Unmasking the Underlying Correlations in Nuclear Reaction Cross Sections

Evaluated nuclear reaction data lies at the heart of the interface between fundamental nuclear physics and the broader world of scientific and societal applications of that nuclear physics, from modeling the life cycle of stars, cataclysmic astrophysical events, and even the backgrounds in precision terrestrial experiments to the production of power and radiological medical therapies and diagnostics. Quantifying the uncertainties of that nuclear data and propagating that uncertainty forward is crucial to enable these applications to determine the confidence level of their predictions, which significantly impacts the reliability of the extracted science and the safety of the society applications. However, these evaluated data only exist for stable or near-stability nuclei, and covariance information within those existent data is sparse at best. Exploiting recent developments in machine learning, we adopt a transformer-style architecture to learn the underlying relationships within nuclear data between neighboring nuclei across the nuclear chart to fill in gaps in reaction libraries. This transformer's success is enabled through a novel representation of the nuclear cross sections' latent data representation learned through an implicit neural representation. We will cover the general features of these machine learning tools and demonstrate their effectiveness in predicting nuclear reaction cross sections.