Item

Abstract

High-performing heterogeneous catalysts are key to a greener chemical industry and future sustainability. In-silico catalyst screening and discovery provide efficient and cost-effective solutions for finding suitable catalysts. Their implementations are commonly driven by the use of quantum mechanical calculations (density functional theory, DFT) to predict catalytic properties. Unfortunately, these calculations are prohibitively computationally demanding, thus incapable of searching the huge chemical space. As an alternative, earlier developed data-driven approaches, e.g., linear scaling relations (LSRs) that bypass fully explicit DFT calculations, have made notable advancements to expedite catalyst discovery on simple catalyst systems, e.g., transition metals (TMs) and monodentate adsorbates. However, given the intrinsic complexity of heterogeneous catalysis, such oversimplified approaches are not applicable for complex catalyst materials and reaction networks in terms of predictive accuracy. The emergence of machine learning (ML) has opened the road to tackling more realistic models of heterogeneous catalysts.
In this publication-based thesis, we seek to develop physics-motivated machine learning models to address the complexity of materials and adsorbates for screening heterogeneous catalysts with a particular focus on transition metal oxides (TMOs) and larger adsorbates that may exhibit mono-, bi- or higher-dentate adsorption motifs at TMs. The ML methods employed range from the Compressed Sensing SISSO method, which seeks descriptors in the form of analytical functions, to Gaussian Process Regression (GPR) with a physics-inspired graph representation. The resulting predictive accuracy that is on par with quantum mechanical calculations, along with great adaptability of these models, make them promising for finding high-performing catalysts across a broad class of materials and complex reaction networks.