Interpretable Machine Learning Models for Practical Antimonate Electrocatalyst Performance.

Area of Science:

• Catalysis
• Materials Science
• Computational Chemistry

Background:

• Predicting catalyst performance under reaction conditions is complex due to in situ surface evolution and solid-liquid interfaces.
• Machine learning offers a potential solution for predicting catalytic behavior, even with limited data.

Purpose of the Study:

• To develop machine learning models for predicting experimentally observed onset potentials.
• To identify key descriptors governing catalytic performance in transition-metal antimony oxides for the oxygen reduction reaction.

Main Methods:

• Utilized density functional theory (DFT) to obtain bulk atomic and electronic structural descriptors.
• Employed human-interpretable genetic programming models incorporating experimental conditions and DFT descriptors.
• Focused on non-precious transition-metal antimony oxide nanoparticulate catalysts for the oxygen reduction reaction.

Main Results:

• Successfully predicted experimentally observed onset potentials using simple machine learning models.
• Identified crucial experimental factors and supplemental bulk electronic/atomic descriptors influencing onset potentials.
• Validated predictions by experimentally confirming that scandium doping in nickel antimony oxide increases onset potential.

Conclusions:

• Machine learning, even with small datasets, can effectively model catalytic performance.
• Macroscopic experimental factors are critical descriptors for predicting catalyst behavior.
• The identified descriptors provide guidance for tuning catalysts to enhance onset potentials.