Improved Representations of Heterogeneous Carbon Reforming Catalysis Using Machine Learning.

Area of Science:

• Catalysis
• Materials Science
• Computational Chemistry

Background:

• Accurate prediction of adsorption energies is vital for understanding and designing catalytic processes.
• Transition-metal catalysis is central to many chemical transformations, including carbon reforming.
• Current methods like density functional theory (DFT) are computationally expensive for large-scale mechanistic studies.

Purpose of the Study:

• To develop efficient and accurate methods for predicting adsorption energies of carbon reforming species on transition metal surfaces.
• To introduce and evaluate novel combined representations for machine learning-based catalysis.
• To assess the performance of these representations against DFT calculations and linear scaling relations.

Main Methods:

• Development of three combined representations, including Elemental Properties and Spectral London Axilrod-Teller-Muto (EP&SLATM).
• Utilizing separate EP and SLATM representations for surface and adsorbate interactions.
• Machine learning regression and tree-based models trained on DFT-calculated adsorption energies.

Main Results:

• The EP&SLATM representation achieved a mean absolute error (MAE) of ~0.18 eV for 68 adsorbates on four metal facets (Cu, Pt, Pd, Ru).
• All three combined representations outperformed traditional linear scaling relations in accuracy.
• Two representations successfully predicted energies using only empirical/experimental molecular structures, bypassing DFT optimization.
• Demonstrated effective "cross-surface" training, requiring only 20% of training data for new catalyst predictions.

Conclusions:

• Combined representations offer a significant advancement in predicting adsorption energies for heterogeneous catalysis.
• These methods drastically reduce computational cost, enabling faster exploration of catalytic mechanisms.
• The developed approach facilitates efficient machine learning model development and transferability across different catalytic surfaces.