FCNN hits 93% accuracy on neutron resonances but fails to generalize

Physicists studying nuclear reactions often rely on R-Matrix codes to fit neutron transmission data, a process that is labor-intensive and heavily dependent on prior evaluations. In this preliminary study, researchers from the team of Nataly R. Panczyk, Athanasios Stamatopoulos, Josef Svoboda, and Majdi I. Radaideh explored whether a fully convolutional neural network (FCNN) could automate the identification of neutron resonances—sharp peaks in transmission spectra that correspond to excited states in atomic nuclei. The model was trained to classify each point in a spectrum as either resonance or non-resonance, tackling the noise and complexity that traditional peak-detection methods struggle with.

The network achieved an impressive 93% classification accuracy across seven spectra, including both evaluated and experimental data. But a deeper analysis revealed this metric was misleading. Despite adding more training data compared to the group's earlier PHYSOR 2026 work, the model failed to generalize to previously unseen isotopes—a critical limitation for practical use in nuclear data evaluation. The authors suggest future approaches should incorporate known physical characteristics of neutron resonances into the model, and explore whether a substantially larger and more diverse training set can yield robust generalization. This work is a promising step toward accelerating post-experimental processing and reducing human bias, but it also serves as a cautionary tale about relying solely on raw accuracy metrics in scientific machine learning.