Leveraging Machine Learning to Understand Heterogeneous Catalyst Systems

Designing heterogeneous catalysts that have improved activity, selectivity and reduced cost are the primary goals in catalysis. To design catalysts with improved activity and selectivity, it is important to have a good understanding of the surface configurations of the catalysts under the reaction conditions. To design catalysts with reduced cost, the ability to find catalysts of similar catalytic properties will be very helpful. In this dissertation, we attempted to tackle these two problems with the help of machine learning (ML). 

To tackle the first problem, we developed the SinlgeNN which is a machine learning potential (MLP) that can be used to predict the energies and forces of catalyst systems. Then, by combining the SingleNN with Monte-Carlo (MC) simulation, we demonstrated that we could study the surface configurations of alloy slabs with or without the presence of adsorbates. Also, since catalysts are usually not slabs in industrial processes, we laid out a systematic way to train MLPs suitable for MC simulations of catalysts in any shape. Therefore, by having a machine learning aided simulation framework for catalysts of different shapes and catalysts under different atmospheric conditions, we are a step closer to being able to simulate the surface configurations of catalysts under the reaction conditions. 

To tackle the second problem, we developed a similarity search method based on the atomic embeddings derived from graph convolutional neural network models. We demonstrated that the method is able to find similar chemical systems with examples from multiple databases. With this method, we opened up a potential new direction for finding alternative catalysts with reduced costs.