Heterogeneous catalysis on metallic nanoparticles

The secret ingredient of most catalysts used in the chemical industries, but also in objects of daily use, such as the catalytic converters in our cars, are tiny metallic particles with diameters in the nanometer range. We take advantage of the fact that the properties of these nanoparticles vary strongly with their size, their shape and the type of metals involved. This allows, at least in principle, also the adjustment of catalytic properties. However, little is known about the behavior of materials in this size regime. The main reason for this lack of information is the still very large number of atoms forming a metal `cluster`, which makes an exact description of such a many-body quantum system impossible. Fortunately, very good results can be achieved by the application of electronic structure methods such as density functional theory. However, the computational effort of the latter technique grows approximately with the fifth power of the system size. One way to deal with this problem is the combination of a highly accurate method for the evaluation of energies for a subset of particles of manageable size with a clever computer program which is able to learn basic features from the data provided. If the training data set is large enough, a reliable prediction of energies for related, but unknown systems can be given in a fraction of the original computational time. Particularly interesting for this task are neural networks, a machine-learning concept inspired by nature, which is based on an abstraction of the central nervous system. This project is dedicated to the development of such a neural network for the simulation of nanoparticles consisting of about 10 to 1000 metal atoms. We focus on the noble metals silver and gold, which are known to be catalytically highly active at this size, and investigate the structures of pure and mixed-metallic clusters. At a later stage, the force fields derived will be used to simulate the adsorption of selected gas molecules. Our long-term goal is to gain new insights in the effects of particle size, shape, inner structure and metallic ratio on the catalytic properties of small bimetallic clusters.

This research project was dedicated to the theoretical investigation of smallest mixed-metallic particles with potential applications as catalysts, i.e. as materials which help to accelerate a certain chemical reaction without being consumed in the process themselves. In order to describe these materials on the atomic level, a large computational effort is necessary due to the numerous interactions between the atoms of a given system. Although the solution of this many-particle problem is known in principle, its actual computation is not possible with conventional computers. Instead, approximations must be made, which simplify these interactions and limit their range. One type of approximation is based on machine learning techniques, a not-so-young paradigm of informatics, which has seen a tremendeous revival in recent years due to big advancements in computer chip design and technology. Originally, we were focusing on the application of artificial neural networks, a certain type of machine learning inspired by biological systems such as the human brain, to the prediction of unknown metal cluster structures given a data base of known atomic structures. This way, the very time- and energy-consuming computation of unknown structures on a quest toward improved materials could be circumvented. Yet, approximately half way into the project, we realized that the bottleneck of this approach will always be the extremely costly production of suitable data sets. Therefore, besides our work on direct predictors for materials properties based on their atomic structures, we also started to investigate those steps of data production which we identified as most time-consuming, and to tackle them with machine learning methods directly. We identified geometry optimizations, so-called transition state searches in molecular simulations, the actual energy evaluation with a method named "Density Functional Theory", and the typically iterative process of finding stable electron density distributions as the most crucial aspects. For all of them, we have developed machine learning approaches which can help improve their performance and will hopefully be useful to our fellow researchers in theoretical materials science.