アイテム詳細

要旨

Green hydrogen is an integral part of the future, renewable energy landscape as clean fuel and bulk chemical of important chemical processes. Hydrogen production in flexible electrolysis cells is, furthermore, perfectly suited to buffer the fluctuating electricity supply of wind and solar parks.Proton Exchange Membrane (PEM) and Alkaline Electrolysis (AEL) are to date the most promising options to challenge the currently cheaper and thus predominant hydrogen production via steam reforming of fossil gas. Although the PEM cell offers some essential advantages compared to the Alkaline Electrolyzer, the roll-out at industrial scale is a long time coming. This is primarily due to the harsh operating conditions and the hence very limited number of potential catalysts. In the acidic and corrosive operating environment iridium oxide (IrO₂) (and to smaller extend ruthenium oxide ( RuO₂)) are the only promising catalysts for the oxygen evolution reaction (OER) at the anode. The low iridium abundance, however, demands a significant reduction of the metal loading in order to make the PEM technology sustainable and profitable. Although efforts to optimize IrO₂/ RuO₂ catalysts have led to a variety of studies the catalyst performance has only improved gradual. A more stringent and systematic optimization requires a more precise understanding of the catalyst surface and related atomic processes. While computer simulations based on Density Functional Theory (DFT) or classical force fields have enabled investigations at the atomic level for decades, both methods are limited for complementary reasons: DFT methods provide a highly accurate and complete description, are, however, limited to small(er) systems due to their enormous computational costs. The much faster force fields can cope with large systems, yet, only provide an incomplete description as they are restricted to a fixed atomic connectivity. Here emerging Machine Learning (ML) methods might close the gap, as they enable a fast, accurate, and complete description of larger systems including catalytic surfaces.
The identification of the (still unknown) IrO₂/RuO₂ surface morphologies via a specifically developed global optimization method is the central topic of this thesis. This method applies a combination of ML and DFT to determine the global energy minimum structure of various IrO₂/ RuO₂ surfaces via an iterative simulated annealing (SA) protocol. Already the first, still manually executed version of this iterative training protocol revealed multiple hitherto unknown surface morphologies (so-called complexions), that are characterized by a reordering of the terminating atom layers. The energetically lowest and thus decisive complexion was then experimentally validated in cooperation with the chair of Prof. Ulrike Diebold at the University of Vienna. Motivated by this success the training protocol has been optimized and quantified to obtain a fully automated ML enhanced global optimization method that provides a gateway to structural optimization of crystal surfaces.