Neural networks for hadron resonance extraction

Abstract

We present a neural network approach to resonance parameter extraction and model classification applied to the pion vector form factor. Five theoretical parametrizations of the ρ(770) resonance of increasing sophistication are considered, ranging from simple Breit-Wigner models to fully analytic implementations derived from Chiral Perturbation Theory. Trained exclusively on synthetic spectra, the networks recover the physical pole parameters with precision comparable to traditional χ2 fits in a single inference step. Extended to simultaneous parameter extraction and model classification, a multitask network consistently identifies the parametrization that best describes the experimental data, in agreement with traditional fitting results, while extracting the model parameters with comparable precision. The resulting fast, single-step inference framework opens the door to future applications in hadronic physics