Nucleare GP Machine learns

Email contact: multiversou@gmail.com This work presents two complementary studies on the use of Gaussian Processes (GPs) to correct nuclear mass residuals and to assess the physical information required for reliable extrapolation. The first study introduces a physics‑motivated feature engineering improvement—Gaussian proximity to magic numbers—which replaces linear distances with exponentially decaying functions that better capture shell effects. The second study performs a “brutal” ablation test, removing all shell-related features and retaining only seven bulk nuclear properties to evaluate whether shell structure can emerge from data alone. Using a dataset of 64 nuclei in the Ca–Sn region (Z = 20–50), with test nuclei ⁷⁸Ni, ⁸⁰Ni, ¹³⁰Sn, and ¹³²Sn excluded from training, the enhanced GP model (v7.0) achieves a MAE of 0.29 MeV in leave-one-out validation. Gaussian proximity reduces the prediction error for ¹³⁰Sn by 45% compared to linear distances, and ARD analysis confirms the high relevance of shell-related features. In contrast, the ablation model—using only bulk features—fails catastrophically on doubly magic nuclei, with a MAE of 1.42 MeV and a maximum error of 2.17 MeV for ⁸⁰Ni. These results provide quantitative evidence that nuclear shell structure does not emerge from bulk observables alone and must be encoded explicitly in machine‑learning models. Overall, the study demonstrates that combining physics‑informed feature engineering with advanced GP architectures yields interpretable and accurate predictions, and that ablation tests serve as powerful computational experiments for probing the physical content of ML models in nuclear theory.