Speaker
Description
Graphene (Gr) is a two-dimensional allotrope of carbon which has been demonstrated as a highly attractive material for many high technology applications. Current large-scale synthesis strategies are, however, costly and usually yield a defect-rich product, void of its outstanding properties. Recently, rapid and high-quality synthesis of Gr has been discovered on liquid Cu. Compared to its solid state, the characterization of Cu’s catalytic properties in its liquid state is a major challenge both experimentally and computationally. For the latter, the main obstacles are the large length and time scales which are necessary to reliably emulate the temporal evolution of the liquid. Corresponding molecular dynamics simulations require large simulation cells and need to span into the nanosecond timescale – an endeavor presently intractable via first-principles methods. Owed to the complexity of the catalytic process, the predictive quality of first-principles methods is, however, a prerequisite. In this work we use computationally efficient machine-learning potentials (MLPs) trained to density functional theory (DFT) data in order to extrapolate first-principles predictability to the required scales. Detailed benchmarking confirms that the MLP captures the involved physics well, accurately reproducing X-ray reflectivity determined graphene adsorption heights [1] and other observables of Gr and liquid Cu. We then apply the MLP to further study the catalytic mechanism of graphene synthesis in order to rationalize the distinct reaction kinetics found experimentally. Furthermore, we investigate the stability of graphene during its separation from the liquid Cu catalyst, an integral part of the overall production process. Our work draws a path for the use of accurately trained MLPs as a multiscale modeling technique to explore previously unchartered computational problems. In that we provide new insight into the domain of liquid metal catalysts which are generally still poorly understood at the atomic scale. Simultaneously, we show how MLPs open a new avenue for quality-control of first principles methods which is important in general to the field of computational chemistry.
[1] ACS Nano, 2021, 15, 6, 9638-9648
