Fully Automated Nanoscale to Atomistic Structure from Theory and Spectroscopy Experiments

Computational investigation into the structural and electronic properties of a material begins with knowledge of the underlying atomistic structure. When investigating novel or non-stoichiometric materials, various experimental spectroscopic techniques can be used to probe the material. However, moving from the spectra to the oxidation state and atomic configuration requires searching a vast structural space, where it is critical to not only match the experimental data but to also minimize quantities such as the energy to ensure structures are physically plausible and realizable. To address this need, we have previously developed the FANTASTX code, a multi-objective evolutionary algorithm which performs structure search for a variety of spectroscopies using genetic algorithm and basin-hopping methods. While FANTASTX has demonstrated success with few-atom systems, a significant challenge in extrapolating to more complex large-scale systems is the presence of near-duplicates within the search space and the significant computational expense of first-principles calculations on large-scale systems. To address these issues, we have extended FANTASTX to automatically incorporate on-the-fly machine learning methods, including both structural fingerprinting and graph neural network methods, to both identify and eliminate structural duplicates prior to processing and replace the use of density functional theory as the geometric relaxation and energy prediction mechanism.