Nuclear mass systematics using neural networks

Abstract

New global statistical models of nuclidic (atomic) masses based on multilayered feedforward networks are developed. One goal of such studies is to determine how well the existing data, and only the data, determines the mapping from the proton and neutron numbers to the mass of the nuclear ground state. Another is to provide reliable predictive models that can be used to forecast mass values away from the valley of stability. Our study focuses mainly on the former goal and achieves substantial improvement over previous neural-network models of the mass table by using improved schemes for coding and training. The results suggest that with further development this approach may provide a valuable complement to conventional global models.

Nuclear mass predictions for the crustal composition of neutron stars: A Bayesian neural network approach

“…In general, we observe a marginal, yet systematic, improvement in σ post relative to the corresponding values obtained in the case of the outer crust. Moreover, our findings appear consistent with earlier results obtained by Clark and Li [30] using support vector machines (SVM) and by Athanassopoulos and collaborators [29]. As in the present case, the approach by Athanassopoulos et al consisted in developing a multilayer feedforward neural network to reproduce the differences between experimental nuclear masses and theoretical predictions provided by the FRDM.…”

“…The application of artificial neural networks to nuclear physics started in the early 90s with the pioneering work of Clark and collaborators [24][25][26][27], Athanassopoulos and collaborators [28,29] and continues to this day with extensions that include beta-decay systematics that are highly relevant to our understanding of r-process nucleosynthesis [30,31]; for a more recent application of artificial neutral networks to binding-energy systematics see Ref. [32].…”

Coefficients and terms of the liquid drop model and mass formula

“…Microscopic Hartree-Fock selfconsistent calculations using mean-fields and Skyrme or Gogny forces and pairing correlations [7,8] as well as relativistic mean field theories [9] have also been developed to describe these nuclear masses. Finally, nuclear mass systematics using neural networks have been undertaken recently [10].The nuclear binding energy B nucl (A,Z) which is the energy necessary for separating all the nucleons constituting a nucleus is connected to the nuclear mass M n.m byThis quantity may thus be easily derived from the experimental atomic masses as published in [11] since : MeV. The fission, fusion, cluster and α decay potential barriers are governed by the evolution of the nuclear binding energy with deformation.…”

Application of Support Vector Machines to Global Prediction of Nuclear Properties

“…The results in Table 1 attest to a quality of performance, in both fitting and prediction, that is on a par with the best available from traditional modeling 10,11 and from MLP models trained by an enhanced backpropagation algorithm. 16,18,24 To emphasize this point qualitatively, we display in Table 2 some representative rms error figures that have been achieved in recent work with all three approaches. (We must note, however, that the data sets used for the different entries in the table may not be directly comparable, and the division into training, validation, and test sets does not necessarily have the strict meaning assigned here.)…”

“…(i) A number of studies 5,14,15,16,17,18,19 suggest that the quality and quantity of the data has already reached a point at which the statistical models can approach and possibly surpass the theory-thick models in sheer predictive performance. In this contribution, we shall present strong evidence from machine learning experiments with Support Vector Machines that this is indeed the case.…”