Exploring Spectrum-based Molecular Descriptors for Reaction Performance Prediction.

Area of Science:

• Chemistry
• Machine Learning
• Spectroscopy

Background:

• Modern spectroscopic characterization is accurate but underutilized in chemical machine learning.
• Extracting meaningful data from spectral information for ML models remains a challenge.

Purpose of the Study:

• To develop an automated method for converting spectral data into usable molecular descriptors for machine learning.
• To apply this method to predict reaction yields in palladium-catalyzed cross-coupling reactions.

Main Methods:

• An optical character recognition (OCR)-based approach was developed to transform spectrum images into readable vectors.
• The method was applied to a dataset of palladium-catalyzed Buchwald-Hartwig cross-coupling reactions involving aryl halides.
• The utility of predicted spectra as an alternative data source for model training was investigated.

Main Results:

• The OCR-based method successfully converted spectrum images into molecular descriptors.
• The developed approach demonstrated effectiveness in a machine learning application for predicting reaction yields.
• Predicted spectra were shown to be a viable alternative encoding source for model training.

Conclusions:

• The proposed method significantly advances the utilization of spectral information in chemical machine learning.
• This technique offers a novel way to encode spectral data, improving the prediction of chemical reaction outcomes.
• The findings pave the way for more sophisticated applications of spectral data in chemistry-related AI.