MGT: Machine Learning Accelerates Performance Prediction of Alloy Catalytic Materials

Area of Science:

• Materials Science
• Computational Chemistry
• Catalysis

Background:

• Deep learning models for materials science often overlook critical active site information.
• Accurate prediction of adsorption energy is crucial for designing efficient catalysts for hydrogen evolution reactions.

Purpose of the Study:

• To develop an advanced deep learning framework for predicting adsorption energy in catalytic materials.
• To improve the focus on active atoms and adsorbates within deep learning models for enhanced catalytic performance prediction.

Main Methods:

• Inputting both overall molecular graphs and masked graphs (excluding fixed atoms) into a Masked Graph Transformer (MGT) network.
• Integrating a nonlinear message-passing mechanism with attention mechanisms and deep tensor products to capture positional information.
• Developing the NLMP-TransNet framework combining Message Passing Neural Networks (MPNN) and Transformer architectures with weight sharing and residual connections.

Main Results:

• The MGT model achieved a low error rate of 0.5447 eV on the OC20-Ni dataset, outperforming existing methods.
• Ablation studies validated the importance of site-specific features for accurate adsorption energy predictions.
• The developed framework demonstrates enhanced learning and prediction capabilities for catalytic materials.

Conclusions:

• Focusing on active site features is essential for accurate adsorption energy prediction in materials science.
• The NLMP-TransNet framework offers a promising approach for accelerating catalyst discovery and materials design.
• This work advances the application of deep learning in predicting catalytic properties for energy applications.