Machine learning-accelerated modelling of clean energy materials

August Edwards Guldberg Mikkelsen

Research output: Book/ReportPh.D. thesisResearch

12 Downloads (Pure)

Abstract

The urgency and severity of the climate crisis calls for the immediate development of green technologies, which can accelerate the transition to a sustainable energy sector. At the scientific level, much of this development is expected to happen through the use of computer simulations, which through their enormous predictive power can facilitate the discovery and understanding of materials relevant to the clean energy transition. This thesis describes the outcome of two research projects focused on the computational modelling of such materials.
The first project investigates the possibility of using machine learning to accelerate molecular dynamics simulations of the catalytically important water/Pt(111) interface. Here the room temperature structure of the interface is studied using an ensemble of neural networks with the aim of clarifying its order. The interface is found to be characterized by a primary adsorption layer of water molecules bonded atop Pt sites, which is coupled to a secondary, weakly adsorbed layer. The order in the primary adsorption layer is studied with a lattice-based approach, and an effective repulsion between the constituent water molecules is found leading to a semi-ordered interfacial structure. It is furthermore shown that these conclusions are outside the scope of ab initio based simulation methods. The neural network based molecular dynamics framework is then used to study the room temperature structure and energetics of water-hydroxyl layers on Pt(111). Here the coverage dependence of the OH adsorption energy is first examined based on a small unit cell, where a near-linear dependence, strikingly different from the predictions of the 0K bilayer model, is found. The structure of water-hydroxyl layers in larger unit cells is then investigated, where the presence of hydroxyls is found to promote water adsorption on the Pt(111) surface at the expense of introducing a depletion region at the interface. Finally, evidence is provided for the formation of water-hydroxyl clusters at low OH coverages.
The second project is focused on the computational modelling of inorganic Janus nanotubes, which constitute a novel class of low-dimensional materials with potential applications for clean energy technologies. A density functional theory-based screening study on the stability of 135 different inorganic Janus nanotubes in their armchair and zigzag forms is conducted, and a wide range of these are hypothesized to be stable at radii below 30Å. The wrapping mechanism is investigated and for isovalent anions, this is found to be governed by the lattice mismatch between the inner and outer atomic layers. Based on the calculated stabilities andradii, MoSTe Janus armchair and zigzag nanotubes are deemed particularly interesting for further studies, and a density functional theory-based investigation of their electronic properties is conducted. The effects of curvature and quantum confinement on their band structures is analyzed, and it is found that the bands are heavily modified by the curvature while confinement effects are negligible. The size dependence of their band gaps is also investigated, and it is found to depend sensitively on the radius of the simulated nanotubes due to the strain sensitivity of the Mo d states in the conduction band.
Original languageEnglish
Place of PublicationKgs. Lyngby
PublisherDTU Energy
Number of pages152
Publication statusPublished - 2021

Fingerprint

Dive into the research topics of 'Machine learning-accelerated modelling of clean energy materials'. Together they form a unique fingerprint.

Cite this