LEARNING AND PREDICTION OF NUCLEAR-STABILITY BY NEURAL NETWORKS

The backpropagation learning algorithm is used to teach layered feedforward networks of model neurons the existing data on nuclear stability and atomic masses. Specific applications include (i) the construction of networks that decide stability, (ii) learning and prediction of nuclear mass excesses and (iii) analysis of the systematics of neutron separation energies. With suitable architecture and representation of input and output data, learning can be accomplished with high accuracy. Evidence is presented that these new adaptive computational systems can grasp essential regularities of nuclear physics including the valley of beta-stability, the pairing effect and the existence of shell structure. Significant predictive ability is demonstrated, opening the prospect that neural networks may provide a valuable new tool for computing nuclear properties and, more broadly, for phenomenological description of complex many-body systems.