Predicting binding motifs of complex adsorbates using machine learning with a physics-inspired graph representation.

Area of Science:

• Catalysis
• Materials Science
• Computational Chemistry

Background:

• Computational screening in heterogeneous catalysis is crucial for discovering new materials and reactions.
• First-principles calculations are computationally expensive, limiting exploration of complex materials and reaction mechanisms.
• Machine learning offers a more efficient alternative for predicting catalytic properties.

Purpose of the Study:

• To develop a data-efficient machine learning model for predicting adsorbate binding motifs and adsorption enthalpies.
• To apply the model to transition metals and their alloys, including complex adsorbates.
• To enable efficient exploration of materials spaces in catalysis research.

Main Methods:

• Utilized a customized Wasserstein Weisfeiler-Lehman graph kernel for feature extraction.
• Employed Gaussian process regression for predictive modeling.
• Incorporated an ensemble uncertainty estimation approach for active learning.

Main Results:

• The model accurately predicted binding motifs and adsorption enthalpies for elemental transition metals.
• Demonstrated good predictive performance on a transition metal alloy, extending beyond the training set.
• Showcased the ability to predict properties for an out-of-domain transition metal with minimal new data.

Conclusions:

• The developed data-efficient machine learning model effectively predicts adsorbate binding and enthalpies in heterogeneous catalysis.
• The approach is applicable to complex materials like alloys and can be extended to new elements.
• The model shows promise for integration into active learning strategies to accelerate catalyst discovery.