Use this URL to cite or link to this record in EThOS:
Title: Oxide chemomechanics by hybrid atomistic machine learning methods
Author: Caccin, Marco
ISNI:       0000 0004 6347 9792
Awarding Body: King's College London
Current Institution: King's College London (University of London)
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Atomic scale phenomena concurring to atomic bond ruptures at a crack tip determine the chemomechanical properties of oxide materials, thus a better understanding of them is instrumental in addressing engineering issues related to the brittleness of oxides. In a fracturing material, the macroscopic stress field couples with the chemical reactions occurring at the crack tip in a bidirectional interplay requiring a concurrent multiscale (QM/MM) computational approach. Due to long range electrostatic interactions, the dynamics of chemically accurate description of the neighbourhood of a breaking bond requires ab initio calculations on several hundred atoms on a timescale of picoseconds. First, to make these prohibitively demanding calculations tractable I developed an ensemble parallel QM/MM computational strategy comprising a novel graph partitioning method for optimal load balancing that is able to efficiently parallellise the workload over hundreds of thousands of cores on supercomputing facilities. Secondly, I present a computational study of crack propagation in two–dimensional silica systems that have recently been experimentally synthesized, which provide ideal and physically observed structures that are key to the understanding of atomic scale phenomena in fracture events in oxides. The atomic structure, either crystalline or amorphous, and the emerging set of free energy barriers to crack advance are the basis to understand the fundamental difference in the crack dynamics observed at a larger scale. Finally, I explored different pathways to make efficient use of the information produced by ab initio calculations by studying machine learning methods capable of predicting local physical observables as a function of the local atomic environment. This includes a machine learning–augmented method to obtain free energy barriers of hybrid accuracy that only require DFT calculations on just a small fraction of the sampled atomic configurations.
Supervisor: Wurtz, Gregory Alexandre ; Kermode, James ; De Vita, Alessandro Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available