Statistical Analysis and Discovery of Heterogeneous Catalysts Based on Machine Learning from Diverse Published Data

Abstract

AbstractThe literature provides insights for catalyst design and discovery. Effective analysis of reported data using machine learning (ML) methods offers the ability to gain valuable information. However, utilizing the literature in this way has obstacles such as lack of compositional overlaps, bias from prior published data, and low sample counts for many elements. The present study describes an ML approach that considers elemental features as input representations instead of inputting catalyst compositions directly. This ML method has the potential for catalyst discovery, including catalytic reactions with limited catalyst composition overlap in the available data. Oxidative coupling of methane (OCM), water gas shift (WGS), and CO oxidation reactions were chosen to confirm the effectiveness of the proposed method by analysis using several state‐of‐the‐art ML methods. Among the ML methods tested, gradient boosting regression with XGBoost (XGB) provided the best results, and prediction accuracy was improved by the proposed approach for all three reaction types. In addition, a quantitative value of “feature importance score” was calculated to evaluate the most influential input variables on catalyst performance. Finally, catalyst optimization was explored using ML as a “surrogate” model, and the top 20 promising candidate catalysts were identified for the OCM reaction based on the optimization. The advantages of ML in catalysis analysis as well as the difficulties and limitations originating from the complexity of heterogeneous catalysis were explored.