Learning correlations in nuclear masses using neural networks

Abstract

There have been great improvements in the predictions of nuclear masses, yet it is difficult to exactly reproduce the measured nuclear mass. It has been suggested that the cause of such discrepancies is due to the negligence of many-body effects in the available theoretical models. The errors in the prediction of the nuclear mass show residual correlations due to the missing physics in the mass models. In the present Letter we have tried to learn such correlations by using the neural networks. We have used a neural network architecture which adaptively learns the linear and nonlinear correlations between the data of different fidelity. We have used the theoretical predictions of finite range droplet model and Hartree- Fock-Bogoliubov models in the input of the neural networks. The present approach show significant improvements in the accuracy of the predictions. It has been clearly presented that the difference between the predictions from the present approach and the experimental data behave more, such as white noise, showing that using the present approach the residual correlations arising due to the missing physics from the available mass models can be learned.