Oxygen storage materials (OSMs), such as pyrochlore type CeO2-ZrO2 (p-CZ), are used as a catalyst support for three-way catalysts in automotive emission control systems. They have oxygen storage capacity (OSC), which is the ability to release and store oxygen reversibly by the fluctuation of cation oxidation states depending on the reducing or oxidizing atmosphere. In this study, we explore high-capacity OSMs by using materials informatics (MI) combining experiments, first-principles calculations, and machine learning (ML). To generate training data for the ML model, the OSC values of 60 metal oxides were measured from the amount of CO2 produced under alternating flow gas between oxidizing (O2) and reducing (CO) conditions at 973, 773, and 573 K. Descriptors were computed by atomic properties and first-principles calculations on each oxide. The support vector machine regression model was trained to predict the OSC at each temperature. The features describing OSC were automatically selected using grid search to achieve practical cross validation performance. The features related to the stability of the oxygen atoms in the crystal and the crystal structure itself such as cohesive energy are highly correlated with OSC. The present model predicts the OSC of 1300 existing oxides. Based on its high predictive power for OSC and synthesizability, we focused on Cu3Nb2O8. We synthesized this material and experimentally confirmed that Cu3Nb2O8 showed a higher OSC than conventional OSM p-CZ. This MI scheme can significantly accelerate the development of new OSMs.
This journal is © The Royal Society of Chemistry.