Systematic Bayesian Optimization for Atomic Structure Calculations of Heavy Elements

TL;DR

This paper introduces a Bayesian optimization method to improve atomic structure calculations for heavy elements, significantly enhancing energy level accuracy and transition data for lanthanides and actinides relevant to astrophysics.

Contribution

The study presents a novel sequential model-based optimization technique to refine the fictitious mean configuration in atomic calculations, achieving higher accuracy for complex multielectron systems.

Findings

Energy level discrepancies reduced from 20-60 ext{ }% to 10 ext{ }% or less.

Transition wavelengths agree within 10 ext{ }% for about 90 ext{ }% of cases.

Method achieves accuracy comparable to or better than ab-initio codes for lanthanides.

Abstract

This study presents a novel optimisation technique for atomic structure calculations using the Flexible Atomic Code, focussing on complex multielectron systems relevant to $r$-process nucleosynthesis and kilonova modelling. We introduce a method to optimise the fictitious mean configuration used in the Flexible Atomic Code, significantly improving the accuracy of calculated energy levels and transition properties for lanthanide and actinide ions. Our approach employs a Sequential Model-Based Optimisation algorithm to refine the fictitious mean configuration, iteratively minimising the discrepancy between calculated and experimentally determined energy levels. We demonstrate the efficacy of this method through detailed analyses of Au II, Pt II, Pr II, Pr III, Er II, and Er~III, representing a broad range of atomic configurations. The results show substantial improvements in the accuracy of the calculated energy levels, with average relative differences to the NIST data reduced from 20-60\% to 10\% or less for the ions studied. Transition wavelength calculations exhibit exceptional agreement with experimental data, with about 90\% of the calculated values falling within 10\% of measurements for Pr and Er ions. While improvements in transition probability calculations are observed, the calculated transition probabilities (log($gf$) values) still show significant discrepancies compared to the experimental data, with root mean square deviations of approximately 1.1-1.4 dex for Pr and Er ions. We extend our optimisation technique to systematic calculations of singly and doubly ionised lanthanides, achieving accuracies comparable to or surpassing those of \textit{ab-initio} atomic structure codes. The method's broad applicability across the lanthanide series demonstrates its potential for enhancing opacity calculations and spectral modelling in astrophysical contexts.