Multitask Gaussian process predictions of low-lying excitation observables and shell evolution near and 28
Phys. Rev. C 114, 054302 – Published 2 November, 2026
DOI: https://doi.org/10.1103/p21l-zm7x
Abstract
The low-lying excitation observables of even-even nuclei, e.g., the excitation energies of the first and states, and , and the reduced electric quadrupole transition probability between the ground state and the first state, , serve as promising experimental probes for nuclear shell evolution in unstable nuclei far from the -stability line. Previous machine learning studies on low-lying nuclear states have typically employed independent single-task models that neglect the intrinsic correlations among , and , potentially leading to unphysical divergences in practical extrapolation. In addition, the manual selection of hyperparameters introduces subjective tuning bias and model dependence. To overcome these limitations, we propose a multitask Gaussian process (MTGP) approach that jointly predicts the three observables while preserving their mutual physical couplings. The MTGP hyperparameters are automatically optimized via Bayesian optimization. The examination of the established MTGP method demonstrates its significant improvement in predictive robustness over previous single-task approaches, both in terms of the prediction accuracy and the description of inherent correlations among the three observables. We then apply the MTGP model to predict unmeasured , and values around the magic shells and 28. Systematic analysis of the MTGP-predicted observables along their respective isotonic and isotopic chains reveals nuanced shell-evolution patterns at and 28: the closure weakens in the neutron-rich region, strengthens at , and weakens again for ; the closure is quenched at , robust at , and shows signs of weakening in and ; the closure persists in but gradually weakens beyond ; and the closure remains strong in and weakens in neutron-rich nickel isotopes beyond . These data-driven results support that the MTGP-based systematics capture the evolution of shell gaps and collectivity, suggesting the MTGP as a potential tool for nuclear structure studies.