Infrared spectroscopy data- and physics-driven machine learning for characterizing surface microstructure of complex...

Area of Science:

• Surface science
• Computational chemistry
• Materials science

Background:

• Characterizing complex materials under reaction conditions is crucial for accelerating materials design.
• Infrared (IR) spectroscopy provides accurate, operando insights into adsorbate/surface interactions.
• Current IR spectral interpretation relies on simplified assignments, limiting detailed structural analysis.

Purpose of the Study:

• To develop physics-driven surrogate models for generating synthetic IR spectra from first-principles calculations.
• To enable detailed surface microstructure inference from experimental IR spectra.
• To advance the characterization of complex interfaces in materials science.

Main Methods:

• Combined data-based approaches with chemistry-dependent problem formulation.
• Developed physics-driven surrogate models to generate synthetic IR spectra.
• Utilized multinomial regression via neural network ensembles to learn probability distribution functions (pdfs) for adsorption sites.
• Applied pdfs to infer surface microstructure from experimental spectra.

Main Results:

• Successfully generated synthetic IR spectra for carbon monoxide on platinum.
• Learned pdfs that describe adsorption sites and quantify uncertainty.
• Demonstrated the inference of detailed surface microstructure from experimental spectra.
• Validated the methodology for broader application to other complex systems.

Conclusions:

• Physics-driven surrogate models can accurately generate synthetic IR spectra.
• This approach allows for detailed surface microstructure characterization and uncertainty quantification.
• The methodology represents a significant step towards closing the materials gap and understanding complex interfaces.