Neural network extraction of resonances in pion-pion scattering

In this work, we present novel methods for resonance extraction from experimental data through physics-informed Monte Carlo replica ensembles of neural networks. Low-energy pion–pion scattering is plagued by poor-quality and even incompatible datasets, with some previous analyses failing to fully propagate uncertainties. Traditional lineshape fits require specific model parameterizations, introducing bias that can distort extracted resonance pole locations. Neural networks provide a highly flexible, near‑model‑independent class of interpolants, capable of fitting scattering data while enforcing physics constraints through custom loss functions encoding fundamental S‑matrix principle. Drawing inspiration from NNPDF methodology, we implement dataset discrimination criteria and retrain neural networks on Monte Carlo replicas to propagate statistical and systematic errors faithfully. We show resonance pole positions in pion-pion scattering below 1GeV, and systematically address data quality and model bias issues. This represents a step toward future model‑independent analysis of more interesting/exotic hadronic states.