Investigating amorphous graphene and graphene oxide using machine-learned potentials

Thesis

The atomic-level structure of amorphous solids, which lack long-range order, has been a long-standing challenge at the crossroads between chemistry, physics and materials science for over a century. This thesis addresses this fundamental question by employing state-of-the-art atomistic simulation techniques to investigate topological disorder in two-dimensional carbonaceous materials: graphene and graphene oxide (GO).

A central theme is the application of machine learning interatomic potentials (MLIPs) to explore complex potential energy surfaces that are inaccessible to conventional methods. First, a novel approach is introduced where machine-learned atomic energies are used to drive bond-switching Monte–Carlo simulations. This method successfully navigates the configurational space of graphene, providing feasible computational models of both continuous random network and paracrystalline (cybotactic) structures and offering a unique energetic fingerprint to distinguish between them where traditional structural descriptors, such as the radial distribution function, fall short.

Next, this thesis details the development of a bespoke, highly accurate MLIP for graphene oxide (GO-MACE-23) using an iterative, on-the-fly training workflow. This potential is used to simulate the thermal reduction of a large-scale GO structural model, revealing the dynamic transformation into reduced graphene oxide (rGO). The resulting rGO structure, which is cybotactic in nature and features pores and embedded oxygen functional groups, shows excellent agreement with experimental observations, a conclusion supported by simulated X-ray photoelectron spectroscopy (XPS) that matches experimental spectra.

Finally, the GO-MACE-23 potential is applied to investigate the structure-property relationships governing the mechanical behaviour of GO and rGO. Uniaxial strain simulations reveal that the mechanical response is highly dependent on the type and distribution of oxygen functional groups. Structures rich in epoxide groups exhibit a more plastic response, while hydroxyl-rich structures are more brittle, providing key insights for designing functionalised graphene materials with tailored mechanical properties for applications in flexible electronics and composite reinforcement.

Collectively, this work demonstrates the power of machine learning in advancing our fundamental understanding of amorphous materials, providing not only new structural insights but also robust computational tools for the predictive design of next-generation carbon-based technologies.

Authors: El-Machachi, Z

• DOI: 10.5287/ora-y0q4vajvg
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: graphene; machine learning interatomic potentials; graphene oxide; amorphous materials
• Subjects: Machine learning; Computational chemistry; Materials