Parts | Days | Selection | Search | Updates | Downloads | Help
A: Fachverband Atomphysik
A 22: Charged ions and their applications
A 22.1: Invited Talk
Thursday, March 17, 2022, 10:30–11:00, A-H1
Optimizing large atomic structure calculations with machine learning — •Pavlo Bilous1, Adriana Pálffy2, and Florian Marquardt1 — 1Max-Planck-Institut für die Physik des Lichts, D-91058 Erlangen, Germany — 2Department of Physics, Friedrich-Alexander-Universität Erlangen-Nürnberg, D-91058 Erlangen, Germany
Atomic structure calculations for heavy atoms and ions are computationally demanding due to the presence of strong electronic correlations. These corrections account for admixture of electronic configurations with excitations to unoccupied (virtual) or partially occupied orbitals. For systems with many electrons the number of such additional configurations becomes exponentially large. In this work, we make an attempt to employ a neural network to select which configurations do influence the physical quantity of interest (e.g. a transition energy or a hyperfine structure constant), and which can be omitted without significant loss of precision. As an example, we consider a highly charged Th35+ ion with the electronic configuration 4f9. This case allows for an electronic bridge scheme [1] relevant for a nuclear clock based on the 8 eV nuclear 229Th isomeric state. In this approach, accurate electronic transition energies are required. The latter were obtained recently in Ref. [2] under usage of massive computational resources. We discuss how the required resources can be reduced by carrying out neural-network-assisted calculations instead.
[1] P. V. Bilous et al., Phys. Rev. Lett. 124, 192502 (2020).
[2] S. G. Porsev et al., Quantum Sci. Technol. 6(3), 034014 (2021).