VOLUME 19 (Supplement)

Model picking beyond the χ² fitting

The study addresses the challenge of interpreting near-threshold enhancements seen in scattering data, specifically focusing on the P_cc̄(4312)⁺ pentaquark signal. While the K-matrix formalism provides a robust framework for maintaining S-matrix unitarity by construction, multiple parametrizations can often produce statistically equivalent χ² fits despite having fundamentally different pole structures. In an attempt to resolve this ambiguity, a supervised deep neural network (DNN) was trained on 40,000 synthetic spectra generated from four candidate K-matrix families. Using a fixed held-out test set and five independent training seeds, the classifier achieved a mean accuracy of 82.61% and mean macro-F1 of 82.37%. The repeated DNNs predominantly assigned perturbed realizations of the experimental spectrum to the two-resonance model with nonresonant K-matrix terms (Class 3). However, this assignment was not stable across classifier families, and performance was sensitive to noise and the analyzed energy interval. The machine-learning result is therefore interpreted as conditional evidence within the four simulated model families and nominal analysis assumptions rather than as a posterior probability or definitive identification of the physical pole structure.