Infrared spectroscopy-based zero-shot learning for identifying reaction intermediates in unseen systems.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Identifying reaction intermediates is crucial for understanding chemical reaction mechanisms.
• Spectroscopic techniques are key for intermediate identification, but AI models require substantial data.
• Transient intermediates often lack sufficient data for traditional AI approaches.

Purpose of the Study:

• To develop a generalizable machine learning model for identifying chemical reaction intermediates using spectra.
• To enable predictions in unseen catalytic systems, overcoming data limitations.
• To apply zero-shot learning for robust spectral analysis in chemistry.

Main Methods:

• A novel machine learning model employing zero-shot learning was developed.
• Spectra-intermediate correlations were analyzed using SHapley Additive exPlanations (SHAP).
• Visual dimensionality reduction was used to understand model generalizability.

Main Results:

• The zero-shot learning model accurately identifies intermediates in unseen catalytic systems.
• The model learns common spectral patterns, mapping diverse systems to analogous digital spaces.
• Effective predictions were achieved even with noisy data or in the presence of solvents.

Conclusions:

• The proposed model offers a robust and interpretable solution for cross-system spectral prediction.
• This work establishes a foundation for using spectral descriptors in real reaction systems.
• The study advances the application of machine learning in spectroscopy for mechanism elucidation.