AI for Hadron Spectroscopy at JLab

Jun 4, 2025, 8:00 AM → Jun 5, 2025, 5:30 PM US/Eastern

CC F224-225 & Zoom (Jefferson Lab)

Gloria Montana (Jefferson Lab), Alessandro Pilloni (Messina U. and INFN Catania), Arkaitz Rodas (ODU/Jefferson Lab), Adam Szczepaniak (Indiana Univeristy/JLab)

Hadron spectroscopy at Jefferson Lab has entered a new era, driven by high-precision data from GlueX and CLAS12 and a growing interest in exotic hybrid states. At the same time, AI and machine learning are emerging as powerful tools to tackle the complexity of this data and deepen our understanding of QCD. This workshop brings together experimentalists, theorists, and data scientists working with JLab data to explore how AI can support hadron spectroscopy. The focus will be on identifying key challenges, sharing ideas, and building collaborations that can shape future research. Through a mix of talks and open discussions, we aim to connect communities and define practical steps toward integrating AI into spectroscopy efforts at Jefferson Lab.

 

Wed, June 4

10:15 AM → 10:30 AM Welcome

10:30 AM → 12:00 PM AI for Hadron Spectroscopy at JLab

10:30 AM Summary talk of NPTwins 2024

Modern nuclear and high energy physics facilities, including CERN, Jefferson Lab, RHIC, and the upcoming Electron-Ion Collider (EIC), are generating exascale of data. This unprecedented amount of data offers an opportunity to answer many fundamental questions in elementary particle interactions, such as QCD in the nonperturbative regime. The NPTwins2024 workshop, held in Genova, Italy in December, 2024, highlighted the emerging role of digital twins, virtual replicas of complex physical systems, in modeling and simulating nuclear and particle interactions. This talk summarizes the methodologies and applications of digital twins presented in NPTwins2024 to bridge the gap between data and physical insight, as well as the associated challenges and future directions.

Speaker: Yaohang Li (Old Dominion University)

NPTwins2024 Summary Talk.pptx

11:00 AM EAIRA: Evaluation of LLMs/RLMs as Research Assistants

Recent advancements have positioned Large Language Models (LLMs) as transformative tools for scientific research, capable of addressing complex tasks that require reasoning, problem-solving, and decision-making. Their exceptional capabilities suggest their potential as scientific research assistants, but also highlight the need for holistic, rigorous, and domain-specific evaluation to assess effectiveness in real-world scientific applications.

This talk describes a multifaceted methodology for Evaluating AI models as scientific Research Assistants (EAIRA) developed at Argonne National Laboratory. This methodology incorporates four primary classes of evaluations. 1) Multiple Choice Questions to assess factual recall; 2) Open Response to evaluate advanced reasoning and problem-solving skills; 3) Lab-Style Experiments involving detailed analysis of capabilities as research assistants in controlled environments; and 4) Field-Style Experiments to capture researcher-LLM interactions at scale in a wide range of scientific domains and applications. These complementary methods enable a comprehensive analysis of LLM strengths and weaknesses with respect to their scientific knowledge, reasoning abilities, and adaptability. Recognizing the rapid pace of LLM advancements, we designed the methodology to evolve and adapt so as to ensure its continued relevance and applicability. This talk describes the current methodology's state. Although developed within a subset of scientific domains, the methodology is designed to be generalizable to a wide range of scientific domains.

Speaker: Franck Cappello (Argonne National Laboratory)

11:30 AM Extracting $x$-Dependent GPDs with Normalizing Flow

Obtaining the $x$-dependent generalized parton distributions (GPDs) is essential for advancing our understanding of hadron tomography. However, this goal has been hindered by the limited sensitivity of most well-known experimental processes, such as deeply virtual Compton scattering (DVCS) and time-like Compton scattering (TCS). In this talk, I will compare these traditional processes with new ones that offer enhanced sensitivity to the $x$-dependence. By employing a pixelated GPD construction using a normalizing flow neural network, we can visualize and quantitatively examine the point-by-point sensitivity encoded in the physical processes.

Speaker: Zhite Yu (Jefferson Lab)

AI_GPD_ZhiteYu.pdf

12:00 PM → 12:15 PM Group photo

12:15 PM → 2:00 PM Lunch break

2:00 PM → 4:00 PM AI for Hadron Spectroscopy at JLab

2:00 PM AI Applications in Hadronic Physics

We consider the task of using AI for hadron spectroscopy using partial wave analysis combined with production models. There are new challenges not seen in similar tasks at the LHC coming from the parameterization of amplitudes and not cross sections directly. We also have the opportunity and challenge of combining data from the full reaction with reactions with one or more, and even all particles missing in the final state. AI surrogates can speed up the unfolding process. The extension to a 22 GeV accelerator is considered. We suggest this work should be carefully positioned to make good use of the High Performance Data Facility (HPDF) and help Jefferson Lab configure HPDF to support related AI applications across DOE.

Speaker: Geoffrey Fox (University of Virginia)

AI Applications in Jefferson Laboratory Hadronic Physics

2:30 PM Hadron spectroscopy analyses with CLAS and CLAS12

In this talk, selected spectroscopy analyses completed with CLAS data and now in progress with CLAS12 will be presented, and the challenges that could benefit from AI/ML techniques will be discussed.

Speaker: Raffaella De Vita (Jefferson Lab)

devita_ai2025.pdf

3:00 PM AI for Unfolding

AI has enabled high-dimensional and unbinned differential cross section measurements for the first time. In this talk, I will discuss state-of-the-art methods and the latest experimental results using these tools.

Speaker: Benjamin Nachman (LBNL)

3:30 PM Amplitude-based analyses at GlueX

The GlueX experiment at Jefferson Lab employs 9 GeV linearly polarized
photons striking a proton target to study the spectrum of light hadrons.
A key focus is the precise measurement of the light-meson spectrum and
the search for exotic mesons. Most spectroscopy analyses rely on
amplitude analysis and require detailed reaction models to accurately
describe the often high-dimensional data. In this talk, I will present
recent results and highlight the challenges encountered in these analyses.

Speaker: Mr Boris Grube (Jefferson Lab)

talk.pdf

4:00 PM → 4:30 PM Coffee break

4:30 PM → 5:30 PM Discussion

Thu, June 5

9:00 AM → 10:30 AM AI for Hadron Spectroscopy at JLab

9:00 AM Amplitude extraction with generative AI

I will present recent progress on extracting the scattering amplitude for elastic pion-pion scattering from cross-section pseudodata using generative models.

Speaker: Gloria Montana (Jefferson Lab)

2025_AIforHadronSpectroscopyAtJLab_GloriaMontana.pdf

9:30 AM Generative-AI for smearing and acceptance unfolding in hadron physics

In particle scattering experiments, detector effects such as smearing and acceptance, distort the measured data, making it challenging to recover the true underlying physics. In this work, we propose an AI-assisted framework that employs Generative Adversarial Networks (GANs) to mitigate the impact of these distortions, enabling accurate reconstruction of the vertex-level distributions information from detector-level data. We demonstrate our approach using the photoproduction of a single $\pi^0$ as a case study, focusing on the reaction $\gamma p \to p \pi^0$. This reaction, depending on two independent kinematic variables, serves as an ideal test case for our methodology. Our generative model is trained on detector-level pseudodata obtained from a parametrization of the CLAS detector, correcting for smearing and acceptance effects. We apply our GAN-based methodology to perform a closure test using different physics models, comparing synthetic data to experimentally simulated pseudodata. Our results show that the framework effectively reconstructs the full kinematic phase space, including regions of phase space that are not directly measured by the detector, while minimizing model dependence in unmeasured regions. The approach is validated through comparisons of generated and measured event distributions, demonstrating a high level of agreement and the ability to improve the accuracy of physics analyses in particle physics experiments.

Speaker: Tareq Alghamdi (ODU)

10:00 AM A Modular Software Stack for A(i)DAPT

A(i)DAPT, or AI for Data Analysis and PreservaTion, is a CLAS group, the goal of which is to re-analyze and improve upon the measurements from past CLAS experiments using machine learning. This project utilizes deep learning techniques, specifically generative adversarial networks (GANs), to improve Monte Carlo methods used in the analysis of data. Producing simulations with generative AI as opposed to traditional simulation software is significantly faster and more computationally efficient. Thus, there is an appreciable demand for machine learning approaches to assist with (or possibly replace) these necessary tasks in nuclear and particle physics. The Jefferson Lab Data Science Group became involved with A(i)DAPT in an effort to help streamline, generalize, and optimize the framework already successfully operating within the group. In this talk, I will focus on the contributions from the Data Science Group towards achieving these goals.

Speaker: Trevor Reed (FIU)

Trevor_Reed_AiDAPT_presentation.pptx

10:30 AM → 11:00 AM Coffee break

11:00 AM → 12:30 PM AI for Hadron Spectroscopy at JLab

11:00 AM Diffusion Models as Surrogate Event Generators for QCD Analysis

Diffusion-based generative models have recently emerged as a powerful alternative to GANs, VAEs, and normalizing flows for learning complex, high-dimensional physics distributions. After briefly introducing the forward–reverse noising process, we demonstrate how a conditional diffusion model can replace the costly Monte-Carlo event generator that maps quantum-correlation-function parameters to observable scattering events in deep-inelastic-scattering simulations, achieving higher fidelity and more stable training than a baseline conditional GAN.

Speaker: Jitao Xu

11:30 AM Moments of two-pion photoproduction and diffusion models

In 2009, the CLAS collaboration reported the first observation of scalar meson photoproduction in the π+π− channel. Because the cross section in this channel is dominated by the vector ρ(770) resonance, the observation of the $f_0(980)$ peak in the mass distribution was not possible. Instead, the resonant S-wave contribution was inferred through subtle interference effects in the moments of the angular distribution.
To further clarify the role of scalar mesons in photon-induced reactions, both careful theoretical modeling and refined experimental techniques are essential. A recent Regge-theory-based study by JPAC marks a significant step forward in modeling, but further developments—such as the proper inclusion of coupled-channel effects—are still needed. On the experimental side, efforts include, among others, addressing the limited phase space coverage of the CLAS detector.
To this end, we introduce a novel strategy to unfold detector distortions in CLAS two-pion photoproduction measurements using a diffusion model. By training the model on simulated events, we enable it to recover the "true" kinematic distributions from reconstructed data, from which moment distributions can then be extracted. Simulations are initially based on pure phase space, but we plan to incorporate the JPAC model, with or without the $f_0(980)$, in the near future.
This approach improves the learning convergence rate—and, desirably, flexibility across different topologies—compared to generative adversarial networks (GANs), and paves the way for extracting moments from real CLAS data.

Speakers: Giorgio Foti (University of Messina & INFN Catania), Lukasz Bibrzycki (AGH University of Krakow)

jlabAI-1.pdf

12:00 PM Autoencoders for Partial Wave Analysis

Partial Wave Analysis (PWA) provides us with a richer understanding of particle scattering phenomena than, for example, cross sections. PWA has been employed for decades in nuclear physics data, but there are several considerable hurdles that make this topic a challenging one, including large parameter spaces and multiple solutions. In this talk, I will present some of the work done within our group, whose focus is applying AI techniques to PWA. Utilizing autoencoders to study nuclear/particle physics reactions has proven beneficial in better understanding these partial waves. I will discuss promising results already achieved by this group, as well as inherent challenges that we’re still facing.

Speaker: Trevor Reed (FIU)

Trevor_Reed_AIPWA_presentation.pptx

12:30 PM → 2:00 PM Lunch break

2:00 PM → 4:00 PM AI for Hadron Spectroscopy at JLab

2:00 PM Towards simulation based inference of hadron structure

In this talk, I will discuss some new technical developments build a differentiable pipeline for event level analysis of hadron structure.

Speaker: Nobuo Sato (Jefferson Lab)

Towards Simulation Based inference of hadron structure

2:30 PM Neural Networks for QCD

In this talk we discuss novel methods for the application of AI to assist with the extraction of physical information from QCD. We discuss neural network architecture and machine learning for the extraction of topological quantities in lattice QCD, and for neural network-enforced unitarity and error propagation in the description of pion scattering data.

Speaker: Dr Wyatt Smith (University of Messina)

3:00 PM Variational Neural Network Approach to QFT in the Field Basis: Results from Klein-Gordon Theory

We present a variational method for solving quantum field theories in the continuum field basis using neural networks. As a benchmark, we consider the free Klein–Gordon model in one spatial dimension, where the ground-state wavefunctional is known analytically. The variational ansatz is implemented using a feed-forward neural network trained to minimize the Hamiltonian expectation value in the Schr¨odinger picture. Our formulation preserves continuum-inspired operator structure by treating the canonical momentum as a functional derivative and discretizes only the momentum domain via Riemann summation. While our numerical results are computed at finite resolution with a momentum cutoff, the approach maintains close contact with the infinite-dimensional Hilbert space
of continuum QFT. We demonstrate accurate reproduction of ground-state observables—including energy, momentum-space two-point correlators, and field expectation values—and provide direct visualizations of the learned wavefunctional. This work establishes a flexible and physically interpretable framework for nonperturbative QFT, paving the way for future applications to interacting theories and gauge fields.

Speaker: Kevin Braga (William & Mary)

Variational Neural Network Approach to QFT in the Field Basis_ Results from Klein-Gordon Theory.pptx

3:30 PM Applications of Density Ratio Estimation in Experimental Hadron Spectroscopy

The solution to many problems can be described by the ratio of the probability densities of two event samples. For example, detector acceptances can be modeled by the ratio of the probability density for detected (accepted) events over that for all events. Similarly, sWeights can be converted to positive definite probabilities via density ratio estimation in order to create machine learning training samples from experimental data. In this talk, I will describe how binary classification can be used towards density ratio estimation. I will then describe the two applications mentioned above in more detail and their potential use cases in hadron spectroscopy.

Speaker: Richard Tyson (Jefferson Lab)

Tyson_AIHadronSpec_June2025.pdf

4:00 PM → 4:30 PM Coffee break

4:30 PM → 5:30 PM Discussion