Nuclear mass refinement via Bayesian neural networks and applications to α-decay of SHE

This work presents a systematic investigation into nuclear mass prediction using Bayesian Neural Networks (BNNs) to refine the results of nine theoretical nuclear mass formulas. The statistical methodology, based on residual analysis, significantly reduced the root-mean-square deviation relative to experimental data from the AME2020 compilation, achieving the accuracy required for modeling the r -process nucleosynthesis. Uncertainty quantification inherent to the BNN approach demonstrated a progressive widening of confidence intervals when moving toward unexplored regions of the nuclide chart. The predictive power of the hybrid model was successfully validated against recent mass measurements for Tc and Zn isotopes obtained via Time-of-Flight Magnetic-Rigidity (ToF-B \(ρ\) ) techniques and the TRIUMF’s Ion Trap for Atomic and Nuclear Science (TITAN). Additionally, the corrected masses were applied to the study of \(α\) -decay in superheavy elements (SHE) using a molecular cluster formalism within a variable mass-asymmetry scheme (VMAS). Coupling the WS4+BNN model with the effective potential barrier improved the consistency of calculated half-lives, reducing discrepancies with experimental data by up to four orders of magnitude, as observed for \(¹⁴⁷Sm\) .