Speaker
Description
This talk presents a method for extrapolating ab initio results obtained within the No-core shell model (NCSM) using an ensemble of neural networks. A central limitation of NCSM calculations is the finite truncation of the model space, characterized by parameters $N_{\max}$ and $\hbar\omega$, which necessitates reliable extrapolation to the infinite-basis limit.
The proposed approach employs an ensemble of neural networks trained to approximate the dependence of nuclear observables—such as binding energies, excitation energies, radii, and quadrupole moments—on the model space parameters. In contrast to traditional analytic extrapolation schemes, the ensemble captures complex nonlinear structures in the data and provides a more flexible approximation framework. The use of multiple independently trained models enables uncertainty quantification through the statistical distribution of predictions.
Particular attention is given to the generalization properties of the models, their robustness, the interpretation of predictive uncertainty intervals and their universalism. The results demonstrate that the proposed method reproduces the expected asymptotic behavior of observables and achieves competitive performance compared to established extrapolation techniques.
These findings highlight the potential of machine learning methods to improve the accuracy and reliability of ab initio nuclear structure calculations, as well as to provide a systematic framework for uncertainty quantification in theoretical predictions.