We present a machine learning framework that combines  sparse principal component analysis (SPCA) and artificial neural networks (ANN) to understand the critical descriptors (e.g., catalyst and support type and reaction conditions) that best explain catalyst performance. SPCA is a powerful tool for dimensionality reduction and visualization that is used to analyze the information content of different descriptor combinations and available data sets. This technique differs from standard principal component analysis (PCA) in that it identifies principal components combine information for only a handful of descriptors (as opposed to all of them, as done in PCA). This sparsification of the principal components enable a crisper analysis of the most critical descriptors. We use PCA to guide input data selection  to train an ANN that predicts reaction rates as a function of the descriptors . We also use PCA to identify predictability limits of the ANN given available experimental data and to propose new experimental date sets that can enhance predictability. In particular, we propose an optimization formulation that identifies experimental conditions that maximize information content in the principal component space. We apply our framework to the analysis of catalyst performance for the water gas shift (WGS) reaction, which is an important industrial reaction that is utilized in the manufacture of ammonia, hydrocarbons, methanol, and hydrogen. The WGS reaction presents an ideal case study since it is perhaps the most widely studied reaction and vast amounts of experimental data are available in the literature. Our framework reveals severe gaps in the explored experimental space reported in the literature and suggests that it is virtually impossible to train a robust ANN to predict catalyst performance using available experimental data. We also demonstrate that the framework is effective at suggesting new experimental conditions that can robustify catalyst prediction performance.

Machine learning framework for the analysis of catalyst performance

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.