Machine-Learning-Enabled Thermochemistry Estimator.

Area of Science:

• Computational Chemistry
• Materials Science
• Catalysis

Background:

• Density functional theory (DFT) is crucial for modeling adsorbates on metals, but computationally expensive.
• Investigating complex molecules on noble metals with DFT is limited by computational resources.
• Accurate thermochemical data is vital for catalyst design and understanding surface phenomena.

Purpose of the Study:

• To develop a novel framework for efficient estimation of thermochemical information at DFT accuracy.
• To overcome the computational limitations of DFT for modeling complex adsorbates on metal surfaces.
• To enable broader exploration of surface species energetics.

Main Methods:

• Utilized molecular encoding, descriptor synthesis, and machine learning techniques.
• Employed simplified molecular-input line-entry system (SMILES) for molecular representation.
• Incorporated group theory-based short-range (nearest neighbors) and long-range (second nearest neighbors) descriptors.
• Applied linear regression and Gaussian process regression for thermochemical property prediction.

Main Results:

• Developed and validated a machine learning framework trained on 459 surface species across Pt(111), Ru(0001), and Ir(111) surfaces.
• Achieved robust performance in reproducing energetics like enthalpies, entropies, and heat capacities.
• Demonstrated significant reduction in mean absolute errors by 48% during training and 19% during prediction compared to classical group methods.

Conclusions:

• The novel machine learning framework efficiently estimates thermochemical information with DFT accuracy.
• This approach significantly enhances the ability to research thermochemistry of complex species on metal catalysts.
• The method integrates first and second nearest neighbor group theory for comprehensive featurization and prediction.