Quantum Computing and Machine Learning Workshop 2025

19 Aug 2025, 09:00 → 23 Aug 2025, 21:30 Asia/Shanghai

To promote the application of quantum computing and machine learning in high-energy theoretical and experimental physics, we will hold a workshop on quantum computing and machine learning at Shandong University, Qingdao, China. Researchers from domestic and international fields related to quantum computing and machine learning are sincerely invited to exchange ideas and discuss on the application of quantum computing algorithms, machine learning, hardware advances, and the use of development platforms.

Look forward to meeting you in Qingdao!

Registration:

• Registration fee: 1500 CNY for regular attendee; 800 CNY for student 

Zoom room:

https://cern.zoom.us/j/69679099518?pwd=iomp2TRbU80Ryd7lPo8ClE10CODM7y.1

poster.jpg

合影.jpg

第三届量子计算与机器学习会议通知v3.pdf

Tuesday, 19 August

Registration

Dinner

Wednesday, 20 August

Session

1 欢迎致辞

Speaker: Xingtao Huang (Shandong University)

2 AI for High Energy Theory

Speaker: Dr Tianji Cai

AI for High Energy Theory.pptx

3 超导量子计算进展

Speaker: 明 龚

4 Physics reach of the CEPC - in the scope of AI enhanced reco & analysis

Speaker: Manqi Ruan (IHEP)

CEPC - Holistic - QCandML.pdf

Coffee break

Session

5 机器学习FPGA加速在STCF触发系统预研中的初步应用

Speaker: Changqing Feng (中国科学技术大学)

机器学习FPGA加速在STCF触发系统预研中的初步应用-V1.10.pptx

6 Machine Learning for real-time data processing at Belle II

Speaker: 启东/Qidong 周/Zhou (山东大学/Shandong university)

QumtAI_Belle2TDAQ_QidongZhou.pdf

7 Development of the neural network algorithm on Versal AI-engine

Next generation high-energy experiments aim to probe unexplored regimes in particle physics. Real-time data processing of terabyte-per-second is a critical technological bottleneck, especially, track reconstruction suffer from degraded efficiency under high-background environment. To address this challenge, we are designing a heterogeneous data processing system that integrates a real-time graph neural network(GNN) algorithm into the Xilinx Versal ACAP, as a potential upgrade for future detector. To demonstrate the feasibility of the system, we implement baseline neural network deployments, including an online deep neural network (DNN) algorithm and an offline GNN algorithm, on the AI Engine of the Xilinx Versal ACAP. As a result, the best latency achieved for the DNN deployment is approximately 2 microseconds per event, while the GNN deployment achieves 1 milli second latency with a 27 ε-nearest neighbor (ε-NN) graph construction input. In this report, we will show the design and performance of the development.

Speaker: 兆志 刘

量子计算2025.pdf

Lunch

Session

8 Quantum & quantum-inspired optimization in high energy physics

Speaker: Hideki (英希) Okawa (大川) (IHEP)

9 正确地编写量子程序

量子计算在凝聚态物理、高能物理、量子化学等诸多领域的计算任务中展现出巨大的潜力。为了应用量子计算技术，研究人员和开发者需要编写量子程序。编程过程可能出现错误，然而量子计算独特的性质使得发现和定位量子程序中的错误存在困难。例如，由于测量导致塌缩使得无法跟踪程序运行中间变量的值；错误对运行结果的影响难以观察等。本报告将关注“正确地编写量子程序”这一越来越重要的主题。报告将首先分享几类典型的量子程序的编程错误，并给出对应的发现和避免的方案；然后简要介绍量子软件工程这一新兴的研究领域；最后介绍高能所计算中心近年来在量子程序开发工具方面的工作和计划。

Speaker: 沛洵 龙 (中国科学院高能物理研究所)

正确地编写量子程序_龙沛洵.pptx

10 Re-discovery of $Z_c(3900)$ at BESIII Based on Quantum Machine Learning

Quantum Machine Learning (QML) is an advanced data analysis technique, which can detect data structures within massive datasets, building models to achieve data prediction, classification, or simulation, with less human intervention. However, the practical viability of QML still remains a topic of debate, requiring more examples of real data analysis with quantum hardware for its further verification.
Based on this background, our research focuses on the application of QML in the re-discovery of $Z_c(3900)$, which was first observed by BESIII collaboration in 2013 while analyzing the decay process of $Y(4260)$. Using the same $525 \ \mathrm{pb}^{-1}$ data collected at $\sqrt{s} = 4.26 \ \mathrm{GeV}$, this study applies Quantum Support Vector Machine (QSVM) method to event selection criteria, using classical cut-based and ML-based analysis strategy as references. A 1-D fit will be applied to the selected dataset to extract the parameters of $Z_c(3900)$ in order to evaluate the selection efficiency. The invariant mass distribution will then be plotted and compared with the results of traditional analysis.

Speaker: Siyang Wu (Shandong University)

11 Quantum Simulations of Particle Scattering in Lattice Field Theories

Speaker: Yahui Chai

QingDao2025.pdf

Coffee break

Session

12 Quantum simulation for real-time dynamics at colliders

Speaker: Ying-Ying Li (IHEP)

13 Quantum simulations of quantum electrodynamics in Coulomb gauge

In recent years, the quantum computing method has been used to address the sign problem in traditional Monte Carlo lattice gauge theory (LGT) simulations. We propose that the Coulomb gauge (CG) should be used in quantum simulations of LGT. Since the redundant degrees of freedom of gauge fields can be eliminated in CG, the Hamiltonian in CG does not need to be gauge invariance, allowing the gauge field to be discretized naively. Then the discretized gauge fields and fermion fields should be placed on momentum and position lattices, respectively. Under this scheme, the CG condition and Gauss's law can be conveniently preserved by solving for the polarization vectors from algebraic equations. Furthermore, we discuss the mapping of gauge fields to qubits and evaluate the associated qubit and gate cost of this framework. We point out that this formalism is efficient for simulating hadron scattering processes on future fault-tolerant quantum computers. Finally, we calculate the vacuum expectation value of the U(1) plaquette operator and the Wilson loop on a classical device to test the performance of our discretization scheme.

Speaker: Li Tianyin (South China Normal University)

tli_CGQED.pdf

14 Computing n-time correlation functions without ancilla qubits

The $n$-time correlation function is pivotal for establishing connections between theoretical predictions and experimental observations of a quantum system. Conventional methods for computing $n$-time correlation functions on quantum computers, such as the Hadamard test, generally require an ancilla qubit that controls the entire system -- an approach that poses challenges for digital quantum devices with limited qubit connectivity, as well as for analog quantum platforms lacking controlled operations. Here, we introduce a method to compute $n$-time correlation functions using only unitary evolutions on the system of interest, thereby eliminating the need for ancillas and the control operations. This approach substantially relaxes hardware connectivity requirements for digital processors and enables more practical measurements of $n$-time correlation functions on analog platforms. We demonstrate our protocol on IBM quantum hardware up to 12 qubits to measure the single-particle spectrum of the Schwinger model and the out-of-time-order correlator in the transverse-field Ising model. In the demonstration, we further introduce an error mitigation procedure based on signal processing that integrates signal filtering and correlation analysis, and successfully reproduces the noiseless simulation results from the noisy hardware. Our work highlights a route to exploring complex quantum many-body correlation functions in practice, even in the presence of realistic hardware limitations and noise.

Speaker: Xiaoyang Wang (P)

computing_n_time_correlation_functions.pdf

15 基于量子计算的SU(2)规范理论中的手性不平衡研究

We implement a variational quantum algorithm to investigate the chiral condensate in a 1+1 dimensional SU(2) non-Abelian gauge theory. The algorithm is evaluated using a proposed Monte Carlo sampling method, which allows the extension to large qubit systems. The obtained results through quantum simulations on classical and actual quantum hardware are in good agreement with exact diagonalization of the lattice Hamiltonian, revealing the phenomena of chiral symmetry breaking and restoration as functions of both temperature and chemical potential. Our findings underscore the potential of near-term quantum computing for exploring QCD systems at finite temperature and density in non-Abelian gauge theories.

Speaker: 国丰 张 (South China Normal University)

基于量子计算的SU(2)规范理论中的手性不平衡研究.pdf

16 Nuclear structure and dynamics calculations on quantum computers

Speaker: 伟杰 杜

山大会议_QcML_v1_杜伟杰_CN.pdf

Dinner

Thursday, 21 August

Session

17 AI usage at ILC reconstruction

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

250819-hfreco-chinaqcai-suehara.pdf

18 面向量子AI的算法研究和设计

Speaker: 东灵 邓

19 分类与异常检测技术在ATLAS实验上的新进展

Speaker: Bingxuan Liu (The Ohio State Univ)

ML_Advances_In_ATLAS_BingxuanLiu.pdf

20 分类与异常检测技术在CMS实验上的新进展

Speaker: Congqiao Li (Peking University)

25.08.21_QCML_CMS_Progress.pdf

Coffee break

Session

21 智慧光源大脑2.0平台

Speaker: 丽娜 赵

赵丽娜LinaZhao - 智慧光源大脑IPSBrain2.pptx

22 粒子物理数据分析AI助手-赛博士

Speaker: Zheng-De Zhang (张正德) (IHEP(高能所))

Dr.Sai-Progress-20250821-v4.0.pdf

23 A New Intelligent LHAASO: Pushing the Limits of AI for High-Energy Astronomy

The Large High Altitude Air Shower Observatory (LHAASO) stands at the forefront of high-energy astronomy, with core physics goals including the identification of galactic PeVatrons, indirect dark matter searches, and cosmic ray origin studies. However, achieving these objectives is challenged by the critical need to process data streams while maintaining real-time background rejection and precise angular resolution to high-energy sources. In this report, we aim to enhance LHAASO’s fundamental performance using AI. ​​We test several distinct model architectures, discuss the origins of performance differences, and probe the boundaries of AI capabilities within LHAASO's data.​

Speaker: Dr Chenyuan Xu (之江实验室)

AI_for_LHAASO_徐宸原_250819.pdf

Lunch

Session

24 Intern-S1: 科学多模态大模型

Speaker: 琦鹏 郭

25 磐石科学大模型

Speaker: 林静 李

26 AI赋能科学研究

Speaker: 辉 陈

量子计算与机器学习研讨会8月21号-陈辉.pdf

27 BigBang：用于科学多任务学习的自回归大模型

Speaker: 恒魁 吴

BigBangTransformer_青岛0821_v1.pdf

Coffee break

White paper discussion

Discussion.pdf

Dinner

Friday, 22 August

Session

28 Neural Network Feild Theory

Speaker: Anindita Maiti

29 AI for discovery: a preliminary glimpse

Speaker: 东山 见 (北京大学)

AI discovery.pdf

30 ML for lattice gauge field simulations

Speaker: Wei Wang (Shanghai JiaoTong University)

ML for lattice gauge field simulations.pptx

Coffee break

Session

31 The Neural Networks with Tensor Weights and the Corresponding Fermionic Quantum Field Theory

In this paper, we establish a theoretical connection between complex-valued neural networks (CVNNs) and fermionic quantum field theory (QFT), bridging a fundamental gap in the emerging framework of neural network quantum field theory (NN-QFT). While prior NN-QFT works have linked real-valued architectures to bosonic fields, we demonstrate that CVNNs equipped with tensor-valued weights intrinsically generate fermionic quantum fields. By promoting hidden-to-output weights to Clifford algebra-valued tensors, we induce anticommutation relations essential for fermionic statistics. Through analytical study of the generating functional, we obtain the exact quantum state in the infinite-width limit, revealing that the parameters between the input layer and the last hidden layer correspond to the eigenvalues of the quantum system, and the tensor weighting parameters in the hidden-to-output layer map to dynamical fermionic fields. The continuum limit reproduces free fermion correlators, with diagrammatic expansions confirming anticommutation. The work provides the first explicit mapping from neural architectures to fermionic QFT at the level of correlation functions and generating functional. It extends NN-QFT beyond bosonic theories and opens avenues for encoding fermionic symmetries into machine learning models, with potential applications in quantum simulation and lattice field theory.

Speaker: GUOJUN HUANG (The University of Hong Kong, Shenzhen)

CVNN-QFT.pdf

32 Machine Learning for Parton-Level Studies of Quantum Entanglement Using pp→ττ

Quantum entanglement is a hallmark feature of quantum mechanics, manifesting as correlations between subsystems that cannot be fully described without one another, regardless of spatial separation. While entanglement has been observed in processes such as $pp\to t \bar{t}$ and thoroughly analyzed in Higgs decay channels ($H\to VV$) at the Large Hadron Collider (LHC), it remains comparatively underexplored in the $pp\to \tau\tau$ system. In this study, we adapt OmniLearn, a foundational model for solving all jet physics tasks, to reconstruct the neutrino information in the final state of $pp\to \tau\tau$ system, which is an essential step toward probing quantum entanglement in this channel. Good neutrino reconstruction has reached now, which is the key to the following steps in the reconstruction level study.

Speaker: Baihong Zhou (Tsung-Dao Lee Institute, Shanghai Jiao Tong University)

25_08_22.pdf

33 Vision / Language Calorimeter: deep-learning-based anti-neutron reconstruction in an electromagnetic calorimeter

Long-lived neutral hadrons, including (anti-)neutron and $K^0_L$ meson, are important probes for physics in the tau-charm energy region. However, most tau-charm facilities do not include dedicated hadronic calorimeters, and their neutral hadron detection must rely on the electromagnetic calorimeter (EMC). Because the EMC's small volume and dense material only partially contain hadronic showers, conventional reconstruction methods face significant limitations. In the talk, we introduce Vision Calorimeter (ViC) and Language Calorimeter (LaC), two deep-learning frameworks inspired by modern architectures from computer vision and natural language processing fields. By leveraging an end-to-end, data-driven approach, ViC & LaC perform unified reconstruction of anti-neutrons, simultaneously identifying particle type, estimating the incident position on the EMC, and inferring the momentum magnitude.

Speaker: Yangu Li (University of Chinese Academy of Sciences)

ViC_LMC_250822.pdf

34 AI-assisted Four Top Quark Reconstruction

Speaker: 翔 陈

AI-asisted_Top_Reconstruction_250822.pdf

Lunch

Session

35 BESIII量能器快速模拟

Speaker: Tong Liu (IHEP)

EMC.pdf

36 1-1 correspondence reconstruction at electron-positron Higgs factories

Particle flow reconstruction has become the standard paradigm for event reconstruction at current high-energy collider experiments. Its ultimate goal is to establish a one-to-one (1-1) correspondence between reconstructed and truth incident particles, while achieving highly efficient particle identification. In realistic experimental environments, however, this ideal correspondence is inevitably compromised by factors such as limited detector acceptance, finite spatial granularity, and intrinsic limitations of reconstruction algorithms.
In this talk, I will take the CEPC as an example to introduce studies on 1-1 correspondence in particle flow reconstruction at future electron-positron Higgs factories. By utilizing truth links, we conduct a detailed analysis and categorization of reconstructed particles. In combination with advanced machine learning techniques, this enables efficient identification and suppression of fake particles—which currently represent a major bottleneck in CEPC analyses. Consequently, the invariant mass resolution for di-jet Higgs boson final states is improved by 15%. In addition, the developed method enables efficient identification of nine particle species. For five charged particle types (electron, muon, pion, kaon, and proton) as well as photons, identification efficiencies exceed 97%. For the three neutral hadron species (long-lived neutral kaon, neutron, and antineutron), identification efficiencies reach 75–80%. These results considerably extend the range of particle species that can be reliably identified compared to conventional methods.

Speaker: Yuexin Wang

Yuexin_1-1_20250822.pdf

37 Study of CNN Algorithms for PID System in STCF

The Super Tau Charm Facility (STCF) is a next-generation electron-positron collider proposed in China, operating at a center-of-mass energy range of 2–7 GeV with a peak luminosity of 0.5×10³⁵ cm⁻²s⁻¹. In high-energy experiments, the identification of high-momentum charged hadrons is crucial for studying various physics processes. At STCF, the particle identification (PID) system based on Cherenkov detection technology includes a barrel Ring Imaging Cherenkov detector (RICH) and an endcap Time-of-Flight detector utilizing internally reflected Cherenkov light (DTOF). Additionally, a Barrel Time-of-Flight detector (BTOF) is designed as a backup system for the RICH. With the development of deep learning, particle identification algorithms have continued to evolve. In this study, convolutional neural network (CNN)-based PID algorithms were developed for the three sub-detectors of the PID system. By taking as input the 2D images converted from Cherenkov photon hit patterns on photomultiplier tubes, along with kinematic information from the tracking system, the model predicts the probability of different particle types. Preliminary results indicate that the algorithm effectively integrates image features with particle kinematic information, achieving outstanding performance in particle identification and offering a promising high-precision solution for PID at STCF.

Speaker: 万琳 林 (华中师范大学)

Study of CNN Algorithms for PID System in STCF.pdf

38 Studying Hadronic Shower Development in HERD CALO with Machine Learning

The High Energy cosmic-Radiation Detection (HERD) facility will be installed as a space astronomy payload on the China Space Station in 2028. The three-dimensional imaging calorimeter (CALO) of HERD comprises about 7500 lutetium yttrium oxy-orthosilicate (LYSO:Ce) cubes, where the topological development of hadronic showers can be measured. Over the years, advancements in deep learning, particularly Convolutional Neural Networks (CNNs) and Transformers, have demonstrated state-of-the-art performance in high energy physics. These deep learning architectures are designed to exploit large datasets to reduce complexity and find new features in data. In this seminar, I will discuss the application of the CNN and Transformer techniques to precisely reconstruct the parameters of a hadronic shower initiated in HERD CALO. Both models are optimized using isotropic proton simulations and demonstrate superior performance over a wide energy range from 30 GeV up to 1 TeV. After some adaptation, these architectures could be applied for different types of calorimeters.

Speaker: Xiao-Fan Tang (IHEP)

39 Application of machine learning method for energy reconstruction on space based high granularity calorimeter

The High Energy Cosmic-Radiation Detection Facility (HERD) is dedicated to achieving several scientific objectives, including the search for dark matter, precise measurement of the cosmic ray spectrum, and gamma-ray sky survey observations.
HERD’s innovative design incorporates a three-dimensional imaging calorimeter with five sensitive faces, significantly enhancing geometric acceptance. However, this design introduces a new challenge for reconstructing particles incident from all directions. This article aims to integrate rapidly advancing deep learning techniques into the reconstruction task. Utilizing simulation data, Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and other deep learning networks are employed to reconstruct the energy of isotropic electrons. Model performance sees a significant boost through the application of end-layer visible energy correction and a “multi-class multi-prediction approach,involvingdifferentmodelstrainedfordistinctenergy ranges. Moreover, recognizing differences between simulation and physical samples, the model is validated using the beam test data.
The model predicts an energy resolution of better than 1% for simulation isotropic electrons ranging from 10 to 1000 GeV. In the case of beam data, the model achieves an energy resolution of 1.3% at 200 GeV, comparable to traditional methods. The results demonstrate the significant potential of deep learning in the reconstruction of three-dimensional calorimeters.

Speaker: UNKNOWN 廖川黎

e-_energyreco@20250822qingdao.pdf

Coffee break

White paper discussion

Discussion.pdf

Banquet

Saturday, 23 August

Session

40 Reconstruction of Electromagnetic Shower Axis in the AMS-02 ECAL using Deep Learning Method

The Electromagnetic Calorimeter (ECAL) in the AMS-02 experiment is a 3D imaging detector and plays a pivotal role in various physics analysis results. The precise reconstruction of electromagnetic shower axis in the ECAL contributes to a better understanding of its performance in particle identification as well as pointing capability of gamma rays. Conventional methods reconstruct the shower axis based on 3-dimensional shower profiles combined with detector-related effects.
In this talk, we present an innovative deep learning technique using Residual Network (ResNet) model to obtain the key parameters of shower axis, including the inclination angle and the incident position. The ResNet model is trained with Monte Carlo simulation and validated in the cosmic ray electron data collected on the ISS. Significant improvement in the angular and spatial resolution covering the energy range from 2 GeV to 1 TeV is observed and presented.

Speaker: Yaozu Xiong (Zhejiang University)

2025-08-23-Qingdao-YaozuXiong.pdf

41 基于机器学习的eXTP卫星PFA载荷地面和在轨数光电子径迹重建

eXTP卫星是我国下一代旗舰级X射线天文台，其中的PFA载荷对能量为2-10 keV的X射线进行可成像的偏振、时变和能谱观测。PFA载荷的气体X射线偏振探测器基于X射线在气体探测器中发生光电效应时光电子出射方向在垂直于X射线入射方向的平面内的角分布来测量偏振。在这类探测器的数据分析中，最关键的就是光电子径迹的重建。由于光电子径迹是通过测量光电子在气体探测器内与气体分子碰撞而产生的电离电子所描绘的径迹而获得的。光电子在运动过程中，除了电离出电子，还会因为和气体分子中的原子核发生库里散射而改变方向，同时还会伴随有俄歇电子的产生，所有这些因素增加了光电子径迹重建的复杂性。传统的光电子径迹重建方法是力矩法，随着机器学习技术的发展，我们开展了基于深度学习的地面和在轨光电子径迹重建方法的研究，本报告将介绍相关研究情况。

Speaker: Dr 维春 姜 (高能所)

基于机器学习的eXTP卫星PFA载荷地面和在轨光电子径迹重建.pdf

42 Towards a foundational jet model: Enhancing generalization with contrastive “gen-reco” pre-training

A foundation jet model aims to achieve optimal performance across all jet analysis tasks while ensuring strong generalization. Building on Sophon, a pre-trained jet classification model, we develop Sophon++, which employs contrastive learning to connect initial, parton-level, and reconstruction-level particles, enabling continuous encoding of generator-level particle configurations into the model’s latent space. While matching Sophon in classification performance, Sophon++ demonstrates stronger generalization through several fine-tuning tasks. This work provides a promising pathway towards a foundation jet model for analysis.

Speaker: Zixun Kou (Peking University)

Sophon++_report_Zixun_kou_20250823.pdf

Coffee break

Session

43 Pile-up events discrimination based on machine learning in JUNO

In the large liquid scintillator detector JUNO, $\beta$-decays of $^{14}\mathrm{C}$ inevitably deposit energy within the detector, producing scintillation light. These photons can potentially overlap with positron ($e^{+}$) signals, forming pile-up events. This pile-up effect can consequently impact the fine reconstruction of the $e^{+}$ signal. We have employed three distinct machine learning models – CNN, Transformer, and KamNet – to discriminate between pile-up events and pure $e^{+}$ events. This report provides a detailed discussion of this work.

Speaker: 肇祥 吴 (IHEP)

Pile-up events Discrimination Based on Machine Learning in JUNO.pdf

44 JUNO上基于深度学习的大气中微子和宇宙线缪子重建

Speaker: 晓晗 谭 (S)

reconstruction_atmosphericNu_and_mu@QCML.pdf

45 Transformer-based Fermi/GBM Background Predictor

As a cornerstone scientific payload aboard NASA's Fermi Gamma-ray Space Telescope, the Gamma-ray Burst Monitor (GBM) is specifically designed to detect and study cosmic gamma-ray bursts (GRBs) and other high-energy transients, such as solar flares, magnetar outbursts, and Terrestrial Gamma-ray Flashes (TGFs). The primary observational challenge for GBM stems from the inherent complexity and highly variable nature of background radiation, which contaminates astrophysical signals and degrades the precision of data analysis. Therefore, accurate modeling of background radiation is critical for effectively extracting signals and enhancing the potential for scientific discoveries. This paper proposes a deep neural network architecture for background spectral modeling of Fermi/GBM detectors. The proposed model achieves efficient estimation of background signals for individual NaI detectors through the following methodologies:
(1) Based on the Transformer architecture, dedicated models with parameter isolation were established for each detector of Fermi/GBM, significantly reducing the number of network parameters
(2) Integrating a median absolute deviation (MAD) preprocessor and Cauchy loss function to address outlier sensitivity in astronomical photon-counting data
(3) Implementing a lightweight training framework using a single NVIDIA RTX 4080 GPU (16GB VRAM), ensuring computational efficiency while enabling end-to-end validation of the model. The experimental results demonstrate that our model achieves accurate predictions for both light curves and energy spectra, reaching the precision required for orbit revisit predictions.

Speaker: Mr 攀 鲁 (中国科学院高能物理研究所)

鲁攀.pptx

46 Transfer learning empowers material Z classification with muon tomography

Cosmic-ray muon sources exhibit distinct scattering angle distributions when interacting with materials of different atomic numbers (Z values), facilitating the identification of various Z-class materials, particularly those radioactive high-Z nuclear elements. Most of the traditional identification methods are based on complex statistical iterative reconstruction or simple trajectory approximation. Supervised machine learning methods offer some improvement but rely heavily on prior knowledge of target materials, significantly limiting their practical applicability in detecting concealed materials. For the first time, transfer learning is introduced into the field of muon tomography in this work. We propose two lightweight neural network models for fine-tuning and adversarial transfer learning, utilizing muon scattering data of bare materials to predict the Z-class of materials coated by typical shieldings (e.g., aluminum or polyethylene), simulating practical scenarios like cargo inspection and arms control. By introducing a novel inverse cumulative distribution-based sampling method, more accurate scattering angle distributions could be obtained from data, leading to an improvement by nearly 4% in prediction accuracy compared with the traditional random sampling-based training. When applied to coated materials with limited labeled or even unlabeled muon tomography data, the proposed method achieves an overall prediction accuracy exceeding 96%, with high-Z materials reaching nearly 99%. Simulation results indicate that transfer learning improves prediction accuracy by approximately 10% compared to direct prediction without transfer. This study demonstrates the effectiveness of transfer learning in overcoming the physical challenges associated with limited labeled/unlabeled data, and highlights the promising potential of transfer learning in the field of muon tomography.

Speaker: Mr 浩辰 王 (合肥工业大学)

muon transfer报告.pdf

muon transfer报告.pptx

47 会议总结

Speaker: 腾 李 (Shandong University)

总结.pdf

Lunch

Free discussion

Dinner