Experimental discovery of novel ammonia synthesis catalysts via active learning

Abstract

The importance of ammonia synthesis under mild conditions is increasing due to growing interest in ammonia for large-scale applications of renewable energy storage and utilization. Being one of the most investigated reactions in heterogeneous catalysis, multi-dimensional literature data are available for this reaction as a base to explore new catalysts. Machine learning (ML) can be applied to develop models using existing literature data. However, ML models developed only from literature data may not be able to efficiently predict or suggest new catalyst formulations without additional experimental data. Herein, we present an active learning (AL) framework for accelerating the discovery of novel ammonia synthesis catalysts initiated by literature data to explore a pre-determined search space based on domain knowledge efficiently. This framework generates and selects features for the ML model to capture the effects of catalyst preparation variables, kinetics, thermodynamics, support, and interactions between Ru, promoter, and the support for data mined from literature. Experimental results showed that the AL framework could discover novel catalysts that exceeded the activity of many state-of-the-art catalysts. AL reduced the number of experiments necessary to reach the best catalyst in the search space by 50%, even when no training data related to the best catalyst exists. Furthermore, AL gave insight into the properties of the catalysts that contribute to higher ammonia synthesis activity.