics enable comparisons with the representation space implicitly generated by an ML model. This paves the way to further dissect a model’s internals and offers hope for discovering new physical laws directly from data.

Speaker: Tianji Cai (SLAC National Accelerator Laboratory)

Grand Design of the Collider Space: How can AI and human work together to mine more physics from the LHC data?

18:00 Workshop Dinner: Rice Cooked in a Bamboo Tube (대나무통밥, 竹管飯)

Venue: https://maps.app.goo.gl/wZic4pDDzU3FvgNp7

Thursday, 27 February ¶

09:00 Coffee Break

Morning Talks¶

Convener: Vinicius Mikuni (NERSC, Berkeley Lab)

18 Machine Learning for Galactic Dynamics: Mapping Dark Matter in the Local Universe¶

Recent advances in machine learning (ML), particularly neural density estimation like normalizing flows, diffusion models, and flow matching, have opened new doors for high-precision, model-independent density estimation. These techniques are highly valuable for galactic dynamics studies, as they allow us to estimate the distribution of stars in phase space (position and velocity) without relying on traditional simplified models. By combining these ML-based stellar density estimates and equations of motion solvers for inferring gravitational fields, we can measure the local dark matter density in a model-independent way. This talk presents new research opportunities in this direction, focusing on modeling objects in our local universe using neural networks and using them for understanding local galactic dark matter distribution. We anticipate that these modern machine learning-based approaches will allow us to fully utilize the potential of current and future astronomical catalogs, significantly improving our understanding of dark matter in the local universe.

Speaker: Sung Hak Lim (IBS CTPU-PTC)

AIHEP_short.pdf

10:30 Coffee Break

19 Bridging GW and HEP: Next Generation AI at the Frontier¶

Gravitational Wave (GW) Physics has entered a new Multi-Messenger Astronomy (MMA) era, marked by increasing detections from GW observatories led by LIGO, Virgo, and KAGRA collaborations. This presentation will introduce the KAGRA experiment and explore the transformative role of machine learning (ML) in GW data analysis — some successful ML key applications, among which glitch identification, will be discussed.demonstrate the transformative role of machine learning (ML) in GW data analysis.

This talk also bridges advancements in computational techniques between fundamental research in Astrophysics and High-Energy Physics (HEP). Innovative solutions for addressing next-generation data analysis challenges will be presented, with a focus on the use of modern ML tools within the ROOT C++ Framework (CERN) and introducing Anaconda HEP-Forge for rapid software deployments. These tools, available as simple libraries, integrate key requirements for astrophysical analysis — such as vector manipulation, KAFKA & other Cloud data transfers, and complex tensor computations—enabling efficient ML training & inference on both CPU and GPU technologies.

Speaker: Marco Meyer-Conde (Tokyo City University)

2025 02 27 Workshop AI+HEP (MEYER).pdf

20 Modern Computational Approaches to Early Universe Modeling¶

In physics, while analytical calculations remain appealing, there are situations where the use of computers becomes indispensable. A straightforward example is the three-body problem, where even the interactions of just three bodies are challenging to solve analytically. This necessity is similarly evident in studies of the early universe. Understanding the dynamics of the inflaton requires numerical analysis through lattice simulations in this context. Accessing high-performance computing (HPC) used to be difficult, but it is now widely available and significantly more powerful. We apply numerical results obtained through HPC to various physical problems, including the dark matter problem. However, practical constraints still hinder our ability to fully simulate the early universe, so we are exploring machine learning as a means to overcome these challenges.

Speaker: Jong-Hyun Yoon (Chungnam National University)

20250227_IBS_YOON.pdf

12:00 Lunch

Afternoon Talks¶

Convener: Manqi Ruan (Institute of High Energy Physics (IHEP), CAS)

21 Physics-Conditioned Diffusion Models for Lattice Gauge Field Theory¶

We develop diffusion models for simulating lattice gauge theories, where stochastic quantization is explicitly incorporated as a physical condition for sampling. We demonstrate the application of this novel sampler to U(1) gauge theory in two spacetime dimensions and find that a model trained at a small inverse coupling constant can be extrapolated to larger inverse coupling regions without encountering the topological freezing problem. Additionally, the trained model can be employed to sample configurations on different lattice sizes without requiring further training. The exactness of the generated samples is ensured by incorporating Metropolis-adjusted Langevin dynamics into the generation process. Furthermore, we demonstrate that this approach enables more efficient sampling of topological quantities compared to traditional algorithms such as hybrid Monte Carlo and Langevin simulations.

Speaker: Qianteng Zhu (RIKEN / Shanghai Jiao Tong University)

DM_U1_AIHEP_Qianteng.pdf

22 AI with Hamiltonian Mechanics: From Predictions to Understanding¶

How do AI and Hamiltonian Mechanics drive each other’s advancement, enabling stronger predictions and offering deeper insights? In this presentation, we explore how AI can not only predict but also understand Hamiltonian dynamics. First, we introduce a robust long-term prediction framework that combines an improved Hamiltonian Neural Network with Bayesian data assimilation. This method achieves high accuracy even in 3D environments, showing that AI can become a powerful tool for real-world applications.

Building on these predictive successes, we now ask whether AI can truly understand Hamiltonian mechanics. We introduce operator learning, a method that allows AI to handle mappings in infinite-dimensional spaces, then apply it to Hamiltonian systems. We test whether AI can produce phase space trajectories from an arbitrary potential function without using equations or solvers. Our results show that, under certain conditions, AI can indeed predict these trajectories. Finally, we discuss current limitations, propose future research directions, and consider how AI might advance scientific discovery.

Speaker: Tae-Geun Kim (Yonsei University)

AI_with_Hamiltonian_AIHEP-EastAsia.pdf

14:30 Coffee Break

23 LeStrat-Net: machine learning in Lebesgue style stratified Monte Carlo¶

In this talk we describe LeStrat-Net: a new algorithm to perform Monte Carlo integration using Lebesgue style stratified sampling and machine learning. We divide the domain of integration based on the height of the integrand, similar to Lebesgue integration. The isocontours of the integrand can in principle create regions of any shape and with many disconnected subregions. We take advantage of the capacity of neural networks to learn complicated functions in order to predict these complicated divisions and preclassify large samples of the domain space. From this preclassification we can select the required number of points to perform a number of tasks such as variance reduction, integration and even event selection. The network ultimately defines the regions with what it learned and is also used to calculate the multi-dimensional volume of each region.

Speaker: Raymundo Ramos (Korea Institute for Advanced Study)

Ramos-AIHEP_EA.pdf

24 AI assist IBP reduction¶

In high-energy physics, calculating Feynman integrals efficiently remains challenging due to the computational demands of traditional methods. We are developing a new approach relying on code generation capabilities of large language models to optimize integration-by-parts (IBP) reduction by combining advanced techniques with classical algorithms. This method shows significant improvements in efficiency and scalability, with reduced memory usage compared to conventional techniques.

Speaker: Zhuo-Yang Song (Peking University)

ibp.pptx

Friday, 28 February ¶

09:00 Coffee Break

25 Guided Discussions¶

26 Closing Remarks¶

12:00 END