Ultra-Peripheral Collisions as a Nuclear-Structure Interferometer with Interpretable Multitask Deep Learning

Research Question and Approach

Probing the spatial structure of atomic nuclei—specifically collective shape deformations and the radial distribution of neutrons (neutron skin)—is a fundamental challenge in nuclear physics. Traditional methods often suffer from parameter degeneracy, where different structural features produce indistinguishable signatures in experimental data. This paper explores the use of ultra-peripheral collisions (UPCs) as a femtoscopic interferometer. In these collisions, the coherent photoproduction of vector mesons (like J/ψ) creates a momentum-space distribution that acts as a diffraction and interference pattern, encoding the nuclear geometry.

The authors propose an "AI-enhanced lens" using a multitask deep-learning architecture (ResNet-34) to map these 2D transverse momentum distributions directly to nuclear-structure indicators. By training the model on simulated data that includes both coherent signals and incoherent backgrounds, the framework learns to isolate features associated with deformation (which influences interference fringes) from those associated with the neutron skin (which influences diffraction bright spots).

The study demonstrates that the multitask approach effectively breaks the degeneracy between deformation and neutron-skin effects. The model achieves high predictive accuracy for quadrupole deformation (β2) and octupole deformation (β3), and successfully classifies neutron-skin types even in the presence of significant incoherent background contamination.

Interpretability analyses (using Pixel-level Discriminative Analysis) confirm that the model physically distinguishes between the two structural sectors: the network focuses on the dark interference fringes to determine deformation, while it prioritizes the bright diffraction spots to identify neutron-skin characteristics. This confirms that the deep-learning model is not merely performing pattern matching but is extracting physically meaningful, localized features from the momentum-space images.