Validating neural-network refinements of nuclear mass models
Nuclear astrophysics centers on the role of nuclear physics in the cosmos. In particular, nuclear masses at the limits of stability are critical in the development of stellar structure and the origin of the elements. In this contribution we test and validate the predictions of recently refined nuclear mass models against the newly published AME2016 compilation. The basic paradigm underlining the recently refined nuclear mass models is based on existing state-of-the-art models that are subsequently refined through the training of an artificial neural network. We observe a significant improvement in the Bayesian Neural Network (BNN) predictions relative to the corresponding "bare" models when compared to the nearly 50 new masses reported in the AME2016 compilation. Further, AME2016 estimates for the handful of impactful isotopes in the determination of r-process abundances are found to be in fairly good agreement with our theoretical predictions. Indeed, the BNN-improved Duflo-Zuker model predicts a root-mean-square deviation relative to experiment of about 400 keV. Given the excellent performance of the BNN refinement in confronting the recently published AME2016 compilation, we are confident of its critical role in our quest for mass models of the highest quality. Moreover, as uncertainty quantification is at the core of the BNN approach, the improved mass models are in a unique position to identify those nuclei that will have the strongest impact in resolving some of the outstanding questions in nuclear astrophysics.