Deep learning approaches for nuclear binding energy predictions: A comparative study of RNN, GRU, and LSTM neural-network models
Phys. Rev. C 112, 024305 – Published 1 August, 2025
DOI: https://doi.org/10.1103/rzjx-9zz1
Abstract
This study investigates the application of deep learning models—recurrent neural networks, gated recurrent units, and long short-term memory networks—for predicting nuclear binding energies. Utilizing data from the Atomic Mass Evaluation (AME2020), we incorporate key nuclear structure features, including proton and neutron numbers, as well as additional terms from the liquid drop model and shell effects. Our comparative analysis demonstrates that the gated recurrent units model achieves a root-mean-square error () of 0.326 MeV, outperforming traditional regression-based approaches, such as those based on the liquid drop model, which typically yield higher errors. To assess model reliability, we validate predictions using the Garvey-Kelson relations, obtaining an error of 0.202 MeV, and further test extrapolation capabilities using the WS, WS3, WS4, DZ28, FRDM2012, and HFB25 models. The extrapolation analysis confirms the robustness of our approach, particularly in predicting binding energies for nuclei near the drip lines. These results highlight the effectiveness of deep learning in nuclear BE predictions, highlighting its potential to enhance the accuracy and reliability of theoretical nuclear models.
