Constructing high-dimensional neural network potentials for molecular dynamics
Year: 2017
Type: dissertation
Abstract: We implement a method for constructing analytic interatomic potentials by fitting artificial neural networks to data from molecular dynamics simulations. The LAMMPS distribution is extended with an algorithm for sampling atomic configurations and energies, and TensorFlow, in combination with the Behler-Parrinello method, are employed to perform the fits. Further, we add a custom potential to LAMMPS that apply trained neural networks in simulations. We construct a neural network potential for Si, and its performance is evaluated by comparing the radial distribution function and various mechanical properties with Stillinger-Weber simulations. The crystal structure of Si is reproduced with small deviations, and the bulk modulus, shear modulus and Poisson ratio are measured to be similar to Stillinger-Weber values.
Author: John-Anders Stende
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2779405689