Learning Transverse Momentum Distributions from Raw Scattering Events via Conditional Diffusion
本文提出了一种条件扩散模型，直接从原始SIDIS事件动量学中学习TMD PDFs，解决了传统方法依赖参数化函数形式和迭代拟合的局限性。
0 citationsRead paper
Selected work
Representative Papers
Multi-Dataset Inverse Problem Solving with Distributed Generative AI
本文提出了一种基于分布式生成AI的框架，用于从多个异构数据集中同时分析并提取未知参数，解决了多数据集逆问题中的计算和异质性挑战。
0 citationsRead paper
S-matrix informed neural networks for amplitude analysis
本文提出S-矩阵信息神经网络解决粒子物理中散射振幅重构问题，直接从数据学习并尊重基本原理，同时开发新数据选择程序。
0 citationsRead paper
Posterior Inference of Hamiltonian Parameters from RIXS Spectroscopy
This study addresses the challenge of posterior inference for Hamiltonian parameters in Resonant Inelastic X-ray Scattering (RIXS) spectra by proposing the first simulation-based inference framework. Integrating a physics-aware Vision Transformer, truncated marginal neural ratio estimation, and conditional flow matching, this approach enables efficient and accurate full posterior inference for nickel compounds under few-shot conditions. The method not only uncovers critical parameter correlations and yields predicted spectra highly consistent with experimental data but also achieves reliable uncertainty quantification. Consequently, this work establishes a novel paradigm for the spectroscopic analysis of complex quantum materials, overcoming longstanding limitations in extracting precise physical parameters from RIXS measurements through advanced probabilistic modeling and domain-informed deep learning architectures.
0 citationsRead paper
Low Latency, High Bandwidth Streaming of Experimental Data with EJFAT
To address the co-design challenge of ultra-low latency and high bandwidth in experimental data stream transmission within edge computing clusters, this paper proposes an FPGA-based end-to-end acceleration architecture. The architecture innovatively integrates hardware-accelerated streaming compression/decompression, fine-grained data sharding and reassembly, UDP packet-level NAT redirection, and high-speed forwarding—enabling seamless, low-overhead integration from edge nodes to compute clusters. Compared to conventional TCP/IP stack implementations, it reduces end-to-end transmission latency by 42% (measured) and achieves >92% bandwidth utilization. Its compatibility with and scalability on U.S. Department of Energy (DOE) scientific infrastructure are validated through cross-domain experiments across JLab–ESnet–LBNL. This work establishes a deployable hardware acceleration paradigm for time-sensitive scientific data processing.
0 citationsRead paper
Recent publications
Latest Papers
Learning Transverse Momentum Distributions from Raw Scattering Events via Conditional Diffusion
本文提出了一种条件扩散模型，直接从原始SIDIS事件动量学中学习TMD PDFs，解决了传统方法依赖参数化函数形式和迭代拟合的局限性。
0 citationsRead paper
Multi-Dataset Inverse Problem Solving with Distributed Generative AI
本文提出了一种基于分布式生成AI的框架，用于从多个异构数据集中同时分析并提取未知参数，解决了多数据集逆问题中的计算和异质性挑战。
0 citationsRead paper
S-matrix informed neural networks for amplitude analysis
本文提出S-矩阵信息神经网络解决粒子物理中散射振幅重构问题，直接从数据学习并尊重基本原理，同时开发新数据选择程序。
0 citationsRead paper
Posterior Inference of Hamiltonian Parameters from RIXS Spectroscopy
This study addresses the challenge of posterior inference for Hamiltonian parameters in Resonant Inelastic X-ray Scattering (RIXS) spectra by proposing the first simulation-based inference framework. Integrating a physics-aware Vision Transformer, truncated marginal neural ratio estimation, and conditional flow matching, this approach enables efficient and accurate full posterior inference for nickel compounds under few-shot conditions. The method not only uncovers critical parameter correlations and yields predicted spectra highly consistent with experimental data but also achieves reliable uncertainty quantification. Consequently, this work establishes a novel paradigm for the spectroscopic analysis of complex quantum materials, overcoming longstanding limitations in extracting precise physical parameters from RIXS measurements through advanced probabilistic modeling and domain-informed deep learning architectures.
0 citationsRead paper
Low Latency, High Bandwidth Streaming of Experimental Data with EJFAT
To address the co-design challenge of ultra-low latency and high bandwidth in experimental data stream transmission within edge computing clusters, this paper proposes an FPGA-based end-to-end acceleration architecture. The architecture innovatively integrates hardware-accelerated streaming compression/decompression, fine-grained data sharding and reassembly, UDP packet-level NAT redirection, and high-speed forwarding—enabling seamless, low-overhead integration from edge nodes to compute clusters. Compared to conventional TCP/IP stack implementations, it reduces end-to-end transmission latency by 42% (measured) and achieves >92% bandwidth utilization. Its compatibility with and scalability on U.S. Department of Energy (DOE) scientific infrastructure are validated through cross-domain experiments across JLab–ESnet–LBNL. This work establishes a deployable hardware acceleration paradigm for time-sensitive scientific data processing.
0 citationsRead paper