Adsorption Enthalpies for Catalysis Modeling through Machine-Learned Descriptors.

Area of Science:

• Heterogeneous catalysis
• Computational materials science
• Machine learning in chemistry

Background:

• Heterogeneous catalysts are complex materials crucial for chemical reactions.
• Predicting adsorption enthalpies of reaction intermediates is key for catalyst design.
• Traditional methods like Density Functional Theory (DFT) are computationally expensive for large-scale exploration.

Purpose of the Study:

• To categorize and compare recent machine learning (ML)-based approaches for predicting adsorption enthalpies.
• To focus on model sparsity, data efficiency, and physical insight derived from ML models.
• To explore the integration of ML predictions with traditional catalysis modeling techniques.

Main Methods:

• Utilized a compressed sensing method (Sure Independence Screening and Sparsifying Operator, SISSO) to identify sparse descriptors for adsorption enthalpy prediction.
• Developed physically motivated primary features (e.g., acid/base properties, coordination numbers) as input for ML models.
• Compared various ML approaches, including hybrid physics-ML and deep neural networks.

Main Results:

• Demonstrated the effectiveness of ML, particularly SISSO, in predicting adsorption enthalpies with high accuracy and lower computational cost.
• Identified key descriptors that algebraically combine primary features to predict adsorption energies.
• Showcased the potential of ML to provide physical insights into catalytic processes.

Conclusions:

• ML-based methods offer a computationally efficient alternative to DFT for predicting adsorption enthalpies in heterogeneous catalysis.
• The developed ML approaches provide valuable physical insights and can accelerate the discovery and design of novel catalysts.
• Integrating ML predictions with established catalysis modeling tools enhances fundamental understanding and design capabilities.