Reaction Profile Forecasting by Artificial Data Generation for Wittig-Type Geminal Bromofluoroolefination

Machine
learning (ML) is emerging as a valuable tool in organic
synthesis for reaction design and prediction. In recent studies, the
ML approach for reaction development using big data with many features
provided the best reaction conditions for optimal yields and stereoselectivities.
However, the preparation of large data sets is often challenging,
especially for nonspecialists such as experimental scientists. In
this study, we developed simple ML models for predicting reaction
profiles of our geminal bromofluoroolefination with a minimal data
set containing only readily accessible features, including 13C NMR chemical shifts of the reacting sites and Verloop’s
Sterimol values. Notably, the model’s efficiency was significantly
enhanced through an underutilized tabular augmentation method. By
fitting the sparse data points to proper sigmoidal curves, we generated
augmented data sets that improved the predicting ability of the feed-forward
neural network (FNN). Furthermore, the combination of this augmentation
technique with a conditional tabular generative adversarial network
(CTGAN) synergistically refined the model’s performance. Our
achievement highlights the utility of tailored augmentation strategies
as a potential solution for the limitations posed by small experimental
data sets in ML-driven reaction development.