Refining Nuclear Mass Predictions in the Ca-Sn Region via Physics-Informed Gaussian Processes: From FRDM Residuals to Data-Driven Corrections

Contatto e-mail: multiversou@gmail.com This note revisits and provides a comprehensive re-analysis of two earlier reports concerning Gaussian-Process corrections to FRDM2012 nuclear mass residuals in the region from Z=20 to Z=50. The work delivers two primary outcomes: Data Correction Restatement: We preserve and restate the identification of two significant transcription errors in the public FRDM2012 database. The reported residuals for Tin-100 (5.20 MeV) and Tin-132 (4.35 MeV) are physically inconsistent with neighbouring isotopic trends. Based on AME2020 systematics, the physically consistent values are corrected to 0.52 MeV and 1.435 MeV respectively. Methodological Retraction: We retract the earlier claim that explicit shell-structure features offer a substantial extrapolation advantage when predicting doubly-magic nuclei. An audit revealed that the previous results were contaminated by test-set leakage, as hyperparameters and feature engineering were tuned via repeated inspection of the test set. Under a strict, leakage-free protocol—where feature selection is frozen via leave-one-out cross-validation on the training pool alone—the performance of the shell-informed model (mean absolute error of 1.41 MeV) is statistically indistinguishable from a bulk-only baseline (mean absolute error of 1.56 MeV). Significance:While the data correction remains a genuine contribution to nuclear mass tables, the pronounced extrapolation advantage originally reported is shown to be an artifact of test-set leakage. Both models exhibit a significant gap between training-pool cross-validation error (around 0.22 MeV) and genuine extrapolation error, highlighting a robust open challenge for future machine learning applications at shell closures. Reproducibility:The complete, self-contained Python script to fully reproduce all tables, statistics, and figures accompanies this deposit.