Cross-material catalyst discovery via deep learning.

Area of Science:

• Materials Science
• Catalysis
• Machine Learning

Background:

• Catalyst discovery is often limited to specific material classes, hindering cross-disciplinary insights.
• Integrating data from different catalyst types (e.g., single-atom catalysts and perovskite oxides) is challenging.

Purpose of the Study:

• To develop a machine learning approach that bridges different catalyst families.
• To enable the prediction of catalytic activity in previously untrained material classes.

Main Methods:

• Utilized co-descriptors derived from single-atom catalysts (SACs) on carbon and bulk perovskite oxide datasets.
• Employed automated statistical and natural-language analyses to select co-descriptors.
• Developed a unified crossbreeding neural network (CBNN) model.

Main Results:

• The CBNN successfully integrated distinct experimental catalyst datasets by identifying shared activity-related chemical features.
• The model accurately predicted oxygen evolution activity for single-atom catalysts on perovskite oxides, a previously untrained class.
• Identified a multimetallic catalyst with superior performance and utilized explainable AI to link descriptors to activity.

Conclusions:

• Cross-material machine learning accelerates catalyst discovery beyond existing design spaces.
• The CBNN approach facilitates the identification of high-performance catalysts by leveraging diverse data.
• Explainable AI provides insights into the atomic contributions driving catalyst performance.