Regensburg 2022 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 43: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 1
O 43.7: Talk
Wednesday, September 7, 2022, 12:00–12:15, S054
Exploring amorphous graphene with empirical and machinelearned potentials — •Zakariya El-Machachi1, Mark Wilson2, and Volker L. Deringer1 — 1Department of Chemistry, Inorganic Chemistry Laboratory, University of Oxford, Oxford OX1 3QR, UK — 2Department of Chemistry, Physical and Theoretical Chemistry Laboratory, University of Oxford, Oxford OX1 3QZ, UK
The structure of amorphous graphene (aG) lacks long range order whilst having short and medium range order yielding a rich and complex configurational space, which is yet to be fully understood. Here we report on an atomistic modelling study of aG using a machine learning (ML) based force field. ML force fields are typically ``trained" on data from highly accurate but computationally costly density functional theory (DFT) computations. Atomistic models created by such ML methods can achieve near DFT accuracy at a fraction of the computational cost. One key assumption is that the global energy can be separated into sums of local energies. The physical interpretation of ML local energies is an interesting research question. We find that local and nearest neighbour (NN) ML energies can inform the generation of aG models from crystalline graphene via a Monte--Carlo bond switching algorithm. Bond switches are introduced as Stone--Wales defects, with the local energies of the defect pair and its NNs used in the acceptance criterion. Established empirical force fields are used in the same way and the resulting structures are studied. Our results provide insight into the modelling of amorphous graphene and into the nature of ML potential-energy models.