15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

A Bayesian Pixel-Based Approach to TMD Imaging with Generative AI

Not scheduled
15m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML for Data Analysis and Theory AI/ML for Data Analysis and Theory

Speaker

Marco Zaccheddu (Argonne National Lab)

Description

The extraction of Transverse Momentum Dependent distributions (TMDs) from experimental observables represents a classic ill-posed inverse problem in hadronic physics. Traditionally, this challenge is addressed by assuming rigid analytical functional forms, which introduce parameterization bias, obscure unconstrained degrees of freedom, and limit the exploration of the full epistemic uncertainty.

In this talk, we present a novel nonparametric pixel-based framework that conceptualizes TMD imaging as a discrete image reconstruction task. Formulated within a fully differentiable scheme integrating TMD evolution under the Collins-Soper-Sterman (CSS) formalism, our approach treats discrete nodal values (or pixels) in impact parameter ($b_T$) space as stochastic variables. This enables the simultaneous, model-independent extraction of partonic distributions and the nonperturbative evolution kernel within a rigorous Bayesian inference scheme. To efficiently and faithfully sample the resulting high-dimensional posterior distribution, we leverage generative AI through a hybrid Normalizing Flow-driven Metropolis-Hastings strategy.

We perform a systematic study of the resolution limits and stability of this inversion process through multi-scale closure tests of increasing complexity, ranging from baseline Gaussian models to unpolarized TMD PDFs and convoluted $F_{UU,T}$ structure functions. Using Singular Value Decomposition (SVD), we formally characterize the epistemic uncertainty and reveal the existence of null TMDs, which are functional components residing in the null space of the integral kernel that remain unconstrained by single-scale observables. Finally, we demonstrate how the synergy between multi-scale data and generative AI breaks these degeneracies, establishing a robust, low-bias foundation for high-fidelity 3D partonic imaging in the upcoming Electron-Ion Collider era.

Author

Marco Zaccheddu (Argonne National Lab)

Presentation materials

There are no materials yet.