Abstract
In this work, the beta-decay halflives problem is dealt as a nonlinear optimization problem, which is resolved in the statistical framework of Machine Learning (LM). Continuing past similar approaches, we have constructed sophisticated Artificial Neural Networks (ANNs) and Support Vector Regression Machines (SVMs) for each class with even-odd character in Z and N to global model the systematics of nuclei that decay 100% by the beta-minus-mode in their ground states. The arising large-scale lifetime calculations generated by both types of machines are discussed and compared with each other, with the available experimental data, with previous results obtained with neural networks, as well as with estimates coming from traditional global nuclear models. Particular attention is paid on the estimates for exotic and halo nuclei and we focus to those nuclides that are involved in the r-process nucleosynthesis. It is found that statistical models based on LM can at least match or even surpass the predictive performance of the best conventional models of beta-decay systematics and can complement the latter.
Introduction
Predicting the half-lives of radioactive nuclei is a cornerstone of nuclear physics and astrophysics, essential for understanding element formation in supernova explosions (the r-process) and exploring the nuclear landscape far from the valley of stability. While experimental data exists for many stable nuclei, thousands of exotic, neutron-rich nuclei remain unmeasured. Traditional "theory-thick" models, such as the Quasiparticle Random-Phase Approximation (QRPA), often struggle to maintain accuracy far from stability, sometimes deviating from experimental values by an order of magnitude. This research investigates whether "theory-thin" statistical models can bridge this gap by learning underlying patterns directly from experimental data.
Methodology and Datasets
The study treats beta-decay half-life prediction as a nonlinear optimization problem using two primary Machine Learning (ML) architectures: multilayer feedforward Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs).
A key innovation in this approach is the subdivision of the dataset into four distinct models based on nuclear parity: Even-Even (EE), Even-Odd (EO), Odd-Even (OE), and Odd-Odd (OO). This acknowledges the fundamental physical reality of nuclear structure. The models were trained using the Nubase2003 evaluation, specifically focusing on 838 nuclides (NuSet-B) that decay 100% via the $\beta^-$ mode.
Key Results and Discussion
The performance of the ML models was evaluated using the Root Mean Square Error ($σ_{RMSE}$) of the $\log_{10}$ half-lives. The ANN and SVM models achieved test set errors of 0.51 and 0.53, respectively. These results represent a significant improvement over earlier neural network attempts and demonstrate that ML models are competitive with, or even superior to, state-of-the-art global nuclear models like FRDM+pnQRPA.
Analysis suggests that while SVMs excel at interpolation—often matching experimental data with high precision—ANNs may offer greater robustness for extrapolation into unmeasured regions. Furthermore, the ML models showed remarkable agreement with experimental data for nuclei involved in the r-process, suggesting high reliability for astrophysical simulations.
Conclusion
Statistical models based on Machine Learning provide a powerful, theory-thin alternative to traditional nuclear models. By avoiding the systematic biases inherent in complex physical approximations, these models can complement existing theoretical frameworks, serving as a statistical check on predictions. Future work will focus on "whole basis" models to mitigate spurious fluctuations across isotopic chains and exploring the potential for ML to uncover new physical insights into beta-decay systematics.
Summary updated