Efficient catalyst screening using graph neural networks to predict strain effects on adsorption energy.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Small-molecule adsorption energies are key to catalytic reaction mechanisms.
• Altering catalyst adsorption energies via strain can improve reaction engineering.
• High-dimensional search spaces make identifying optimal strain patterns computationally challenging.

Purpose of the Study:

• To develop a predictive model for catalyst adsorption energy response to surface strain.
• To overcome the limitations of traditional density functional theory (DFT) for strain engineering.
• To identify promising catalyst materials and strain patterns for enhanced catalytic performance.

Main Methods:

• Training a graph neural network (GNN) to predict adsorption energy changes under strain.
• Generating training data by randomly straining and relaxing Cu-based binary alloy catalysts from the Open Catalyst Project.
• Evaluating model performance against ensemble linear baselines on unseen test data.

Main Results:

• The GNN model accurately predicts adsorption energy responses for 85% of unseen strain patterns.
• The model outperforms ensemble linear baseline methods.
• Cu-S alloy catalysts were identified as promising candidates for strain engineering in ammonia synthesis.

Conclusions:

• A machine learning approach, specifically a GNN, can efficiently predict strain effects on catalyst adsorption energies.
• This method facilitates the discovery of novel strain engineering strategies to break scaling relations and enhance catalyst performance.
• The developed approach offers a pathway to accelerate catalyst design for important chemical transformations.