Machine learning-assisted dual-atom sites design with interpretable descriptors unifying electrocatalytic reactions

Area of Science:

• Materials Science
• Computational Chemistry
• Electrochemistry

Background:

• High-throughput catalyst screening is vital for renewable energy technologies.
• Interpretable machine learning offers potential for catalyst design but faces challenges.
• Predicting catalyst activity and selectivity across multiple reactions remains complex.

Purpose of the Study:

• To develop an interpretable descriptor model for unifying catalyst activity and selectivity prediction.
• To accelerate the discovery of efficient catalysts for electrocatalytic reactions.
• To enable rapid identification of optimal catalysts using easily accessible intrinsic properties.

Main Methods:

• Developed a physically meaningful feature engineering and selection/sparsification (PFESS) method.
• Created the Atomic Property (A), Reactant (R), Synergistic (S), and Coordination (C) effects (ARSC) descriptor.
• Applied the ARSC descriptor to predict performance for O2/CO2/N2 reduction and O2 evolution reactions.

Main Results:

• The ARSC descriptor successfully decouples key effects on the d-band shape of dual-atom sites.
• The model rapidly identifies optimal catalysts, significantly reducing the need for extensive density functional theory calculations.
• Co-Co/Ir-Qv3 identified as optimal bifunctional electrocatalysts for oxygen reduction and evolution.

Conclusions:

• The ARSC descriptor provides a universal and interpretable approach for catalyst design.
• This work advances intelligent catalyst design in high-dimensional systems with physical insights.
• The findings pave the way for accelerated development of catalysts for renewable energy applications.