arXiv Science⌕ Search

arXiv · 2611.18066

Product units in gated recurrent networks enhance prediction of nuclear masses and charge radii

Abstract

The prediction of nuclear properties using machine learning can complement theoretical models and advance the exploration of poorly known domains of the nuclear chart. We propose a machine learning technique for predicting nuclear masses and charge radii based on gated recurrent units (GRU) that integrates learnable product-unit mappings into the candidate state to model multiplicative dependencies among neighboring nuclei. Computations are performed in the complex domain to jointly capture amplitude and phase dynamics. Across three data partitions and three training seeds per partition, AM-PU-GRU achieves a mean mass interpolation RMSE of 0.2698 MeV and a temporal RMSE of 0.1905 MeV. Configuration comparisons support the integration of learned products into gated recurrence. Without mass inputs, MI-PU-GRU predicts charge radii with an RMSE of 0.019947 fm, 14.40% below the fixed WS* prior. Matched radius controls further show how candidate fusion shapes the benefit of learned products. Paired errors, nuclide-resolved maps, and inference measurements characterize predictive gains and computational cost. This article is an extended version of our ICCS 2026 conference paper (doi:10.1007/978-3-032-29921-5_10).

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Ziyuan Li, Paulo S. A. Freitas, John W. Clark, Babette Dellen. 2026-11-13. Product units in gated recurrent networks enhance prediction of nuclear masses and charge radii. https://arxiv.science/abs/2611.18066

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Proton and neutron elastic scattering at low energies on He targets from \textit{ab initio} optical potentials

We construct and discuss $\textit{ab initio}$ nucleon-nucleus optical potentials at low energies for $^{3,4,6}$He targets. In this work, we use the $\textit{ab initio}$ SA-NCSM/GF approach that combines the $\textit{ab initio}$ symmetry-adapted no-core shell model with the Green's function technique to construct optical potentials, and extend this formulation to proton scattering and targets with nonzero spin. We show that these optical potentials reproduce experimental differential cross sections and phase shifts for proton and neutron elastic scattering remarkably well. The $\textit{ab initio}$ SA-NCSM/GF approach provides nonlocal, energy dependent and dispersive optical potentials, suitable for the astrophysically relevant regime of low energies and for exotic nuclei, where experiments are difficult and data is often unavailable.

nucl-th↗

Nuclear mass calculation based on anisotropic kernel ridge regression and $α$-decay chain recursion

Many short-lived or rarely produced nuclei defy direct mass spectrometry, but their masses are constrained by the $α$-decay energies connecting them through decay chains. We combine anisotropic kernel ridge regression (AKRR), an extended Garvey--Kelson relation (GK16), and $α$-decay chain recursion: AKRR learns the residuals between experimental $α$-decay energies and WS4 predictions, GK16 removes residual local fluctuations, and the corrected energies convert a measured anchor mass into the target mass by exact recursion---replacing absolute-mass extrapolation with interpolation of relative decay energies plus exact algebra. The only data-adjusted quantities are the three kernel hyperparameters, fixed by leave-one-out cross validation. For the 22 nuclei whose masses were measured only after AME2003, the $Q_α$ RMSD falls from 169~keV (WS4) to 108~keV (AKRR) and 65~keV (AKRR+GK16), and the mass-excess RMSD reaches 86~keV, versus 348 and 211~keV for WS4 and AKRR. Mass excesses are further predicted for the 122 nuclei with measured $Q_α$ but no directly measured masses.

nucl-th↗

A Predictive Nuclear Effective Field Theory from Wilsonian Renormalization Group Principles I: Medium-mass nuclei and nuclear matter

The predictive power of nuclear effective field theory (EFT) from few- to many-body systems is substantially restricted by regulator dependence. Wilsonian momentum-shell renormalization group (RG) offers a natural framework for cutoff-independent predictions, but its many-body implementation requires regulating single-particle momentum modes, which breaks Galilean invariance. RG scaling analysis shows that this symmetry breaking is encoded in irrelevant operators whose effects can be included explicitly or absorbed into higher-body interactions. Guided by Wilsonian RG, we construct a regularization scheme for chiral EFT contact interactions and predict binding energies of nuclei up to $^{40}$Ca using interactions calibrated only to $A\leq3$ systems. The predicted energies vary at the percent level and agree with experiment for $300\leqΛ\leq450$~MeV. The same scheme reduces the overbinding of symmetric nuclear matter obtained with conventional relative-momentum regulators. Our results suggest that some model dependence observed in nonrelativistic many-body calculations originates from treating EFT vertices as static, Galilean-invariant potentials rather than as field operators regulated consistently across particle-number sectors.

nucl-th↗