Data-driven understanding of disordered materials

Thesis

Machine learning (ML) interatomic potentials have become invaluable tools for atomistic simulations, achieving near quantum-chemical accuracy at extended length and time scales. For disordered materials, both slow quenching from the melt and large cells are precisely what is needed for realistic structural models, which make ML potentials especially useful for this application. This thesis describes methodological developments for further accelerating ML-driven simulations and their application to two important amorphous materials: silicon and molybdenum trisulfide.

First, we show how one ML potential model can be used to to train another, referred to as a teacher–student approach. We use an accurate, but computationally expensive model to generate training data for much faster potentials. Without the need for additional quantum-mechanical computations, extensive training datasets can be easily generated, and we find this improves the fast potentials.

Secondly, we use these fast potentials to prepare ultra-large structural models of amorphous silicon, through which we investigate rare coordination defects in the mostly fourfold random network using reliable statistics on thousands of individual defect examples. We combine structural descriptors and ML local energies to categorise over-coordinated defects into three types—a revision of the established ‘floating-bond’ model. We study defect-defect interactions and find that strain induced by defects on their neighbours leads to an energetic driving force for aggregation.

Thirdly, we parametrise an ML potential using a broad Random Structure Search database to model amorphous MoS₃, whose structure has long been debated. Constrained MD simulations discriminate between the main propositions for the structural units in MoS₃: trinuclear Mo clusters or extended Mo chains. Our results support the existence of triangular Mo₃ motifs, which agree better with total scattering data and are lower energy according to DFT than models based on chains. In large-scale MD simulations, we observe a tendency for triangles to align, which produces local nematic order in the amorphous phase. This could explain the easy decomposition of a-MoS₃ to form crystalline MoS₂.

Authors: Morrow, JD

• DOI: 10.5287/ora-ryn026xvm
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: disordered materials; GAP; machine learning; MTP; ACE; MLP; amorphous; MLFF; silicon; MoS3; interatomic potential; force field
• Subjects: Materials; Modelling; Network materials (Materials); Chemistry, Inorganic