Nuclear mass predictions based on convolutional neural networks

Abstract

The precise determination of nuclear masses is essential for understanding atomic nuclei and for applications in astrophysics and nuclear energy. Traditional models like the liquid drop model, with a root mean squared error of σ = 3.94 MeV, fail to meet the accuracy of 100 keV required for nuclear astrophysics research. This work introduces a novel approach by implementing a convolutional neural network (CNN) and leveraging the spatial structure of the nuclide chart. Two models, I3 and I4, are trained and tested on the AME2016 database, achieving values of σ = 0.67 MeV and σ = 0.49 MeV, respectively. Extrapolating to the new nuclei of the AME2020 database, they hold values of σ = 0.64 MeV and σ = 0.57 MeV, demonstrating strong generalization capabilities and proving that CNNs constitute a powerful tool for accurate nuclear mass predictions