Transferability of Machine Learning Models for Predicting Raman Spectra.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Theoretical prediction of vibrational Raman spectra aids experimental interpretation.
• Machine learning (ML) offers efficiency and accuracy in predicting Raman spectra.
• The transferability of ML models across different molecules is not well understood.

Purpose of the Study:

• To develop a strategy for predicting Raman spectra of large alkanes using ML models trained on smaller alkanes.
• To assess the accuracy and extrapolation capability of ML-based polarizability models.
• To evaluate the transferability of ML models using descriptor space analysis.

Main Methods:

• Trained ML-based polarizability models on smaller alkane molecules (up to nine carbon atoms).
• Predicted Raman spectra and polarizabilities for larger alkanes, specifically n-undecane (11 carbon atoms).
• Utilized descriptor space analysis to evaluate model transferability.

Main Results:

• The developed polarizability model accurately predicted spectra for n-undecane, showing good extrapolation.
• The strategy avoids extensive first-principles calculations for larger systems.
• Descriptor space analysis confirmed the potential for accurate and efficient predictions using limited training data.

Conclusions:

• ML models trained on smaller molecules can effectively predict vibrational Raman spectra for larger alkanes.
• This approach offers a balance of efficiency and accuracy, reducing computational cost.
• The study validates the transferability and extrapolation capabilities of ML models in Raman spectroscopy.