Ackland, Graeme

Hermann, Andreas

Huxley, Andrew

Martinez-Canales, Miguel

Kirsz, Marcin

2024-03-26T14:36:33Z

2024-03-26T14:36:33Z

2024-03-26

Computational studies of condensed matter phases by molecular dynamics are limited by the lack of accurate and efficient interatomic potentials. The high level theories, such as density functional theory (DFT), provide accurate potential energy surface description but lack required efficiency for large scale problems. On the other end of the spectrum are empirical potentials which are fast but often not accurate enough. The emergence of new machine learning methods for the development of interatomic potentials aim to bridge this gap. This thesis presents the development of machine learning library for interatomic potentials. The Ta-dah! software is capable of generating machine-learned potentials for mono- and multi-component systems. The library provides wide range of atomic local environment descriptors and its modular structure allows quick implementation of new ideas. The library is fully interfaced with LAMMPS molecular dynamics software. The standard use of Ta-dah! involves training with data generated from DFT packages such as VASP and CASTEP. It also incorporates a training method for learning interatomic potentials from high level quantum mechanical theories, such as coupled cluster. The method allows to harvest existing databases of high quality quantum chemistry calculations to build interatomic potentials based on methods which, in principle, can exceed that achievable by density functional theory. The library is deployed to develop efficient and accurate interatomic potentials to study various systems. The thesis highlights molecular dynamics calculations with a new potential for molecular nitrogen, based on quantum chemistry data. Phase-coexistence and free energy calculations with this potential are used to describe the melt curve and several different crystal phases. This enables calculation of the phase diagram up to 10 GPa. The potential is also applied in the to study of the proposed “Frenkel Line” in the subcritical and supercritical regions.

https://hdl.handle.net/1842/41668

http://dx.doi.org/10.7488/era/4391

en

The University of Edinburgh

Pruteanu, C. G., M. N. Bannerman, M. Kirsz, L. Lue, and G. J. Ackland. “From Atoms to Colloids: Does the Frenkel Line Exist in Discontinuous Potentials?” ACS Omega https://pubs.acs.org/doi/full/10.1021/ acsomega.2c08056.

Pruteanu, C. G., M. Kirsz, and G. J. Ackland. “Frenkel Line in Nitrogen Terminates at the Triple Point.” The Journal of Physical Chemistry Letters 12, 47: (2021) 11,609–11,615. https://pubs.acs.org/doi/10.1021/acs. jpclett.1c03206.

interatomic potentials

computational efficiency

machine learning algorithms

nitrogen

coupled cluster

Developing methods to machine-learn potentials with application to nitrogen

Thesis or Dissertation

Doctoral

PhD Doctor of Philosophy