Uncovering electronic and geometric descriptors of chemical activity for metal alloys and oxides using unsupervised machine learning

We show that unsupervised machine learning (ML) using principal-component (PC) analysis provides a straightforward pathway for developing accurate and interpretable electronic-structure descriptors of the chemical catalytic properties materials. demonstrate approach by finding chemisorption metal alloys surface oxygens on metals oxides. In both cases, PC yield ML models predict material's with competitive accuracy compared built established descriptors. Importantly, interpreting patterns captured each descriptor via signal reconstruction suggests potential design motifs future allows us to identify links between geometric properties. Ultimately, we route find materials readily connect structure composition.