Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.676923
Title: On-the-fly machine learning of quantum mechanical forces and its potential applications for large scale molecular dynamics
Author: Li, Zhenwei
Awarding Body: King's College London
Current Institution: King's College London (University of London)
Date of Award: 2014
Availability of Full Text:
Access through EThOS:
Access through Institution:
Abstract:
Material simulation using molecular dynamics (MD) at the quantum mechanical (QM) accuracy level has gained great interest in the community. However, the bottleneck arising from the O(N3) scaling of QM calculation has enormously limited its investigation scope. As an approach to address this issue, in this thesis, I proposed a machine-learning (ML) MD scheme based on Bayesian inference from CPU-intensive QM force database. In this scheme, QM calculations are only performed when necessary and used to augment the ML database for more challenging prediction case. The scheme is generally transferable to new chemical situations and database completeness is never required. To achieve the maximal ML eciency, I use a symmetrically reduced internal-vector representation for the atomic congurations. Signicant speed-up factor is achieved under controllable accuracy tolerance in the MD simulation on test case of Silicon at dierent temperatures. As the database grows in conguration space, the extrapolative capability systematically increases and QM calculations are nally not needed for simple chemical processes. In the on-the-y ML force calculation scheme, sorting/selecting out the closest data congurations is used to enhance the overall eciency to scale as O(N). The potential application of this methodology for large-scale simulation (e.g. fracture, amorphous, defect), where chemical accuracy and computational eciency are required at the same time, can be anticipated. In the context of fracture simulations, a typical multi-scale system, interesting events happen near the crack tips beyond the description of classical potentials. The simulation results by machine-learning potential derived from a xed database with no enforced QM accuracy inspire a theoretical model which is further used to investigate the atomic bond breaking process during fracture propagation as well as its relation with the initialised vibration modes, crack speed, and bonding structure.
Supervisor: De Vita, Alessandro ; Kermode, James ; Molteni, Carla Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.676923  DOI: Not available
Share: