Active learning with non-ab initio input features toward efficient CO₂ reduction catalysts.

Area of Science:

• Materials Science
• Computational Chemistry
• Catalysis

Background:

• Conventional chemisorption models rely on d-band center theory, often requiring costly *ab initio* calculations for materials screening.
• Accurate prediction of adsorption energies and catalytic activity is crucial for designing new materials.

Purpose of the Study:

• To propose and validate a computationally inexpensive method for predicting chemisorption energies and catalytic activity.
• To develop a machine learning model using alternative descriptors for large-scale materials screening.

Main Methods:

• Utilized d-band width from muffin-tin orbital theory and electronegativity as descriptors for chemisorption.
• Combined these descriptors with machine learning algorithms, specifically neural network (NN) and kernel ridge regression (KRR).
• Employed an active learning algorithm to enhance prediction accuracy.

Main Results:

• Achieved a mean absolute deviation error of 0.05 eV for CO adsorption energy prediction on alloy systems using active learning.
• Demonstrated the model's ability to handle diverse coordination environments across different crystal facets ((100), (111), (211)).
• Identified Cu3Y@Cu* as a promising catalyst for electrochemical CO2 reduction to CO, with a significantly lower overpotential compared to Au catalysts.

Conclusions:

• The proposed descriptor set (d-band width + electronegativity) combined with machine learning offers a cost-effective alternative to traditional methods for materials screening.
• The active learning approach significantly improves prediction accuracy.
• The developed model shows practical applicability in identifying efficient catalysts for important chemical transformations like CO2 reduction.