Simulation of multi-shell fullerenes using Machine-Learning Gaussian Approximation Potential

TL;DR

This paper demonstrates the simulation of multi-shell fullerenes, or buckyonions, using a machine-learning potential trained on DFT data, enabling the study of their formation, structure, and energetics.

Contribution

It introduces a novel application of a machine-learning Gaussian Approximation Potential to simulate buckyonions with large atom counts, validated against DFT calculations.

Findings

Buckyonions formed by clustering and layering from outer to inner shells.

Inter-shell cohesion partly due to delocalized π electron interactions.

Energy differences within 0.02 - 0.08 eV/atom compared to DFT results.

Abstract

Multi-shell fullerenes "buckyonions" were simulated, starting from initially random configurations, using a density-functional-theory (DFT)-trained machine-learning carbon potential within the Gaussian Approximation Potential (ML-GAP) Framework [Volker L. Deringer and Gabor Csanyi, Phys. Rev. B 95, 094203 (2017)]. A large set of such fullerenes were obtained with sizes ranging from 60 ~ 3774 atoms. The buckyonions are formed by clustering and layering starts from the outermost shell and proceed inward. Inter-shell cohesion is partly due to interaction between delocalized $\pi$ electrons into the gallery. The energies of the models were validated ex post facto using density functional codes, VASP and SIESTA, revealing an energy difference within the range of 0.02 - 0.08 eV/atom after conjuagte gradient energy convergence of the models were achieved with both methods.