Practical Deep-Learning Representation for Fast Heterogeneous Catalyst Screening.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Binding site and energy are crucial for high-throughput catalyst screening, correlating with activity and selectivity.
• Machine learning models for binding energy prediction can accelerate catalyst screening but often require extensive computational resources.

Purpose of the Study:

• To present a simple, versatile representation for deep learning models to accelerate catalyst screening.
• To eliminate the need for density functional theory (DFT) calculations by utilizing labeled unrelaxed surface geometries.

Main Methods:

• Developed a labeled site approach for unrelaxed bare surface geometries, applicable to any deep learning model.
• Implemented ensemble learning to reduce model bias by combining multiple predictions.
• Applied the approach and ensemble learning to a crystal graph convolutional neural network (CGCNN) model on a dataset of alloy catalysts for CO2 reduction.

Main Results:

• Achieved mean absolute errors of 0.116 eV for CO and 0.085 eV for H binding energy on unrelaxed structures.
• Outperformed the best literature method (0.13 and 0.13 eV) which requires costly geometry relaxations.
• Demonstrated that the model effectively learns chemical information related to the binding site.

Conclusions:

• The proposed labeled site approach and ensemble learning significantly accelerate catalyst screening by removing the need for DFT calculations.
• This method provides accurate binding energy predictions, outperforming existing approaches that rely on computationally expensive relaxations.
• The model's ability to learn site-specific chemical information highlights its potential for discovering novel catalysts for CO2 reduction.