Sokratis Trifinopoulos

Nuclear physics has long relied on phenomenological models to predict nuclear binding energies, while high-precision AI models are hard to trust when extrapolating. An interpretable neural network taking only proton and neutron numbers as input spontaneously forms a double-helix geometry in its internal representations, reproduces macroscopic terms such as the liquid-drop model in a principal-component decomposition, and rediscovers the local Jaffe factorization proposed in 1969 and since overlooked. Applied as a local correction to the WS4 model, this factorization brings the root-mean-square binding-energy error below 0.12 MeV without relying on a black box.

Nuclear physics has long relied on phenomenological models to predict nuclear binding energies, while high-precision AI models are hard to trust when extrapolating. An interpretable neural network taking only proton and neutron numbers as input spontaneously forms a double-helix geometry in its internal representations, reproduces macroscopic terms such as the liquid-drop model in a principal-component decomposition, and rediscovers the local Jaffe factorization proposed in 1969 and since overlooked. Applied as a local correction to the WS4 model, this factorization brings the root-mean-square binding-energy error below 0.12 MeV without relying on a black box.