Machine Learning Spectroscopy Using a 2-Stage, Generalized Constituent Contribution Protocol.

Area of Science:

• Computational chemistry
• Machine learning applications in spectroscopy

Background:

• Accurate prediction of molecular absorption spectra is crucial for various chemical applications.
• Existing machine learning methods often require large datasets for reliable spectral prediction.

Purpose of the Study:

• To develop a highly accurate and efficient protocol for predicting molecular absorption spectra.
• To reduce the data requirements for spectral prediction models.

Main Methods:

• Integration of Bayesian neural networks (BNN) with corrected group contribution (CGC) methods.
• Development of a molecule contribution (MC) method to interpret mixing rules for CGC.
• Utilizing a small training dataset (<100 samples) for initial model training.

Main Results:

• Accurate prediction of maximum wavelengths for single molecules with <100 training samples.
• Achieved <2% mean square error for full ultraviolet spectra prediction with <500 samples.
• High accuracy in predicting spectra of mixtures using the MC-CGC approach.

Conclusions:

• The combined CGC-MC-BNN protocol offers a significant advancement in spectral prediction accuracy and efficiency.
• This data-driven approach, integrated with chemical principles, shows promise for broader molecular property prediction tasks.