Interpretable AI reveals hidden structure in nuclear masses

Predicting nuclear masses with high precision remains a central challenge in nuclear physics, with implications ranging from stellar nucleosynthesis to determining the boundaries of the nuclear landscape. Approaches using artificial intelligence (AI) have recently achieved remarkable accuracy on this task, but their usefulness is limited by their black-box nature and by the difficulty of assessing their reliability when extrapolating beyond measured nuclei. Here, we present an interpretable AI model that matches the best existing predictive precision while revealing the physical structure underlying nuclear binding energies. The model learns internal representations that form a double helix, in which links between proton and neutron embeddings correspond to the most stable nuclei within each isotopic chain. Analysis of its learned features shows that the AI autonomously recovers established macroscopic components such as the liquid-drop model and rediscovers a long-forgotten physical principle-localized Jaffe factorization-that accounts for much of its performance gains. These findings demonstrate that interpretable AI can both reproduce and extend nuclear models, uncovering latent physical symmetries directly from data.