Predicting ruthenium catalysed hydrogenation of esters using machine learning

Abstract

Catalytic hydrogenation of esters is a sustainable approach for the production of fine chemicals, and pharmaceutical drugs. However, the efficiency and cost of catalysts are often bottlenecks in the commercialization of such technologies. The conventional approach to catalyst discovery is based on empiricism, which makes the discovery process time-consuming and expensive. There is an urgent need to develop effective approaches to discover efficient catalysts for hydrogenation reactions. In this work, we explore the approach of machine learning to predict outcomes of catalytic hydrogenation of esters using various ML architectures – NN, GP, decision tree, random forest, KNN, and linear regression. Our optimized models can predict the reaction yields with reasonable error for example, a root mean square error (RMSE) of 11.76% using GP on unseen data and suggest that the use of certain chemical descriptors (e.g. electronic parameters) selectively can result in a more accurate model. Furthermore, studies have also been carried out for the prediction of catalysts and reaction conditions such as temperature and pressure as well as their validation by performing hydrogenation reactions to improve the poor yields described in the dataset.